Online, it looks like every firm is mid-transformation, every practice has an AI stack, and you’re the last one to the table.

Scroll the feeds for an hour and you’d think a solo who hasn’t automated intake is already a fossil.

Then I close the laptop and talk to actual firm owners, and most of them still haven’t opened the tool once. Not resistant, not behind, just busy running a law practice. The online conversation isn’t describing the profession. It’s talking to itself.

I’m not above this, to be totally clear. I write about legal AI, I read about it constantly, and that immersion warps my sense of what’s normal in exactly the same way it warps yours. The reason I’m talking about it up front is that the bubble has a cost, and the cost doesn’t stay online. It lands on how firms like yours decide to market themselves.

The gap is quietly steering firms wrong

Here’s what happens inside the bubble: you see a hundred firms and vendors announcing their AI capabilities, you feel the pressure not to be the one left out, and you reach for the same language everyone else is using.

It feels like keeping up.

It’s actually the start of the problem, because that pressure pushes the whole profession toward the exact marketing move that stopped working the moment everyone made it at once.

“AI-powered” is already marketing wallpaper

Here’s the tell: “AI-powered” has become wallpaper. Every firm says it. Every vendor says it. It’s slid into the pile of phrases that used to mean something and now mean nothing, right next to “trusted,” “results-driven,” and “client-focused.” Those words didn’t die because they were false, they died because everyone claimed them at once, and a claim that everyone makes carries no information.

When your prospective client reads “AI-powered firm,” their eyes slide right past it, the same way yours slide past “trusted advisor.” You didn’t communicate anything. You filled space.

And the inflation is accelerating. Two years ago, “AI-powered” at least signaled that a firm was paying attention. Today it signals that a firm can read the same posts as everyone else. Its half-life as a differentiator is nearly gone, and firms are pouring it onto their homepages right as it hits zero.

The buyers who matter see through it first

It’s worse than just neutral, though, because the buyers most worth winning are the ones most likely to see through it.

In-house counsel and sophisticated repeat clients have sat through the same vendor demos you have. They’ve heard “AI-powered” from every pitch that crossed their desk this year. They know it’s a label anyone can print, with no bar to entry and nothing behind it by default.

So when you lead with the adjective and can’t immediately back it with something concrete, you don’t read as innovative, you read as one more firm repeating what it heard, which is the opposite of the impression you were trying to make.

You’re spending your credibility to sound current and ending up sounding generic to the exact people whose opinion moves the most business your way.

Drop the label, show the receipt

So here’s the move, and it’s simple: stop describing yourself with an adjective and start describing what you actually did.

Not “we’re an AI-powered firm,” but “we turned a three-day document review into an afternoon, and we handed the client the time back instead of billing for it.”

Not “we leverage cutting-edge AI,” but “we now answer new inquiries in minutes instead of the next business day.”

One is wallpaper. The other is a specific, checkable claim a skeptic can’t wave away, because it names the task, the change, and the result. A sophisticated buyer can interrogate it, and the fact that it survives interrogation is the entire point.

And notice you can do all of this without touching confidentiality. You’re not naming the client or the matter. You’re describing the shape of the work and the size of the change, which is yours to talk about. “A commercial client,” “a stack of contracts,” “a three-day task done in an afternoon” gives a prospect everything they need to be impressed and gives away nothing you’re obligated to protect.

The discipline that replaces the adjective

Every time you’re tempted to write “AI-powered” somewhere in your marketing, stop and ask what the AI actually did: what task, how much faster, what the client got out of it. Then judge your own answer honestly.

If you can’t answer, you don’t have a marketing claim yet. You have a word, and the word is doing the job of hiding the fact that you haven’t done the thing.

If you can answer, then you don’t need the word at all, because the specifics do the work the adjective was only pretending to do.

Either way, the adjective loses. The only reason to keep reaching for it is that it’s easier than doing the work or describing it, and “easier than the work” is exactly why it stopped meaning anything.

Be the firm that kept the receipts

The bubble will pop, in the sense that “AI-powered” as a differentiator is headed for being worth nothing, the way “has a website” is worth nothing today. When it does, it’ll take down the firms that only ever had the adjective. It won’t touch the ones that kept the receipts and did the work, because a specific result a client can verify never depended on the label being scarce in the first place.

So be the second kind of firm.

Do the work, describe it plainly, let the specifics carry the weight, and you’ll still be standing when “AI-powered” means about as much as “trusted” does today, which is nothing at all. Trust me.

And you have a reply ready to go. Not the one you’ll post: the real one, the one composing itself in your head right now. It’s accurate. It’s specific. It names what that client actually did, the calls they didn’t return, the documents they never sent, the deadline they blew and then hung around your neck. It would feel incredible to publish.

I’ve been in that chair. There’s a particular way you type when you’re angry, hitting the keys a little harder than the job requires.

Here’s what’s worth knowing before you press send: the only person who wants you to send that reply is the person who left the review. Nobody else is in the argument. Everybody else is just reading.

Your reply isn’t for the reviewer

It’s for the next person who reads it, and that person doesn’t know either of you.

They weren’t there. They can’t tell who’s lying. All they see is two people in a disagreement, and the only question they’re really answering is which one of you sounds like someone they’d want handling the worst month of their life.

You’re not trying to win. You’re demonstrating, in public, how your firm behaves when something goes wrong, and that tells a prospective client far more about you than anything on your website ever will. Write it like you know that.

A perfect record was never the asset

Read more

What you need is a simple system that turns more of your good leads into paying clients. If you do that with AI, fantastic, but you don’t need to use AI to accomplish this.

This article is compiled from podcast episodes over the years and some of my internal documents. It may sound a little bit more bland than what I typically write, but we are actually trying to build something today.

This is the six-step build I’ve used across ~20 different platforms over the last few years. The specifics change from system to system, but the theory doesn’t, and the order matters more than the software you pick.

Why these 6 steps? Why not the whole build, Toby? Well, in more than a decade working on these systems, what I’ve learned is that there are a handful of places where we tend to get jammed up. These are the six key items that I think will help get you unblocked, they are usually the ones that people miss.

Why is this worth the effort?

