Friday, September 25, 2026

Five times this week I told my AI it was wrong

Five times this week I told my AI it was wrong

I sat down Friday morning with a report my AI had written about my own business, and the first thing on it was a problem that did not exist.

It told me one of our products was on sale for twenty-seven things when only seven of them had been built. That is the kind of sentence that makes your stomach drop, because if it were true it would be the worst thing on the page and it would be my name on it.

It was not true. The seven were the new ones. The other twenty already existed and had been pulled in from elsewhere, which is how every single one of these products is assembled and has been for months. It took me one line to say so. It had taken the machine a paragraph, a table and a confident heading to be wrong.

I want to be honest about my first reaction, because it was not clever. My first reaction was to believe it. It was well written. It had specifics in it. It cited a file. Everything about the shape of it said this has been checked.

That is the whole problem, and it is why I am writing this instead of something more useful.

The other four

Once I started counting I got to five before lunch, and they were all from this week.

One. It went looking for how many of my contacts work in the car repair trade, to decide whether that was a market worth building for. It came back with six. The real number was one. The other five were the letters t i r e sitting inside the word entire, and inside the word retirement. I had a morning of quiet enthusiasm about a market that turned out to be financial advisors.

Two. Same exercise, different trade. It reported thirty-seven contacts in manufacturing, which would have made it the biggest opportunity on the list. The real number was four. The rest were a business coach who works with manufacturing leaders, somebody in manufactured housing, and, my favourite, a note the machine had written itself a week earlier saying you cannot quickly manufacture reviewers. It counted its own sentence as evidence of a market.

Three. It needed the current web address for our community. There are two, an old one and a new one, and the old one was retired three weeks ago. It searched every file we have, found the old address two hundred and eighty times and the new one far less often, and concluded the popular one was the right one. It had run a poll. The documents all voted. The documents were all from August.

Four. It needed the next available number to file something under. A tracking document said number two hundred and forty-nine was free. Two hundred and forty-nine was not free. Something had been filed there days earlier and the tracking document had not been updated, which is the entire reason you do not trust a tracking document.

Five was the one I opened with.

Read that list again and notice what is not in it

Not one of those needed technical knowledge to catch.

You do not need to know anything about how these systems work to know that entire contains tire. You do not need to understand anything at all to think, hang on, thirty-seven seems like a lot. The question that caught the last one was five words long and it was not a clever five words. It was that is not how those work.

Every correction this week came from somebody who knew the business, not from somebody who knew the technology. That is not a small point. It is the whole point, and it is the opposite of what most people assume when they sit down in front of one of these things for the first time.

What I got wrong for about a year

I used to think that checking AI output meant being able to out-think it. That to say no, that is wrong I needed to already know the right answer, and if I did not already know the right answer then who was I to argue.

That is completely backwards and it kept me quiet for longer than I would like to admit.

You do not need the right answer. You need to notice that something is off. Those are different skills and the second one you already have, because you have been doing it to junior staff, to vendors, to bank reconciliations and to your own numbers at eleven at night for your entire career. The feeling you get when a set of books looks fine and something is still wrong, before you can say what, is exactly the muscle this needs. Nobody told you that was transferable. It is.

Why you do not push back

Here is the part I think most people will recognise and nobody says out loud.

It is uncomfortable to argue with something that sounds certain. Especially when you are new to it and it is not. The output arrives fluent, formatted, complete, with headings and a summary and a confident tone, and pushing back on it feels a bit like putting your hand up in a meeting to question the one person who clearly knows what they are talking about.

That is a confidence problem wearing a technical problem's coat. And it gets worse the less experienced you are, which is exactly backwards from where it should sit, because the people who most need to challenge it are the ones who feel least entitled to.

The three questions

I do not have a clever system for this. I have three questions I ask every time now, and the first one does most of the work.

Where did that number come from? Not is it right. Where did it come from. Made it recalculate the manufacturing count and cut it from thirty-seven to four.

What would make this wrong? Asked of anything confident. It forces the thing to argue against itself, and the good ones will.

Show me the thing, not the summary of the thing. The summary is where the errors go to hide, because a summary is polished by definition. Four of this week's five survived right up until somebody asked to see the underlying list.

One more thing

I am not telling you this to make anybody nervous about using these tools. I use them all day and they did an enormous amount of good work this week that I have not mentioned once, because correct work is not interesting to write about.

