In April, Deloitte US quietly announced changes to its employee benefits that will take effect on 1 January 2027.

The changes for that group are severe: paid parental leave cut from 16 weeks to 8, annual paid time off reduced by 5 to 10 days depending on tenure, pension accruals ending after 2026, and the elimination of a $50,000 reimbursement for adoption, surrogacy, and IVF.

The cuts don’t apply to everyone in Deloitte’s 181,000-person US workforce. They apply to one specific group.

Two months earlier, Deloitte had overhauled its entire job architecture. The old titles (analyst, consultant, manager, and everything above them) were replaced with a new four-part categorisation: Center, Core, Project, and Domain. The Center employees are the ones taking the benefits cut. Center means internal support roles: administration, IT, and finance.

The Big 4 firm just did something most companies will do more quietly over the next three years. It made explicit which of its employees it intends to keep investing in, and which it doesn’t.

What Deloitte said, versus what Deloitte signalled

Deloitte’s public framing is straightforward. It’s a benefits review during economic uncertainty, aligned with a new job architecture designed for the AI era. Business Insider and Forbes both broke the story with sourced internal documents. Nothing about the announcement was leaked. The changes were communicated internally as part of a broader restructuring.

The public framing is technically accurate. It’s also incomplete.

Two things are happening at Deloitte at once.

The firm is investing more in the people whose skills are becoming more valuable because of AI. Consultants who can use AI to accelerate client work. Project leads with judgment AI can’t replicate. Domain experts whose knowledge becomes more valuable when AI does the drafting.

At the same time, it’s pulling support out from under the people whose roles AI is starting to compress. Administration, IT support, and finance operations. The Center employees. The ones who used to have long careers at the firm precisely because their work was internal, stable, and hard to automate.

Nobody at Deloitte is going to frame it this way in a press release. But look at the mechanics. The work is getting automated, the benefits are getting cut, and the titles that used to mark a career path are gone.

That’s what it looks like when a company starts sorting its own people by whether AI made them more valuable or less necessary.

This is happening at every company. Just less visibly.

Deloitte’s move is unusually explicit. Most companies won’t publish a new four-category workforce architecture and cut benefits for one category in the same quarter. They’ll do it quietly, through selective hiring freezes, non-replacement of departing employees, quiet erosion of promotion tracks in “internal” functions, and gradual redirection of L&D budget toward the categories the company wants to grow.

If you work in finance right now and are trying to understand your own position, the Deloitte announcement is a useful mirror. Not because your company is planning identical benefit cuts, but because the same underlying sorting is likely already happening.

The question isn’t whether the sorting is real. It is. The question is which side of it you’re on.

Two profiles of finance professional right now

Roughly, there are two archetypes of finance professional emerging, and they’re doing very different things day-to-day.

The finance person AI makes more valuable spends most of their time on judgment, business context, and communication. They understand the operational side of the business well enough to catch when the numbers don’t match what’s actually happening on the floor. They can defend a forecast to a sceptical board with a mix of data and instinct. They can explain to a non-finance executive why a proposed decision looks fine on the P&L but ignores something structural. AI accelerates all of this work, but the judgment underneath is theirs.

The finance person AI makes less necessary spends most of their time on production. Producing reports that go into a folder. Producing reconciliations that nobody questions. Producing schedules that AI can now generate in a fraction of the time. The output is real work, and it’s necessary work, but the productive value has been dropping for two years and will keep dropping.

Neither profile is bad at their job. Both are technically competent. The difference is what percentage of the day is spent on judgment work versus production work. And that percentage is the single strongest predictor of whether AI is going to make someone more valuable or less necessary over the next three years.

What to do about it

The Deloitte announcement is a useful test. Read the description of what Center employees do at Deloitte, and honestly assess how much of your current week looks like that description.

If it’s a lot, you have time to change the composition of your work, but not unlimited time. Not by learning AI as a topic. By reallocating the hours in your week from production toward judgment. Fewer reports you produce yourself, more reports you review and challenge. Fewer reconciliations you complete manually, more strategic conversations with the head of operations or the CEO about what those reconciliations reveal.

This connects to what we covered in why finance is becoming the skill that makes AI useful and the four things CFOs who become CEOs do differently. Same pattern from a different angle. The finance professionals AI is making more valuable are the ones who look like the CFOs of five years from now. The ones AI is making less necessary are the ones who look like the transaction processors of the past decade.

Deloitte just made the split visible. Every finance function will make the same split, quietly, over the next 24 months.

The uncomfortable question is which side of it you’ll be on. The more uncomfortable one is whether you’re doing anything about it this week.

See you next Tuesday.

Adam

Financial controls, accounting, and audit ranked first. Project management ranked second. Financial planning, analysis, and modeling ranked third. Program management office leadership and process optimisation rounded out the top five.

AI didn’t appear anywhere in the top 15.

That’s a surprising result at first read. In a year when every conference keynote, every board meeting, and every trade publication is talking about generative AI as the transformation of the century, the world’s largest executive search firm is saying that companies aren’t primarily hiring for AI skills. They’re hiring for finance skills.

The explanation is on the second page of the report, but it’s easy to miss. AI is no longer being treated as a standalone skill. It’s an expectation built into every function. Companies have stopped asking “who understands AI?” They’re asking “who can use it to make our numbers sharper, our data cleaner, and our operations tighter?”

The answer keeps coming back to finance people.

Why finance is the skill

Three things make finance the right function for this moment.

Finance has the discipline of measurement. AI implementations that fail (which is most of them, as we’ve covered before) fail because nobody defined success, nobody tracked the numbers, and nobody could tell you three months in whether the tool was actually saving what the vendor promised. Finance teams are the ones in the building who know how to define, track, and defend a number over time. That skill, which was table stakes for accountants, has become the missing ingredient in AI transformation.

Finance has the control framework. AI, deployed properly, needs governance. Who approves what, what’s automatic, what still requires human review, what triggers an escalation. The frameworks for this already exist inside finance functions, in the form of internal controls over financial reporting. Nobody else in the company thinks this way natively. HR doesn’t. Sales doesn’t. Operations doesn’t. Finance does.

Finance has the business context. The best AI implementations happen where someone understands both the data underneath and the business decisions the data supports. A Controller with fifteen years of experience understands margins, working capital, revenue recognition, cost drivers, and the specific ways the company makes and loses money. Combine that with AI capability and you have someone who can drive real transformation. Give the AI capability to someone without that context and you have a tool that produces impressive dashboards nobody uses.

The interim CFO signal

The Heidrick report includes another data point that reinforces the argument.

Interim CFO roles now account for 51% of all interim C-suite requests. More than any other function. When companies hit uncertainty, leadership transitions, or transformation, the first call is to a finance leader.

That’s a market signal about who companies trust to steer through complexity. Not the AI expert brought in to advise. Not the consultant with the roadmap. The CFO or interim CFO who can hold the numbers steady while everything else changes.

For any Controller or Finance Director reading this, that’s the seat you’re aiming for. And the market is telling you clearly that the demand is at its highest in years.

What this means for your career

If you’re a Controller thinking about the next five years, the Heidrick data suggests two things worth acting on now.

First, resist the temptation to become an “AI person” who happens to work in finance. The market doesn’t need more AI generalists. It needs finance people who use AI well. Your value is the finance depth. Your differentiator is being able to apply AI on top of that depth, not the other way around.

Second, invest in the categories the Heidrick index calls out specifically: advanced analytics, data management, and process optimisation. These are what boards are looking for when they promote from finance. And they’re all skills you can build inside your current role, without needing permission or a training budget.

The bigger point

AI created the urgency. But the people companies need most to act on it are the ones who understand the data, the controls, and the business well enough to make AI actually work.

Finance is becoming the skill that makes everything else useful.

This connects to what we covered in why AI is becoming the new baseline in finance hiring and the four things CFOs who become CEOs do differently. Same pattern from three different angles. The Controllers and Finance Directors who position themselves inside this shift now will be the CFOs and CEOs the trade press writes about three years from now.

See you next Tuesday.

Adam

If you’ve been reading the CFO trade press this year, this is starting to look like a pattern. Because it is one.

According to the Crist Kolder 2025 Volatility Report, 10.26% of sitting CEOs at Fortune 500 and S&P 500 companies in 2025 came directly from the CFO role. That’s the highest level of CFO-to-CEO promotions in a decade, up from 7.1% in 2024. Every single one of those promotions last year was internal. The board didn’t need to go outside. The CFO in the room was already the answer.

Boards are looking at finance chiefs differently than they were even three years ago. And the CFOs getting the call share a set of behaviours that separates them from the ones who stay in the seat forever.

Four patterns show up over and over. If you’re a Controller reading this now, they’re worth studying before you get the CFO title, not after.

They walk out of rooms that went the wrong direction

Most CFOs sit through a bad decision and try to fix it in the follow-up email.

The ones who become CEO interrupt the meeting. They say “we’re about to make a mistake” while everyone else in the room is nodding along. That takes conviction, not consensus. It also takes the willingness to be wrong in public if the room ends up disagreeing with them.

Boards remember the CFO who stopped a bad decision in real time, especially when nobody else in the room wanted to hear it. Because that behaviour is exactly what a board needs from a CEO. Boards don’t promote the CFO who reports what happened after the fact. They promote the CFO who prevented what shouldn’t have happened at all.

They know the business better than the operators

Most CFOs can explain what happened last quarter.