Making it easier for a good lead to become a client is the first way your firm makes money.

A lead who’s a good fit, agrees to hire you, and then drops out before they pay is pure loss: you’re the only one who loses in that situation. So intake is the last place you can afford to drop the ball, and in 2026 there’s a new way to drop it, which is speed.

A lead you contact within five minutes is 21 times more likely to convert than one you reach after thirty (LEXGRO 2026), and most prospects now say they pick the firm that answers first, not the one with the best credentials (Aloware 2026). The fix isn’t hiring another receptionist. It’s restructuring the system so it doesn’t break when the load gets heavy.

One rule before the steps: don’t rip out your whole intake and replace it overnight.

That’s a recipe for disaster.

Build it gradually.

Phasing the rollout isolates the problems and makes the whole thing go smoother.

Step 1: Map the system you already have, then design the one you want

Read more

He searches for a lawyer, taps your ad, and gets a white screen. He waits one second, then two, then three. Then he backs out and taps the firm listed below you.

Here’s the part that should bother you: you were billed for that click. Your ad did everything you asked of it. It won the auction, it beat the firm down the street, and it earned the tap from a man with a fresh injury and a phone in his hand. Then it handed him to a page that wasn’t there yet.

You pay at the click, not at the page

Google charges you the moment he taps, and your page loads afterward. So a slow site isn’t lost traffic: it’s traffic you already bought and paid for and then set on fire.

Most firms file website speed under IT, somewhere behind the new case management system and the printer that keeps jamming. But it isn’t an IT problem; it’s a line item you’re already paying and not collecting on.

The stat everyone links is old, and it’s too kind to you

You’ve seen this one, probably from whoever sold you your website: a one-second delay costs you 7% of your conversions. It’s real, and it’s from 2008. It came out four months after the App Store opened, which tells you most of what it didn’t measure.

So it’s fair to ask whether it still holds. It does, and the honest answer is that it was generous to you.

In April 2026, Shopify ran the numbers across its entire commerce ecosystem: every actively selling store, not a hand-picked case study. For every 100 milliseconds a site got slower, conversion dropped about 3.5%. A site loading in 2.5 seconds converted roughly 30% worse than one loading in 1.5.

Put that next to the old number. One second was supposed to cost you 7%, and on 2026 data it costs something closer to 30%. The stat everyone’s been recycling for eighteen years isn’t wrong: it’s conservative by about four times. Machines got faster, people got far less patient, and the gap has been widening against you the entire time.

Read more

He doesn’t search for a lawyer. He asks for one.

He opens an AI assistant on his phone and types the way people actually talk now: “I just got served over a construction contract, do I need a lawyer, what kind.” Then, “what does that cost.” Then, “who’s good in Sacramento.”

What comes back is three firms, with a sentence about each.

He reads the three sentences and calls the second firm, because that sentence said they’d walked somebody through exactly this and picked up the phone on a Saturday.

Here’s the part worth sitting with: not one of those sentences came from a law firm’s website. They came from what those firms’ clients had already said about them.

Reviews used to close the deal, now they open the door

For twenty years the job had two halves: be findable, then be persuasive. Reviews lived in the second half. A prospect found you, worked their way down the page, read a few reviews, and either called or didn’t. They were the closing argument.

That isn’t where they sit anymore. When a prospect asks a machine instead of searching, the reviews get read first, by the machine, and they decide whether you make the list at all. They’ve gone from the closing argument to the thing that gets you into the room.

And if you don’t make the list, nothing else you built matters. Your homepage, your practice-area pages, the video you paid too much for: none of it loads, because nobody ever gets there.

Read more

They see a bot on a competitor’s site, decide they need one too, and bolt a generic widget onto the homepage. Then they watch it greet every visitor with the same chirpy “How can I help you today?” while it answers almost nothing and converts even less. The bot becomes wallpaper. People learn to ignore it the way they ignore a cookie banner.

The mistake is treating a chatbot as a feature. It’s not a feature. It’s a worker, and a worker needs a job. You wouldn’t hire a person and then say “stand near the door and do something helpful.” You give them a defined role with a clear definition of done. A no-code AI bot is the same. Assigned to a specific, repeatable job, it’s leverage that runs at 2 a.m. and never gets tired. Assigned to the wrong job, or to no job at all, it costs you the exact leads you built the site to capture.

So the question is never “should we have a chatbot.” The question is “what job am I assigning it to, and is that a job a bot should do.” Here are five jobs where the answer is yes, and two where the answer is a hard no.

1. After-hours lead capture and qualification

Your best leads don’t all arrive between 9 and 5. Someone with a real problem searches at 11 p.m., lands on your site, and finds a contact form and a phone number that rings into voicemail. By morning that person has filled out three more forms on three more sites. The first one to respond usually wins, and you were asleep.

This is the single strongest case for a no-code bot. The job is narrow and rule-driven: greet the after-hours visitor, ask three or four qualifying questions, capture name and contact and the nature of the problem, and drop it into your CRM or inbox before the visitor leaves. Tools like Tidio and Intercom handle this out of the box with a visual flow builder, no developer required. You’re not asking the bot to be brilliant. You’re asking it to catch what would otherwise fall on the floor.

The win is measured in leads recovered, not conversations had. If your site gets even five serious after-hours inquiries a week and you were converting none of them, a bot that captures three of those is a real number you can put against its cost.

Read more

A firm spends a year locked onto the competitor across town.

When the other firm hits 100 Google reviews, this firm pushes to 101, then 130.

When the competitor publishes an 800-word post on car-accident claims, this firm publishes 2,000 words with better photos and more backlinks.

Every profile post, every service page, every ad: matched and beaten.

On paper, they’re winning. Skirmish after skirmish, they’re ahead.

Then someone on the team finally runs the actual market math for their main practice area. Not “how are we doing against them,” but “how big is this market, really, and how much of it is anyone capturing?” The number lands like a gut-punch: the competitor they’d white-knuckled for twelve months held a sliver of the real market. The rest of the room, the majority of people who needed exactly what this firm does, was sitting uncontested the whole time. Nobody was fighting for it. They’d spent a year winning a fight in an empty corner while the room itself went unclaimed.