I am telling you because there is a version of this where somebody reads a confident paragraph, does not want to look slow by questioning it, and lets it through. That is the only failure mode that actually matters, and it is not a technology failure. It is a nerve failure.

So question it. It will not be offended. It is, if anything, slightly better than most people at being told it is wrong, which is the single most underrated thing about the whole business.

Thursday, September 24, 2026

Who trains the next reviewer?

I had two conversations this week that I could not stop thinking about, and the annoying part is that neither person was trying to make a point. They were both just describing their own Tuesday.

The first was a senior accountant at a firm in Brooklyn. Senior accountant is a strange promotion. Nobody throws a party. What actually happens is that the difficult files start appearing on your desk, and one day you look up and realise you have stopped being the person who prepares the work and started being the person who checks somebody else's.

Those are not the same job. Preparing is procedural. Checking is diagnostic. You look at a finished reconciliation and you know, in about four seconds, roughly where the problem is going to be hiding. Nobody teaches you that. It shows up, or it does not.

The second was a man who runs offshore accounting teams for Western firms. He had been thinking about what happens when the machine assembles the file and a person signs it, and he said it better than I have managed to:

"The valuable version of my team is not the one that does reconciliation, it is the one trained to audit a prepared reconciliation fast and catch what is wrong in it. That is a different hiring and training problem than the one I have been solving for. Pods built for production do not automatically become pods built for verification."

He is right. And I have been sitting with it for two days because there is a second half he did not say and I did not think of until later.

The problem is not that the two skills are different.

The problem is how the second one has always been made.

You already know this. You have just never had to say it out loud.

Think about how you actually got good at review.

You did not read about it. There was no session on it. You built several hundred reconciliations, a healthy number of them were wrong, and somebody more senior sent one back with a single question and no other comment, which is somehow worse than being shouted at. And then you did it again. And somewhere in year three or four you started seeing the error before you found it.

That instinct is a residue. It is what is left over after enough repetitions of the boring work, done badly, then corrected in public.

The production years were not a toll you paid on the way to judgment. They were the machine that produced it.

Which brings me to the question I have not been able to answer.

If the machine does the first four hundred reconciliations, where does the fifth-year reviewer come from?

Your current seniors are fine. They already have it, they earned it the old way, and nobody can take it off them. The people who would have become your seniors in six years are the ones I would worry about, because they will not have had the six years. They will have had six years of checking work that was mostly right, which teaches you almost nothing, because the thing that teaches you is being wrong in front of somebody who noticed.

The comfortable answer is wrong and I want to say so before someone else does

The comfortable answer is that we keep juniors doing manual work in order to season them, like a cast iron pan.

That is a bad answer. It is also never going to happen. No firm is going to pay for deliberately slower work as a character-building exercise, and no junior is going to sit quietly through three years of data entry while the person at the next desk finishes at four.

The real answer is less tidy, and it is this. If the training used to be a by-product, it now has to be on purpose.

Somebody has to design it. Not buy a tool. Design a way for people to build the judgment that the grunt work used to hand them for free, in a world where the grunt work has left the building.

I do not have a finished answer to that, and I would be a bit suspicious of anyone who says they do eighteen months in. What I do have is the part that comes first, which is knowing what you are actually training.

The good news, and it is better than I expected

Here is where I changed my own mind this week.

Reviewing a junior and reviewing an agent turn out to be the same skill.

Both hand you work that looks finished. Both are confident. AI is wrong the way a very polite person is wrong: fluently, in a complete sentence, with excellent formatting. And both are right most of the time, which is exactly what makes the occasional miss dangerous, because you stop looking properly.

In both cases the question you are asking is identical. Not "is this wrong" but "where would wrong hide in this."

So if you have ever reviewed another person's work, you already have the beginning of it. You are much further along than you think, and considerably further along than somebody who is brilliant with the tools and has never once had to sign their name underneath numbers they did not prepare.

That is the whole reason our training is built around judgment rather than tooling. Not because the tooling is unimportant, but because the tooling changes every few months and the judgment does not. AI does the typing. You make the judgment calls.

And it is why the hardest belt is not the one where you build something clever. It is the one where you decide what not to automate, and where to stop trusting what comes back.