The ones who become CEO can tell you why it happened, what’s about to break, and which single assumption the entire three-year plan collapses on if it turns out to be wrong. They’ve walked the floor. They’ve sat with customers. They understand margin pressure as a story about operations, pricing, and competitive positioning, not just as a number on a variance report.

Motorola Solutions’ CFO Jason Winkler put this simply in an interview earlier this year: “finance is not a spectator sport.” The CFOs who become CEO approach the function with an operations mindset, fixing problems rather than reporting on them. They understand the business the same way a good COO or CCO does, but with the numbers to back it up.

Boards promote the CFO who sees around corners, not the one who reads the scorecard best.

They make their own seat easier to fill

This is the one that most CFOs get wrong. And it’s the one that separates them the most clearly.

Most CFOs protect their position by being irreplaceable. They hold the key relationships, run the critical processes, control the reporting no one else can produce. Being indispensable feels safe. It also guarantees you never leave the chair, because the company literally cannot afford to move you.

The CFOs who become CEO do the opposite. They already have a successor ready before anyone asks. They hand off high-visibility work to the deputy who’s ready for it. They document their thinking so someone else can defend a decision six months from now without needing to call them. And they make sure the audit committee, the board, and the CEO know exactly who could sit in the CFO chair if the current one moved up.

It feels counterintuitive. Replaceability sounds like the opposite of a promotion strategy. But boards will never move someone up if the company can’t survive without them in the current chair. The Deloitte 2026 CFO readiness study found this pattern explicitly: leaders whose calendars are filled with tasks only they can perform may be signalling a lack of readiness, not the opposite.

Replaceability is a promotion strategy. Adopt it early.

They make the people around them better

Every forecast that passed through them came out sharper. Every meeting they led forced clearer thinking. Every team they touched started operating at a higher standard without being asked to.

This one is harder to measure and easier to feel. But boards notice. When the CEO asks the CFO about performance in an area, and the CFO’s team can defend the numbers with the same clarity the CFO would, that’s data. When the sales lead says “our forecasting got tighter after finance stepped in,” that’s data. When the head of ops mentions that the CFO helped her think through a supplier decision that finance didn’t need to be in, that’s data.

Boards are looking for the CFO who raised the performance of everything around them, not just their own.

What this means if you’re a Controller

The four patterns are worth studying now, not when you get the CFO seat, because they take years to build. Interrupting a bad decision requires reputation you earn over time. Knowing the business better than the operators requires walking the floors and sitting with customers, which nobody schedules for you. Making yourself replaceable requires deliberately handing off work that you enjoy doing. Making the people around you better requires a specific kind of investment in your team that isn’t in your job description.

None of these show up in the annual review. All of them show up in the promotion conversation.

The Controllers who eventually become CFOs, and then CEOs, are the ones who started doing all four while they were still in the Controller chair. As we covered in Why most Controllers never become CFOs, the trap is being too good at the job the company doesn’t want to change. The four patterns above are exactly how you break that trap.

CFO-to-CEO promotions just hit a decade high. The traits above are not a coincidence.

See you next Tuesday.

Adam

On 16 July 2026, the UK’s Financial Reporting Council fined PwC $4.4 million for serious failings in the audits of Babcock International, the FTSE 100 defence contractor.

The original fine was $7.4 million, reduced by 40% because of PwC’s cooperation.

This is the same firm that has invested close to $1 billion in an AI-native audit platform with Microsoft. The entire pitch of that platform is that AI catches what humans miss. It examines every transaction rather than a sample. It flags patterns humans overlook. And it is intended to be the standard way PwC audits every client by 2028.

The Babcock audit was not complex. Cash pooling arrangements were presented on a net basis instead of a gross one. Aircraft-related costs were misclassified. Goodwill impairment was not adequately challenged. Long-term contract accounting was not tested against the underlying evidence. The FRC’s own words: PwC failed to critically challenge management’s accounting choices.

Basic things that should have been caught. Weren’t.

This is not the first time

In March 2023, the FRC fined PwC close to $8 million for failures in Babcock’s 2017 and 2018 audits. Same client. Same category of failures: professional skepticism, challenge of management, adequate audit evidence.

Between the 2023 fine and the 2026 one, PwC changed the engagement partner, published statements about strengthened audit quality, and continued investing in the AI platform that was supposed to make this kind of failure impossible.

None of it worked. The next audit year failed for the same reasons.

That is not a technology problem. That is a governance and judgment problem, wearing an AI budget as a distraction.

What AI can and cannot do in the audit room

The AI in PwC’s platform is good at things humans are bad at. Processing every transaction. Matching patterns across millions of records. Flagging anomalies that a spot check would never surface. All of that is real, and the $1 billion investment probably does deliver on those specific capabilities.

The AI is not good at what actually failed in Babcock.

Professional skepticism is the willingness to ask why the CFO’s team has explained something a particular way. It is the mental muscle that says “this looks fine on the numbers, but the explanation doesn’t hold together.” It is the decision to escalate an uncomfortable observation rather than let it slide. And it is the moment when the auditor asks a follow-up question that the client’s team was hoping nobody would ask.

None of that is a data problem. It is a human problem, requiring courage, judgment, and the discipline to keep asking difficult questions when everyone in the room wants the audit to be finished.

AI processes what it is given and matches it against what it expects to find. It does not ask why the data looks like this. It does not ask what has been omitted. It does not notice when the client is quietly steering the conversation away from something. And it does not have the standing or the reputation on the line to press the CFO on a matter of judgment.

Those are still exclusively human capabilities, no matter how much money PwC or anyone else invests in the technology layer.

What this means for Controllers and Finance Directors

The lesson is not that AI is useless in the audit function. It clearly isn’t. The lesson is where the responsibility remains yours.

Three things AI cannot replace in a well-run finance function.

The sense that something is off. When you know your business well enough to feel that a number is technically correct but tells a story that isn’t quite right, that instinct comes from years of context, not from data. AI has none of that context. It processes what you feed it. If the underlying process is opaque, the AI will confidently produce output that is defensible on paper and wrong in substance.

The decision to escalate before the auditor asks. The Controllers who are trusted over decades are the ones who bring issues to the CFO or the audit committee themselves, before someone external raises them. That decision is uncomfortable. It requires you to admit that a process broke on your watch. AI cannot make that call for you. And if you rely on the AI to catch things before you do, you will find out later that the AI missed the thing you should have raised months earlier.

The difficult conversation with the CEO. When something needs to change (a revenue recognition treatment, a bonus accrual, a related party disclosure), the person who has to sit down with the CEO and explain why is still human. AI can build the model that shows what needs to change. It cannot have the conversation that gets the change accepted. That skill, the political capital and the courage to spend it, is what separates Controllers who last from those who don’t.

Why this connects to everything else in the newsletter

This connects directly to the four things Controllers must settle before AI touches AP and to why 80% of AI-in-finance pilots fail.

Every time AI fails in finance, the story looks like a technology problem and turns out to be a process or judgment problem. Every time.

PwC did not fail the Babcock audit because their AI wasn’t good enough. They failed it because the humans in the room did not challenge management with the professional skepticism the standards require. A $1 billion AI platform does not fix that. It just runs faster over the gap.

The next audit failure at a Big Four firm will not be a technology failure either. It will be another human failure, discovered faster and at greater scale. The AI will process every transaction and produce a beautiful report. And somewhere in the middle of it, someone will still fail to ask the one question that mattered.

That question is what the audit standard actually cares about. And it is still yours to ask.

See you next Tuesday.

Adam

Sarah Friar, CFO of OpenAI, published a framework last month for how CFOs should measure whether AI is worth the money.

You can read the whole thing but the basic question she asks is one sentence: does the value of the work AI completes grow faster than the cost of producing it?

I’ve been asking myself something adjacent lately. Not just “how do you measure AI ROI,” but why does the CFO of OpenAI feel the need to publish a beginner-level scorecard on this in mid-2026?

The honest answer is that most finance teams still don’t know what their AI spend is producing. Friar didn’t publish four steps because they’re clever. She published them because the market has been running billions through AI budgets without a baseline for measurement. The scorecard is the response to that gap.

Which puts every CFO reading it in an uncomfortable position. Because to use Friar’s framework, you have to answer four questions honestly. And some CFOs I speak to can answer none of them.

Why measuring AI ROI is still broken

Before the framework matters, the diagnostic matters.

The 84% adoption to 7% impact gap we covered in a recent issue isn’t only because CFOs are deploying agentic AI on Stage 2 foundations. It’s also because they cannot report on what “working” looks like for their AI spend.

Three things are usually broken.

The finance team never wrote down what a completed AI task actually looks like. They deployed the tool for “invoice processing” or “reconciliation” or “reporting drafts” without defining the specific outcome that would count as success. So when the team is asked “is it working?”, the honest answer is a shrug.

The full cost of AI isn’t in one line item. It’s tokens, compute, human review time, rework when the model gets it wrong, the internal team supporting the vendor, and the operational disruption during rollout. Most CFOs are tracking the licence fee and calling that the cost. That number is off by a factor of two to five.

The dependability metric doesn’t exist. Nobody is tracking how often the AI output is usable without human intervention. If your team is editing every output, the tool is a draft machine with a salary attached. But nobody flags this because it’s invisible in the reporting.

Three broken measurements. And this is before we get to the changes-over-time problem, which is where most AI investments quietly turn into legacy spend.

The four questions from Friar’s scorecard

Here they are, translated into the ones I would ask any finance team defending their AI budget in the next quarterly review.

