Over the past few months, we have built and piloted an effective and secure AI platform for Heritage Trust. This post is about what we learned, which I hope is helpful to other professionals who want to do the same thing without having to become a systems architect.

Written for firms without a CIO

Most of what gets published about AI governance is written by and for organizations with a chief information officer, a security team and a procurement department. If you are a smaller firm where “IT” means the managed services provider who resets your passwords, that material tells you nothing you can act on, and it leaves the impression that doing this properly is out of reach.

But that’s no longer correct. The old model of professional software was capital: buy licences, buy a server, pay someone to look after it. The infrastructure you can build now works the opposite way. You pay per question asked, and when nobody is asking, you pay close to nothing. There is no server sitting in a closet depreciating. A small firm running a modest number of queries a month is operating at a scale where the infrastructure cost is genuinely minor next to what you already spend on your practice management subscription.

Ask where the file goes

The question I hear most is “is it secure?” The AI vendors all say yes.

The more direct question is: where does the text of my client’s file physically go, who is capable of reading it, and how long does it sit there?

Ask that and the various options become more clear. A free consumer chatbot with a marketing page about encryption is one answer. A model running inside infrastructure your firm controls is a different answer, and a different conversation with your regulator.

Whose building does the work happen in

When you use a typical AI product, you are sending your document to someone else’s building for processing. It goes out, something happens, an answer comes back. You are trusting their policies about what happens in between.

What we did instead is closer to hiring a very capable contractor to come work inside our own office, on our own filing cabinets, under our own security guard. The AI model runs inside our own cloud account. We hold the keys. The building is ours, and it is in Canada.

The same technology in two places. Only one of them can be drawn for a regulator.

In practice this means our platform sits in Amazon’s Canadian region, in an account belonging to the company. All the storage, all the logs, all the generated documents live there. Staff sign in with the same work credentials and multi-factor authentication they already use for email. Nobody installs anything. From a user’s chair it is a search box, which is the point.

The difference between using AI and using it defensibly is mostly a question of whose building the work happens in. For a smaller firm, that choice costs almost nothing extra. Running a model inside your own cloud account rather than through a consumer product is a configuration decision, not a capital project. Firms skip it because nobody told them it was available, not because they could not afford it.

You already own most of the pieces

Walk through what a typical small professional firm already has.

You almost certainly have Microsoft 365 or Google Workspace, which means you already have an identity system, and if you have turned on multi-factor authentication you already have the login control. You have a practice management system with a documented way for other software to read from it. You have a document library. You have an accounting package.

That is most of the foundation. What firms are missing is the connective layer, and the decisions about what that layer is allowed to touch.

Two practical consequences. First, ask your existing vendors what they already offer before you commission anything, because your practice management provider may have shipped half of what you want. Second, when you do bring in outside help, insist that everything is built inside an account your firm owns.

We worked with Loop8, an AWS partner, on the engineering. Everything they built sits in our account rather than theirs, which is the arrangement I would push for regardless of who you hire. Same code, same cost, and if the relationship ever ends you keep the keys, the logs and the ability to bring in someone else. Ownership of the account is the most valuable term in the engagement and it is usually free.

The setting most firms never check

One detail is worth checking before you sign anything, because it is easy to miss and awkward to discover after you have made promises about it.

When you run these models inside a cloud account, the provider often turns on something called cross-region inference by default. Translated: to keep the service fast and available, your request may be handled by hardware in another country, usually the United States, even though nothing is stored there and it is encrypted the whole way.

We checked for this at the design stage. Nothing is saved, nothing is used to train anything, and it is a fraction of a second of transit. But it is a fraction of a second of transit across the border, and if you have told a client that their data never leaves Canada, you want to know about that before they ask rather than after.

So we documented it properly. We described exactly what crosses, in what form, under what contract, and for how long. We drew it on our diagram in a different colour so nobody has to hunt for it. Then we had it formally accepted as a known, bounded exception, with a standing rule that any expansion of it requires a fresh decision by a human being.

That is a governance control rather than a technical one, which is the kind of thing our profession is already good at. It also costs nothing but attention, which is a control a small firm can execute as well as a large one.

It reads. It cannot write.

The most reassuring design choice we made is also the least glamorous. Our platform can read from the systems of record. It cannot write to them.

It reads the practice management system, the document library, the estate administration platform and the accounting system, and it respects whatever permissions each staff member already has, so nobody sees anything through the AI that they could not see by opening the file themselves. But it cannot change a matter, cannot alter a ledger, cannot touch a document.

This answers a whole category of anxiety at once. When people worry about AI, a lot of what they are picturing is silent corruption: something quietly editing a file, and nobody noticing until it matters. Make the system structurally incapable of writing and that fear has nowhere to live. The worst case becomes a bad answer on a screen, which a human reads and rejects.

Read-only is also the cheapest safeguard on this list, because it is a permission setting rather than a piece of engineering. Ask for it by name.

Where we deliberately used no AI at all

This is my favourite part, and probably the most useful one for the accountants reading.

We have a monthly obligation to produce a list of names for sanctions and AML screening. It is exactly the sort of tedious extraction job that people assume is a perfect AI use case.

We used no AI for it at all. That job runs on plain deterministic rules: pull the active parties, deduplicate them, apply the exclusion criteria, flag the exceptions, produce the workbook. A language model is built to produce fluent, plausible output, and fluent and plausible is the wrong quality for a compliance list. A name silently dropped or subtly altered on a screening list is not a quality problem you notice later. It is a regulatory failure.

We went further and attached a standing condition to it. Every run has to reconcile line by line against its own summary memo before the list goes anywhere. If it does not reconcile, the automation suspends itself and we go back to doing it by hand.

Knowing where the model is not allowed to be turned out to be at least as valuable as knowing where it helps. If you are evaluating AI for your practice, do that inventory first. Anywhere the answer must be exactly right rather than approximately useful, keep the machine out.

Nothing goes out the door unread

Our quarterly matter status and financial report drafts are generated automatically and delivered into a review folder. A person reads them, corrects them, and approves them. Nothing is ever sent to a client automatically.

I realize this sounds like it gives back most of the efficiency. It does not. Reviewing a draft that already has the numbers pulled, the deadlines listed and the missing items flagged is a fundamentally different task from building it from nothing. We kept the professional judgment exactly where it was and removed the assembly.

Every answer the system produces is also cited back to the specific system and record it came from, so verification is a click rather than an act of faith. And every request is logged: who asked, what they asked, which sources were read, what was generated. When our regulator asks how a particular output came to exist, we answer from records instead of from memory.

Draw the picture

If you do nothing else, do this one.

Before we wrote a line of anything, we produced two one-page diagrams: what the system is made of, and where data moves. That exercise was more valuable than any policy document, because a diagram cannot hedge. You either know where the client data sits or you have a blank spot on the page, and the blank spots are where the work is.

Our data flow on one page. The single amber box is the cross-border path. Everything else stays in Canada.

Ours went to a supervisory review, and having a single page that showed the boundary, with the one cross-border path marked in a colour you cannot miss, changed the tone of that conversation entirely. We showed our work rather than asking to be trusted.

A firm of any size can produce that page. It is an afternoon with a whiteboard and someone who knows how the systems connect.

How a small firm actually starts

If you want to move on this in the next quarter, here is the order I would go in.

Pick one workflow, not a strategy. Choose something you already do repeatedly and do not enjoy, where the output gets reviewed by a person anyway. Status summaries, intake, file review. Resist anything framed as “AI for the firm.”

Write the diagram before you write the cheque. One page: which systems it reads, where it runs, where the output lands, who approves it. If a vendor cannot fill that page in, you have learned something valuable for the price of a meeting.

Insist on your own account. Your firm owns the cloud account and the tenant. Not the consultant, not the vendor.

Read-only from day one. No writing back to the systems of record until you have lived with it for a while, and probably not then.

Put a human gate on anything that leaves the building. Nothing client-facing goes out unreviewed. This is the control that lets you sleep.

Keep the manual process running in parallel. For anything with a compliance dimension, do it both ways for a few cycles and reconcile. If they disagree, the automation stops and you go back to manual until you understand why.

Budget partner time, not just money. The scarce resource is the professional who can say what a good answer looks like and spot a wrong one, not the developer. That person needs real hours, and if they do not get them the project produces something impressive and useless.

Do not ask whether it is secure. Ask whoever is selling it to you to draw you the picture, and see whether they can tell you which side of the border each box sits on.

The full architecture. Green is inside our own account in Canada, amber is the one documented exception.

I am happy to share genericized versions of both diagrams with anyone who wants a starting point. Reply and I will send them along. The build was done with Loop8, who handled the AWS side while we determined what processes inside the business to automate.

I sat in on the AI session at the STEP Canada National Conference in Toronto this month, Intelligent Counsel: Harnessing AI for the Future of Estate and Tax Practice. The live demo walked through a workflow many of us will recognise: feed a trust deed in, get structured practitioner notes out, then turn those into a client-ready slide.

A few points in the talk are worth holding onto, and one in particular that I think every Canadian lawyer, accountant, and advisor should sit with. Here is what deserves your attention.

Hallucinations did not go away. They got harder to catch.

A few years ago, the failure mode was obvious: the tool would invent a case that simply did not exist. Embarrassing, but easy to catch. You search for the citation, find nothing, and you know you have a problem.

The failure mode now is more dangerous precisely because it is more plausible. The tool will hand you a real case, correctly cited, and then attach a quotation that was never in it. The scaffolding is real. The substance is fabricated. If you only verify that the case exists, you will be fooled.

Good rule of thumb: treat AI output the way you would treat work from a junior you do not yet trust. You do not assume their research is wrong, but you also do not file it without checking. When something looks unfamiliar, you go and verify it. That instinct, the one you already have as a professional, is exactly the one to apply here.

The practical takeaway: verify the content, not just the existence, of anything the tool cites. A correct citation wrapping a fabricated quote is the trap that is catching people right now.

The confidentiality rule is simpler than people think

A lot of anxiety in our profession around AI is really anxiety about client confidentiality, and the session reduced it to a clean rule of thumb:

Enterprise tools generally treat your inputs as confidential by default. Public consumer tools generally treat your inputs as training data by default.

Before you paste anything from a client file into a tool, you need to know which side of that line it sits on. Check your own enterprise policy, because the defaults differ and they change. But the framing is helpful: the question is never “is AI safe,” it is “what does this specific tool do with what I give it.”

For those of us bound by professional duties of confidentiality, this is not optional homework. It is the homework.

An Important Reframe

We are all comfortable asking whether we are meeting our duty of care if we use AI carelessly. Fair question. But one of the speakers flipped it: are you meeting your duty of care to your client if you are not using the best tools available to you?

That is uncomfortable, and it should be. The duties of competence, confidentiality, and fiduciary care do not disappear when AI enters the picture. But “competence” is not static. If a tool can catch an issue in a client’s fact pattern that you would otherwise miss, declining to use it is itself a choice with professional consequences.

The honest position sits between the two questions. You are not meeting your duty if you let a tool draft unsupervised and ship its work unchecked. You are also, increasingly, not meeting it if you refuse to engage with tools your peers are using to serve clients better. Both failures are real.

A genuinely useful way to start, if you have not

For readers who experimented with these tools a couple of years ago and have not returned since: the session’s strongest practical advice was to go back, and to go back with a specific first move.

Do not start by asking the tool to generate something new from scratch. You have no basis yet to judge whether it is any good. Instead, take something you have already drafted, a memo, a letter, a planning summary, and ask: what have I missed? What in this client’s fact pattern have I not accounted for?

You already know the right answer, so you can immediately judge the quality of what comes back. When it surfaces something real, your confidence builds on evidence rather than hype. When it produces nonsense, you have lost nothing. It is the lowest-risk way I know to calibrate your trust in a tool, and it mirrors how you would break in a new member of your team.

The session framed the broader path as crawl, walk, run, sprint. Reviewing your own drafts is the crawl. Tax scenario modelling and cross-document consistency checks are the walk and run. None of it requires you to leap.

Context is Critical for Success

If there is one technical lesson worth internalising, it is this: the quality of what you get out is driven almost entirely by the quality and completeness of what you put in.

We hold the context. We understand the family dynamics, the history between the beneficiaries, the reason a particular clause is worded the way it is. The tool knows none of that unless we tell it. A disappointing output is very often not the tool failing, it is us failing to give it what it needed to work with. If you would not expect a new associate to produce good work without a proper briefing, do not expect it of the tool either.