That’s the trap, and good firms fall into it constantly: beating the firm next door and knowing how big the room actually is are two completely different questions, and only one of them tells you whether the fight was worth having.

Your competitors are a free resource, not a target

I’m not going to tell you to ignore your competition. Their marketing is one of the most useful free resources you have. Everything they do in public, the keywords they rank for, the reviews they collect, the ads they’re running right now, is sitting out in the open for you to study. It’s a sneak peek into their playbook. It’s all public, you’re not breaking a single rule, and it’s about as close to a cheat sheet for your next campaign as you’ll ever get.

A couple of those plays are worth naming, because they still work:

• Keyword gaps: find the searches they rank for and you don’t, then go be the better answer for those searches.

• The sub-niche snipe: comb their reviews for the service clients keep asking about that nobody in your market actually offers, then offer it. I’ve always called this sniping a sub-niche, and it still holds.

But here’s what all of those tactics have in common, and it’s the thing nobody says out loud: they’re all just methods of extracting data. Each one is you pulling information out of a competitor’s public footprint. That data is only worth having if you do something with it besides matching them move for move. Matched move for move, competitor research curdles into reactive one-upmanship: they got to 100 reviews so we need 101, they wrote 800 words so we write 2,000. You can win every one of those and still lose, because you never asked the only question that pays: how big is the actual opportunity, and is this even the slice worth chasing?

How big is the room?

That question has a real answer, and it’s more concrete than it sounds. There’s a framework marketers use, total addressable market, serviceable market, and obtainable market: three glorious pieces of jargon that sound a little like a VC’s last words. Strip the jargon and it’s just three nested circles. First, there’s everyone in your area who could ever need your service. Then there’s the realistic slice you could actually serve, and there’s the piece you can genuinely capture right now, given your size and your competition. The whole game is pushing that innermost circle outward, and you cannot do that if the only number you track is your review count versus the guy down the street.

The discipline underneath this is simple: only spend money on a marketing activity if you have a specific plan to make revenue from it. Chasing a competitor’s review count, backlinks, or top of page position is not a plan. Claiming an uncontested slice of the market that you sized on purpose is.

The biggest empty room in 2026 is the AI answer

Here’s what’s changed since I first made this argument, and it’s the part most firms haven’t caught up to. In 2024, the fight was for the top of the Google results page. In 2026, there often isn’t a top of the page the way there used to be. A prospect asks ChatGPT or Google’s AI “who’s a good estate planning attorney near me” and gets back a short list of names, no ten blue links in sight. That’s the new empty room. Almost nobody in legal is deliberately competing to be the firm the AI names, which means the single most useful competitive-intelligence exercise you can run this month costs nothing: open the AI tools your clients use, ask them for a lawyer in your practice area and city, and see who comes up. If it isn’t you, that’s your gap for 2026. The old arms race was more words and more backlinks. The new one is being the clean, specific, quotable source an AI actually pulls into its answer.

The firms that win aren’t the ones grinding hardest against the competitor next door. They’re the ones who know how big the room is, know which corners are empty, and point their budget at the space nobody else is fighting for. Beat the market, not the firm beside you.

She isn’t searching the way you picture it. She’s not typing “custody lawyer” into Google and scanning ten blue links. She’s talking to an AI assistant in full sentences, the way people actually do now: “do I need a lawyer if my ex filed for full custody,” then “what happens to my kids while the case is going on,” then finally, “custody lawyer near me who will actually fight for me.”

She’s terrified and she’s shopping at the same time, and the question underneath every one of those searches isn’t really a legal question. It’s this: is my kid going to be okay, and will this person fight for us?

Now look at what she finds.

The first firm’s page opens with the founder’s headshot, his law school, and the year the firm was established. The second leads with “over 20 years of combined experience” and a row of award badges. The third promises “aggressive representation” in a serifed font laid over a stock photo of a gavel. Every one of them is talking. None of them are talking to her.

She bounces off all three. Then she lands on a fourth, and the first thing she reads is a paragraph that could have been written about her exact night: the papers on the counter, the fear, the not-knowing what happens to the kids while it all plays out, and what this firm actually does in the first two weeks to protect a parent’s position. She reads it twice. That’s the firm she calls in the morning.

The mistake is that you made yourself the hero

That’s the whole disease, and I see it constantly. So often it seems like some lawyers really wanted to be doctors, with their diplomas hung on the wall behind the desk. There’s a romance to it: the credentials, the pedigree, the wall of framed proof that you’re smart. But you’re not in business to be admired; you’re here to help people and make money, and the person on the other end of that midnight search does not care where you went to law school if you’re getting the job done.

Here’s a fast way to know whether your firm has this problem: if prospects regularly ask where you went to law school, your marketing and sales process is doing a rough job. They’re asking because nothing you put in front of them answered the question they actually had, so they fell back on the only proxy for “is this lawyer any good” that they know how to check.

The fix has a name, but it isn’t mine: it’s the client story, and I first learned it from Ned and Lee at Rosen Institute. The idea is simple to say and hard to do: make the client the hero of the story, not you. You’re the guide who gets them through it.

Two things changed, and both make the mistake more expensive

I’ve been saying this for years: two things have shifted recently, and both raise the cost of getting it wrong.

The first is who reads your website before your prospect ever does. When that woman asked an AI assistant for a custody lawyer, something read every one of those firm pages and decided which ones to surface and summarize back to her. Answer engines don’t quote “20 years of experience.” They pull the specific, situation-matching language that actually answers the question a person asked. Credential copy is invisible twice over now: the human skims past it, and the AI never surfaces it in the first place.

The second is that AI is now writing a lot of the bad copy. A large share of firm websites in 2026 were drafted, or at least first-passed, by an AI tool. Left to its defaults, AI writes exactly the wrong thing: founding year, credentials, “dedicated and aggressive representation,” generic everything. Unless you steer it hard toward the client’s story, the machine will happily mass-produce the number-one mistake for you, faster than you ever could by hand.

Read more

A firm owner has a great consultation.

The prospect is a perfect fit, the call runs long in the good way, and everyone hangs up expecting a signed engagement letter by Friday.