1. Define what “done” looks like on this specific workflow. In numbers, with a date.

Not “AI for accounts payable.” That’s a category. Try: “By 30 September, the AP team is capturing invoice data from vendor 1 through 45 with less than 3% exception rate, and processing time per invoice is under 90 seconds average.”

If you can’t write a sentence like that for the workflow, the AI on that workflow isn’t measurable. Which means it isn’t defensible, and it will be the first thing your CFO cuts when budget pressure comes.

2. Calculate the real cost. All of it.

The licence fee, tokens, compute, integration costs, human review time (measured in headcount hours), rework when the AI gets something wrong. And the cost of the internal team’s attention during rollout, which is real even if it doesn’t hit an invoice.

Most CFOs I speak to are underestimating by two to three times. The uncomfortable moment is when they add it up honestly and realise the “savings” from AI don’t cover the fully loaded cost. That’s not a reason to stop. It’s a reason to know.

3. Track dependability, meaning percentage of outputs the team uses without editing.

If the team is editing every result, you don’t have an AI tool. You have a first-draft generator that also happens to cost twenty thousand pounds a year.

The number worth tracking is: of 100 outputs the AI produces, how many go straight through to a completed workflow without a human touching them? If the answer is below 60%, the tool isn’t ready for the workflow. If it’s below 30%, someone in finance is being paid to babysit software.

4. Keep tracking, because everything changes underneath you.

The models get updated, the workflows evolve, the team learns and starts expecting more, and the vendor changes their pricing. What worked in Q2 might be silently underperforming by Q4.

Most finance functions measure ROI once at implementation and never again. That’s how you end up paying for a tool nobody uses six quarters later.

What most CFOs will actually do

The uncomfortable read of Friar’s scorecard is that a lot of AI budgets in finance are surviving right now because nobody is asking these four questions. The tool is on the balance sheet. The team says it’s helping. The vendor sends impressive dashboards. Nobody has done the honest math.

The CFOs who take Friar’s framework seriously will do something disruptive. They will audit their existing AI spend against these four questions. In a lot of cases, the answer will be: we bought this without knowing if it works, and we’re not measuring it now.

That’s the uncomfortable moment. It’s also the moment the finance function starts to lead on AI instead of watching it happen.

The tool: a one-page audit for your next quarterly review

Before your next quarterly finance review, take one hour and answer the four questions honestly for each AI tool your function pays for. One page. One tool per page.

For each tool, write down what “done” looks like on the specific workflow, in numbers and with a date. Calculate the real fully-loaded cost, including human review time and rework. Note the dependability number as a percentage of outputs used without editing. And list what has changed in the last 90 days that means the measurement needs updating.

If you can complete the page for every AI tool in the function, you have measurement. If you can’t complete it for even one, that’s where the CFO’s attention needs to go.

This connects directly to why 80% of AI-in-finance pilots fail and to why CFOs are jumping straight to autonomy. Different symptoms, same root. AI keeps failing in finance not because of the technology, but because of the measurement the finance team never built underneath it. Friar just gave us the four questions. Whether the market answers them honestly is the actual test.

See you next Tuesday.

Adam

The story goes: AI won’t survive the audit. Models are black boxes. Auditors can’t examine what they can’t explain. Therefore AI in accounts payable is a compliance risk waiting to detonate.

The premise is fair, but it points at the wrong black box. AI is not what makes accounts payable un-auditable. Manual AP has been a black box for decades. Nobody could ever answer why a particular invoice was paid on the 14th, for that amount, approved by that person. It arrived in an inbox, sat there, was approved by whoever was at their desk, and cleared by someone using knowledge they had never written down. AI inherits that opacity. It doesn’t create it.

Which means the answer isn’t to reject AI. It’s to build the foundation the AP function should have had all along, in the order Controllers can actually deliver it.

What’s about to change

PwC has spent close to a billion dollars building a next-generation audit platform. The intention, published in trade press, is to audit every client this way by 2028, examining every transaction rather than a sample. Sampling has hidden a lot of messy AP for a long time. Reading the whole population in near real time removes that cover.

Ardent Partners’ 2025 numbers put the average invoice exception rate at 18.4%, average time per invoice at 8.2 days, average cost per invoice at $9.84. Best-run teams do it for roughly four-fifths less. The gap is process, not technology. And AFP’s 2026 survey reports that 76% of US organisations were hit by attempted or actual payment fraud last year, most of it flowing through the same undocumented cracks that let duplicate payments and miscodings through.

The teams that will do well under real-time audit are the ones where the process is written down, testable, and defensible before the AI arrives. That means the Controller has real work to do before the vendor demo.

Here are the four things to settle first, in order.

1. Decision rights

Most Controllers I speak to have never actually mapped decision rights in AP.

The organisation runs on convention. The finance director approves up to a certain amount without board sign-off. The AP lead can release payments under another amount without dual authorisation. The specific individual who handles the top 20 suppliers has quiet authority over exceptions nobody has written down. It works, until someone leaves, or the AI is asked to enforce rules the humans never articulated.

The work: for every category of AP decision, write who has authority, what the limit is, what the escalation trigger is if the limit is exceeded, and what documentation is required. One page. Reviewed by finance director and CFO.

You’ll be surprised how many “obvious” decisions turn out to be conventions with no policy behind them. Those are the decisions the AI will be forced to guess at, and the auditor will flag first.

2. Accountability by name, not by team

The second gap is the one that kills AI in accounts payable in production even when the decision rights are clear.

When something goes wrong (a duplicate payment, a miscoded invoice, a supplier paid twice against the same order), the honest answer to “who owns this” cannot be “finance” or “AP” or “ops.” It has to be an individual whose name goes on the incident.

Most organisations avoid this because naming a specific person for a specific class of failure feels harsh. It isn’t. It’s the difference between a function that learns from its mistakes and a function that repeats them. And when AI is executing decisions inside AP faster than any human can catch, the named owner is the only person with the standing and the motivation to fix the underlying rule when the model gets something wrong.

The work: for each of the top ten failure modes in your AP process (duplicate payments, unmatched invoices, out-of-policy approvals, and so on), name the individual accountable. Not the team. Not the role. The individual.

3. Oversight, escalation, and what actually triggers reporting

This is where most governance projects on accounts payable fail quietly.

Organisations design well-staffed committees that meet weekly. They produce dashboards. They review exceptions. And they never define what actually causes something to surface in the first place.

An AP process that includes AI has to answer three questions in writing. What gets reported, to whom, on what cadence? What specifically triggers an escalation above that cadence? And when the escalation happens, who is on call?

The three questions look basic. They aren’t. Most functions I’ve seen have vague answers to all three, and the ambiguity costs them the audit. The auditor reads “escalations are reviewed monthly by the finance leadership team” and asks: define escalation. Define review. And when there was an issue in March, who did what within 24 hours?

The work: write two paragraphs. First paragraph, the reporting cadence and content. Second paragraph, the escalation triggers, ownership, and on-call rotation. Two paragraphs. This is where most Controllers can produce the most immediate governance improvement with the least effort.

4. Change control

The fourth thing is the one most Controllers underestimate.

Once the AP process is documented, decision rights mapped, and accountability named, the function has a real governance foundation. And the moment anything changes, that foundation starts to erode unless there is a formal way for changes to be evaluated and integrated.

A new supplier onboarded. A new integration to a payments system. A new coding rule requested by the plant in Berlin. Each of these is a change to the AP process. Without change control, people quietly route around governance. The AI, still executing the old rules, starts producing errors nobody can explain.

The work: define what constitutes a “change” to AP process, and define the evaluation route. Who assesses the change, who signs off, and how it gets incorporated into the documented process. Most finance functions have a change control process for the ERP. Almost none have one for the AP operational workflow, which is where AI actually lives.

The example that made it click

A real example from a multi-site business with a large vehicle fleet. Their monthly fuel invoice ran to 30 pages of refuels across dozens of vehicles and cost centres, and used to take two days of manual reconciliation across two disconnected systems.

The audit problem there wasn’t the speed. It was that hundreds of purchase orders sat permanently open in their operational system because closing them was a manual step the team often skipped.

Once the four foundations were in place (decision rights, accountability, escalation, change control), the AI infrastructure did something the team had struggled to do for years. Every order closed automatically. Every invoice posted into the ERP with full accounting dimensions. Not because the AI was smarter. Because the process it inherited was finally written down.

That’s what auditable AP looks like in practice.

Why this connects to everything else in the newsletter

This connects directly to why AI fails when CFOs jump straight to autonomy and to why 80% of AI-in-finance pilots fail. Same pattern.

Every time AI fails in finance, the story looks like a technology problem and turns out to be a process problem. Every time. The vendor is not the enemy. The auditor is not the enemy. The gap between “how the work actually gets done” and “how the work is documented” is the enemy. AI just makes it visible faster.

By 2028, most companies will be audited transaction by transaction. Better to put the process in now than have an auditor find out you never had one.

For readers who want the audit angle developed further, I wrote a longer version of this argument in Accounting Today earlier this month.

See you next Tuesday.

Adam

The CFO reads about agentic AI. The CFO watches a vendor demo of autonomous close, autonomous reconciliations, self-healing reporting. The CFO signs the contract. Six months later, the team is exhausted, the auditor is nervous, and the CFO is quietly writing off another AI initiative.

Then they blame the vendor, or the model, or their team.

The numbers now make this pattern impossible to hide. Gartner surveyed 183 CFOs in June 2025. Eighty-four percent said they had implemented or were planning AI in finance. Only seven percent reported high or very high impact on operations. A separate RGP survey of 200 US finance chiefs put clear measurable impact at 14%. The Journal of Accountancy reported in April 2026 that 60% of finance teams are actively piloting AI, and only 7% report strong workflow results from those pilots.