Where this leaves us

The closing remark from the floor was the one that got a laugh and also happened to be right: maybe we should stop calling it “hallucination” and start calling it “lying.” The euphemism makes the failure sound softer than it is. Whatever we call it, the obligation is the same. Every output gets reviewed by a human before it goes anywhere near a client. Every output gets reviewed by a human before it goes anywhere near a client.

None of this is reason for alarm, and none of it threatens what we actually do. The judgment, the relationships, the reading of a room full of family members who do not agree with each other, that is still ours, and it is becoming more valuable, not less. The tools are good at the grunt work. We are good at the part that matters.

So if you have been on the fence: open whatever your firm permits, take one thing you have already written this week, and ask it what you missed. Start there.

I write about tech tools for Canadian professionals. If you found this useful, you can subscribe at techtools.substack.com.

I have always been a collector of information.

Long before “second brains” and personal knowledge management became fashionable, I was clipping articles, saving research, hoarding notes, and hunting for better ways to connect ideas.

As a lawyer, entrepreneur, writer, and lifelong avid reader, I have spent much of my career circling the same question:

How do you capture what you know, organize what you are learning, and turn it into meaningful work?

For years, the answer was a different tool for each part of the process.

In the olden days (five years ago?), Evernote was my digital filing cabinet. It was where I collected information, clipped web pages, and built a searchable archive of everything I might one day need.

When I needed to structure my thinking, I turned to Workflowy. Its elegant outlining system was perfect for organizing ideas, drafting book outlines, mapping presentations, and working through complex problems. Workflowy helped me think.

And when it was finally time to write, I opened Microsoft Word. That is where articles, blog posts, speeches, and book manuscripts took shape.

Each tool did one job well - Evernote helped me collect, Workflowy helped me organize, and Word helped me create.

Then I discovered Notion.

One Place for Everything

What drew me to Notion was simple: it folded all three jobs into a single environment.

Today, Notion is where I collect information, store research, organize ideas, manage projects, track goals, draft articles, and write books. My books, blog posts, podcast episodes, and course outlines now begin their lives inside Notion.

In many ways, it has become the home of my intellectual life.

The Notion Dashboard

What truly transformed the experience was discovering the work of productivity expert August Bradley.

I adopted a version of his dashboard system several years ago and have been refining it ever since. From a single page, I can see active projects, writing initiatives, content ideas, reading notes, tasks and priorities, meeting notes etc.

And everything is connected.

A note from a book I am reading can become an idea for a Substack article. That article can become a podcast episode. That episode can become a chapter in a future book.

Nothing gets lost.

The Real Value Is Not Storage

Most of us already have far more information than we could ever use. Collecting is not the hard part. Finding it again, at the moment you need it, is.

The real power of a system like Notion is not storage. It is retrieval. It is connection.

It is rediscovering an article you saved three years ago precisely because it suddenly matters to the book you are writing today. It is seeing the relationship between two ideas that would otherwise stay buried in separate notebooks and folders.

For anyone who works with knowledge, that is where the value lives.

For Fellow Information Workers

If you are a writer, lawyer, consultant, entrepreneur, executive, researcher, or anyone whose work depends on learning and thinking, I would encourage you to explore Notion if you have not already.

I will be honest about the trade-off. There is a real learning curve, and you may spend more time customizing databases than you would care to admit. But once you build a system that fits the way you actually think, it stops being a productivity tool - itbecomes an operating system for your ideas.

Looking back, I can see I was always searching for the same thing. Evernote helped me collect. Workflowy helped me organize. Word helped me create.

Notion finally brought all three together. Check it out and share what you think.

Artificial intelligence is moving quickly, but for regulated professionals, the question is not simply, “How can we use AI?”

The better question is:

How can we use AI without compromising professional judgment, confidentiality, regulatory obligations, or client trust?

At Heritage Trust Company Inc., we are beginning to answer that question through a carefully scoped proof of concept. We are working with Loop8, a technology implementation partner, to build an internal AI integration platform for our company.

Loop8 is helping us develop a secure AWS-based platform that connects to our approved internal systems on a read-only basis and uses AI to help staff search, summarize, and organize matter information across those sources.

Their website is www.loop8.io.

I wanted to share what we are doing because I think this is where the legal industry needs to focus. Not on vague AI hype. Not on replacing professional judgment. Not on asking AI to “be the professional.”

The real opportunity is more practical and, in many ways, more powerful.

AI can help professional teams find information faster, reduce administrative friction, prepare better first drafts, and preserve institutional knowledge. But only if it is implemented with the right controls.

The problem we are trying to solve

Like many legal and fiduciary organizations, our information is spread across multiple systems.

At Heritage Trust, relevant matter and client information may sit in Microsoft 365 and SharePoint, Clio, Estateably, QuickBooks, email, documents, checklists, and internal notes.

Each system serves a purpose, but they do not naturally speak to each other.

That creates predictable friction.

Staff may need to move from one system to another to understand the status of an estate, trust, or client matter. They may need to check one place for documents, another for tasks, another for financial snapshots, another for client information, and another for compliance-related data.

That slows down service. It makes reporting more time-consuming. It increases the risk that information will be missed. It also creates vulnerability when institutional knowledge sits in individual staff members’ heads rather than in a system that can be searched, audited, and reviewed.

This is not just an efficiency issue. In a regulated fiduciary environment, fragmented information can become a governance issue.

What we are building

The proof of concept is designed around three practical use cases.

1. Global matter search

The first use case is a global matter search agent.

In plain English, staff will be able to ask a question like:

“What is the current status of this estate?”

The system will then search across approved internal systems and generate a unified, source-cited summary.

The summary is intended to identify:

• current matter status
• key documents
• outstanding tasks
• approaching deadlines
• recent activity
• missing information
• gaps, deficiencies, or risks
• suggested next steps for review

The key feature is source citation.

In a legal or fiduciary environment, an AI-generated answer cannot simply sound confident. It must be verifiable.

If the system says a step has been completed, staff need to know where that information came from. Was it from Clio? SharePoint? Estateably? A QuickBooks matter snapshot? A specific record or document?

That is critical. The goal is not to have AI invent conclusions. The goal is to help staff locate, synthesize, and verify information more efficiently.

2. Quarterly status and financial report drafts

The second use case is quarterly report drafting.

Preparing status and financial reports can be a slow and detailed process. Information has to be collected, checked, organized, formatted, and reviewed.

The AI platform will generate draft quarterly reports for internal review. These reports may include matter status, completed steps, outstanding steps, financial summary information, upcoming deadlines, missing items, and suggested next actions.

But the word “draft” is important.

Nothing will be automatically sent to clients, beneficiaries, regulators, or third parties.

A human will review, correct, approve, and send anything external.

This is one of our core design principles. AI can assist with information assembly and first drafts. It should not make fiduciary decisions or send professional communications without human review.

3. Monthly AML client-name export

The third use case is compliance support.

The system will compile and deduplicate active client and beneficiary names from our internal systems and generate an Excel file for upload to our existing sanctions screening tool.

This is not using AI to make AML decisions.

It is using automation to reduce the manual work involved in preparing a clean client-name list.

That distinction matters. We are not outsourcing compliance judgment to AI. We are automating a repetitive data assembly step so our compliance process can be more efficient and less prone to manual omission.

Why read-only access matters

One of the most important decisions we made was to proceed with read-only access.

The platform will be able to read approved information from connected systems, but it will not be able to write back to them.

That means it will not create, edit, delete, approve, or transmit records in our source systems.

This applies to our key systems, including Microsoft 365 and SharePoint, Clio, Estateably, and QuickBooks.

For QuickBooks specifically, we are using read-only access for matter snapshot purposes. The AI platform may be able to reference financial information, but it will not be able to alter accounting data, approve payments, initiate transactions, or change financial controls.

For a trust company, that is essential.

We are not changing our two-person EFT approval process. We are not changing wire-transfer voice verification. We are not automating payments. We are not removing human review from financial workflows.

The purpose of QuickBooks access is visibility and reporting support, not transaction authority.

Data residency and AI inference

Another important issue is data residency.

The platform is being deployed on AWS in the Canada Central region. Storage, logs, credentials, generated outputs, and the core AWS environment are intended to remain in Canada.

For the AI reasoning component, we are using Amazon Bedrock and Claude. The selected approach uses a newer Claude model through an AWS-approved cross-region inference profile. In practical terms, that means prompts and responses used for matter-search summaries may be processed through a documented inference pathway outside Canada.

We are making that choice deliberately and documenting it for technical and compliance review.

This is one of the issues I think lawyers need to understand more clearly.

It is not enough to ask, “Is this AI tool secure?”

Better questions include:

• Where is the data stored?
• Where is the data processed?
• Are prompts retained?
• Are outputs retained?
• Can the model provider train on the data?
• What contractual protections apply?
• What audit logs exist?
• Who has access?
• What permissions apply?
• What can the system do?
• What is the system technically prevented from doing?
• What remains subject to human review?

Those questions are more important than the marketing language around any AI product.

The governance controls we are building in

Because Heritage Trust is a regulated trust company, this project is as much about governance as it is about technology.

The proof of concept includes several safeguards:

• read-only access to connected systems
• no write-back to source systems
• no automatic client communications
• no automatic regulatory submissions
• source-cited outputs
• staff login through existing Microsoft 365 credentials
• MFA enforcement
• role-based access
• respect for existing source-system permissions
• audit logs showing who asked what, when, and which sources were accessed
• controlled storage of generated outputs
• human review before anything external is used or sent
• technical review by our IT provider
• documentation for regulatory and compliance assessment

The principle is simple.

AI should support the professional team. It should not quietly bypass professional judgment, compliance obligations, or internal controls.

Why this is a proof of concept

We are intentionally starting with a proof of concept rather than a large enterprise transformation project.

The purpose is to test technical feasibility, practical usefulness, governance controls, and workflow fit before making a larger commitment.

The current project is structured as a short implementation with defined phases:

• foundation and data connectivity
• global matter search
• report and compliance automation
• compliance hardening, user acceptance testing, and handoff

At the end of the process, we should have a better understanding of what works, what needs refinement, and what should or should not be scaled.

That is important. AI implementation should not be treated as a leap of faith. It should be tested against real workflows, real data, real users, and real governance requirements.

The funding structure

One practical aspect of the project is the funding.

The proof of concept fee is approximately $30k. Loop8 is applying for AWS Partner Funding of $25,000. If that funding is approved, Heritage Trust’s net cost for the proof of concept will be approximately $5k, plus AWS usage costs.

If the funding is not approved, we can either proceed at the full proof of concept fee or walk away without further obligation.

That funding structure makes the project easier to justify. It allows us to test the value and governance of the system before deciding whether to invest more heavily.

For small and mid-sized professional services firms, this is worth noting. AI adoption does not necessarily need to begin with a massive software transformation. A focused proof of concept can be a more disciplined way to proceed.

What this project is not

It is equally important to be clear about what this project is not.

It is not replacing lawyers, trust officers, or fiduciary judgment.

It is not allowing AI to make decisions for clients or beneficiaries.

It is not automatically sending emails or reports.

It is not automatically submitting anything to regulators.

It is not changing payment controls.

It is not allowing AI to alter accounting records or source data.

It is not a full AI transformation of the business.

It is a focused operational proof of concept designed to test whether AI can safely help us locate information, summarize matter status, prepare first drafts, support compliance data preparation, and reduce repetitive administrative work.

Why this matters for professionals

My view is that the most valuable near-term use of AI in legal and fiduciary work is not asking AI to “be the professional.”

It is using AI as a governed information layer across the systems professionals already use.

Most legal organizations have a knowledge fragmentation problem. Information is spread across email, practice management systems, document folders, accounting tools, spreadsheets, intake forms, notes, and the memories of individual staff members.

That creates real operational risks:

• staff waste time searching for information
• matter knowledge is fragmented
• reporting takes too long
• compliance processes become overly manual
• onboarding new staff is harder
• quality control depends too much on individual memory
• institutional knowledge is vulnerable when people leave
• professionals spend too much time assembling information and not enough time exercising judgment

AI has the potential to help with that, but only if implementation is disciplined.

The legal sector should be skeptical of AI tools that promise too much, obscure data flows, or blur the line between assistance and judgment.

But we should also be careful not to dismiss AI entirely because of those risks.