The same problem, from two different chairs

If you work on the books, this is a career question, and the honest version is uncomfortable. The part of your work that is easiest to hand over is the part you are currently fastest at. The part that is hardest to hand over is the part you probably undersell, because it feels like something anybody could do. It is not. I promise you it is not.

If you run a company, this is not a career question at all. It is a workforce design problem on a very long fuse.

You will get the productivity. That part is real and it shows up fast. What shows up slowly, and only if somebody builds it deliberately, is the bench underneath your current senior people. Nobody notices a broken training pipeline in year one. You notice it in year five, when the person who should have been ready is not, and you cannot make up the ground quickly because the ground was made of repetitions that do not exist any more.

The firms that come through this well will not be the ones with the best tools. They will be the ones who worked out early how to make reviewers on purpose.

Both doors are open and both are free

If you work on books, the nine-belt AI Native Foundations course is free for life. It normally costs $1,000 per person. Twenty-seven worked examples built on the real work rather than a case study, and the belts that matter most for everything above are the ones about judgment and knowing when to stop trusting the output.

And you can gift that same course, free for life, to every client you have. Uncapped. Not a trial and not a discount code. Your clients are one rung down the same ladder, and it is a reason to be in touch that is not an invoice.

https://humansplusagents.ai/legacy-sbc

If you run a company with ten or more employees, the entire first year of the $25,000 programme is free. All nine courses, a complete AI-Native Company Transformation Plan, monthly CEO peer-group forums, and Foundations training for every employee you have.

That last part is the one that matters here. Training everybody in the building is how you rebuild the bench on purpose rather than hoping it happens by accident. Ten employees is a real threshold rather than a suggestion, and year two is $25,000. Sign up free first, then request CEO access from the offer bar inside your account.

https://humansplusagents.ai/legacy-sbc-aia

Nothing expires. No card. Nothing renews.

We are all working this out at the same time, and the people who will figure it out first are the ones already doing the work rather than reading about it. If you have a view on where the next generation of reviewers comes from, I would honestly like to hear it. I am still short of an answer.

The community is free and this is mostly what we talk about: https://www.facebook.com/groups/1024923770013348

And if you would rather start with something small, we build free agents one job at a time, yours to keep, no course required first: https://aiaccounting.legacysbc.com/claim-agent

Wednesday, September 23, 2026

We already paid for the lessons

The question I get asked most about what we are giving away is not what is in it. It is why.

Why is a nine-belt course that costs $1,000 per person now free. Why can a bookkeeper hand that same course to every client they have, for life, with no cap on the number. Why does an agent get built for someone who has never given us a penny and may never give us one.

People ask it politely, but what they mean is what is the catch, and that is a fair thing to want to know before you spend five hours on anything.

Here is the real answer, and it is simpler than the suspicion.

We already paid for these lessons. Charging you for them a second time would not make them any truer.

What three years actually bought

Around 10,000 business leaders across 1,000 companies have come through this training over the last three years.

That number is not a credential. I am not putting it here to impress anyone. It is here because of what it actually means, which is that we have now watched several thousand intelligent, busy, capable people try to start using AI in real work, and we have watched most of them get stuck in the same four or five places.

Not different places. The same ones. Over and over, in different industries, at different sizes, with different budgets.

That is the only thing we have that is genuinely hard to get, and it is not knowledge. It is pattern. You cannot read your way to it and you cannot buy it. You get it by being in the room while a few thousand people struggle, and by paying attention to what they struggle with.

So here is what it taught us.

Lesson one: nobody needs to know how it works. They need to know when to stop trusting it.

The instinct when people feel behind on AI is to go and learn the machinery. Prompts, models, tokens, which tool is best this month.

Almost none of that changes your Tuesday.

What changes your Tuesday is knowing when the answer in front of you is wrong. The people who get value out of these tools fastest are not the ones who understand them best. They are the ones who stopped accepting the first response.

That is why the training is built around judgment rather than tooling, and why the line we keep repeating is AI does the typing, you make the judgment calls. Five hours will not make you an expert at the typing. It is enough practice at judging the output to stop taking it at face value, and that is the part that has ever mattered.

Lesson two: build the boring thing first

Everybody wants to start with the impressive project. The dashboard. The thing they could show someone.

The people who actually keep going started with something dull. The reconciliation exception list. The document chase. The schedule they rebuild from scratch every single month and quietly resent.