The 77-point gap between adoption and impact is not evidence that AI fails in finance. It’s evidence that CFOs are jumping over the step that makes AI actually work.

The step CFOs skip

McKinsey’s 2025 AI readiness index puts most finance functions at Stage 2 or Stage 3 of maturity. Agentic AI, the kind vendors are selling right now, requires Stage 4. Teams that score below 30 on the readiness index fail more than 70% of the time.

Stage 4 isn’t a technology tier. It’s a foundation.

Documented workflows that at least 80% of the team actually follows. Data that can be exported clean from the ERP in under five minutes without manual cleanup. Governance policies for what the AI is allowed to touch and what still requires a human. Tested escalation procedures for when the AI gets it wrong.

Most finance teams don’t have any of these. Not because they’re bad. Because nobody has ever asked them to build these things, and there was no visible cost until now.

Then a vendor arrives selling agents. The CFO signs. And the team gets asked to plug a Stage 4 tool into a Stage 2 organisation.

That’s the story of the 77-point gap. Not vendor overpromising, though that happens too. Not model limitations, though those are real. It’s finance organisations trying to deploy the newest layer of AI on foundations that were never built for it.

What the 7% did differently

The CFOs who saw high impact aren’t using better technology. They’re using technology on a better foundation.

The categories where AI is working right now are narrow and specific. Accounts payable automation, at 37% adoption in finance, is producing 60 to 90 day payback for companies processing 500 or more invoices a month. Report generation is at 57% adoption with consistent user satisfaction across mid-market and enterprise teams. Anomaly detection at 34%, with fewer false positives than the old rules-based systems it replaces.

These are the workflows where the vendors have been in production long enough to learn from real data, and where the finance team was already doing the work in a documented way. The gap between the demo and the pilot is small enough to survive.

The workflows where AI is not working yet are the strategic ones. FP&A forecasting. Autonomous close. Self-directed treasury monitoring. These are Stage 4 use cases, and most finance teams aren’t ready for them. The CFOs who deployed them anyway are the ones sitting in the 77-point gap.

The eight questions before the next demo

There’s a version of the McKinsey framework, combined with what Consero Global published in their 2026 CFO report, that fits on one page. Before the next AI vendor conversation, run these questions honestly.

Can your ERP export clean, complete data in under five minutes without manual cleanup?

Are your top three finance workflows documented and followed by at least 80% of your team?

Do you have a written policy on what the AI is allowed to post to the general ledger without human review, and what still requires it?

If the AI makes a mistake at 3am on a Tuesday, is there a documented escalation procedure and someone on call?

Have you defined success in numbers with a date, before signing any contract?

Can the AP lead, the FP&A lead, or the close lead describe the workflow the AI will change without opening the tool?

Do you know what your current baseline is on the metric the AI is supposed to move?

If you dropped the AI tool six months after go-live, what specifically breaks and how long can you operate without it?

The count of no’s is the diagnostic. Six or more, you’re at Stage 2, and no agentic tool is going to work here until the foundation is built. Three to five, you’re at Stage 3, and one narrow pilot might work if scoped carefully. Two or fewer, you’re at Stage 4, and it’s time to make the vendor prove impact.

Most CFOs I speak to answer no to six or seven of the eight. That’s the honest starting point.

What the CFOs closing the gap actually did

The CFOs who are in the 7% and 14% didn’t find better tools. They fixed their data. They documented their processes. They put governance in place before anything ran autonomously. Then they picked one narrow workflow that already had a clean before-and-after metric, deployed AI on it, measured, and only expanded when the numbers actually moved.

Most of that work is free. None of it ships in a vendor contract. It just takes honest answers to the eight questions and someone willing to clear calendar time to deal with what those answers reveal.

This connects directly to why 80% of AI-in-finance pilots fail and to the first AI workflow every Controller should pilot. Same pattern, different angles. Pilots die when the foundation isn’t there. Agents fail when the maturity isn’t there. And in both cases, the CFO who won got there by resisting the temptation to buy their way past the foundation work.

The 84% who adopted are just early. The 7% who saw impact figured out the actual sequence. The 77-point gap is where most of finance is currently sitting, waiting for the vendor to solve a problem that isn’t the vendor’s to solve.

See you next Tuesday.

Adam

It’s the kind of line that could easily read as one more headline about AI. It isn’t. What Friar actually said is closer to a hiring standard, and it’s the first time I’ve seen someone in her seat put it in those exact terms.

Codex isn’t really the point. Excel is the interesting part.

Excel took decades to become a baseline expectation in finance. There was a stretch when it was optional, then a stretch when it was preferred, then a stretch when the CFO stopped asking because the assumption had settled. Nobody sends a job description today saying “must know Excel.” It’s inside the water.

AI is going through the same shift. Just much faster. And most finance teams are still hiring like it’s 2022.

Why this signal matters

You could file Friar’s comment under “another AI hype quote from a tech CFO.” I don’t think that’s the right filter. She wasn’t marketing OpenAI. She was answering a hiring question.

The context around it is worth spelling out. Deloitte’s 2026 Finance Trends report surveyed more than 1,300 finance leaders globally. AI and automation came out as the top skill development priority for the second year running, ahead of traditional finance competencies. That’s not signal from one CFO. That’s the market forming out in the open.

The candidate who knows the numbers cold, closed last quarter clean, understands consolidation across three subsidiaries, and has never opened an AI tool is starting to look like the candidate five years ago who was very good at bookkeeping but couldn’t build a pivot table. Still competent, still capable, but no longer the profile most CFOs are optimising for.

Nobody in that position has been told this yet. Most job descriptions haven’t changed. Most hiring managers are still asking the same questions they asked in 2023. That’s the quiet part.

For hiring managers: the interview loop changed

The defensive move is to add “familiarity with AI tools” to the next job description and move on. That’s what most teams will do. It doesn’t cost anything and it looks modern.

The more useful move is to rewrite the interview loop itself. Not to filter for AI expertise, which most finance candidates don’t have yet. But to filter for the disposition that gets fluent quickly.

What has the candidate tried on their own? What did they get wrong? What did they change about their workflow after they tried it? Those questions surface something different from “have you used Copilot.” And they surface it in a way that predicts adaptability, which is what the next three years are really going to reward.

The Controllers and Finance Directors who are going to still be hiring five years from now are the ones asking questions in interviews that they weren’t asking twelve months ago.

For candidates: one honest project is enough

If you’re a Controller or a senior accountant reading this and thinking about your next move, the signal points somewhere specific.

You don’t need to become an AI expert. You need one honest project in the last twelve months where you tried a tool, got something wrong, adjusted, and can describe what you learned. One project is enough. Two is generous. Zero is where the credibility gap starts.

Most Controllers have zero right now. The ones who move ahead in the next two years will be the ones who fix that gap this quarter, not the ones who wait for a mandatory training program that isn’t coming.

What to do in the next thirty days

If you take one thing away from Friar’s comment, this is what I’d suggest.

Pick one workflow inside your function this month. Not the biggest one. Not the highest impact. Just one you touch weekly and one where you have some data to work with. Spend two hours applying an AI tool to it. Whatever your organisation already has a licence for.

Get something wrong. Fix it. Take notes on what you learned.

That’s the whole exercise. It sounds small because it is. But it’s the exercise that separates the finance professional who talks about AI at conferences from the one who quietly, over the next twelve months, becomes the person the CFO trusts to lead the next thing.

This connects directly to why 80% of AI-in-finance pilots fail and to the three conversations piece from a couple of weeks ago. Teams that don’t understand AI don’t scope AI projects well, and Controllers who don’t understand AI won’t be the ones the CFO asks to lead the next initiative.

Friar drew the line. It’s on the table now. Whether the rest of finance treats it as noise or as signal is up to each Controller reading this.

See you next Tuesday.

Adam

The close update, the variance, the audit prep, the cash position, the vendor selection, the board pack review. All necessary. All transactional. All about the work that has to happen now. None of them about the Controller, what the Controller is becoming, or what the Controller could be trusted with next.

That’s why most Controllers stay Controllers. Not because they aren’t capable. Because the conversations that move a Controller toward CFO never happen unless the Controller initiates them. The CFO is too busy. The annual review isn’t enough. And the Controller who waits for the CFO to bring it up is waiting for a meeting that nobody scheduled.

Three conversations, in the next ninety days, change that calculation. Each takes thirty minutes. None of them lives in the recurring 1:1. Each one signals to the CFO something the Controller cannot communicate any other way.

1. The structural conversation

“If you could rebuild this finance function from scratch with no constraints, what would you change first, second, and third?”

This question does three things at once. It pulls the CFO out of operational mode. It signals that the Controller is interested in the function as a system, not just their corner of it. And the answer tells the Controller exactly where the next year of strategic projects is going to live.

Most CFOs have never been asked this directly. The answer comes out half-formed because they have been thinking about it for months without anyone to think alongside. When the Controller asks the question, listens carefully, and follows up with two or three specific clarifications, the CFO starts to see them differently. Not as the person who runs the close. As the person they can think out loud with.

Follow-up: at the end of the conversation, ask which of the three changes the Controller could help unblock. Then deliver on at least one before the next quarter ends.

2. The development conversation

“What would you trust me to lead in the next twelve months that you don’t trust me with today, and what would you need to see from me to get there?”