The better approach is to start with narrow, high-friction problems and build controls around them.

My early takeaway

For regulated professionals, I think the right AI strategy looks something like this:

• start with a defined business problem
• use read-only access wherever possible
• preserve existing professional and financial controls
• require source citations
• maintain human review
• document data residency and processing
• involve IT and compliance early
• test before scaling
• measure usefulness against real workflows
• treat AI governance as seriously as the technology itself

That is the model we are trying to follow at Heritage Trust.

I will share more as we move through the process, including what works, what does not, and what we learn from implementing AI in a regulated legal and fiduciary environment.

For now, my conclusion is this:

AI will not replace professional judgment.

But carefully governed AI may help professionals spend more time exercising judgment, and less time searching for the information they need to do their jobs well.

Custom AI usually sounds out of reach for small and mid-sized professional firms. The pricing alone tends to end the conversation.

But there is a path we have been working through at Heritage Trust that changes the math, and I think more firms in regulated industries should know about it.

The short version: AWS has a funding program that can offset the cost of cloud migration and modernization, including the foundation you would need for a secure custom AI system. For lawyers, trust officers, accountants, financial advisors, and other regulated professionals who need AI to be auditable and defensible rather than just clever, that matters.

Here is what we are exploring.

What we are actually trying to build

At Heritage Trust, we are looking at whether AWS could become the cloud foundation for a custom AI system that connects across the systems we already use every day. That includes Clio, Estateably, Microsoft 365, SharePoint, Outlook, Excel, intake forms, estate documents, trust records, compliance files, and client reporting workflows.

The goal is not a chatbot. It is a secure AI layer over the business. For regulated professionals, the threshold questions are whether the system is secure, auditable, permissioned, source-linked, and defensible. A consumer AI tool does not pass that test. A custom system inside our own controlled cloud environment might.

The funding angle: AWS MAP

AWS runs a program called the Migration Acceleration Program, or MAP. It is designed to help organizations migrate and modernize workloads on AWS through a three-stage framework: Assess, Mobilize, and Migrate and Modernize. It bundles AWS tools, partner expertise, training, and AWS investment intended to reduce risk and offset initial migration costs.

What makes MAP interesting is that AWS funds a meaningful portion of those costs as part of its client acquisition strategy. AWS says its service investments can include AWS service credits or partner investments to help offset one-time migration expenses, including labour, training, and running parallel environments.

In practical terms, a firm can work with an AWS partner to:

• Assess its current IT and cloud environment

• Identify duplicated hosting costs

• Build a migration business case

• Modernize legacy infrastructure

• Create a cleaner data architecture

• Offset eligible migration and build costs

• Lay the technical foundation for custom AI

That is very different from buying another AI subscription.

Why this matters for small and mid-sized firms

Most firms in our category are already paying for a long list of overlapping services: managed IT, remote hosting, Microsoft 365, practice management software, document storage, legacy servers, cloud applications, cybersecurity tools, manual reporting processes, and Excel-based compliance workflows.

The real question is whether some of that spend can be rationalized into a better cloud architecture that also supports AI. Depending on scope, AWS approval, and the partner structure, MAP funding may cover or substantially offset the assessment, migration, proof-of-concept, and professional services costs.

And here is the part most AI conversations skip: AI strategy is really data strategy.

If your documents, emails, matter records, client information, and compliance files are scattered across disconnected systems, AI will not magically solve that. It will more likely expose how messy the data environment actually is.

So the mature sequence looks like this:

1. Assess the current environment

2. Identify duplicated spend and technical debt

3. Build the AWS migration and modernization case

4. Create a clean cloud and data foundation

5. Build a secure AI proof of concept

6. Start with read-only intelligence

7. Move slowly into workflow automation with guardrails

The AI is the last step, not the first.

Why Bedrock matters

For regulated work, Amazon Bedrock does the heavy lifting on the security side.

Bedrock is AWS’s managed service for building generative AI applications. AWS says it gives customers control over the data used to customize foundation models, encrypts data in transit and at rest, supports AWS Key Management Service, and allows identity-based policies to control what users and roles can do.

Two things in particular stand out: AWS says customer data is not shared with model providers and is not used to improve base models. Bedrock can also connect through AWS PrivateLink, allowing private connectivity from a virtual private cloud without exposing traffic to the public internet.

This is the technical distinction that matters. There is a real difference between:

• Pasting client information into a public AI tool, and

• Building an AI workflow inside a controlled AWS environment with encryption, IAM permissions, private connectivity, audit logs, and source-system controls.

Bedrock also supports monitoring and logging through Amazon CloudWatch and AWS CloudTrail, which gives you governance, auditability, API activity monitoring, and compliance review.

None of this makes compliance automatic. A firm still needs privacy review, vendor due diligence, data residency analysis, cybersecurity controls, access policies, retention rules, staff training, and human review. But it gives you a foundation to build those controls on, instead of working around their absence.

Why Claude on AWS is interesting

AWS supports Anthropic’s Claude models through Bedrock. AWS describes Claude as available for enterprise use cases including reasoning, document synthesis, extraction, classification, summarization, coding, and legal document review.

Practically, that means a firm may be able to use a leading AI model while still building inside AWS’s security and governance framework. The stack starts to look like this:

• AWS account

• Amazon Bedrock

• Claude or another approved foundation model

• IAM permissions

• AWS KMS encryption

• PrivateLink

• VPC architecture

• CloudTrail logs

• CloudWatch monitoring

• S3 storage and logs

• Connectors to firm systems

• Retrieval augmented generation

• Source-linked outputs

• Human-in-the-loop workflow approvals

It is a lot of acronyms, but each one is doing a specific job in the compliance picture.

The questions every firm should ask first

For professional services business, these are the questions I would want answered before implementing AI:

• Is the AI system inside our controlled cloud account?

• Is client data used to train the model?

• Is data shared with the model provider?

• Can we choose the geographic region where data is hosted?

• Can we restrict access by role?

• Can we log API calls and user activity?

• Can the AI cite the source document?

• Can we keep the first version read-only?

• Can we require human approval before any action is taken?

• Can we preserve records of AI-generated outputs?

• Can we disable or constrain agentic actions?

• What is covered by AWS funding?

• What remains out of pocket?

• What are the ongoing AWS consumption costs?

• What happens if funding is not approved?

If you cannot answer most of those with confidence, you are not ready to deploy.

A staged model for professional services

The AI model for regulated firms will likely evolve in three stages.

Stage 1: Read-only intelligence. The AI can search across firm systems, summarize matter status, locate documents, extract structured information, and cite sources. It cannot do anything on its own.

Stage 2: Assisted workflow. The AI can draft emails, prepare reports, create checklists, generate screening exports, and flag missing information. A person reviews before anything is sent or saved.

Stage 3: Controlled agentic workflow. The AI can perform limited tasks, such as saving a report to SharePoint or creating a draft record, but only within defined permissions, audit logs, and human approval rules.

That is the responsible path for regulated work. Skipping to Stage 3 is how firms end up in front of their regulator.

Why I am watching AWS closely

A few things have to line up for custom AI to make sense for a firm our size.

MAP funding may make the cloud modernization piece financially realistic. Bedrock may make the AI architecture secure and auditable. Claude on Bedrock may provide the model capability.

We are working with an AWS partner, Loop8, that understands professional services workflows and is helping us with business process, cloud architecture, AI, and compliance.

That combination is what makes this interesting: MAP funding, AWS cloud migration, Bedrock security, Claude model access, source-linked matter intelligence, workflow automation, auditability, compliance architecture, reduced technical debt, and better data strategy.

The future of AI in professional services is not a better chatbot. It is the right infrastructure underneath it.

Today, the team from Estateably visited us at our Heritage Trust offices in West Vancouver for an in person training and strategy session. Sitting across the table from Alexander Wulkan (Co founder and COO), Paola Stolz (Customer Success Manager), and Dalia Waldston (Team Lead, Customer Success), I was reminded of something I keep coming back to in this newsletter: the right technology, adopted thoughtfully, can fundamentally change how professionals deliver their services.

In today’s post, I will walk you through what Estateably is, how we actually use it at Heritage Trust, and why I think it deserves a closer look from any Canadian professional involved in estate or trust administration. This is not a sponsored post. We are a paying customer, and this reflects our experience using the platform at Heritage Trust.

What Is Estateably?

Estateably is a cloud based estate and trust administration platform built in Montreal by cousins Ari Brojde and Alexander Wulkan. Founded in 2018, Estateably now says it serves more than 1,000 firms and 5,000 professionals across North America.

At its core, Estateably is designed to replace the highly manual, paper heavy processes that have defined estate administration for decades. If you have ever worked on an estate file, you know the reality: repetitive data entry, stacks of court forms, spreadsheet based accounting, and a constant risk of missing a deadline or a document. Estateably brings all of that into a single, secure, web based platform.

The platform covers the full lifecycle of estate administration, from the initial intake of client and estate information through to final distribution. Key capabilities include automated generation of court formatted probate forms and letters (with province specific workflows, forms, and accounting tools for Canadian estate administration), fiduciary accounting, asset and liability inventory management, task management with customizable checklists, bank and investment statement import using AI, and notice to creditors publishing.

Why Heritage Trust Chose Estateably

Heritage Trust is a boutique, independent BC trust company. We are regulated by the BCFSA, and we administer estates across the province. With a lean team, efficiency is not a luxury for us. It is a necessity.

Before Estateably, estate administration at Heritage Trust relied on Clio for file and practice management, manual checklists, and precedent documents saved in SharePoint. We still use Clio. It remains our core practice management platform for client records, time tracking, billing, contacts, and communications. But every court form had to be populated by hand from our precedents. Every accounting had to be built from scratch. Task tracking meant working through checklists that required constant manual updating, and our precedent library in SharePoint, while functional, meant a lot of copying, pasting, and reformatting for each new file.

Estateably did not replace Clio. It filled a gap that Clio was never designed to fill. Clio is excellent practice management software, but it is a general platform built primarily for law firms across all practice areas. It does not generate BC probate forms. It does not produce fiduciary accountings. It does not import bank statements or calculate date of death valuations. Estateably does all of those things, and the integration between the two reduces duplicate data entry by allowing us to create Estateably matters from Clio and carry key client and estate information across.

Here is what our workflow looks like now, with Clio and Estateably working together.

How We Actually Use It

File Setup and Intake. When Heritage Trust is appointed as executor or trustee of an estate, the first step is creating a new file in Estateably. We enter the deceased’s information, the details of the Will, the beneficiaries, and the key contacts (lawyers, accountants, financial institutions). Estateably’s shareable estate intake feature (available in BC and Ontario) allows us to gather some of this information directly from family members or referring lawyers through a secure online form, which saves significant back and forth.

Inventory and Asset Tracking. We use Estateably to log every asset and liability in the estate. The platform helps us track date of death values, holdings, and estate inventory in one place. For a trust company that administers multiple estates at any given time, having a single, reliable place to track inventory across all files is invaluable.

Bank Statement Import. One of the features we find most useful is the AI powered bank and investment statement import. Rather than manually keying in every transaction from a bank or brokerage statement, we upload the PDF directly into Estateably. The platform reads and extracts the transactions and places them into the accounting table. This alone saves hours per file.

Task Management. Every estate has a long list of required steps, many of which are jurisdiction specific. Estateably provides customizable, province specific task lists that we can tailor to each file. We use these to track progress, assign tasks to team members, and ensure nothing falls through the cracks. This is especially important for us given the regulatory expectations on a trust company.

Court Forms and Document Generation. Estateably has a large library of court forms and template documents. Its current public materials say the platform includes more than 3,000 court approved forms across Canadian and U.S. jurisdictions. With one click, we can generate the probate application documents, notices to creditors, and other forms required for a BC estate. The data we have already entered into the file populates automatically. This eliminates the risk of transcription errors and dramatically reduces the time spent on document preparation.

Fiduciary Accounting. At the end of an estate administration, the executor must pass their accounts to the beneficiaries (and sometimes to the court). Estateably generates court ready accounting reports directly from the transaction data in the file. For Heritage Trust, where accurate and compliant accounting is both a professional obligation and a regulatory requirement, this is one of the most valuable features of the platform.