Boring work is repeatable, and repeatable is the whole game. An impressive one-off proves nothing to you. A dull job that comes round every month and now takes twenty minutes instead of three hours proves everything, because you feel it again in four weeks.

It is also why our agent library is built the way it is. Not one big platform. One job at a time, each one named by an actual person who was sick of doing that actual job by hand.

Lesson three: everyone goes through five stages, and nobody skips the second one

Indifference. Fear. Acceptance. Confidence. Agency.

Is this even useful for my job? What if I get it wrong? I can see where this helps. I know what to check. I can improve this work.

I have watched people move through those five in that exact order, and I have never once seen anybody skip the fear. Not the confident ones, not the technical ones, not the ones who were enthusiastic in the first session.

That matters for how you teach it. If your training assumes people arrive at stage three, you lose everyone standing at stage two, and stage two is where most of the profession is standing right now. So we built for the fear rather than around it. The worked examples say out loud that you verify the matches and the totals, and that nothing posts itself. That is not a disclaimer. It is the answer to the question people are actually asking.

Lesson four: training one person changes almost nothing

This is the expensive one, and it took us longest to accept.

Send one person on a course and they come back enthusiastic, get busy, and are back to their old workflow inside a month. Not because the training failed. Because they went back into a team that works the old way, and one person cannot hold a new way of working on their own.

It is why the CEO package includes free Foundations training for every employee, not just the person who signs up. It is the biggest item on that list and the one people skim past. It is also the only version that works.

Lesson five: the person who gives something away is the one who gets the most from it

This one surprised me and it is the reason the bookkeeper offer is shaped the way it is.

When a bookkeeper hands a client free lifetime access to a real course, the client gets something genuinely useful for nothing. But the bookkeeper gets something better: a reason to have a conversation that is not an invoice, a renewal, or a chase.

That is worth more than a newsletter and more than most marketing anyone in this profession has ever paid for. So the gift is uncapped, permanently, deliberately. Not a trial. Not a discount code. Give it to every client you have if you want to.

So what do we actually do

We teach. The nine-belt AI Native Foundations course, twenty-seven worked examples built on reconciliation, chasing missing documents and assembling month end for review, rather than a case study about a fictional company.

We build. One agent at a time, for one job somebody named. Free, yours to keep, no conditions on it.

We sit with people. The community, the peer-group forums, and the ordinary answering of questions from people who are stuck at stage two and think they are the only one.

And we consult, for the businesses that want somebody to do the building with them rather than learn to do it themselves. That is the part that costs money, and I would rather tell you that plainly than let you find it on page six.

Which is why the rest is free

It is not philanthropy and I am not going to dress it up as philanthropy. Some proportion of people who do the course and find it genuinely useful will later want help building something, and they will come back. Most will not, and that is fine, and nobody is going to chase you.

But the lessons above are already paid for. They cost us three years and several thousand people's patience. Handing them to you costs us nothing, and the alternative is watching an entire profession pay for them again, one firm at a time.

I would rather we all got further, faster.

Everything that is open right now

The nine-belt course, free. Normally $1,000 per person. If you work on books it is free for life, and you can gift it free for life to every one of your clients: humansplusagents.ai/legacy-sbc Everyone else, including CEOs: humansplusagents.ai/legacy-sbc-aia

If you are the CEO of a company with ten or more employees, your entire first year of the full programme is free. All nine courses, a complete AI-Native Company Transformation Plan, monthly CEO peer-group forums, and free Foundations training for every employee you have. That is a $25,000 package. Ten employees is a real threshold rather than a suggestion, and year two is $25,000. Sign up free first, then request CEO access from the offer bar inside your account.

A free AI agent, built for you. One job you actually do, yours to keep, no conditions attached: claim one here. Tell us the job when you claim it, and be specific. The more precise you are, the more useful the thing we build.

The community, which is free, open, and stays that way. Nothing is sold in it: the AI for Accounting group

Nothing on that list expires. No countdown and no closing date.

Come and get the lessons. We have already paid for them.

Yvonne Razo is the founder of Legacy Small Business Consulting, CFO of Simple Academy, and co-founder of AI for Accounting.

Tuesday, September 22, 2026

What just became free

I want to do this as plainly as I can, because the numbers involved are the kind that make people assume there is a catch.