Almost no Controller asks this question. They wait for the CFO to volunteer it during the annual review. The annual review is the worst possible context for this conversation, because it’s tied to compensation and the CFO is in performance-evaluation mode, not coaching mode.

The honest version of this conversation, asked outside the review, gets a different answer. The CFO will name something specific. Usually something they have already been thinking about but never said. “I’d like you to own the board reporting cycle, but I don’t see you flexing on cross-functional negotiations yet.” That sentence is gold. The Controller now knows exactly what to demonstrate in the next two quarters.

The Controllers who get promoted have had this conversation. The Controllers who don’t, never have.

Follow-up: ask the CFO to give specific feedback in two weeks on whether they are seeing the change.

3. The business conversation

“What is the CEO worried about right now that finance could help with, and is anyone in our function actually working on it?”

This is the hardest conversation to start because it requires the Controller to step outside their function. But it is the most signal-generating one of the three. It tells the CFO that the Controller is starting to think like a CFO.

The answer is usually surprising. CEOs worry about things that are not in the board pack: sales pipeline that doesn’t smell right, talent that is about to leave, a regional bet that nobody has stress-tested. The CFO knows these things and is rarely able to redirect resources from finance to address them because the function is too busy with the close.

A Controller who picks one of the CEO’s worries and figures out how finance can contribute, without being asked, is a Controller who has stopped being a Controller in the head of the CFO.

Follow-up: take the issue the CFO names, talk to one cross-functional leader about it (head of sales, head of operations), and bring back a short note to the CFO with what you learned. Two pages, ideally.

What these three have in common

They are all questions, not statements. The Controller is asking, not telling. The questions force the CFO to think in front of the Controller. That is more important than any specific answer the CFO gives.

They are also questions that compound. The structural conversation generates project ideas. The development conversation generates a specific skill to build. The business conversation generates a way to demonstrate that skill. Done over a quarter, the three conversations turn the Controller from “the person who runs the close” into “the person the CFO is starting to think about as a successor.”

This is the same shift covered in Why most Controllers never become CFOs. The three habits in that piece (translate the work into business language, build credibility outside finance, commit to a number before the close is final) are reinforced and made visible to the CFO precisely through these three conversations.

The tool: book three walks

The mistake most Controllers make is trying to have these three conversations inside the recurring 1:1. The 1:1 has its own agenda, its own urgency, its own list. The conversations get squeezed and the answers stay shallow.

The Controllers who get this right book the conversations as walking 1:1s. Thirty minutes, outside the office, no slides. One walk per month over the next quarter. The first walk gets the structural conversation. The second gets the development conversation. The third gets the business conversation.

Three walks. Ninety days. Career-changing if you actually do them.

See you next Tuesday.

Adam

The CEO has been thinking about something all day. They want a number. Not the kind of number that comes from the close. The kind that comes from a forecast, from a scenario, from a model that someone in FP&A built, except your company doesn’t have an FP&A team yet, or the FP&A team is one person who’s already at 110%, so the CEO does what every CEO does. They ask the person they trust most in finance.

That’s you. The Controller.

You stay an extra two hours and put something together. You qualify it heavily, because the data isn’t clean and the assumptions are rough. The CEO reads it on their phone at 9pm and uses your number in a conversation the next morning. Now your number is in circulation as if it were a real forecast. And next week, the CEO sends you another Slack.

This happens in almost every scaling company. The pattern is identical regardless of industry. And it puts the Controller in the worst possible position: doing work they are not staffed for, that distracts them from work they are accountable for, with data they can’t fully defend.

Here is the honest framing. The problem is rarely the CEO. It’s the structure underneath the CEO. And handling it well is one of the more underrated skills in the path from Controller to CFO.

Why this happens

Three patterns explain almost every instance.

The first is that the FP&A function in the company is immature or non-existent. In companies under £50M revenue, FP&A is usually a fractional role at best, often folded into the Controller’s brief. The CEO doesn’t know to ask anyone else because there isn’t anyone else.

The second is that the Controller is too capable for their own good. They have the data, the modelling skills, the patience. So they do the work. And every time they do, they reinforce the CEO’s habit of asking them for it.

The third is that the CEO can’t see the distinction. To a non-finance person, accounting and FP&A look like the same function. Both involve numbers, both come from finance people. The fact that one is historical and audit-grade and the other is forward-looking and explicitly approximate is a distinction that takes a finance background to feel intuitively.

Knowing which of the three you’re in changes what you should do about it.

What doesn’t work

Three responses look reasonable and fail in practice.

The first is refusing outright. “That’s not my job.” This sounds principled and lands like obstruction. The CEO needs an answer. If you don’t help, they get one from someone less qualified or they make one up. Either way the company is worse off and you’ve burned political capital.

The second is doing the work and complaining about it privately. The Controller stays late, hits send, then vents to a peer about how the CEO doesn’t respect the function. The CEO sees the work happening on time and assumes there’s no problem. The pattern continues forever.

The third is doing the work badly. Some Controllers, frustrated by the unreasonable ask, deliver a half-modelled answer with the implicit message of “fine, but don’t expect quality.” The CEO reads the half-modelled answer, uses it anyway, and the next time gives the Controller half the time. The dynamic gets worse, not better.

What works

Three moves. Done together, they change the dynamic in a quarter or two.

Move 1. Deliver fast directional answers with explicit caveats.

When the CEO asks at 6pm, send back something short by 9pm. Not a model. A directional answer with the caveats written into the answer itself. “Based on current pipeline and historical close rates, the directional answer is X. This is not a forecast and I would not put it in front of the board without three more days of work. For a quick decision, X is the right zone.”

This answer costs you twenty minutes, not two hours. It satisfies the CEO’s actual need (a number to use in tomorrow’s conversation) without committing you to defending it later. And the explicit caveat trains the CEO over time to distinguish between a directional answer and a forecast.

Move 2. Use every request as ammunition to build FP&A.

Every time the CEO asks for forward-looking work, the Controller logs it. Date, request, time spent, decision the CEO made with the answer. After a quarter, the log becomes a one-page memo to the CFO: “In the last quarter, the CEO asked me for fifteen pieces of FP&A-grade analysis. I spent forty-three hours on this work outside my mandate. Five of the requests changed CEO decisions. Here is the case for a half-time FP&A hire.”

A half-time FP&A hire is much easier to approve than a full one. And the memo shifts the conversation from “the Controller is overwhelmed” to “there is a structural gap costing the function forty-three hours a quarter.”

Move 3. Document the line in language the CEO understands.

Send the CEO a short note, once, that defines the difference between accounting work (historical, audit-grade, slow, defensible) and FP&A work (forward-looking, modelled, fast, explicitly approximate). Don’t make it long. Two paragraphs. Frame it as helpful context, not as a complaint.

Most CEOs read the note and start qualifying their own requests. “I know this is more FP&A than accounting, but can you give me a directional answer by tomorrow?” That qualification alone changes everything. It tells the Controller what kind of answer is expected, and it tells the CEO that they are asking for something approximate, not something they can put in a board pack.

What this signals about you

The Controllers who handle this well don’t just survive the dynamic. They use it to advance.

They train the CEO into a better requester. They build the case for an FP&A function. They demonstrate, in front of the CEO, that they can think structurally about the finance organisation, not just their own role. And the next time there’s a conversation about who could run a broader finance remit, the CEO has data on the Controller’s strategic thinking that the Controller never had to advocate for.

This is the same dynamic as in the chart of accounts question piece. The Controllers who get the next seat are the ones who design the function to be robust without them. Handling the CEO’s FP&A requests well is exactly that kind of work, done in public, in front of the executive who picks the next CFO.

The tool: the response template

The next time the CEO Slacks you at 6pm asking for FP&A work, the response is this.

“Here’s the directional answer based on current data: X. Three caveats: this isn’t a forecast, the assumptions are A, B, C, and if you need a defensible number for the board I need three days. For a decision tomorrow, X is the right zone.”

Save that template. Adapt it. Send it within three hours of the request.

By the third or fourth time you send a version of it, the CEO starts asking for what they actually need, with the framing built in. Which is the only way out of the loop.

See you next Tuesday.

Adam

The close took too long because of the ERP. The reporting is unreadable because of the ERP. The AP team is drowning because of the ERP. The auditor’s findings will be ugly because of the ERP.

Sometimes the ERP really is the problem. More often, the ERP is doing exactly what it was built to do, and the problem is that the finance function has been asking it to do six other things it was never designed for.

The data on ERPs is brutal regardless. Panorama Consulting’s 2025 ERP Report puts the average cost overrun on implementation at 189% across industries, rising to 215% in manufacturing. Gartner predicts that 70% of ERP implementations will fail to achieve their original business goals by 2027. Only 30% of ERP projects are delivered on time and on budget. And once installed, only 7% of organisations use their ERP as-is. The other 93% pay for customisations, often forever.

Those numbers describe the cost of implementing the ERP. They don’t describe the bigger cost, which is what happens after the ERP is live and the team starts using it for work it cannot do well.

This is the pattern I see across most finance teams, regardless of which ERP sits underneath. Not a vendor problem. A category problem. And it’s almost never the ERP’s fault.

What the ERP was actually designed to do

An ERP is a system of record. It exists to keep one definitive version of the financial truth for the company. General ledger. Statutory consolidation. Regulatory reporting. The audit trail. The chart of accounts. These are the things an ERP genuinely does well, and the things you really do not want to do anywhere else.