Clio Integration. The integration between Clio and Estateably is straightforward: we can create an Estateably case directly from an existing Clio matter, and key client and estate information carries across so we are not re entering data. Clio remains our system of record for time tracking, billing, correspondence, and document storage. Estateably handles the estate specific work: probate forms, inventory, accounting, and administration task lists. For referring lawyers who also use Clio, this integration reduces friction and makes collaboration more efficient.

What It Does Not Do

Estateably is not a replacement for legal judgment, and it does not replace our broader matter management stack. Like any purpose built platform, it also requires onboarding discipline and internal consistency in how files are set up. If your team is not committed to entering data accurately and following a standardized process, no software will fix that. And while the platform continues to expand its capabilities, it remains an estate administration tool, not a complete practice management system. That is why we still use Clio alongside it.

Who Should Consider Estateably?

If you are a lawyer, trust company, accounting firm, or certified executor advisor involved in estate or trust administration in Canada, Estateably is worth evaluating. It is purpose built for this work. It is not a general practice management tool that has been adapted for estates. It was designed from the ground up for estate administration, and that specificity shows in every feature.

A few things to keep in mind:

It is a professional platform. Estateably is not designed for individual executors to use on their own. It is built for the professionals who guide executors through the process.

It is cloud based. Your data is accessible from anywhere, which matters for firms with remote or hybrid teams.

It is jurisdiction aware. The forms, task lists, and workflows are tailored to each Canadian province. This is a significant advantage over U.S. centric tools or generic document management systems.

The learning curve is manageable. Like any new software, there is a period of adjustment. But the onboarding support is strong, and the platform is designed to be intuitive for anyone with estate administration experience.

Resources

If you want to learn more, here are some helpful starting points:

Getting Started

• How to Set Up Your Estateably Account

• Creating Your First File

Core Features

• Using Task Management

• Adding Transactions

• Importing Bank Statements

• Fiduciary Accounting: Features and Overview

• Inventory: Managing Assets and Liabilities

• Publishing a Notice

• Shareable Estate Intake (BC and Ontario Only)

Clio Integration

• Integrating with Clio

• Creating a Case from a Clio Record

• Clio on Estateably (App Directory)

Videos

• Product Overview

• 5 Minute Demo

• Investment Statement Import Walkthrough

• COO Series with Alex Wulkan (Playlist)

Connect with the Estateably Team

• Estateably on LinkedIn

• Alexander Wulkan on LinkedIn

• Paola Stolz on LinkedIn

• Dalia Waldston on LinkedIn

Other Platforms Worth Evaluating

Estateably is the platform we chose, and it is the one I can speak to from direct experience. But the estate administration technology space is growing, and I want this newsletter to be genuinely useful to professionals who are evaluating their options.

One important caveat: most of the estate administration platforms on the market are U.S. centric. If you are a Canadian professional, your shortlist is shorter than you might expect. Here is how the landscape breaks down.

Canadian or Canada Relevant Options

Dye & Durham Estate-A-Base is probably the most obvious Canadian competitor. It is an estate administration product focused on Ontario, includes estate accounting, and automates estate documents on a subscription basis. If your workflow is heavily Ontario probate focused, this is worth evaluating. The limitation is its provincial scope.

EstateExec (estateexec.com) is a lighter weight option with a Canadian edition that provides province specific task lists, automated accounting, bank import, PDF reports, and collaboration features. It is marketed to executors, lawyers, and other users. I would treat it as more suitable for solo practitioners or smaller matters than for a trust company operating at institutional scale, but it is a useful tool to be aware of.

U.S. Centric Platforms Worth Knowing About

These platforms are strong in the U.S. market but do not currently offer meaningful Canadian jurisdiction support. I include them here because they are well known in the estate administration space and may be relevant for firms with cross border practices.

Trustate (trustate.com) is a cloud native platform covering estate planning, trust funding, asset and liability searches, and probate and trust administration. It is built around all 50 U.S. states.

EstateWorks (estateworks.com) is a configurable workflow platform for trust and estate teams, emphasizing custom setup, tasks, due date rules, and reporting. U.S. based.

EstateMin (estatemin.com) is a newer, purpose built probate platform aimed at law firms, focusing on intake through court form submission. Currently U.S. positioned.

The Lackner Group 6-in-1 (lacknergroup.com) is a long established estate and trust administration system used by law firms, accounting firms, and banks. Strong on estate administration, fiduciary accounting, and tax document preparation in one database. U.S. oriented.

Thomson Reuters ONESOURCE Fiduciary Accounting (tax.thomsonreuters.com) is less of a full practice workflow tool and more of a serious fiduciary accounting engine for trusts and estates. Worth considering if accounting rigour is your primary buying criterion. U.S. focused.

FASTER Systems (fastersystems.com) focuses on court accounting, estate tax, and gift tax software for law firms, trust companies, banks, and accounting firms. Also U.S. oriented.

A Practical Distinction

A lot of comparison pages also surface Clio, MyCase, Actionstep, CARET Legal, and LEAP in this category, but those are generally broader law practice management systems rather than tight, estate specific substitutes. They can come up in procurement conversations, but they are not the closest like for like alternatives if your core need is fiduciary accounting plus probate and trust administration workflows.

Where Estateably Fits

For Canadian trust companies and estate law firms, the realistic shortlist today is Estateably and Estate-A-Base as the primary purpose built options, with EstateExec as a lighter weight alternative. If you also have U.S. matters, the cross border picture opens up considerably. The right choice depends on your jurisdiction, your team size, and how much you value Canadian specific court forms, task management, and fiduciary accounting. For Heritage Trust, Estateably was the clear choice based on those criteria.

Final Thoughts

Technology in estate administration is still in its early stages in Canada. Many firms continue to rely on manual processes that have not changed in 20 or 30 years. Estateably is one of the tools that is actively changing that, and I can say from direct experience that it has made a meaningful difference in how Heritage Trust operates. If you have questions about our experience with the platform, feel free to reach out.

I have been adopting new technology in professional services for over twenty years. I was an early adopter of virtual servers, paperless offices, and VoIP systems back when most firms were still debating whether to move away from fax machines. Every wave of technology brought the same conversation: this will change everything, and also, this will never work here.

AI is different. Not because the conversation has changed, but because the speed has.

In late February, a Substack post rattled the stock market. A few days later, one of the most powerful financial firms in the world published a rebuttal. AI business advisor Allie K. Miller published an excellent breakdown of both reports in her newsletter AI with ALLIE, and it got me thinking about what this debate means specifically for Canadian professionals. Her analysis is worth reading in full. Here is my take on what matters most for our audience.

What Happened

Citrini Research published “The 2028 Global Intelligence Crisis” on Substack. It is a speculative piece written as a fictional financial memo from June 2028, looking backward at an economic collapse driven by AI displacing white-collar workers. It is not a forecast, but it spooked markets enough to send IBM’s stock down 13%.

Citadel Securities responded with “The 2026 Global Intelligence Crisis,” an institutional research note grounded in current labour market data and economic indicators. Their message: the data does not support the panic.

The Case for Alarm

Citrini’s scenario targets knowledge workers specifically. They argue that AI will strip out the friction that many business models depend on. When things that used to be slow and expensive become fast and cheap, the companies that charged for that complexity lose their reason to exist.

In their scenario, unemployment hits 10.2%, the mortgage market cracks as high-earning professionals default, and tax revenue collapses because governments are built to tax human income. GDP still rises on paper, but the gains flow to the owners of AI infrastructure, not to households. They call this “Ghost GDP.”

The Case for Calm

Citadel counters with numbers. Unemployment is at 4.28%. Software job postings are up 11%. New business formation is surging. They point out that data centres take years to build, energy supply is limited, and if companies tried to automate at the pace Citrini describes, the cost of compute would spike above the cost of simply paying a human. Physics and economics impose a ceiling.

They also invoke Jevons paradox: when a resource becomes more efficient to use, people historically use more of it, not less. Productivity gains tend to create new demand and new industries rather than permanent unemployment. In 1930, Keynes predicted that productivity growth would have us working 15-hour weeks by now. He was right about the productivity. He was wrong about what we would do with it. We consumed more.

Where I Come Down

I have watched technology adoption inside law firms, trust companies, and financial institutions for two decades. Here is what I know from that experience: Citadel is right that the adoption bottleneck is real. Getting a new system into a professional services organization involves change management, internal politics, procurement, training, budget approvals, and regulatory caution. That friction is genuine and it is slow.

But here is what I think both reports miss.

Everyone is focused on AI getting smarter. That matters. But the more consequential shift may be AI getting itself adopted faster. Not the model improving its own reasoning, but AI helping organizations integrate AI. Reducing the cost of connecting systems. Shortening learning curves. Handling implementation work that used to take months of consulting. Building automations that nobody asked for but everyone needed.

Citadel’s entire argument depends on the adoption curve being slow. If AI starts bending that curve, their comfortable timeline compresses. And from what I am seeing in practice, that process has already begun.

I spent years pushing firms toward paperless workflows. The resistance was never about the technology. It was about habits, comfort, and trust. AI is starting to lower those barriers in ways that previous technologies could not, because it meets people where they already work. It does not require a system migration or a six-month rollout. It sits inside the tools professionals already use.

That changes the adoption math.

How to Prepare

Regardless of which scenario you find more persuasive, the practical steps are the same.

Build real AI workflows now. Not casual experimentation. Actual integration into your daily practice. The gap between professionals who treat AI as a curiosity and those who treat it as infrastructure is growing every week. I write about specific tools in this newsletter because I believe that hands-on familiarity is the single best thing you can do for your career right now.

Diversify your skills and income. If 100% of your income depends on one type of knowledge work, you carry concentration risk. Look for adjacent skills. Build something on the side. Create an income stream that is not purely your time.

Double down on what AI still does poorly. Professional judgment. Relationship management. Navigating ambiguity. Understanding the human context behind a file or a transaction. In estate work, for example, the legal analysis is only part of the job. The rest is working with families in difficult circumstances, reading between the lines, and exercising discretion. Those skills are not going anywhere soon.

Lower your fixed costs. Both scenarios involve turbulence. A larger financial buffer and lower monthly obligations put you in a stronger position no matter what.

Watch the infrastructure. Data centre construction, energy projects, and who is financing them tell you where the money thinks the world is going. These projects take years, so they give you a window into what is coming.

Talk to people outside your field. AI will hit different sectors at different speeds. If your professional circle is limited to lawyers and trust officers, you will miss the signals coming from industries that are further along the curve. Some of the most useful conversations I have had about AI have been with people in completely different sectors.

Go to the Source

Both reports are worth reading in full. Citrini Research published “The 2028 Global Intelligence Crisis” on Substack. Citadel Securities published “The 2026 Global Intelligence Crisis” on their website. The assumptions underneath each argument matter more than the conclusions, because that is where you can start forming your own view.

Stop Typing: Why Wispr Flow is the Dictation Tool That Actually Works

If you are like most professionals I know, you spend a significant portion of your work day typing, including drafting emails, sending messages, preparing file notes, and drafting and editing documents. Your fingers are doing the heavy lifting for your brain, and they often cannot keep up.

I have been testing a tool called Wispr Flow that I believe deserves a place in your tech stack to significantly streamline your work flow. It is an AI-powered voice dictation app that works seamlessly across applications on your computer and phone. This is not the clunky, error-riddled dictation tool you may have abandoned years ago. It does not have a steep learning curve or high price point, and it actually works.

What Is Wispr Flow, Exactly?

Wispr Flow is a voice-to-text tool available for Mac, Windows, iPhone, and Android. You install it, configure a few crucial settings, and whenever you want to type something, you simply start talking instead.

Here is what makes it different from the native dictation tools built into your operating system: Wispr Flow does not just transcribe. It uses AI to clean up your speech in real time, removing filler words, adding proper punctuation and capitalization, and formatting the text so it reads like polished writing rather than a raw transcript.

Even more impressively, if you change your mind mid-thought ("We should meet Tuesday... actually no, let's do Thursday afternoon instead"), the AI course-corrects and outputs only the final intended version: "Let's meet Thursday afternoon." Wispr Flow calls this "course correction." You can ramble, backtrack, and think out loud, and what appears on screen is clean, coherent text.

It works natively inside Outlook, Word, Teams, Gmail, Slack, Google Docs, Notion, iMessage, and essentially anywhere you can place a cursor. No copying and pasting required. You speak, and the text appears right where you need it.