Two things are now available through us, at no cost.

The first is the AI Native Foundations Black Belt, a complete nine-belt course that normally costs $1,000 per person. It is free. If you are a bookkeeper, it is free for life, and you can give it to every one of your clients for life as well.

The second is for CEOs of companies with ten or more employees. The entire first year of the full programme, which lists at $25,000, is free. That includes all nine courses, a complete AI-Native Company Transformation Plan, monthly CEO peer-group forums, and free Foundations training for every employee in your company.

That is the announcement. The rest of this is me explaining exactly what it covers, what it does not, and why anyone would give it away, because I would want all three before I signed up for anything.

There are two doors, and they lead to different places

If you are a bookkeeper or you work on books:

humansplusagents.ai/legacy-sbc

Free for life. Nine belts, and 27 worked examples built on bookkeeping work: reconciliation, chasing missing documents, assembling month-end. Not generic AI training with an accounting sticker on it. The reconciliation example walks through a prepared exception list linked back to the source rows, and it is explicit that you verify the matches and the totals, and that nothing posts itself.

Then there is the part I did not expect. You can gift the same course, free for life, to every one of your clients. Not a trial. Not a discount code. The whole nine-belt course, permanently, to as many clients as you have.

If you have been looking for a reason to ring a client you have not spoken to properly since the last filing, that is a reason. The page even suggests the opening line: "I can give you free lifetime access to a practical nine-belt AI course. What would you like AI to help you do?"

If you are anyone else:

humansplusagents.ai/legacy-sbc-aia

Same nine-belt Foundations course, free, open-ended access. This door is for everybody. Owners, CEOs, CFOs, operations people, team leads, and the large number of people who are not the boss of anything and simply want to be better at the work in front of them. It is also the door for anyone thinking about building something of their own in this space.

And behind that door, for a specific group, is the bigger thing.

The $25,000 part, and the line it draws

If you are the CEO of a company with ten or more employees, your first year of the full programme is free. I am going to list what that is, because "$25,000 value" means nothing until you see it broken up:

A complete AI-Native Company Transformation Plan for your business
Monthly CEO peer-group forums
All nine courses, not just Foundations: AI Agent, AI Role OS, Intelligence OS, Governance OS, Execution OS, Company Transformation OS, AI-Native Company Management OS and Enterprise Value OS
Free AI Native Foundations training for every single employee you have

That last one is the one people skim past, and it is the biggest item on the list. Not the training for you. The training for everyone who works for you, at no cost, this year.

I run a company and I sit in the CFO seat at another one, so I know exactly what line that lands on and what it normally costs to put a whole team through anything. This is the only part of the package I would have had to argue for, and it is the part being given away.

Now the line. Ten or more employees is a real threshold, not a suggestion. Most of the people reading this run practices smaller than that, and I am not going to blur it and let you find out later. If you have fewer than ten employees, the Foundations course is still yours free, with open-ended access. The CEO package is not.

And so you have it in writing now rather than in month eleven: year two is $25,000. The first year is the gift. The Foundations course stays free either way.

What you actually come out with

Nine belts, in order, each one building on the last. Grey through Black.

You start by telling your AI what you want. Twelve goal exercises and twelve interactive tools you build for yourself, around your own goals rather than a case study about a fictional company. Then problem solving. Personalised learning. Creative thinking. Critical thinking. Better decisions. Coaching from your own transcripts. Project planning. And finally directing a complete multi-stage workflow.

The stated outcome is the honest version of what this kind of training can give you: the ability to choose an outcome, use AI to pursue it, and judge the result for yourself.

That last clause is the whole thing. AI does the typing. You make the judgment calls. Five hours of anything will not make you an expert at the typing. What it gives you is enough practice at judging the output to stop accepting the first answer, which is the only part that has ever mattered.

There is a section on the bookkeeper page I keep going back to, which maps the journey as five stages: indifference, fear, acceptance, confidence, agency. Is this useful for my job? What if I get it wrong? I can see where it helps. I know what to check. I can improve this work. I have watched people move through those exact five stages in that exact order, and I have never seen anyone skip the second one.