Friday comes and goes.

So does the next one.

By week three the owner has quietly written the lead off: not serious, shopping around, probably went with someone cheaper.

Here’s what actually happened:

The prospect got busy, got scared, got one competing quote, and then waited to be led. Nobody led. There was no follow-up email recapping the plan, no second touch, no nudge back toward the decision. The awareness worked, because they found the firm. The interest worked, because they booked the call. The decision was nearly made. And then the firm, which had spent real money getting this person to raise their hand, had built nothing to catch them at the one-yard line.

That’s the most expensive pattern in legal marketing, and it comes from one specific misunderstanding: firms think a funnel is a way to get more leads, when a funnel is actually the whole system that carries a stranger all the way to becoming a client who sends you more clients.

Most firms build the top of it and quit halfway down.

What a funnel actually is

Strip away the jargon and a funnel is just the path a person takes from never having heard of you to hiring you, and beyond. Built well, it’s the difference between hoping people find you and hire you on the spot, and having a steady pipeline where trust gets built before the phone ever rings. It matters more for law firms than for almost any other business, because legal decisions are heavy. Nobody impulse-buys a divorce lawyer. The road from “I might need help” to “here’s my retainer” is long and paved with fear, which means your funnel can’t be a quick-sale pipeline. It has to nurture the whole way down.

Most firms do build the first stretch, and they build it competently. For awareness, they show up where people are looking: content that answers the real questions (”what do I do after a car accident,” “how do I start a business in my state”), some Google Ads, maybe social. For interest, they capture the lead with a guide or a free consult and drop them into an email sequence. For the decision, they stack up the social proof, because testimonials are still the single most powerful trust-builder you have, worth more than any paid award you can buy. So far, so normal. This is the half of the funnel everyone builds.

Read more

That isn’t hypothetical. In Mata v. Avianca, a federal judge in the Southern District of New York sanctioned two lawyers and fined their firm $5,000 in 2023 after they filed a brief built on cases ChatGPT had fabricated, then stood behind those fake cases when the court asked them to produce the opinions. The cases didn’t exist. The lawyers hadn’t checked.

That story hit a nerve, because it touched every fear a firm owner has about this technology:

• Will I be embarrassed in front of a judge?

• Will I increase my malpractice exposure?

• Will I get a bar complaint?

• Will I make a mistake that lives online forever?

The lesson from Mata isn’t “never use AI.” That’s the easy read, but it’s wrong. The real lesson is narrower and more useful: never use a plain chatbot as if it were a researcher.

The problem was never that AI is dumb

The old workflow was seductive: ask ChatGPT a question, get a fluent, confident answer, and let the polish stand in for reliability. It doesn’t.

A hallucination is what happens when the model fills a gap with something plausible. It doesn’t stop and say “I don’t know.” It gives you an answer that reads well enough to survive a quick skim, especially when you’re busy or asking about something you don’t already know cold. That’s the dangerous part: the answer doesn’t sound unsure. It sounds like a confident associate who never checked the source.

I’ve watched lawyers distrust AI for exactly the right reason and then use it the wrong way anyway. They’ll ask a chatbot “what are the best marketing channels for a family law firm?” or “what are my competitors doing in estate planning?” and then treat the reply as research. It isn’t research. It’s a guess in a nice suit.

What actually changed

The shift worth paying attention to isn’t from bad AI to perfect AI. It’s from summarizing chatbots to research agents.

A summarizing chatbot answers from its own memory plus whatever you paste in. That’s genuinely useful for brainstorming, rewriting, and getting unstuck. It’s not enough when the whole point is to know whether something is true.

A research agent does a different job: it searches, reads real sources, compares them, cites them, and hands you a trail you can inspect. That doesn’t make it perfect. It can still misread a source or overstate a point. But it moves the work from “trust me” to “check me,” and for a firm owner that’s the difference that matters. You don’t need an oracle. You need a faster way to reach sources and patterns you can verify yourself.

The two research jobs in your firm

Here’s the distinction most people miss: there are two completely different research jobs inside a law firm, and the AI tools split along that exact line.

The first job is the actual practice of law. That means case law research, drafting, discovery, contracts, matter analysis, and that’s really the home of the legal vertical AI tools:

• Harvey: enterprise and big-firm legal AI for research, drafting, and due diligence.

• CoCounsel (Thomson Reuters, formerly Casetext): legal research and drafting wired into Westlaw and Practical Law, with linked citations.

• Lexis+ with Protégé: LexisNexis legal AI built around its own case law and Shepard’s citation checks.

• Clio Work: Clio’s legal AI workspace for matter analysis, strategy, and drafting, with a built-in citator.

• Vincent AI (vLex): legal research, fifty-state surveys, and litigation intelligence, cited to primary sources.

• Paxton AI: an all-in-one legal assistant for research, drafting, and document review, priced for small and midsize firms.

There are new tools that could be on this list every day, and the reason they get on this list is because they’re trustworthy enough for casework, and that’s precisely because they’re tied to a real legal authority: those are databases, citators, primary law. If you’re doing actual legal research for a legal matter, that’s where you go. You are still going to need to verify the authority yourself.

But that’s not the research I’m talking about here.

Read more

The firm ranking first had 58 reviews. The firm at third had 129. The firm at fourth had 170. And one firm in that same pack had 515 reviews and still wasn’t number one. The firm with the fewest reviews of the group was the one winning.

That’s the part most attorneys miss: Google’s local results aren’t a straight reputation contest. Reviews matter. They build client trust, they lift conversion, they tell Google your firm is real and active. But reviews don’t automatically make you relevant for every search.

Google is trying to answer a specific question, from a specific person, in a specific place. Someone who searches “estate planning attorney” gets firms that look unmistakably like estate planning attorneys. Someone who searches “elder law attorney” gets elder-law signals. Someone who searches “Medicaid planning attorney” gets Medicaid-planning signals. That sounds obvious until you see how most firm marketing is actually built: the firm says it handles estate planning, elder law, probate, Medicaid planning, and guardianship, but its Google Business Profile says “estate planning,” its homepage says “estate planning,” and its directory listings say “estate planning.” The other services are thin, missing, or bundled onto one catch-all page.