The ERP was never designed to be a fast operational tool. It was designed to be a slow, reliable, defensible one. When you ask it to approve invoices in real time, route exceptions to the right person, surface anomalies before audit catches them, run a flux analysis between two minutes and two pages, model cash forecasts that update overnight, or build dashboards your CEO can read on a phone, the ERP does those badly. Because nobody designed it to do them well.

So your team works around it. Excel spreadsheets, email approval threads, shadow databases, a small army of bolt-on tools. Each workaround was added because the ERP couldn’t do that one thing well, and each workaround is now a system the team has to maintain. Multiply that by every workflow your function runs, and you get the real cost of the ERP. Not the licence fee. The 47 workarounds you built around it.

Three patterns I see in finance teams that are fighting their ERP

1. They use less than 20% of what their ERP can do

Walk into any mid-sized finance function and ask them to demo their ERP. You will see them open three modules. Maybe four. They use general ledger, accounts payable, accounts receivable, and sometimes fixed assets. Everything else (procurement, projects, budgeting, treasury, advanced reporting, workflow automation) was paid for, configured at implementation, and quietly abandoned within eighteen months.

The ERP vendor charges for all of it forever. The team uses a fraction of it. The auditor questions why. The CFO writes off the cost as “we’ll use it eventually.” Eventually never comes.

2. They keep trying to make the ERP do operational work

Try to do anything in real time inside a typical ERP. Approve an invoice. Match a payment. Reclassify a journal. Even the simple actions take longer than they should, because the interface was designed for an accountant in 2003 who would batch-process at the end of the day, not for a finance team in 2026 that needs answers in the next ten minutes.

The team that accepts this and reaches for specialist tools moves forward. The team that keeps insisting “the ERP should do this” stays stuck, builds a customisation, pays for it forever, and discovers two years later that the customisation breaks every time the ERP vendor pushes an update.

3. They confuse replacing the ERP with fixing the function

Every Controller I’ve ever spoken to has had a conversation about replacing the ERP. Almost none have actually done it. And almost all of them learned, somewhere in that conversation, that the new ERP would not fix the underlying problem.

Because the underlying problem usually isn’t the ERP. It’s the operational workflows on top of the ERP that have been duct-taped together for years. Replace the ERP, you keep the workflows. The workarounds get more complicated, not less. The function moves sideways, not forward.

What actually works

The Controllers I see making real progress on this stop trying to fix the ERP. They accept the ERP for what it is and reframe their job around three principles.

Use the ERP for what the ERP is good at. General ledger, system of record, regulatory reporting, statutory consolidation. These are the things the ERP genuinely does well, and the things you don’t want to do anywhere else. Stop fighting it on these.

Use specialist tools for everything else. Accounts payable automation, approval workflows, cash forecasting, FP&A modelling, dashboards, supplier portals, bank reconciliation. Each of these has a specialist category of tool that integrates with the ERP and does the job better, faster, and at a fraction of the operational cost. The integration is the price you pay. The productivity gain is the return.

Audit your ERP licence at every renewal. Look at the modules you’re paying for that the team doesn’t use. Look at the customisations you maintain that nobody can defend. Take the conversation to the vendor and ask for a narrower contract. They will fight you. Sometimes they will give in. Either way you learn something useful about what you’re actually paying for.

The shift is psychological as much as technical. The Controllers who get further are the ones who stop treating the ERP as the system that should do everything, and start treating it as the system of record around which the actual finance function is built. The function is the workflows on top of the ERP. Not the ERP itself.

The question to ask before your next ERP conversation

When your IT lead, your CFO, or your ERP vendor proposes the next round of ERP investment (a new module, a customisation, an upgrade, a replacement), ask one question first.

“What workflow are we trying to make faster, and is the ERP the right place to make it faster?”

If the answer is “we want the ERP to do X better,” and X is operational (not consolidation, journals, or statutory reporting), the answer is probably no. The ERP is not where X lives. X lives in a specialist tool that integrates with the ERP. The money goes further there, the team adopts it faster, and the productivity gain shows up in the next close, not in three years.

If the answer is “we want the ERP to consolidate Y entities faster” or “we want better statutory reporting”, then yes, the ERP is the right place. Invest.

This connects directly to the AI pilots piece from a few weeks ago. The same pattern that kills AI pilots in finance (building inside the ERP, expecting the ERP to do operational work, ignoring data readiness) is the pattern that makes Controllers feel like their ERP is broken. The solution is the same too. Pick the narrow workflow. Use the specialist tool. Let the ERP do what only the ERP can do.

See you next Tuesday.

Adam

RAND Corporation’s 2025 analysis put the overall AI project failure rate at 80%. Financial services specifically came in at 82%. MIT’s NANDA study, published in mid-2025, found that 95% of GenAI pilots in enterprise produced zero measurable impact on the P&L. Whichever study you trust, the conclusion is the same. Most AI initiatives in finance never deliver what was promised when the contract was signed.

I run Dost. Our customers are finance teams using AI for accounts payable and the close, so I see this from both sides. The pattern in our own conversations matches the data. The pilots that fail, fail for predictable reasons, and they fail in the first six months because that is when the gap between the vendor demo and the real finance function becomes impossible to hide any longer.

Here is what I see, every quarter, in the implementations that go sideways. And what the 5% of pilots that actually work do differently.

1. The demo runs on clean data. The pilot runs on yours.

Every AI vendor demos with a curated dataset. Clean invoices, well-formed delivery notes, supplier names that match across systems, GL accounts that are tagged correctly. The demo runs at 95% accuracy because the data was selected to make that happen.

Your finance function does not have curated data. You have suppliers who change their invoice templates every quarter. You have multiple subsidiaries with overlapping vendor lists. You have GL accounts that were renamed three CFOs ago and never cleaned up. The same model that hit 95% in the demo lands at 70% on your data, and 70% is below the threshold where the team trusts the system enough to use it.

By month three, the clerk is checking every result manually because they cannot trust the model. By month five, the project is quietly killed because “we are not seeing the productivity gains.”

The model was never the problem. The data was. And nobody surfaced this before the contract was signed.

2. The CFO bought it. The AP team operates it.

The CFO sees the pitch. The CFO signs the contract. The CFO expects the productivity story to land. The AP team finds out three weeks later that they now have a new system to learn.

If the AP lead was not in the room during vendor selection, the pilot will struggle. Not because the AP team is resistant to change, but because they know things about how the close actually works that the CFO does not. The unusual supplier who always sends three invoices for one order. The intercompany allocation that needs manual intervention every December. The accrual that one specific accountant has been doing one specific way for eleven years.

The pilots that work are the ones where the operational team co-owned the vendor decision. The pilots that fail are the ones where the operational team was told what they were going to use.

3. Nobody defined what success looks like

“We want to automate AP” is not a goal. It is a wish.

The pilots that succeed start with a number. Reduce manual invoice processing time by 60% within six months. Cut the number of unmatched invoices from 800 per month to under 200. Get the close from day seven to day five.

The pilots that fail start with vibes. “We need to do something about AI.” “Our peers are doing it.” “The board wants to see progress on automation.”

When the success metric is fuzzy, the team finds reasons to declare success during the demo and reasons to declare failure during the pilot. The CFO who insists on a specific, measurable, time-bound goal before the contract is signed is the CFO whose pilots actually deliver.

4. They built when they should have bought

MIT’s NANDA study found that purchased AI tools from specialist vendors succeed about 67% of the time. Internal builds succeed about one-third as often.

This isn’t because internal teams are worse. Building accounting-grade AI is genuinely hard, requires domain expertise that most internal teams don’t have, and competes for budget with every other internal initiative for years. By the time the internal build is “ready”, a specialist vendor has shipped four major updates and accumulated thousands of customer deployments’ worth of training data.

If your CIO is telling you they can build what the vendor is selling, ask them when. If the answer is more than six months, you are not building, you are stalling.

What the 5% that work have in common

The pilots that survive past six months and start delivering value share four things.

They pick a narrow workflow with a clear before-and-after metric, not a broad transformation. They put the end user in the room during vendor selection, not just the executive sponsor. They buy from a specialist vendor instead of waiting on their internal team’s roadmap. They define what good looks like before the pilot starts, in writing, with a number and a date.

That is the formula. Most finance teams do one or two of the four. Almost none do all four. The ones who do are the ones who write the AI-in-finance case studies that everyone else cites.

The question to ask before signing the next pilot

There is one question that, in my experience, predicts whether an AI-in-finance pilot is going to work before the contract is even signed.

“What does the team who’ll actually use this think about it?”

If the CFO can answer with specifics, the pilot has a chance. If the answer is some version of “we haven’t asked them yet” or “they’ll be fine with it”, the pilot will fail. Not maybe. Will.

The pilots that fail are the ones where the people who do the work were the last to know. The pilots that succeed are the ones where the people who do the work were the first to know.

Run that test on every AI conversation in your function this quarter. The next time a CFO or a Head of Finance asks you whether some AI tool would work, ask back: “What does the team who would use it think?” If they don’t know, the answer is, it won’t work, no matter how good the demo was.

This connects directly to the close-process piece from a few weeks ago. The bottlenecks where AI works are the ones where the operational team understood the gain, agreed it was worth the disruption, and committed to the change. The ones where it doesn’t work are the ones where the disruption was bought before the value was clear.

See you next Tuesday.

Adam

Not because the board doesn’t care. Because the memo wasn’t written for the way a board actually reads.

A non-executive director who sits on three boards reads somewhere around two hundred pages of board materials per quarter, often the night before the meeting, often on a plane. They don’t read top to bottom. They scan, looking for the answer to one question.