What it is not

I would rather say this now than have you discover it:

Implementation is separate. This is training. If you later want somebody to build the thing for you, that is a different conversation with a different price on it.
Certification is separate. Access does not award it. You would have to meet the programme requirements.
AI software subscriptions are not included. Whatever tool you use is your own arrangement, and you should be using tools approved for your work and checking what information you are allowed to put into them.

Why would anyone give this away

Because it starts a relationship, and I would rather tell you that than pretend it is philanthropy.

The people behind this have trained around 10,000 business leaders across 1,000 companies over the last three years. The bet is straightforward: some proportion of people who do the course and find it genuinely useful will later want help building something, and will come back. Most will not, and that is fine, and nobody is going to chase you.

That is the whole commercial logic. There is no card, nothing renews, and the Foundations course does not quietly expire in ninety days.

For bookkeepers there is a second layer, and it is worth naming. When you gift the course to a client, they get something real for nothing, and you are the person who gave it to them. That is a better reason to be in touch than a newsletter.

How to start

Create a free account on whichever page fits you. First name, last name, work email, and your LinkedIn profile. That is it.

If you are the CEO of a company with ten or more employees, sign up first, then request the CEO access from the offer bar inside your account. It does not happen automatically, and it is easy to miss. The request goes to their support team and to us.

Bookkeepers and anyone working on books: humansplusagents.ai/legacy-sbc

Everyone else: humansplusagents.ai/legacy-sbc-aia

Everything else on the table right now

This is not the only thing we have open, so here is the whole list in one place.

A free AI agent, built and yours to keep. If a whole course is not where you want to start, claim an agent instead. One job you actually do, automated, built for you. No conditions attached to it, and you keep it whether or not you ever do anything else with us.

The community. The AI for Accounting group is free, open, and nothing is sold in it. If you want to see what people are actually building before you commit an hour to anything, that is the room to sit in.

The nine-belt course, free. Which is what this whole post was about.

Now the thing I want to say plainly, because I have spent a lot of this month writing about deadlines. Nothing on that list expires. There is no countdown, no closing date, and nothing changes if you come to it in November instead of this afternoon. I would rather you started when it is genuinely useful to you than because a timer told you to.

If you are trying to work out whether it is worth your time, do not start from the price. Start from the sentence.

"The job I am sick of doing by hand is ______."

Or, if you are thinking about who else this is for:

"The first person I would help with this is ______."

If you can finish either one, you already know what you would build in the first week. If you cannot, the course is still free, and it will still be free when you can.

Yvonne Razo is the founder of Legacy Small Business Consulting, CFO of Simple Academy, and co-founder of AI for Accounting. The community is free and open, and stays that way.

Monday, September 21, 2026

Reasons not to come

I have spent three weeks asking people to register for a free week of AI training.

Today I want to do the opposite, because I have now watched enough of these to know who finishes and who quietly disappears on Wednesday, and I would rather say the unhelpful part out loud than have another forty people sign up for something that was never going to suit them.

So here is the honest list. If any of these is you, do not come. I mean it, and I will not think less of you for it.

Do not come if you want to watch.

There is no lecture. Steve puts a real piece of work in front of you and you build something with it while he is there. If you were planning to have it on in another tab while you clear your inbox, you will get precisely nothing from the hour and you will feel vaguely worse about AI at the end of the week than you did at the start.

Do not come if you do not have one specific job in mind.

This is the big one. "I want to learn AI" is not a goal, it is a mood. The people who finish the week arrive with one tedious, recurring, specific job that they privately resent, and they leave with the thing that does it. The people who drift off on Wednesday arrived to learn AI in general and discovered on Tuesday that general is not a thing you can build.

If you cannot finish the sentence "the job I am sick of doing by hand is ______", stay where you are until you can.

And if you are coming at this from the other direction, because you want to build an AI consulting business rather than fix your own back office, the sentence is different but the rule is the same. Finish this one instead: "the first person I would help with this is ______". Naming an actual human being, not a market.

Do not come if you cannot protect the hour.

Everybody is busy, so that is not what I mean. I mean that if you already know, today, that Wednesday and Thursday are going to get eaten, then you are registering for two days of a five day thing. Yes, it is recorded. I will also tell you plainly that most people do not watch the recording. I have the numbers and they are not flattering to any of us.

Do not come if you are trying to work out who you can let go.