Then the owner wonders why the firm ranks for one search and vanishes on the rest. The answer usually isn’t that Google hates the firm: it’s that Google doesn’t have clean, consistent evidence that the firm is relevant for the other searches. That’s the pattern I want to show you: not theory, not a generic checklist, but a real anonymized teardown.

The firm that wins one search and loses the next

One firm in that market is a genuinely strong estate-planning practice: 4.8 stars, nearly 60 reviews, and a spot near the top of the map pack for “estate planning attorney.” Google clearly understands it as a credible estate-planning option, and it’s already done enough right to compete.

But when I checked “elder law attorney,” the same firm was buried around number eleven. Same for “Medicaid planning attorney.” Same firm, same city, same reputation. Completely different visibility.

Read more

And one of them talks back.

The ad writes itself now

Start with the format Google calls Conversational Discovery ads. Here’s how it works: when someone asks AI Mode a question, Google’s Gemini model reads that specific query and builds the ad on the spot, tailored to what they actually asked. You don’t write the headline for “how do I file for divorce in my state.” Gemini writes it, per search, from the assets and signals you’ve given it. A companion format, Highlighted Answers, drops your ad directly inside the AI’s recommendation when Google decides you’re the most relevant option.

There’s a catch worth knowing. Alongside your ad, Gemini generates its own independent explainer that weighs and synthesizes the options, and it can compare you against what it finds elsewhere. You don’t write that part, and you can’t switch it off. The AI is now a participant in your ad, not just the machine that delivers it.

Read more

Think about what billing actually is: a lawyer does the work, and at some point later that work gets written down, described, priced, packaged into an invoice sent, and then eventually, hopefully, paid. Every step in that chain is a potential place for value to escape. Time can go uncaptured. Descriptions go vague. Invoices can go out late. A chunk gets written off because nobody can defend the line item. By the time cash lands in the account, the gap between what you earned and what you collected can be real, and it can be bigger than you think.

That gap is the entire reason boutique small firms are moving to AI-powered billing. It’s not because it’s modern, and it’s not because a vendor showed them a slick dashboard; it’s because it plugs the leaks in a process that was bleeding the money the whole time, and the firm never had a report that could prove it, or a way to stop the bleeding. That time is over.

A quick word before the breakdown: what follows is about a category, not one company, and a few tools are racing into this space right now. But I’ll be direct with you: I’ve looked closely at one of them, Ajax (joinajax.com), and it does almost exactly what this article describes. I have no financial relationship with them, and they don’t know I’m writing this— I just think it’s genuinely good. I’ll come back to it at the end.

The leak nobody tracks

Here’s the number every partner is trying to avoid looking at: the gap between hours worked and hours billed, then the second gap between hours billed and dollars collected. Industry surveys of small firms have put realization, the share of worked time that actually turns into collected revenue, somewhere in the 80s on the high end and far lower on the low end. The exact figure varies by firm and source. Even with variance, the direction here is not in dispute; a meaningful slice of the work your lawyers actually do never becomes money.

Walk back through the chain, and you can see where each dollar dies. A lawyer takes a six-minute call, means to log it, but then gets pulled into the next thing and never does. That time is gone forever. It was earned, and it will never be billed. Multiply one forgotten entry per lawyer per day across a year, and I know... it’s enough to make any of us vomit. We’re talking about thousands of dollars simply evaporating because logging happened from memory at the end of the day or week.

Then there are the entries that do get logged but go out late. An invoice sent 60 days after the work is an invoice the client half-remembers and is going to be slower to pay or have questions about. That slows the entire process down and sucks your staff out of doing things that actually earn more money. The longer the gap between the work and the bill, the longer the gap between the bill and the cash, and the higher the odds that client disputes align with what they no longer recall agreeing to.

Read more

That’s really the entire misunderstanding, and almost everyone starts there. I did, and I know a lot of people like me who did as well. People still treat AI like a search engine, where a few keywords get you a pretty good result, but it’s not a search engine. It’s the most literal, fastest-working, most context-starved staffer you’ve ever managed. It’ll do exactly what you asked, but if what you asked was kind of vague, you’ll get kind of vague back.

Prompt engineering, which has been all the rage for the last couple of years, is actually a terrible name for what fixes this problem. It sounds like something that requires a computer science degree when it really doesn’t.

For someone who is actually running a business, prompt engineering is just learning to brief your AI the way you’d brief a sharp new hire on their first day. You would never walk up to a new employee and say “Write a client email” and then just walk away. You would actually tell them who the client is, what just happened, and what you want the email to accomplish, and probably how long the email should be. Do that with AI, and your outputs will change completely.

Vague Mush in, Vague Mush out

Let me show you the failure, because you’ve probably lived it. I know I have.

Weak prompt: “Write an email to a client who hasn’t paid their invoice.”

Here’s roughly what comes back:

Dear Valued Client,

I hope this email finds you well.

We wanted to reach out regarding your outstanding balance.

We value our relationship and would appreciate your prompt attention to this matter.

Please let us know if you have any questions.

It’s grammatically perfect and completely useless. It doesn’t know the amount, the relationship, your tone, or whether this is the first nudge or the fourth. So it hedges on all of it and produces the beige average of every collections email ever written.

The model didn’t give you a bad email. It gave you the only email it could write with the information you provided. You asked a stranger to write a letter about people they’ve never met. You got a letter about nobody.

The fix is a structure you already know from managing people. Five parts: context, role, task, format, and an example of good. You don’t need all five every time, but the more of them you supply, the sharper the output.

Here’s the same email request, briefed properly:

“You are an office manager at a small law firm in your city. You’re writing to a long-standing client who is normally reliable but is 21 days late on a $4,200 invoice. This is the second reminder; the first was a soft note two weeks ago. The tone should be warm but clearer than last time because the balance is now genuinely overdue. Keep it under 150 words. Do not use phrases like ‘I hope this finds you well’ or ‘valued client.’ Reference that we have worked together for years and that this is unlike this person, so the ask will land more as a favor between partners, not a demand.”