What does the CFO want me to know, and what does the CFO want me to decide?

If the memo doesn’t answer that in the first two hundred words, it dies, and the discussion in the room becomes whatever the loudest director happens to raise.

This is the piece that nobody teaches Controllers, but every CFO learns it the hard way. Here is the structure that works, the line your CEO is hoping to see in the first paragraph, and three things that should never go in a board memo no matter how true they are.

The structure that survives every board pack

Jeff Bezos banned PowerPoint at Amazon in 2004 and replaced it with what he called the narratively structured six-page memo. The whole executive team reads in silence at the start of the meeting, then discusses. In his words: “There is no way to write a six-page narratively structured memo and not have clear thinking.”

A board memo for a private company is not six pages. Three to five is the right length for a board pack. But the same principle holds. A narrative memo forces the writer to think. A slide deck lets the writer hide behind transitions and bullets.

The structure that survives every board pack I’ve seen has six sections, in this order. The first hundred words carry the headline and the recommendation. The next section covers context, describing what has changed since the previous meeting. After that comes the decision required, with options if there are any. Then the key numbers, restricted to the ones that drive the decision, not every number you have. Then the risks, each one named and paired with a mitigation. The final section is what happens next, with named owners and dates.

The only section most directors will read carefully is the first one. Everything else is for the people who want to go deeper or who get asked a follow-up question in the meeting. Write the first hundred words for the director who will skim. Write the rest for the director who will pull on a thread.

The line your CEO wants to see in the first paragraph

There is one line every CEO is hoping to see when they open the CFO’s section of the board pack. It is not a number. It is not a chart. It is a position.

The line takes this shape:

“I recommend we [specific action] by [specific date or trigger] because [specific reason].”

One sentence. A clear recommendation, a clear deadline, a clear rationale.

What most Controllers write instead is something like: “Q3 revenue came in below forecast, driven primarily by softer demand in the European segment. The team has prepared a range of scenarios for Q4, detailed below.”

That sentence describes reality. It doesn’t recommend anything. Reading it, the CEO doesn’t know whether the CFO wants to cut costs, raise prices, double down on Europe, or wait. The CEO has to read the rest of the memo to find out, and three times out of four the recommendation isn’t even in there.

This is the same shift we covered in Why most Controllers never become CFOs. The Controllers who become CFOs are the ones who lead with their recommendation, defend it with the numbers, and accept that being wrong about a position is recoverable. Being invisible isn’t.

The three things to never put in a board memo

A board memo is a discoverable document. In an audit. In a litigation. In a regulator’s review. In a future executive search where someone digs through old materials. Whatever you put in writing can be read back to you later by someone with no context for why you wrote it.

Three categories of content that should not appear in a board memo, no matter how true they are.

First, named criticism of another executive. If the head of sales is missing their numbers, the memo can describe the impact in finance terms (”commercial performance is below plan”) without naming or implying that the head of sales is the problem. That conversation happens verbally, with the CEO, before the board meeting. It does not happen in writing in the board pack.

Second, predictions you can’t defend with a model. “I expect Q4 to recover” is not a defensible statement. “Based on a pipeline coverage ratio against historical conversion, Q4 revenue is modeled at the midpoint of the range below, with the assumptions detailed in the appendix” is. If you cannot show the model behind the prediction, take the prediction out.

Third, promises with a single fixed number. “We will land Q4 at £18M” is a sentence that will get screenshotted and read back to you if you miss it by four hundred thousand. “We are modelling Q4 between £17.4M and £18.6M, with the midpoint at £18M, contingent on close rates in the top fifteen deals” is a forecast that survives.

These are not about being evasive. They are about being precise about what you know and what you don’t. The Controllers who get trusted with bigger memos are the ones who never put a sentence in writing that they cannot defend three years later in front of an auditor.

The one question to ask before sending

Chris Tottman writes a newsletter called CFO Secrets, which is one of the more honest accounts of the boardroom written by someone who has actually been in it. He proposes a question that I have started using on every memo I write, and which I now teach to the Controllers and Finance Directors I work with.

Before any section goes in the board pack, ask: “Why am I telling you this?”

If the answer is some version of “because the CEO will be in the room and I want to look prepared”, the section isn’t ready. The section is ready when the answer is one of three things:

The board needs to make a decision and this section gives them what they need to make it. Or the board needs to be aware of something that will affect a future decision. Or the board has previously asked for an update on this topic.

Any section that fails that test should not go in. Brevity beats comprehensiveness, because a five-page memo that the board reads is worth more than a fifteen-page memo that nobody finishes. This is also the test that separates Controllers who get trusted with board materials from those who don’t, and it connects directly to the chart of accounts question we covered two weeks ago. A Controller who can answer “why am I telling you this” cleanly for every section of a memo is a Controller the CFO can hand the pen to.

What to do before your next board memo

If you write the first draft of the board memo your CFO presents, try this once before your next board pack.

Draft the recommendation first, in one sentence, before you write anything else. Then write the rest of the memo as a defence of that recommendation. Then run the WAITYT question through every paragraph. The paragraphs that fail the test, cut.

You will end up with a memo that’s shorter than the one you would have written. And shorter, in board memos, is almost always better.

See you next Tuesday.

Adam

The question is some version of this.

“Walk me through how you’d close the books next month if half your team left tomorrow.”

It doesn’t have to be those exact words. The shape of the question is what matters: a hypothetical that forces the Controller to expose how dependent the close is on specific people, specific habits, and specific workarounds that exist nowhere on paper.

The CFO is not testing resilience planning. The CFO is testing something much harder to fake. They want to find out, before they have to find out the hard way, whether the Controller actually knows how the close works, or whether the Controller is the close.

The answer to that question, in the first 90 days, tells the CFO almost everything about whether they will be working with this Controller in a year.

Why this question, and not a technical one

A new CFO inherits a finance function they did not build. They walk in on day one knowing the company’s numbers from the outside (the board pack, the investor deck, the audit file) and almost nothing about how those numbers actually get produced.

That gap is what scares them. Because if the Controller is the only person who knows why journal 4471 always needs a manual reversal, or why one specific accrual is booked one week earlier than the policy says, or which two people in AP know which suppliers are unreliable about delivery notes, then the CFO is exposed. They are betting their reputation on a process they cannot see and a single point of failure they cannot replace.

So the CFO probes. Not with a technical question (technical questions are answered well by every Controller), but with a structural question that forces the Controller to either reveal that the function is well-designed and documented, or reveal that the function is held together by the Controller’s own memory.

The three categories of answers

There are roughly three ways Controllers answer.

The first kind of answer is the optimistic one. “We’d manage. The team is great. We’d reprioritise the close steps and probably push out a day or two but we’d get there.”

This sounds reasonable and it is the answer that ends most Controllers’ careers in that company within 18 months. It tells the CFO that the Controller has not seriously thought about what would actually happen. It tells the CFO that the function is fragile and that the Controller is unaware of the fragility. The CFO walks away polite, and quietly starts looking for a different partner.

The second kind of answer is the defensive one. “We have documentation. The process is in (system). Any reasonable accountant could pick it up.”

This is worse. Because the CFO has already opened the system. The CFO has already looked at the documentation. They know it’s incomplete or out of date, and the Controller’s confident reassurance tells them the Controller either hasn’t checked recently or is hoping the CFO won’t.

The third kind of answer is the honest one. “Here is what would actually break. The intercompany reconciliation depends on Maria who has been here eleven years and holds the supplier mapping in her head. The AP exception queue would fall behind by a week because Tom is the only one fast enough at coding the construction segment invoices. The cash flow forecast would be a mess because Priya owns the relationship with treasury and they don’t trust me yet. Here is the plan I would propose to fix each of those vulnerabilities in the next two quarters, prioritised by risk.”

That answer earns trust on the first delivery. It tells the CFO that the Controller knows the function from the inside, has thought about its weaknesses, and is already proposing solutions. It also signals something deeper: that the Controller is not personally invested in the function staying fragile. The Controller wants the function to be robust, even if that makes the Controller more replaceable.

That last part is what most Controllers miss. The CFO is partly testing whether the Controller is willing to make themselves less indispensable, because indispensable Controllers are a CFO’s nightmare.

What this means for you

If you are a Controller and you’ve recently had a new CFO arrive, the question is probably coming. It may have already come and you may have answered it without realising. If you can recall the conversation, replay it honestly. Which of the three answers was yours?

If a new CFO hasn’t arrived yet but is on the way, you have one preparation worth doing. Take an hour this week and write down, on one page, the answer to the question. Not a polished memo. The real answer.

The format that works:

For each of the five biggest close steps, identify the single person whose absence would cause the most damage, what specifically they hold that nobody else does, and what the plan is to reduce that risk in the next two quarters. Then identify the two or three workarounds that exist in the process today (the manual reversal, the early accrual, the unofficial supplier list) and whether each one is worth fixing or worth documenting and accepting.

That page is two things at once. It is the prep work for the question when it comes. And it is the thinking exercise that converts a Controller from someone who runs the close into someone who designs the close to be run by other people.

This connects directly to the three habits from a couple of weeks ago in Why most Controllers never become CFOs. Designing the close to be robust without you is exactly the kind of work that gets noticed outside finance, and the CFO who asks the question is doing it because they are trying to figure out if you are someone who could do their job in three years.

The question to ask yourself before next Monday

Before the question gets asked, ask yourself the version of it that matters most.

If I were the new CFO walking in, would I trust me?

If the answer is yes, the conversation when it comes will be a non-event. You’ll give the honest answer, the CFO will nod, and the relationship will start on the right foot.