You will be disappointed, and more importantly you will be wrong. Everything I have seen says the same thing: the typing compresses and the judgment does not. If you arrive looking for a headcount answer you will spend five hours getting an answer you did not want, from a room full of people who are using the same tools to take on more work rather than less.

Do not come if you want to be told AI is overhyped.

You will not get that. You also will not get the opposite. Both of those camps tend to leave unhappy, because the week is about a much less exciting middle position: these things are genuinely useful and they are confidently wrong often enough that someone competent has to stay in the chair.

And do not come if you are already doing this well.

If you have already got these tools handling your repetitive work, and you know from experience where they lie to you, then you do not need the week and I am not going to pretend otherwise. Go and show somebody who does not. That is a better use of your Tuesday than sitting through something you have already worked out.

Who it is actually for

What is left, after all that, is not one person. It is five, and they get very different things out of the same five hours.

The owner. Somebody running a practice or a business who does the same handful of jobs every month and has a nagging sense that the way they are doing them is about to look old.

The CEO. Somebody who is not going to do this work personally but has to decide what the organisation does about it, and who would rather form that view by building something once than by reading another vendor's deck.

The CFO. Somebody who has to put a number on it. You will not get a business case handed to you, but you will get the only thing a business case is actually made of, which is a realistic sense of what these tools do well and where they quietly fail.

The person who is not the boss. Somebody with a job who wants to be better at it, or who leads a team and wants to bring them with her. I want to be direct about this one, because nobody else will be: your employer is probably not going to train you in this. Most organisations are still deciding what their policy is, and while they decide, the people who get good at these tools are getting good on their own time. Being the person on the team who can do this, and who knows where it goes wrong, is a real advantage right now. In about two years it will be an ordinary expectation and worth nothing. That window is open and it is not going to stay open.

And if you lead a team, the same thing applies with a sharper edge. You cannot supervise work you have never done. If your people start using these tools and you have never built anything with one, you will not be able to tell their good output from their plausible output, which is the only part of the job that will still be yours.

And the person starting something. This is the one I did not expect and now see every week. People who want to build an AI consulting practice of their own, some of them between jobs, some of them still in one. The week is genuinely useful to them, possibly more than to anyone else, because you cannot sell this to somebody else until you have built one thing with your own hands and watched it be wrong.

What those five have in common is short. Willing to be a beginner in front of other people for an hour. And able to be irritated by a machine without concluding the machine is useless, because you will be irritated by it, probably on day two.

That is the whole bar. It is not a technical one, and the people who clear it come out of the week with something working and a much better sense of when to argue with it.

AI does the typing. You make the judgment calls. Five hours will not make you an expert at the first half. It gives you enough practice at the second half to stop accepting the first answer, which is the only part that has ever mattered.

The part with a date on it

Monday 28 September to Friday 2 October. One hour a day, live at 3 PM Central, every session recorded.

Registration closes Friday 25 September, which is four days from today.

There was going to be a week in October. There is not any more, so this is the last one, and I would rather say that plainly than dress it up.

No card, nothing renews, and nothing is sold to you during it.

Register here: https://humansplusagents.ai/legacy-sbc-aia

You do not need to own anything, run anything, or have anyone's permission to come.

If you read the list above and recognised yourself in it, close this and get on with your Monday. Genuinely. I would rather have thirty people who finish than a hundred who register out of a vague feeling that they ought to.

And if you read the list and thought "none of those is me, I know exactly what I would build" or "I know exactly who I would help first", then you are the person I wrote the other three weeks of posts for, and you have got four days.

Yvonne Razo is the founder of Legacy Small Business Consulting, CFO of Simple Academy, and co-founder of AI for Accounting. If a whole week is not realistic, claim a free agent instead and keep it either way. The community is free and open, and will stay that way long after the last free week has gone.

Friday, September 18, 2026

Things I have said out loud to my computer this week

I work alone a lot of the day, which means nobody is around to hear me talking to a machine.

I have decided this is fine. I have also decided to write some of it down, because looking at the list I realised it was not a list of complaints. It was a transcript of someone doing their job.

Here is the week, more or less in order.

"That is not what I asked."

Said at least daily. Usually about something that was, in fairness, quite good. Just not the thing.

"Where did you get that number?"

The single most important sentence in this whole business. I say it to AI. I say it to staff. I have said it to a bank. It is the question that separates a professional from an audience.

"I did not ask you to be confident."