Why Professionals Should Care

You might think, “I type fast enough.” Perhaps you do. But consider this: the average professional types roughly 40 to 45 words per minute. A fast typist might hit 90. With voice dictation through Wispr Flow, you can comfortably produce 120 to 220 words per minute of clean, formatted text. That fundamentally changes how quickly you can clear your desk.

For practitioners specifically, here is why this matters:

We are a multilingual jurisdiction. Wispr Flow supports over 100 languages and detects language switches automatically. If you are drafting a message in English but need to reference a concept in French, Mandarin, or Punjabi for an international file, you can switch mid-sentence without touching the settings.

Administrative drag consumes too much professional time. Routine emails, meeting follow-ups, and team delegations do not require esoteric legal theory, but they consume hours due to the mechanical friction of typing. Voice dictation compresses that time dramatically.

Our working environments demand discretion. Wispr Flow includes a “Whisper Mode” that lets you dictate quietly, barely above a murmur, allowing you to work in a shared office or co-working space without broadcasting confidential matters. You do not toggle this on or off. You simply speak quietly, and the AI adapts. The accuracy when whispering is not quite as high as normal speech, but it is remarkably close and more than adequate for correspondence and notes.

Why Wispr Flow Is Worth Your Time

Before getting into the how-to, it is fair to ask whether this is worth learning or whether it will disappear in a year or two. The real cost of a new tool is the time it takes to build it into your workflow, so durability matters as much as features.

Wispr Flow looks like a relatively low-risk bet because the signals point to sustained adoption, not just a brief surge of interest. The company reports rapid growth alongside strong retention, and independent comparisons tend to place it near the top of dictation tools, with high out-of-the-box accuracy. That combination matters because voice tools only become “daily drivers” when they are reliable with minimal setup. Add in visible market validation, including a top ranking in Product Hunt’s AI dictation category, and the overall picture is of a product that is actively used and being built for the long term. Nothing is guaranteed, but if you invest time learning it, you are more likely building on a tool that will remain supported and keep improving.

The Exact Setup Guide for Professionals

To get the most out of this tool without running afoul of professional confidentiality obligations, for example the Law Society of British Columbia's Code of Professional Conduct, alongside provincial privacy legislation like BC's Personal Information Protection Act (PIPA) and the federal Personal Information Protection and Electronic Documents Act (PIPEDA), you need to set it up correctly from day one. Do not just install it and start talking. Follow these specific steps:

Step 1: Download, Install, and Grant Permissions

Navigate to wisprflow.ai/get-started. Mac users drag the app to the Applications folder; Windows users run the installer.

Launch the app and create your account (or sign in with Apple or Google). You will immediately be prompted to grant Microphone and Accessibility permissions. Do not skip this. Flow cannot insert text into your active windows without Accessibility access.

Mac: Go to Apple Menu, then System Settings, then Privacy and Security. Under both Microphone and Accessibility, toggle Wispr Flow ON.

Windows: Go to Start, then Settings, then Privacy and Security, then Microphone, and allow desktop apps to access your microphone.

If it seems like nothing is happening when you try to dictate, permissions are always the first thing to check.

Step 2: The Mandatory Privacy Lockdown

Before dictating a single word regarding an estate or trust file, click the gear icon in the Wispr Flow app to open your Settings, and navigate to Data and Privacy.

1. Enable Privacy Mode: Toggle this ON. This enforces zero data retention. Wispr will retain absolutely no audio, transcripts, or edits after the initial processing. This is the single most important setting for professionals handling confidential information.

2. Disable Context Awareness: Toggle this OFF. Context Awareness allows the app to scan the content of the window you are typing in to better understand what you are responding to. It improves accuracy, but for professionals handling confidential documents, it means the app is reading content on your screen and sending it to external servers. Disable it.

3. Opt out of model training: Ensure “Help improve Flow” is unchecked so your data is excluded from AI training.

4. Review Auto-add to Dictionary: This feature watches the text field after Flow pastes text, detects corrections you make, and adds corrected terms to your dictionary. If you handle sensitive text and are uncomfortable with that monitoring, disable it under Settings, then Personalization, and add dictionary terms manually instead.

5. Review the full Data Controls page at wisprflow.ai/data-controls before you begin using the tool for client or patient work.

I go into more detail on the privacy considerations below. These five settings are the minimum configuration for any Canadian professional before using the tool in practice.

Step 3: Set Up Your Languages and Writing Styles

During onboarding, the app will ask you to select your preferred languages for dictation. If you work in both English and French, add both now using the searchable language picker.

You will also be asked to configure personalized writing styles for different contexts: personal messages, work messages, emails, and other. For each category, you can choose from styles like casual, formal, very casual, or excited. The app shows example text for each to help you decide. Take a moment to get these right. The style settings have a meaningful impact on output quality. If you set your email style to “formal,” the output will be closer to what you actually want to send to clients and co-counsel. If you leave it at the default, you will spend time editing the tone.

Step 4: Hardware and Audio Configuration

In Flow Settings, navigate to the Audio or Microphone tab. Do not leave this on “Auto-detect.” Click the dropdown and explicitly select the specific headset or external microphone you are using. The closer the mic is to your mouth, the better the transcription accuracy. A headset microphone will give you noticeably better results than the microphone built into your laptop, especially in a room with any background noise.

Click Save and close after making your selection; the settings do not save automatically.

Step 5: Train the AI (Dictionary and Snippets)

This is the step most people skip, and it is the one that makes the biggest difference to your day-to-day experience.

In Settings, navigate to Personal Dictionary. Add industry-specific terms you use constantly. For estate and trust professionals, this might include: WESA, Trustee Act, TEP, CRA, Henson trust, RRSP, TFSA, RRIF, T4, T3, T1, prima facie, res judicata, inter vivos, testamentary, holograph, codicil, probate, letters of administration, clearance certificate, representation grant, Wills Variation Act, PIPEDA, PIPA.

The dictionary syncs across all your devices, so once you add a term on your Mac, it will be recognized on your iPhone and Android as well. Over time, Flow also picks up terms you use frequently and adds them automatically, but seeding the dictionary on day one dramatically improves the early experience.

Navigate to the Snippets tab to create voice-activated text expanders. Click “Add New” and set a trigger word paired with the full text you want inserted.

Here are some examples:

Trigger: “clearance certificate” expands to your standard paragraph explaining the CRA clearance certificate process and timeline to beneficiaries.

Trigger: “calendar link” expands to your full Calendly or booking link.

Trigger: “standard disclaimer” expands to your firm’s standard engagement or disclaimer paragraph.

Trigger: “scheduling” expands to something like “I have availability on Tuesday and Thursday afternoons next week. Would either of those work for you?”

When you are dictating and say your trigger word, Flow automatically inserts the full snippet text. So mid-dictation, you might say: “Thanks for reaching out. I would be happy to discuss this further. Calendar link.” And Flow will output the sentence followed by your complete booking link.

For teams on the Teams or Enterprise plans, you can create shared snippet libraries so everyone in your firm uses the same standardized responses, clauses, disclaimers, and templates.

The Mechanics: Exactly How to Use It

Understanding the precise hotkeys and visual cues is the key to making the tool feel like a natural extension of your workflow. When Flow is active, you will see a small floating icon (the Flow Bubble) or a waveform animation indicating it is listening.

Mode 1: Push-to-Talk (The Daily Driver)

Use this for 80% of your tasks: emails, Slack messages, quick file notes, team delegations. It works exactly like a walkie-talkie.

Mac: Press and hold the Fn key.

Windows: Press and hold the Ctrl + Windows keys.

Click into your text box. Press and hold your hotkey. Wait a fraction of a second to see the visual waveform appear, then start speaking. Do not dictate punctuation. The AI adds periods, commas, question marks, and paragraph breaks based on your cadence and context. When you finish your thought, release the key. Flow processes your speech and pastes the formatted text instantly.

If you start dictating and want to cancel, press Esc while still holding the hotkey. The text will not paste, though you can find the transcript in Flow’s “Recent activity” panel if you want it later.

If the default hotkey does not work for you (some Mac users find the Fn key awkward), go to Preferences in the app and customize it to a different key combination. Pick something you can press easily with one hand while the other is free or on the mouse.

Mode 2: Hands-Free Mode (For Complex Drafting)

Holding down keys for five minutes while dictating a long memo gets tiring. Hands-Free mode leaves the microphone open.

Mac: Double-tap Fn (or press Fn + Space).

Windows: Press Ctrl + Windows + Space.

Click into your Word document, Google Doc, or practice management software. Execute the shortcut. Release your hands from the keyboard. Speak freely, pace your office, and brain-dump your analysis. When you are done, tap your hotkey once more to stop listening and paste the transcript.

You can also activate Hands-Free mode by left-clicking the Flow bar at the bottom of your screen.

While there is no strict recording time limit, it is best practice to break exceptionally long dictations into sections. Dictate a few paragraphs, tap your hotkey to paste them and review the structure, then start a new session for the next section. This becomes second nature after a day or two.

Mode 3: Command Mode (Editing With Your Voice)

Command Mode turns Wispr Flow from a dictation tool into an active AI writing assistant. It requires a Pro plan.

Use your mouse to highlight text you have already written. Hold your Push-to-Talk hotkey (Fn on Mac or Ctrl + Windows on Windows) and give a direct instruction:

“Make this more formal.” (For turning a casual draft into a client-ready email.)

“Summarize this in two sentences.” (For condensing a lengthy paragraph.)

“Turn this into bullet points.” (For reformatting prose into a list.)

“Translate to French.” (For bilingual communications.)

“Format this into organized notes with headings.” (For structuring raw brain-dump notes after a call.)

Release the hotkey. Flow evaluates the highlighted text, rewrites it according to your instruction, and replaces the selection in place.

You can also trigger Command Mode proactively by highlighting text and saying “Hey Flow” followed by your command. Command Mode works most reliably in text editors like Google Docs, Notion, and Word. It is less consistent in chat apps like Slack, though basic commands usually work there too.

Tips for Getting the Best Results

Speak in complete thoughts, not fragments. The AI produces better output when you give it full sentences rather than choppy phrases. Think of it as explaining something to a colleague rather than dictating to a transcriptionist.

Do not say punctuation out loud. You do not need to say “period,” “comma,” or “new line.” The AI handles this. If you do say them, it will sometimes transcribe them as words rather than punctuation marks.

Say “new paragraph” when you want an explicit break. Flow inserts paragraph breaks naturally based on pauses and topic shifts, but if you want to force one at a specific point, saying “new paragraph” works.

Break long dictation into segments. Rather than dictating an entire ten-paragraph memo in one session, work in chunks of two to four paragraphs. This reduces the chance of anything getting lost and gives you a natural review point between sections.

Use it for first drafts, not final copy. The output is remarkably clean, but treat it the way you would treat a first draft from a strong associate. Read it through before sending. The AI occasionally misunderstands a word, drops a nuance, or makes a formatting choice you would not have made. A 30-second review is faster than retyping the whole thing.

Concrete Workflow Examples

Morning Email Triage

Instead of typing responses to 30 emails, click Reply in Outlook. Press and hold your hotkey, and speak: “Hi Sarah, thank you for sending over the revised draft report. I’ve had a chance to review the changes, and the updated financial summaries look excellent. I really appreciate how quickly you turned this around. Before we circulate this to the wider team, let's grab 15 minutes on Tuesday afternoon to ensure we are fully aligned on the next steps.” Release the key. Review. Send. Repeat for the next email. What used to take 45 minutes of typing now takes 15 minutes of speaking.

Post-Call File Notes

You just finished a 30-minute strategy call with co-counsel. Open your practice management software, execute the Hands-Free shortcut, and lean back. Dictate the high-level takeaways, delegated action items, and structural decisions while the conversation is still fresh. Do not worry about structure. Once pasted, highlight the raw text, hold your hotkey, and command: “Format this into organized notes with headings for key decisions, action items, and deadlines.”

Drafting a Firm Update or Team Directive

Open Word. Execute the Hands-Free shortcut. Talk through your directives for the team’s weekly operations update. Speak naturally and do not worry about polish. Once pasted, highlight the text, hold your hotkey, and command: “Make the tone encouraging but direct, and format the deliverables as a numbered list.”

Quick Slack and Teams Messages

These are where Push-to-Talk shines. Click into the Slack message field, hold your hotkey, say: “Hey team, the client approved the revised timeline so we are good to proceed. I will send the updated schedule by end of day.” Release. Send. Flow automatically adjusts its tone for chat applications, making the output more conversational than what it produces for an email.