Now the model has what a real assistant would have. It knows who it is (the role), what the situation is (the context), what you want done (the task), how long and in what voice (the format), and what to avoid. The email it returns will name the amount, acknowledge the relationship, escalate slightly from the first reminder, and skip the corporate filler you explicitly banned. You went from a generic template to something you can send after a ten-second read.

That’s the whole trick. Notice there’s nothing technical in that prompt. No code, no jargon. It reads like instructions to a competent person, because that’s exactly what it is.

The four moves that do most of the work

You don’t need to master twenty techniques. Four of them carry the load, and an owner who uses only these will beat almost everyone fiddling with the chat box.

Give it a role. “You are a paralegal screening intake calls” produces sharper output than no role at all, every time. The role tells the model which version of itself to be. The difference between “summarize this” and “you are a litigator; summarize this deposition for the three facts that help my client” is the difference between a book report and a useful brief. One line, large effect.

Give it an example of good. This is the most underused move by far. If you have a client email you love, paste it in and say “match this voice.” If you have a marketing blurb that sounds like you, show it. The model is extraordinary at pattern-matching and nearly hopeless at guessing your taste from nothing. One good example teaches it more about what you want than three paragraphs of description ever will.

Tell it what to leave out. Most owners only say what they want. The bigger gains often come from saying what they don’t want. “No exclamation points. Don’t mention price. Don’t sound salesy. Don’t use the word ‘solutions.’” Negative instructions cut the generic filler that makes AI output read as AI output. The banned-phrases line in that collections email did more for it than anything else.

Make it ask you questions first. This one feels strange until you try it. End your prompt with “Before you write anything, ask me up to five questions that would make your answer better.” Suddenly the model interrogates you the way a good hire would, and the answers you give become the context you forgot to provide. It surfaces the gaps in your own brief. I use this for anything that matters, and the questions it asks are frequently better than the prompt I started with.

The traps that will burn you

The output is going to sound confident. That confidence is the single most dangerous thing about these tools, and it has nothing to do with whether the content is right.

AI states wrong things in the exact same authoritative tone it uses for right things. There’s no flicker, no hedge, no tell. It’ll invent a statistic, cite a case that doesn’t exist, quote a law that was never passed, and present all of it with the smooth assurance of an expert. People have sent fabricated figures to clients and filed briefs with made-up citations because the output read as polished and they assumed polished meant true. Polished means polished. Nothing more.

So, the rule is simple and it’s not optional: anything the model produces that contains a specific fact, a number, a name, a date, a citation, or a claim about the outside world must be verified by a human before it leaves your building. You can absolutely use AI to draft or to summarize or to rephrase or to organize your own material, but don’t use it as a source of truth about the world or about specific information without actually checking. The drafting is where it really picks up the pace, but the fact-asserting is where AI ends careers.

The related trap is letting it invent specifics to fill a gap. If you ask for a case study and give it no real details, it’ll fabricate a client, a result, and a glowing quote, because you asked for a case study and it has no way to say “I don’t have that.” The fix lives in your brief: “Use only the details I provide. If you need information I haven’t given you, ask for it rather than inventing it.” That one sentence prevents most of the confident fiction.

The prompt is a skill, and skills compound

Here’s what separates the owner who gives up in week two from the one who quietly doubles their output: the first one keeps typing vague requests, keeps getting mush, and keeps blaming the tool. The second one learns to brief well, and every brief they write makes the next one faster, because the moves stack and the good prompts get saved and reused.

You already know how to do this. You’ve spent years learning to delegate to people, to give context, to set expectations, to say what good looks like. Every bit of that translates directly. The owner who can brief a new hire into doing excellent work can brief AI into doing it, and AI doesn’t need a salary, doesn’t sleep, and doesn’t call in sick. The skill you’re building isn’t technical; it’s managerial, and you have most of it already. The ones still typing two-word requests into the box are managing a brilliant new staffer by handing them a sticky note and walking away.

It’s not Google. It’s YouTube.

YouTube is the second-largest search engine in the world, behind only Google, which owns it. More than two and a half billion people use it every month, and a growing share of them treat it exactly like Google: 57% of adults now say they search YouTube for information the same way they’d search a traditional search engine.

Here’s the shift that should get your attention. When younger people have a question, video is increasingly the first place they go, not the last. A 2026 Sprout Social study found that 41% of Gen Z turn to social and video platforms first when they’re looking for information: ahead of traditional search engines, which came in at 32%. Search isn’t dying. It’s splitting, and a big piece of it is moving to video.

Now think about what that means for your law firm.

People don’t want to read about their legal problem; they want someone to explain it.

Picture the person who needs you. They were just served divorce papers by their spouse. Their parent fell at home, and now they are urgently trying to arrange care. They’re staring at a DUI and a court date they don’t understand.

They’re scared, and they have one question they need answered right now: what do I do?

That person is not in the mood to read a 2,000-word blog post. They want to watch a human being look into a camera and walk them through it in plain English. One in six people researching a legal question online specifically search for video, and of those, 79% go to YouTube to find it.

These questions are video-shaped: “how do I file for divorce,” “will I lose my house if I file for bankruptcy,” “how do I get guardianship of an aging parent,” “can I bring my spouse to the U.S.” Every one of them is a search someone is making today, and every one of them is a video you could own.

Read more

It’s the brief.

The brief doesn’t feel like a problem because it feels like the job. A client describes what they want on a call or in a tangled email thread. Then a strategist disappears for a few hours and turns that mess into a structured document the creative team can actually use. Audience, objective, deliverables, tone, mandatories, success metrics. That document is necessary. The hours spent producing it are mostly wasted.

Here’s the part nobody is saying out loud: writing the brief isn’t the strategic work. It’s just translation. The client already knows what they want, more or less. The strategist isn’t inventing the answer; they’re just converting raw input into a structured format. This type of translation, the slow act of moving meaning from one shape into another and formatting it, is exactly the thing machines have gotten really good at this year.

The hidden cost of the brief

Run the math on a single brief and the leak becomes obvious. A senior strategist or account lead spends two to four hours turning a client conversation into a usable brief. That’s real time at the most expensive rate in the building. Multiply it across every project, every month, and the agency is spending a meaningful slice of its highest-value hours on formatting.