If the answer is no, you have the next 90 days to change it. The page-long exercise above is where to start.

See you next Tuesday.

Adam

I’m one of those vendors. We build AI for accounts payable at Dost, which means the close is the part of the finance function I look at most days. So this is the honest map. Five places where AI is working in 2026, and three where it’s still a science project no matter what the demo shows you.

Read it before your next vendor call.

The 5 bottlenecks AI actually solves

1. Invoice capture and data extraction

This is the most mature application. AI reads supplier invoices regardless of format, layout, language, or whether they arrive as PDF, email, photo, or scanned image. It extracts vendor name, invoice number, line items, totals, and tax codes without template setup and without OCR rules.

In a real close, this turns a two-week accounts payable backlog into a real-time queue. The team that used to spend the first three days of every month catching up on invoice entry now starts the close already current.

This is the bottleneck where AI works. If a vendor demos this and it fails on real invoices from your actual suppliers, walk out.

2. Three-way matching

AI matches the three documents (invoice, orders, delivery notes) at the line-item level, not just at the header. It flags discrepancies in quantity, price, or missing delivery confirmation, and queues only the exceptions for human review.

The change is in which jobs the accounts payable team actually does. Old workflow: open every invoice, find the matching order, find the matching delivery note, reconcile manually. New workflow: open the exceptions queue and resolve the small percentage of invoices that didn’t match cleanly.

For most finance functions, three-way matching is the single biggest source of close delay. This is where AI changes the shape of the month.

3. Coding and GL classification

AI learns from history. After processing a few hundred invoices, it can predict which GL account, cost centre, and project a given supplier invoice should map to, with confidence scores attached. The clerk approves or corrects. Over time, accuracy climbs.

This compounds. A team that was spending two days on coding at month-end ends up spending two hours, because the AI is right most of the time on standard transactions and the human only intervenes on the unusual ones.

4. Duplicate detection and anomaly flagging

AI catches duplicates that humans miss. Same supplier, slightly different invoice number, same amount, same date. It also catches unusual patterns. A supplier that has never invoiced over £5k suddenly invoicing £45k. Invoice splitting that breaks approval thresholds.

This isn’t about replacing the human auditor. It’s about giving them a queue of twelve flagged anomalies to investigate instead of three thousand transactions to scroll through.

5. Reconciliation matching

Bank reconciliation, intercompany matching, supplier statement reconciliation. AI handles the fuzzy matches that humans would have to investigate manually. A payment for £4,237.50 that hits the bank as £4,237 with a £0.50 fee, for example. It surfaces only the items that genuinely don’t reconcile.

This is unglamorous and high value. The team that used to spend two days a month chasing bank discrepancies spends two hours.

The 3 bottlenecks AI doesn’t solve

These are the ones the demo doesn’t show you. Or rather, the demo shows you a clean version that doesn’t survive contact with your actual finance function.

1. Judgement calls on accruals and estimates

AI can flag what should be accrued. It can’t decide how much. A complex revenue recognition estimate, a contingent liability assessment, a deferred revenue adjustment based on contract interpretation: these all require human judgement informed by policy, materiality, and context that the model doesn’t have.

Any vendor who tells you their AI does this is selling you a problem.

2. Period-end cross-functional negotiations

Some of the hardest accruals require a conversation, not a calculation. These are conversations: procurement confirms whether the goods were delivered, sales confirms whether the revenue was earned, legal weighs in on the disputed invoice. AI can flag these as items pending. It can’t resolve them.

The close gets faster everywhere except here, which means month-end stops being about data entry and starts being about negotiation. That’s a different skill, and it’s where Controllers earn their seat (more on that in last week’s piece on Controller to CFO).

3. New supplier onboarding and risk decisions

Whether to accept a new supplier, on what terms, with what payment days, after what KYC checks. These are decisions, not extractions. AI can run the KYC checks. AI cannot make the decision.

This matters because vendors often demo “AI vendor onboarding” as if it’s automated. It isn’t. The data work is automated. The decision is yours.

The question to ask your next AI vendor

The next time a vendor pitches you AI for any part of the close, ask one question before any other.

Show me where your model is wrong, and how often.

A serious vendor has an answer. A confidence threshold per task. An accuracy band. A list of edge cases their model doesn’t handle. They will show you because they want you to trust the system in the cases where it’s right, and to know when to override it.

If the answer is “our AI doesn’t make mistakes” or “accuracy is in the high 90s, we don’t break it down further,” the demo is theatre. Walk out.

What this means for your next twelve months

The five bottlenecks above are solved. The three are not. That means the close, in 2026, splits into two layers. The layer that AI runs with human oversight. And the layer that still belongs to you.

The Controllers who use the next year to keep fighting for the bottom layer will lose ground. The ones who use it to take ownership of the top layer (the judgement calls, the negotiations, the supplier decisions) become the person the CFO cannot replace.

See you next Tuesday.

Adam

That is the trap of the role. The skills the controllership rewards are the skills that keep you in it. The system that pays you optimises for keeping you exactly where you are.

This is why most Controllers spend a decade waiting for the CFO seat and never get it. Not because they aren’t good enough. Because they have become too valuable in a place the company doesn’t want disturbed.

If you’re reading this, you already feel some version of it. You run the close. You catch errors before audit does. You’ve been told you’re being “developed” for years. Then the seat opens. The board picks the divisional CFO, the treasurer, or the senior FP&A leader. You stay where you are. Again.

The way out isn’t more technical skill. You already have that. The way out is three habits that move you out of the trap, plus one question you can use on Monday to start practising them. I run Dost. Our customers are Controllers and CFOs, so I watch this pattern repeat every week. It’s worth writing down.

1. They reframe their work in the CEO’s vocabulary

When Luca Maestri joined Apple in 2013, he came in as VP of Finance and Corporate Controller. Fourteen months later he was CFO. He kept the seat for a decade.

Look at what he actually did once he had it. He led Apple’s pivot from a hardware company to a services business, the shift that changed how Wall Street values the company today. He oversaw one of the largest capital return programmes in corporate history. He kept Apple lean while peers in tech splurged on splashy projects.

None of that is what a Controller’s job description prepares you for. Maestri got the seat, and kept it for ten years, because he was already thinking about Apple in those terms before he had the title. The Controller chair was where he sat while he proved it.

A Controller’s instinct is to report. “We closed in five days. Variance pack is clean.” A CEO doesn’t need that. A CEO needs: “the close gave us numbers in time to commit to the Q4 hiring plan, two weeks before the board call.” Same close. One sentence sounds like a finance update. The other sounds like a business outcome. That gap is the whole career.

2. They build credibility outside finance, before they need it

Kyle Epley took over as CFO of Caterpillar earlier this month. Nearly three decades at the company. He held the corporate controller role for several years. He also held divisional CFO roles, leading finance for individual business segments, before stepping up to the corporate seat.

The corporate controller experience didn’t get Epley the CFO chair. The divisional CFO time did. His path put him in business reviews with operations leaders across Caterpillar’s segments for years. He learned what the COO worried about before the COO said it out loud. By the time the corporate CFO seat opened, the rest of the leadership team had already been working with him at close range for years.

Most Controllers don’t have a divisional CFO seat available to them. That’s fine. The principle works in less formal ways. The Controllers who get the call have spent years walking out of finance to have the conversations that don’t happen inside finance. They volunteer for the cross-functional project nobody wants. They take the secondment to a business unit. They have a quarterly coffee with the head of sales, not to lobby for anything, but to understand how sales actually thinks about deal slippage.

When the seat opens, the board doesn’t pick the best technician. They pick the one the leadership team has already accepted as a peer.

3. They commit to a number before the close is signed off

A Controller is trained for accuracy. The job rewards getting it right. Getting it right means waiting until reconciliations are done, accruals booked, intercompany tied.

A CFO is paid for judgment. The job rewards getting it directionally right, fast, with incomplete information. By the time the numbers are clean, someone in the leadership meeting has already made the call. The Controller who insisted on waiting for the final variance is no longer the person whose voice carries.

This is the hardest jump because it asks the Controller to violate the instinct that got them this far. The first time you sit in a Wednesday leadership meeting and have to commit to a topline forecast before the deferred revenue cut has landed, it feels reckless. It is not reckless. It is the job.

The Controllers I see making the jump practise it in low-stakes settings first. They give a directional answer to the COO on a Thursday call before they’ve reconciled the underlying entries. They commit to a number on a Monday pipeline review knowing the close hasn’t been signed off. They get comfortable being approximately right, and adjusting publicly when more information lands.

That muscle is what makes a Controller credible to the rest of the leadership team. It is also what makes a Controller credible to themselves as a future CFO.

The question to ask before next Monday

Most Controllers will read this, agree, and report next Monday’s close the way they always have. The three habits are real, but they need a starting point small enough to actually use.

Here is the smallest one. Before any meeting where you’ll be reporting finance work to people outside the function, ask yourself a single question.

Because of this work, what can leadership do today that they couldn’t do yesterday?

If you have an answer, that’s your headline. Lead with it.

If you don’t, you’re about to give a finance update, not a business outcome. Stop and reframe before you walk in.

Try it once. Once is usually enough to notice the difference in how the room reacts.

The choice

You probably already have the technical skill. That isn’t what’s keeping you out of the seat.

What’s keeping you out is the trap of being too good at a job the company doesn’t want disturbed. The way out is in the three habits above. The starting point is the question above.

Use it once before next Tuesday.

See you then.

Adam