Dictating on the Go: iPhone and Android

Mobile is where Wispr Flow removes the most friction, because typing on a phone keyboard is universally painful.

iPhone Setup and Usage

Wispr Flow on iPhone works as a third-party keyboard that you install alongside your regular keyboard. It requires iOS 18.3 or later.

Step 1: Download and install. Get Wispr Flow from the App Store. Open it once and create your account (or sign in if you already have a desktop account). Your personal dictionary and settings will sync automatically.

Step 2: Add the Flow keyboard. Go to your iPhone Settings, then General, then Keyboard, then Keyboards, then Add New Keyboard, and select Wispr Flow. You will be prompted to grant Full Access. This is required for the keyboard to function, as it needs to send your voice data to Wispr’s servers for processing.

Step 3: Grant microphone permission. The first time you try to dictate, Flow will ask for microphone access. If you accidentally deny it, go to Settings, then Privacy and Security, then Microphone, and toggle Wispr Flow on.

Step 4: Configure privacy settings. Open the Wispr Flow app and enable Privacy Mode. The same privacy settings from the desktop app apply here.

How dictation works on iPhone. Open any app where you can type: iMessage, WhatsApp, Gmail, Slack, Outlook, Notes, or anything else with a text field. When the keyboard appears, tap the globe icon to switch to the Wispr Flow keyboard.

Wispr Flow on iPhone uses what it calls “Flow Sessions.” When you tap Start Flow on the keyboard, the app briefly opens the full Wispr Flow app to activate a microphone session, then bounces you back to where you were. This is a workaround for iOS limitations that restrict third-party keyboards from accessing the microphone directly. It feels a bit fiddly the first time, but you get used to the rhythm quickly.

Once a Flow Session is active, long-press the Flow Bubble to dictate. Press and hold to speak, release to paste. In the Flow app settings, you can configure how long sessions last: immediately, 5 minutes, 15 minutes, 1 hour, or always on. If you are going to be dictating repeatedly, set this to 15 minutes or 1 hour so you do not have to re-activate the session each time.

Quick Capture Notes. You can also use Wispr Flow to capture notes without being in any specific app. Open the Flow app directly, dictate your thought, and it saves to your recent activity. These notes sync with your desktop app.

Quick Capture via the Action Button (iPhone 15 Pro and iPhone 16 models)

If you have a newer iPhone with the physical Action Button, Wispr Flow offers a particularly elegant workflow for capturing fleeting thoughts without unlocking your phone.

1. Open your iPhone’s native Settings app and tap Action Button.

2. Swipe through the visual options until you reach the Shortcut screen.

3. Tap the blue “Choose a Shortcut” button.

4. Search for “Wispr” or scroll to the Wispr Flow app icon, and select the shortcut titled “Take a Flow note.”

Once mapped, press and hold the Action Button on the side of your phone. You will feel a haptic click. Speak your thought naturally: “Note to file: we need to review the tax deadlines before our call with the accountants on Thursday.” You do not need to look at your screen, unlock your device, or open an app. Flow captures the audio, processes it through the AI, and saves a clean text note. Because the mobile app syncs with the cloud, that note appears in the Notes tab of your Wispr Flow desktop app the next time you sit down at your computer.

Known limitations on iPhone. Some users report that long-form dictation on iPhone can occasionally lose transcriptions, particularly when using Bluetooth audio devices like AirPods. The app works best in relatively quiet environments. If you experience issues, dictate in shorter bursts rather than extended monologues.

Android Setup and Usage

Wispr Flow for Android launched in February 2026 and uses a different interface from the iPhone version. Instead of a keyboard replacement, the Android app uses a floating bubble that sits on top of other apps. It requires Android OS 13 or newer.

Step 1: Download and install. Get Wispr Flow from the Google Play Store. Open the app after installation.

Step 2: Complete onboarding. The app walks you through a sequence: privacy settings first (where you should enable Privacy Mode), then an educational screen, then permission requests.

Step 3: Grant permissions. This is the most critical step. Wispr Flow on Android needs two permissions beyond the microphone:

“Display over other apps” permission, which allows the floating Flow Bubble to appear on top of whatever app you are using. Without this, you will not see the bubble.

Accessibility permission, which allows Flow to detect which app you are in and paste text into any text field. During setup, you will see an explanation screen describing what this permission enables and your ability to revoke it at any time. Tap “Agree” and you will be taken to Android’s native accessibility settings to toggle Wispr Flow on.

Important: During accessibility setup, you may see an option for a Wispr Flow “shortcut.” Do not turn the shortcut on. You only need to enable “Use Wispr Flow” under Accessibility. Enabling the shortcut instead can cause the bubble to appear but not function properly.

If you skip accessibility permissions during setup, the app will work with limited functionality: no floating bubble and no system-wide dictation. You can grant the permission later through the app’s home screen or through your phone’s Settings, then Accessibility, then Wispr Flow.

How dictation works on Android. Once set up, the Flow Bubble appears as a small floating icon whenever you tap into a text field in any app. Dictating is simple:

Tap-to-Dictate: Tap the Flow Bubble once to start recording. Speak freely without holding anything. When you are finished, either tap the close button on the bubble to stop and paste the text, or simply pause and let the auto-detect realize you have stopped speaking.

The bubble includes visual indicators: waveform animations show it is actively listening, and state transitions are communicated through colour changes and haptic feedback. The floating bubble approach is arguably more elegant than the iPhone’s keyboard method, because you never need to switch keyboards. The bubble just appears whenever you are in a text field.

Transcription history on Android. After each dictation, the app saves your transcription text and audio locally on your device. You can access past transcriptions through the app’s home screen. If a transcription is inaccurate, you can tap the report button to flag it directly to Wispr’s support team.

Current Android limitations. The Android version is brand new and you may encounter rough edges. The experience is not yet fully optimized for tablets and foldable devices. As of launch, unlimited dictation on Android is free for all users. Wispr has indicated that pricing plans will be added later.

Mobile Tips for Both Platforms

Use earbuds or a headset with a built-in microphone. The closer the mic is to your mouth, the better the accuracy, especially outdoors or in noisy settings. Wired earbuds with a mic can be more reliable than Bluetooth for dictation.

Dictate in shorter bursts on mobile. Rather than dictating a five-paragraph email in one go, break it into a paragraph or two at a time. This reduces the chance of lost transcriptions and gives you a natural review point.

Enable Privacy Mode before dictating anything sensitive. Especially important on mobile, where you may be dictating in public spaces or on shared networks.

Two Technical Realities to Consider Before You Start

1. System Performance

The AI is exceptional, but the desktop application is not lightweight. Users report that Wispr Flow consumes approximately 800 MB of RAM even when idle, and approximately 8% CPU in the background. The Windows version, which is Electron-based, tends to be heavier than the Mac version. There is also a noticeable startup delay when you first launch the application. If you routinely have large PDFs, practice management software, and multiple browser tabs open simultaneously, you may notice the performance drag. This is not a dealbreaker for most people, but it is worth knowing.

2. Remote Desktops (Citrix, RDP, VDI)

Many Canadian professionals, particularly in larger law firms and financial institutions, work through remote desktop environments like Citrix Workspace (i-worx), Microsoft Remote Desktop (RDP), or other VDI platforms. There is an important limitation: Wispr Flow cannot be installed inside most virtual desktops. It must run on your local machine.

Here is how it works in a remote desktop setup:

1. You dictate on your local machine, where Flow is installed.

2. Flow transcribes your speech and copies the cleaned-up text to your local clipboard.

3. Flow attempts to paste using Ctrl + V (Windows) or Cmd + V (Mac).

4. The remote desktop session may or may not receive the paste, depending on whether your organization has enabled clipboard sharing.

The success of that final step depends entirely on your remote desktop configuration. Many Citrix and RDP setups support clipboard redirection, but some organizations disable it for security reasons. If clipboard sharing is locked down, Flow’s text will land on your local clipboard but will not make it into the application running inside your remote desktop.

Before investing time in setup, check with your IT department about whether clipboard sharing is enabled. For more detail, Wispr has published official guidance on remote desktop compatibility.

If your entire workflow lives inside a Citrix or VDI environment, test with the free plan and your specific setup before committing to a paid subscription.

Privacy and Security: What You Need to Know

I have already walked you through the mandatory privacy settings above. This section explains why those settings matter and gives you the full picture so you can make an informed decision for your practice.

The Early Privacy Controversy

Any honest review of Wispr Flow needs to acknowledge what happened early in its history. When the tool was newer, technically sophisticated users noticed some troubling behaviour: the app kept re-adding itself to startup items even after users removed it, network monitoring showed constant outbound connections even when the app was not actively in use, and the original privacy policy permitted using customer content for AI model training without making this clearly opt-in. The Context Awareness feature, which reads content from your active app window, was not well explained.

When a Reddit thread about these issues went viral, the user who raised the concerns was initially banned from the community. To the company’s credit, Wispr’s CTO eventually stepped in publicly, acknowledged the problems, apologized, and committed to changes.

Since then, the company has made significant improvements. AI training on user data is now opt-in and off by default. The Data Controls page clearly explains what is collected, how it is used, and how to control it. Privacy Mode with zero data retention is available on all plans, including the free tier. The episode is a fair reason to approach the tool with your eyes open, but the current state is substantially improved.

How Your Data Actually Flows

Let me be specific about what happens when you dictate, because “cloud-based” can mean different things.

Your voice audio is sent to external servers for processing. All transcription happens in the cloud. The company uses a combination of models including Meta’s LLAMA and OpenAI’s Whisper to process your speech. Your voice data leaves your device and is handled by third-party AI infrastructure. Wispr states it has agreements with all providers ensuring your data is not retained by those providers or used to train their models.

Context Awareness collects screen content. If you enable Context Awareness, the app reads limited text from your active app window to improve transcription accuracy. For example, if you are replying to an email, it might read the email thread to better understand names and context. This is powerful for accuracy but means the app is reading content on your screen. For professionals handling confidential documents, this should be disabled.

Without Privacy Mode, data may be retained for 30 days. Standard plans retain audio and transcripts for up to 30 days before automatic deletion. During this period, if you have not opted out, this data may be used to improve Flow’s features and models.

With Privacy Mode enabled, nothing is stored. Zero data retention. No audio, no transcripts, no edits after processing. The only thing still collected is aggregated usage statistics like total word count. This is the setting every Canadian professional should use.

All data is encrypted using industry-standard TLS during transit and at rest on servers.

Current Security Certifications

SOC 2 Type II compliant across all plans, including the free tier. This is an independent audit verifying that the company safeguards your data across security, availability, and confidentiality principles.

HIPAA compliance available on all plans with acceptance of a Business Associate Agreement (BAA). While HIPAA is a U.S. framework, it signals a level of data handling rigour relevant for Canadian healthcare professionals operating under PIPEDA. I enabled this on my account.

ISO 27001 certification in progress and available on Enterprise plans.

Enterprise controls including SSO/SAML, admin dashboards, and enforced zero data retention at the organization level.

What This Means for Different Professional Contexts

General business use (emails, Slack, internal communications). Wispr Flow is a reasonable choice. Enable Privacy Mode, disable Context Awareness, and review the Data Controls page. The security posture is comparable to other cloud-based productivity tools you likely already use.

Lawyers handling client-privileged information. Proceed with caution. Your obligations under your law society’s rules of professional conduct require reasonable steps to protect client confidentiality. Enable Privacy Mode immediately. Disable Context Awareness. Consider whether you are comfortable with voice data being processed through third-party cloud servers even with contractual protections in place. For highly sensitive matters, you may want to type rather than dictate. For routine correspondence and internal notes, Flow with Privacy Mode enabled is likely within the bounds of reasonable practice, but make that assessment based on your own obligations and risk tolerance.

Accountants and financial advisors. Similar considerations. Enable Privacy Mode. Be thoughtful about dictating specific client financial details, account numbers, or tax information. For general correspondence and advisory writing, the tool is practical with appropriate privacy settings.

Healthcare professionals. HIPAA compliance with a BAA is available. Canadian healthcare privacy obligations under PIPEDA (and provincial statutes like Ontario’s PHIPA or Alberta’s HIA) are broadly similar in their requirement to protect personal health information. Enable Privacy Mode and disable Context Awareness before using Flow for any patient-related content.