That time is invisible because it’s never billed cleanly and never questioned. It hides inside “project setup” or gets quietly absorbed as overhead. But it’s the single most common form of billable-hour leakage in an agency: senior people doing work that requires seniority to start but not seniority to finish. The judgment is in the first ten minutes. The next three hours are typing.

Agencies already know this, and they also know that the cost of this compounds in an even worse way: a tedious brief gets rushed. The strategist only captures about 80% of what the client said and then reconstructs the rest of it from memory. The mandatory thing that the client mentioned one time doesn’t actually make it on to the page. Nobody is ever going to really notice this until the creative team delivers work that ends up violating the brief, and then we have a problem.

Read more

That silence is the most expensive thing your firm does, and almost nobody measures it. You spent hundreds of dollars to make that person call. You spent an hour of attorney time to close them. Then you handed them off to a process that nobody owns, and the first impression of your firm becomes a missing email and a document request that never went out. The relationship started warm and is cooling before the matter even opens.

Onboarding isn’t intake. Intake is the work of turning a stranger into a client. Onboarding is everything that happens in the two weeks after the signature, and it’s a different animal entirely. Intake is part judgment and part sales. Onboarding is almost none of either. It’s a checklist.

The dropped handoff is a process problem, not a people problem

Here’s what onboarding actually contains: send the engagement letter for signature. Collect the retainer. Request the specific documents this matter type needs. Open the matter in your case management system. Set up the client in billing. Schedule the kickoff call. Send the welcome packet that tells the client what happens next and who their point of contact is. Add the key dates to the calendar. None of that requires a law degree. All of it requires that somebody do it, in order, without dropping a step.

And that’s exactly where it breaks. The paralegal who handles onboarding is also handling three active matters and covering the front desk. The engagement letter goes out a day late because it was waiting on a template tweak. The document request never goes out at all because everyone assumed someone else sent it. Two weeks later the attorney opens the file to start work and the client has uploaded nothing, because nobody ever told them what to upload. The matter hasn’t started. The clock the client thinks is running isn’t running.

Nobody is bad at their job here. The job is just badly suited to a human. It’s repetitive, it’s rule-based, it has a dozen steps that each have to fire in sequence, and it happens for every single new client while the same person is doing higher-value work. That’s the definition of a task you should hand to a machine. Not because the machine is smarter; because the machine never forgets step seven.

What an onboarding agent actually does, step by step

Strip away the hype and an autonomous onboarding agent is a worker that runs the same sequence for every new client, the moment the trigger fires, without anyone reminding it.

Read more

That sentence is the whole problem.

Most owners think AI adoption is a decision they get to make on their own timeline. It’s not a decision sitting in front of you. It’s already happening inside your firm, every day, with nobody watching it.

A second-year associate is short on time and pastes the facts of an active matter into ChatGPT to get a first draft of a motion. Your marketing person is generating blog posts and email copy through whatever free tool they found. A paralegal is summarizing a deposition transcript by uploading it to a public model because it saves an hour. None of these people asked permission. None of them think they’re doing anything dangerous. And in most small firms, none of them are wrong to assume that, because no one ever told them otherwise.

The risk isn’t a regulation that might arrive in two years. It’s what’s happening this week with no guardrails.

That framing matters because it changes what you’re actually trying to fix. You’re not preparing for a hypothetical future where the bar association hands down a rule. You’re managing a tool your staff is already using in ways you can’t see, and the exposure is real today. “AI safety” sounds like a policy binder somebody bills you to write. For a small firm, it’s not; it’s a short set of operational rules, the kind you can put in place this week, that draw the line between the AI use that helps your firm and the AI use that can cost you a client or a malpractice claim.

RELATED: GenAI is Threatening Confidentiality in Law Practices

The three risks, in plain terms

Forget the abstractions. There are three concrete ways careless AI use hurts a law firm, and you’ve probably already brushed against at least one of them.

1. Confidential client data leaves the building
   When someone pastes privileged facts into a free public chatbot, that information is now sitting on a server you don’t control, under terms of service most of your staff have never read. Some consumer AI tools retain inputs and use them to train future models. The free tier and the paid tier often have different data policies, but the person on your staff using the free version has no idea which one they’re on. Picture a paralegal pasting a settlement memo, party names and dollar figures intact, into a public model to “clean up the language.” That’s confidential client information handed to a third party without consent. You’d never email that memo to a stranger. This is the same act, just with a friendlier interface.

2. The model invents things, and the invention reaches a filing
   Large language models produce text that sounds authoritative whether or not it’s true. They generate case citations that don’t exist, complete with plausible reporters and page numbers. This isn’t a rare glitch. It’s how the technology works when it doesn’t know an answer. Lawyers across the country have already been sanctioned for filing briefs with fabricated citations that a chatbot produced and nobody checked. The danger in your firm is the same: an associate trusts a confident-sounding draft, the fake citation survives a rushed review, and it lands in front of a judge with you or your firm’s name on it. The cost there isn’t measured in hours. It’s measured in sanctions, in malpractice exposure, and in a reputation that took years to build. This kind of risk isn’t new, but it’s more dangerous today than it was even a year ago, which is exactly why it has to be watched instead of assumed away.

3. Unreviewed AI output goes straight to a client
   This one is quieter and more common. A client intake email, a status update, a first-draft engagement letter, written by AI and sent without a human reading it closely. Most of the time, it’s fine; no harm, no foul. The problem is the times it isn’t, when the AI confidently states something about the client’s matter that’s wrong, or strikes a tone that reads as careless, or includes a detail from a different case because the model blended its context (the last one is real, and bad). Your client doesn’t know AI wrote it, but they know your firm sent it. Every unreviewed piece of content is a small bet that the model got it right, and you’re placing that bet without knowing it.

The protocol fits on one page

Here’s the part that surprises owners; the fix isn’t expensive and it isn’t complicated. You don’t need a consultant or a forty-page governance document. You need a one-page set of rules that everyone in the firm reads and follows. Five things go on that page.

1. Name the approved tools
   Pick the specific AI tools your firm is allowed to use and write them down. The key distinction is between consumer tools and business or enterprise tools that contractually agree not to train on your inputs. The paid business tiers of the major providers, ChatGPT Team or Enterprise, Claude for Work, Microsoft Copilot under a business agreement, generally offer data terms that keep your inputs out of training. The free consumer versions often don’t. If a tool isn’t on your list, it isn’t approved. This single rule closes most of the data-leakage hole, because it moves your staff off the tools that retain their inputs and onto ones that don’t.

2. Draw the bright line on client data
   State plainly what never goes into a public or consumer model: client names, case facts, privileged documents, settlement figures, anything that identifies a matter. If staff need AI help with a real document, it goes through an approved business-tier tool only, or it gets anonymized first through an approved process. This rule needs to be simple enough that a stressed associate at 9 p.m. remembers it and is stopped by it without having to think too hard. “No client facts in public AI” is a sentence anyone can hold in their head.

3. Require a human gate on anything that leaves the firm
   No AI-generated work product reaches a client, a court, or opposing counsel without a person reviewing it. Every citation gets verified against the actual source, not the model’s description of the source. Every client-facing message gets read by a human who owns it. The AI drafts, but a lawyer is accountable. That order never reverses.

4. Decide who owns this
   Pick one person responsible for the firm’s AI rules. In a small firm that’s usually the managing partner or the office manager, not a new hire. Their job is narrow: keep the approved-tools list current, answer questions when staff are unsure, and update the page when something changes. A rule with no owner is a rule nobody follows.

5. Tell everyone it exists
   The protocol does nothing sitting in a drawer. Send it to every person in the firm, including the ones you don’t think are using AI, because they probably are. Make it part of onboarding. One short meeting where you walk through the page is enough. The goal isn’t to scare anyone off AI. The goal is to make the safe path the obvious path.

“Wait and see” is the expensive option

The instinct for a lot of owners is to hold off. Let the technology settle, let the rules firm up, let someone else go first. That instinct is wrong here, and the reason is simple; waiting doesn’t pause AI use in your firm. It only pauses the rules. Your staff keeps using these tools the entire time you’re deliberating, and every day without guardrails is a day of unmanaged exposure.

The math is lopsided. Writing the one-page protocol costs you an hour or two (yes, you can and should use AI to help you draft the list) and one short meeting. The thing it prevents, a privilege breach, a sanctioned filing, a malpractice claim from a confident piece of AI fiction that nobody checked, costs you a client at minimum and your firm’s standing at worst. You’re weighing an afternoon against an event that can end relationships you spent a decade building. There’s no version of that trade where waiting wins.

Guardrails aren’t the thing that slows AI adoption down; they’re the thing that lets you speed it up without driving off a cliff.

Do the small version now.

Not the perfect version.

The small one.

Open a document this week, write down which tools are approved, write the one sentence about client data, name the human-review rule, assign an owner, and send it to your staff.

It won’t be complete, and it doesn’t need to be.

A one-page protocol that everyone actually follows beats a perfect policy that arrives next quarter, because the exposure isn’t next quarter.

It’s today, in the chat window your associate has open right now.

They see a number from 1 to 10. They want it higher. They ask their PPC manager to “fix it.” The PPC manager raises the bids, swaps the headlines, hits the keyword again. The number maybe nudges. Maybe not.

That’s not how Quality Score works. Quality Score isn’t a grade. It’s a diagnostic.

A doctor doesn’t look at your temperature and write a prescription for “lower temperature.” She looks at the temperature, then orders the test that explains it. The temperature is the symptom. The test is the diagnosis. The prescription comes after.

Your Quality Score is the temperature. The three components beneath it are the X-ray.

The three views of the X-ray

Google Ads gives you three diagnostic components for every keyword:

1. Expected click-through rate. The probability that someone searching this query will actually click your ad.

2. Ad relevance. How tightly your ad copy matches the searcher’s intent.

3. Landing page experience. What happens after the click. Speed, content match, navigation, trust signals.

Each one comes back with one of three readings: Above Average, Average, or Below Average.

Below Average is the X-ray showing you a fracture. Average means it’s not broken but it’s not strong. Above Average means that part of the funnel is doing its job.

Read more

But most of you missed how much smarter the tool got in the last six months. The way you use the tool stayed the same, but the tool itself actually developed.

That’s the conversation everyone is missing. The conversation about AI has been stuck on the wrong question because everyone is asking which model is the smartest, which app has the best interface, and which prompt gets the best answer. Those are 2023 questions. Everyone is still assuming the unit of work is still a single exchange between you and a chatbot. That’s not where we are today.

The shift this year is not about smarter chatbots, and it’s not about copying and pasting faster. 2026 is about agentic workflows. Once you see the difference, you cannot unsee how much of your week you have been wasting.

The single-shot trap

Here’s what single-shot AI use actually looks like in a business: you paste in a customer email into a chatbot and then you ask it to draft a reply. You read the reply, maybe tweak it, and paste it into your email client and then hit send. Then you go over and log the interaction in your CRM, and then you set a reminder to follow up with the person. When the next email comes in, you start the entire process over.

AI only did one slice of that, and you did all the pieces around it that got everything to work.

Sure, this already feels productive because the one slice it did used to take you a lot longer. Step back and count the steps here; the drafting was maybe 20% of the work. The other 80% was you ferrying information between systems, deciding what happens next, and clicking buttons. You automated the easy part and kept all the connective tissue for yourself.

That is why “asking AI things” plateaus. You can get marginally better answers with better prompts, but you are still the bottleneck. Every multi-step process still runs through your hands. You are a very expensive router for information that should be moving on its own.

What an agentic workflow actually is

An agentic workflow is a chain. AI does several steps in sequence, on its own, without you approving each one. It pulls the data, analyzes it, produces the output, and routes that output to wherever it needs to go. You set up the chain once and define the rules. After that, it runs.

The word “agentic” sounds like science fiction. The reality is mundane, which is exactly why it works. Think of it as the difference between a power drill and an assembly line. A chatbot is a power drill. Useful, faster than a screwdriver, but you still stand there operating it for every single screw. An agentic workflow is the line. The work moves through stations and comes out finished.

Read more