Teams and firms. The Enterprise plan with enforced Privacy Mode at the organization level is the appropriate choice. This ensures that every user operates under zero data retention by default, removing the risk that an individual user forgets to enable it.

The Bottom Line on Privacy

The fundamental trade-off is this: cloud processing is what makes Wispr Flow so accurate and fast. If you need fully offline, on-device voice dictation where no data ever leaves your machine, this is not the right tool. But if you are comfortable with the same cloud-based architecture that powers your email, document storage, and video conferencing, then Wispr Flow with Privacy Mode enabled is a reasonable addition to your workflow, provided you configure it thoughtfully using the steps outlined above.

What Does It Cost?

Wispr Flow displays pricing in your local currency, so Canadian users will see CAD. Here is the structure in USD for reference:

Flow Basic (Free). This is a genuinely useful free tier, not a crippled demo, but note that the limits are platform-specific. You get 2,000 words per week on Mac and Windows, and 1,000 words per week on iPhone. As of their February 2026 launch, Android users currently get unlimited free dictation. This is enough to handle a moderate volume of emails and to properly evaluate whether the tool works for your practice.

Flow Pro (US $12/month billed annually, or US $15/month billed monthly). Unlimited dictation, AI Command Mode, cross-device sync, and early access to new features. This is the plan for anyone who writes extensively as part of their work.

Flow Teams (US $10/user/month billed annually). Everything in Pro, plus shared dictionaries, shared snippets, centralized billing, and admin controls. Ideal for firms that want consistent terminology and streamlined management across their team.

Flow Enterprise (Custom pricing). SOC 2 Type II, HIPAA with BAA, SSO/SAML, enforced zero data retention, and enterprise API access. For organizations with advanced compliance requirements.

All new accounts start with a 14-day free trial of Flow Pro, no credit card required.

Troubleshooting Common Issues

Flow is not hearing me. Check that microphone permissions are granted. On Mac: System Settings, then Privacy and Security, then Microphone. On Windows: Settings, then Privacy and Security, then Microphone. Also confirm accessibility permissions are enabled.

Text is not pasting into my app. This almost always means accessibility permissions are incomplete. On Mac: System Settings, then Privacy and Security, then Accessibility, and toggle Wispr Flow on.

Dictation keeps failing or cancelling. Make sure you have a stable internet connection. If you are on a VPN, try disabling it temporarily to test. Some corporate firewalls block the connection. If dictation works with the VPN off, ask your IT department to whitelist Wispr Flow’s servers.

It is not recognizing a specific name or term. Add it to your personal dictionary manually. The AI will start recognizing it immediately.

The floating toolbar is distracting. You can turn it off in Flow app settings by toggling off the Flow bar. You can still activate dictation using your keyboard shortcut.

The Bottom Line

Voice dictation will not replace your keyboard entirely. Quick edits, keyboard shortcuts, and some types of precise formatting work are still faster with your fingers. But for the heavy lifting of getting high-level thoughts into text, speaking is fundamentally faster and often yields a better result because it bypasses the self-editing friction that slows us down while typing.

If your time is highly valued, the math is straightforward. Dictating at three to four times your typing speed clears your administrative backlog, preserving your energy and focus for the strategic, high-value file work that actually requires your full attention.

Give the free plan a try for a week. I think you will be surprised by how quickly the keyboard becomes a secondary tool.

Have you tried Wispr Flow or another voice dictation tool? I would love to hear your experience. Leave a comment below or reply directly to this email.

If you found this useful, please share it with a colleague who spends too much time typing.

On January 12, 2026, Anthropic released something that quietly changes the game for knowledge workers: Claude Cowork.

If you have not heard of it yet, here is the short version. Cowork is an autonomous AI agent that runs on your computer and can actually do work, not just talk about it. You give it access to a folder on your machine, describe what you need done, and it plans, executes, and completes multi-step tasks while you do other things. You come back to finished files.

This is not a chatbot. You are not going back and forth, copying and pasting outputs into documents. Claude reads your files, creates new ones, edits existing ones, and can even work across multiple applications through connectors to tools like Google Drive, Gmail, Slack, and Microsoft 365.

Cowork is built on the same foundation as Claude Code, the terminal-based coding tool that became one of Anthropic’s most successful products. Developers loved Claude Code, but Anthropic noticed something interesting: people were using it for everything except coding. Vacation research. Slide decks. Email cleanup. Expense reports from receipt screenshots. Cowork takes that same underlying agent capability and wraps it in an interface anyone can use.

The catch: Cowork is currently available only on macOS, and only for Claude Max subscribers ($100 to $200 per month). Anthropic has confirmed Windows support is coming, with a target of mid-2026. A waitlist exists for users on other subscription tiers.

Want to see it in action? There is a demo video on YouTube.

Why Cowork Matters

The shift here is fundamental. For years, AI assistants have been sophisticated typewriters. You ask a question, you get text back, and then you do the actual work of turning that text into something useful: copying it into a document, formatting it, saving it, integrating it with your other files.

Cowork eliminates that middle step. The AI does not just generate content. It completes tasks. It saves files. It organizes folders. It creates deliverables you can actually use.

For knowledge workers, this is the difference between having a research assistant who hands you notes and having one who hands you a finished memo.

The technical backbone is MCP, or Model Context Protocol, an open standard Anthropic introduced for secure two-way connections between AI and external systems. This is what allows Claude to operate within the context of your actual documents and workflows rather than in an isolated chat window.

The Legal Plugin

Which brings us to the news that has legal tech in a state of mild panic.

On January 30, Anthropic open-sourced 11 specialized plugins for Cowork, covering functions like marketing, finance, sales, and customer support. One of them is designed specifically for in-house legal teams.

Bob Ambrogi has the full breakdown in his latest piece on LawNext: Anthropic’s Legal Plugin for Claude Cowork May Be the Opening Salvo In A Competition Between Foundation Models and Legal Tech Incumbents.

According to Anthropic, the legal plugin handles core in-house workflows:

• Reviews contracts clause-by-clause against your negotiation playbook, flagging issues with red, yellow, and green ratings and suggesting redlines

• Triages incoming NDAs and routes them for standard approval, counsel review, or full legal review

• Generates contextual briefings on topics, incidents, or daily updates

• Drafts templated responses for common requests like discovery holds and data subject access requests

It integrates with Slack, Box, M365, Jira, and Egnyte.

Market Reaction

The Guardian, Bloomberg, and Reuters all reported significant stock selloffs following the plugin release. RELX, Wolters Kluwer, and Thomson Reuters all took hits.

The concern is straightforward. For years, legal AI vendors have built products by wrapping foundation models with proprietary playbooks, clause libraries, and workflow integrations. They assumed the model layer would remain neutral infrastructure that they could build on top of.

Anthropic just demonstrated that foundation model companies can ship their own workflow products directly to enterprise customers. And because the plugin is open source, legal ops teams can customize it without waiting on a vendor roadmap.

This is not just about one plugin. It is about the competitive dynamics of the entire legal tech market. Anthropic is no longer just selling API access. It is delivering a packaged solution that combines model, interface, and workflow automation.

The plugin also ships alongside tools for sales, marketing, finance, and customer support. For corporate legal departments that handle contracts originating from those functions, the distribution channel is already built in.

Privacy and Client Data

Before legal departments rush to adopt Cowork, there are serious questions about client data protection that need to be addressed.

The most fundamental issue: anything you share with Claude is sent to Anthropic’s servers for processing. While Anthropic states that Enterprise plan data is not used for model training by default, consumer and Pro accounts operate under different rules. As of late 2025, consumer accounts have training enabled unless users explicitly opt out. The naming is confusing here. “Claude Pro” sounds like a business tier, but it is actually a consumer account with consumer privacy treatment. The same applies to Team accounts at $30 per month. Only Claude Enterprise or API access with a Zero Data Retention agreement provides true business-grade protection.

For lawyers, this creates immediate ethical concerns. Client confidentiality is not optional. If you are using a consumer or Pro account to review contracts or triage NDAs, that client data may be training future versions of the model unless you have specifically disabled that setting.

There are also active security vulnerabilities to consider. Within days of Cowork’s launch, security researchers at PromptArmor demonstrated that attackers can exploit prompt injection to exfiltrate files from Cowork to their own Anthropic accounts. The attack works by hiding malicious instructions in documents that appear harmless, using techniques like white text on white backgrounds or microscopic font sizes. When Claude processes the document, it follows the hidden instructions and sends your files elsewhere. Anthropic acknowledged the underlying vulnerability months before Cowork launched but did not remediate it.

Anthropic’s own safety guidance tells users to “avoid granting access to local files with sensitive information, like financial documents” while simultaneously marketing Cowork as a tool for organizing exactly those kinds of files.

For in-house legal teams considering Cowork, here is what responsible adoption looks like:

Use Enterprise accounts only. Consumer, Pro, and Team accounts do not provide adequate privacy protection for client data. Enterprise plans include SSO, audit logs, custom data retention policies, and explicit exclusion from model training.

Create dedicated working folders. Do not grant Claude access to your entire file system. Create isolated folders for Cowork tasks and move only the files you need for a specific task into that folder.

Assume prompt injection is possible. Do not upload documents from untrusted sources. If you download a file from the internet or receive one from an opposing party, do not feed it directly to Cowork without review.

Verify your data retention settings. Even on Enterprise plans, confirm with your administrator exactly how long data is retained and where it resides.

Understand the audit gap. Cowork stores conversation history locally on users’ computers. This data is not subject to Anthropic’s standard retention policies, cannot be centrally managed by administrators, and is not captured in audit logs or compliance exports. For regulated industries, this is a significant governance gap.

The legal plugin may be impressive, but the infrastructure around it is still a research preview with known security issues. Proceed accordingly.

Looking Ahead

Legal tech incumbents still hold significant advantages. They have deep proprietary datasets, years of domain expertise, and established customer relationships. But they now face a competitor that can iterate quickly, ship open-source tools, and leverage direct relationships with enterprise buyers across multiple departments.

The question is whether Anthropic and other foundation model companies will continue expanding into vertical workflows, or whether this plugin remains a limited experiment.

For legal tech vendors, the pressure is on to demonstrate value beyond what a foundation model can deliver out of the box.

Read Bob Ambrogi’s full analysis for the complete picture.

This post draws on reporting by Bob Ambrogi at LawNext. Thanks to Ernie Svenson for surfacing the article.

Last month, a Canadian litigator told me she’d stopped using ChatGPT entirely. “I don’t trust it,” she said. “It makes things up.”

She’s half right.

ChatGPT will make things up. It will confidently cite cases that don’t exist. It will quote statutes that were repealed in 2019. It will tell you the CRA deadline is March 31st when it’s actually April 30th.

But here’s what she’s missing: configured correctly, ChatGPT becomes a different tool entirely. One that cites sources. One that admits uncertainty. One that works within your documents rather than inventing facts outside them.

This guide shows you how to set it up, and more importantly, how to use it without putting your licence at risk.

The Governing Principle

Let’s get this out of the way:

You may delegate tasks, but you cannot delegate responsibility.

Your Law Society, CPA body, or securities regulator doesn’t care that “the AI did it.” Neither will your E&O insurer. Neither will the judge.

Every citation must be verified. Every factual claim must be checked. Every piece of advice must be yours.

That’s not a limitation of AI. That’s been true since articling students and junior accountants existed. The difference is that AI can do certain things faster, and fail in ways that are harder to catch.

This guide is about catching those failures before they become your problem.

Part 1: Configure ChatGPT Before You Touch a Client File

Don’t skip this. Five minutes of setup prevents months of regret.

Step 1: Upgrade to a Paid Version

Free ChatGPT is not appropriate for professional use. Upgrade to ChatGPT Plus, Team, or Enterprise before doing any client work.

The paid versions offer stronger privacy controls, better data handling terms, and access to the features that make professional use defensible. ChatGPT Team and Enterprise provide additional safeguards, including the assurance that your data won’t be used for training by default.

If your firm handles sensitive client information, use Team or Enterprise. The cost is trivial compared to the liability.

Step 2: Stop Training the Model on Your Work

Go to Settings → Data Controls and toggle off “Improve the model for everyone.”

This prevents your conversations from being used to train future versions of ChatGPT. It does not make your chats invisible. They’re still retained for operational and legal reasons, and anyone with your account credentials can see them.

But it’s a baseline.

Step 3: Review Your Shared Links

ChatGPT lets you share conversations via links. These are convenient and dangerous.

They can be forwarded. They can be accessed by anyone with the URL. They exist outside your control.

Rule: Do not use shared links for privileged or confidential material. Ever.

Go to Settings → Data Controls → Shared Links and revoke anything you wouldn’t want opposing counsel to see.

Step 4: Use Temporary Chat for Sensitive Matters

Temporary chats don’t appear in your history, don’t use or create memories, and aren’t used for model training.

Use Temporary Chat for anything you wouldn’t want discoverable in your chat history.

That’s most client work.

Step 5: Enable Search (and Understand the Tradeoff)

Search is the single most important feature for professional use. It forces the model to cite sources, gives you links to verify, and reduces hallucinations dramatically.

Keep Search on by default for anything outside your provided documents.

But here’s the catch: when Search is enabled, ChatGPT rewrites your query and sends it to search providers. If Memory is on, remembered details can be incorporated into that query.

Translation: if you ask “What are the tax implications for my client John Smith who received a $2.3M severance from ABC Corp?”, some version of that query may travel to Bing.

So: always strip identifying details from your Search queries. More on this below.

Step 6: Install Custom Instructions That Enforce Discipline

Go to Settings → Personalization → Custom Instructions.

What would you like ChatGPT to know about you?

I am a regulated professional in Canada. I use ChatGPT for drafting, organization, summarization, issue spotting, and initial analysis. I do not treat it as a source of truth.

How would you like ChatGPT to respond?

1. Use plain English unless I ask otherwise.

2. Never invent laws, regulations, cases, standards, tax rules, accounting treatment, policy language, quotations, numbers, or citations.

3. If you are uncertain, say so explicitly, and list what must be verified.

4. If I provide documents or excerpts, use only those materials and say when the record is silent.

5. End every substantive response with a Verification Checklist identifying items that require independent verification.

These instructions won’t prevent all hallucinations. But they shift the default from “confident assistant” to “cautious assistant who flags uncertainty.”

That’s the assistant you want.

Part 2: The Two Failure Modes That Damage Careers

Failure Mode 1: Hallucinations

The model doesn’t know what it doesn’t know. When it lacks information, it fills gaps with fiction that sounds plausible. Citations to cases that don’t exist. Statutes with fabricated section numbers. Tax deadlines that are simply wrong.

The more confident the output sounds, the more dangerous it is.

Your defence:

Constrain the model to your documents. Require citations to provided materials. Use Search for anything outside your record. Verify every citation, quotation, and factual claim.

Failure Mode 2: Context Limits

ChatGPT can only hold so much information at once. As conversations grow, earlier material gets pushed out. Even uploaded documents can fall outside active context.

Your defence:

Work in slices. Summarize incrementally. Build chronologies and analyses in stages, confirming accuracy at each step. Don’t assume it “remembers” the lengthy agreement you uploaded an hour ago.

Part 3: The Two Lane Workflow

This is the core of defensible AI use.

Lane A: Confidential Work Within Your Documents

Goal: Use only the documents you provide. No external research.

Use cases: Summarizing a client file. Drafting based on provided facts. Converting narrative into a checklist, matrix, or chronology. Rewriting and polishing internal drafts.

Prompt rule: “Use only what I paste or upload. Do not speculate beyond the record. If the record is silent, say so.”

Lane B: Research Requiring External Verification

Goal: Use Search to verify current law, regulations, standards, or market facts.

Use cases: Confirming CRA filing deadlines. Checking current OSFI guidance. Verifying provincial limitation periods. Finding recent case law developments.

Prompt rule: “Use Search. Provide sources. Prefer primary sources. Flag uncertainty.”

Critical: Strip identifying details from all queries. Remove client names, employers, SINs, policy numbers, and unique fact patterns. Use placeholders like “Client A,” “the insured,” or “the taxpayer.”

The Two Chat Workflow: When You Need Both

This is your safest approach when confidential analysis requires external verification.

1. Confidential Chat (no Search): Do your work with real client facts, constrained to your documents.

2. Research Chat (Search enabled): Ask questions without client details to verify law, standards, and current guidance.

3. Merge: Bring the verified findings back into your confidential analysis, without copying client identifiers into the research chat.

This keeps Search enabled for verification while protecting confidential information from leaving your control.

Part 4: Prompts That Actually Work

The structure that produces usable work product:

1. Role: What perspective should the model adopt?

2. Objective: What outcome do you want, in one sentence?

3. Inputs: What sources may it use?

4. Constraints: No speculation. No invented authority. Flag gaps.

5. Output format: Table, outline, email draft, chronology.

6. Verification requirement: End with items to verify.

Prompt 1: Defensible Summary

You are my professional assistant. Using only the documents I uploaded, produce a structured summary with headings for Issues, Key Facts, and Open Questions. Cite each material statement to document name and page. Do not speculate beyond the record. End with a verification checklist.

Prompt 2: Chronology Builder

Based solely on the attached documents, prepare a dated chronology of events. Cite each entry to document and page. Do not speculate beyond the record. If a date is unclear, flag it.

Prompt 3: Correspondence Draft (Without Inventing Substance)

Draft an email to the client explaining [the issue] in plain language, proposing next steps, and identifying information I’ll need from them. Use only the facts below. Do not add facts. Provide a short internal note listing risks and assumptions.

Prompt 4: Research Query Without Client Details

Use Search. Provide sources. I need the current rule under [statute/regulation] regarding [issue]. Summarize the rule, exceptions, and any recent amendments. Cite primary sources. If sources conflict, explain the conflict.

Part 5: The Verification Checklist You Must Use Every Time

Post this beside your screen. Use it before finalizing anything.

1. Authority Check Every statute, regulation, case, standard, or policy citation verified against the primary source?

2. Fact Check Every number, date, threshold, and factual claim verified against the record or a reviewed external source?

3. Scope Check Output stayed inside the provided record when it was supposed to?

4. Confidentiality Check No client identifiers placed into Search queries or shared links?

5. Completeness Check Identified missing facts that would change the analysis? Asked followup questions?

6. Professional Judgment Check This is your work product, reflecting your judgment and your professional obligations

Quick Setup Checklist

Copy this. Use it today.

☐ Upgrade to ChatGPT Plus, Team, or Enterprise

☐ Turn off “Improve the model for everyone”

☐ Review Shared Links and revoke anything sensitive

☐ Use Temporary Chat for confidential matters

☐ Enable Search by default

☐ Strip identifying details from all Search queries

☐ Install Custom Instructions that forbid invention and require verification

☐ Work in slices and summarize incrementally

☐ Verify every citation before it leaves your hands

The Bottom Line

ChatGPT is not your associate. It’s not your articling student. It’s not even your paralegal.

It’s a tool that can draft faster than you can type, spot issues you might miss, and organize chaos into structure. But it will also lie to you with complete confidence.

Your job is to use it where it excels and catch it where it fails.

Configure it right. Constrain it to your documents. Verify everything.

And remember: you can delegate the task. You cannot delegate the blame.

Have questions about using AI in your practice? Reply to this email or drop a comment below.

Welcome to Tech Tools for Canadian Professionals, your practical guide to using modern technology tools without the hype, the overwhelm, or the risk.

I’m Nicole Garton, President & General Counsel at Heritage Trust, and I’ve spent the past few years watching Canadian professionals split into two camps: those treating new tech tools as a dangerous curiosity, and those using them as a configurable competence engine.

Here’s what I believe: Technology will not replace professionals. But professionals who understand these tools will outpace those who don’t.

The Real Problem Isn’t Technology. It’s Standing Still

Anthropic co-founder Ben Mann said something in late 2025 that stuck with me: “This is as normal as it’s going to be. It’s going to be much weirder very soon.”

He’s right. And for Canadian professionals, whether you work in estates, tax, law, accounting, or financial planning, that shift is already here.

Your hesitation makes sense. Accuracy matters. Confidentiality is non-negotiable. Client trust is everything. These aren’t abstract values. They’re professional duties with real consequences.

But here’s the shift: Those risks aren’t reasons to avoid these tools. They’re your checklist for using them responsibly.

Because avoiding engagement with modern technology is increasingly becoming its own professional risk.

What This Newsletter Covers

I will focus on practical technology tools that Canadian professionals can use immediately:

• AI assistants like ChatGPT, Claude, and Google Gemini for drafting, research, and document analysis

• Research tools like Perplexity and NotebookLM that provide citations you can verify

• Document workflow tools from PDF review to meeting transcription to presentation creation

• Automation platforms that handle repetitive tasks across your existing systems

• Enterprise solutions for firms ready to implement technology at scale

All with a focus on the controls you need to use these tools responsibly within Canadian regulatory rules and professional standards.

The Three Levels of Technology Competency

I’ve organized tech skills into three practical levels. Most Canadian professionals should aim to master “Beginner” competencies first, then build toward “Intermediate” as comfort grows.

Beginner User

What you can do:

• Draft emails, reports, and client communications in consistent voice

• Analyze documents and extract key terms, obligations, and action items

• Pull current information with citations for research briefs

• Generate presentation materials and simple visuals

• Transcribe and summarize meetings

• Polish writing with grammar and clarity tools

Tools to start with:

• One AI assistant (ChatGPT or Claude)

• Perplexity for research

• NotebookLM for document analysis

• Grammarly for writing polish

Intermediate User

What you can do:

• Build custom assistants for recurring workflows

• Create reusable prompt libraries and templates

• Conduct sophisticated research with source tracking

• Automate document generation from templates

• Use collaboration tools for real-time co-drafting

• Design client-friendly handouts and presentations

Tools to add:

• ChatGPT Projects or Claude Projects

• Canva or Craft for document creation

• Gamma or Beautiful.ai for presentations

• Workflow automation basics

Advanced User

What you can do:

• Build interactive tools and dashboards

• Deploy multi-step automated workflows

• Connect tools across email, documents, and CRM systems

• Implement governance controls for firm-wide use

• Create custom applications without traditional coding

• Manage risk systematically across all technology use

Tools to explore:

• Zapier or Make for automation

• Enterprise platforms (Microsoft Copilot, Google Workspace)

• Custom application builders

• Advanced governance and audit tools

Your First Week Action Plan

Don’t try to learn everything at once. Start small:

Pick one tool. ChatGPT or Claude. Free tier is fine.

Try one simple task. Summarize an email. Rewrite a paragraph. Extract key points from a document.

Set up privacy controls. Review data settings. Disable training on your content if available.

Create one rule. “I will anonymize all client information before uploading anything.”

Save one reusable prompt. Pick a task you do weekly and write a structured prompt for it.

That’s it. Build from there.

The Risk Control Reality

Let me be direct: concern about accuracy and confidentiality is completely rational.

But these aren’t reasons to avoid technology. They’re the framework for using it safely.

The critical risks:

Accuracy: These tools can generate plausible but completely false information, including fake case citations and incorrect legal interpretations.

Client Confidentiality: Most consumer tools may use your inputs for training or store them indefinitely.

Professional Responsibility: You remain fully liable for all work product. Technology doesn’t change that.

Transparency: Outputs can lack clear sourcing, making it difficult to verify or defend your work.

Your baseline controls:

• Never paste confidential client information into consumer tools

• Anonymize everything (remove names, addresses, account numbers, unique identifiers)

• Verify every legal citation against authoritative sources

• Treat all outputs as first drafts requiring human review

• Document your verification process for high-stakes work

These controls aren’t optional. They’re how you use technology competently and ethically.

What to Expect

This newsletter will share practical guidance on technology tools for Canadian professionals:

• Step-by-step tutorials on specific tools and features

• Canadian-specific guidance on privacy and professional responsibility

• Real workflows with anonymized examples

• Risk mitigation checklists and verification protocols

• Enterprise tool reviews for firms ready to scale

No hype. No predictions about the future. Just tools you can use today with controls you can actually implement.

One Last Thing

Technology is evolving rapidly. What’s impressive today will be standard in six months and outdated in a year.

The professionals who succeed won’t be the ones with the most technical knowledge. They’ll be the ones who:

• Start experimenting now with low-risk tasks

• Build verification habits into their workflows

• Stay informed without getting overwhelmed

• Implement practical controls they can maintain

This newsletter exists to help you do exactly that.

Welcome aboard.

—Nicole

Detailed resource: aitoolscanada.ca

Have a question or topic you’d like covered? Reply to this email—I read every response.