This one I say with real feeling. The tone is always so certain. The certainty is not attached to anything.

"Yes, but which client."

Because it will happily give you a beautiful, general, entirely useless answer about all clients everywhere.

"No."

Just that. No elaboration. It is surprisingly effective and deeply satisfying.

"Show me the source or take it out."

This is the one I would tattoo on the inside of my eyelids if I could. If it cannot tell you where something came from, it does not go in a file with your name on it.

"Nobody talks like that."

Said about a client email it drafted for me, which used the word "leverage" twice and "synergies" once. I do not employ people who write like that and I am not going to start now.

"You made that up, didn't you."

It had. It was very sorry about it. Being sorry does not unfile a return.

"Stop apologising and just fix it."

See above.

"Okay. That one was actually good."

Said quietly. Grudgingly. Once, possibly twice.

And finally, the one I will defend to my last breath:

"Thank you."

I have said thank you to it more than once this week. I am aware of how that sounds. I do not wish to discuss it.

One more confession, and then I will stop

I call it Claudette when I am happy with it.

I call it Claude when I am not.

I did not decide to do this. I noticed I was doing it, somewhere around Wednesday, and then I could not stop noticing. Claudette gets the good work, the draft that came back better than I had any right to expect, the thing I actually used. Claude gets the made-up figure and the email full of synergies.

I am not going to defend this either. But I will say that if you have ever caught yourself using a different tone of voice with a machine depending on whether it just helped you or just wasted twenty minutes of your life, you already understand the entire relationship. It is the same voice you use with a talented junior who is having an off week.

Here is what I noticed when I read the list back

Not one of those sentences is about typing.

Every single one is a judgment call. Is that the right number. Is that the right client. Does that sound like a human being wrote it. Where did that come from. Is that good enough to go out with my name on it.

That is not a list of things going wrong. That is a transcript of supervision, which happens to be the actual job now.

I have watched a lot of people meet these tools for the first time, and the ones who struggle are almost never the ones who cannot write a clever prompt. They are the ones who accept the first answer. They read something confident and well-formatted and they assume confident and well-formatted means correct, because for their whole working life those two things have travelled together.

They do not travel together any more. That is the entire adjustment.

AI does the typing. You make the judgment calls.

And the uncomfortable truth is you cannot learn to make those calls by reading about them. You learn it the way you learned to review a junior's work, which is by looking at a great deal of work that is nearly right and developing a nose for the nearly.

Which brings me to the thing with a date on it

We run a free week of live AI training with Steve Cunningham, and it is the last one.

Monday 28 September to Friday 2 October. One hour a day, live at 3 PM Central, every session recorded.

There was going to be a week in October. There is not any more. So this is genuinely it, and I would rather say that plainly than dress it up.

The good news, and the reason I am writing this on a Friday instead of panicking: we have had so much coming in this week that registration stays open right through to Friday 25 September. You have a full week. Nobody is being hustled.

What actually happens in it: Steve puts a real piece of work in front of you and you build something. Not a webinar about AI. Not forty slides on the future of the profession. You bring one job you are tired of doing by hand and you leave with the thing that does it, and more to the point, with some idea of when to argue with it.

No card, nothing renews, and nothing is sold to you during it.

Register here: https://humansplusagents.ai/legacy-sbc-aia

One suggestion, since I have been on both sides of this. Do not come to learn AI. Come with one specific, tedious, recurring job that you privately resent. That is the difference between the people who finish the week and the people who drift off on Wednesday.

And if anyone catches you saying thank you to it, send them this.

Thursday, September 17, 2026

The quiet ones registered. The enthusiastic ones did not.

I have been keeping score this week, and the result is not what I expected.

Three people told me they were signing up for the free week. Not vaguely. One wrote "I am going to sign up, thank you." Another said "will do." A third said "sure." All three meant it at the moment they typed it. I have no doubt about that.

None of them has registered.

Two other people registered. Neither of them said a word to me about it first. One of them signed up four hours after she heard the week existed, and I only found out because the confirmation landed in my inbox. She never told me she was thinking about it. She just did it and got on with her day.

That is a small sample and I am not going to pretend it is a study. But I have watched this shape often enough now to stop treating it as a coincidence.

Saying you will do a thing is not the first step toward doing it. For a lot of us, it is the thing we do instead.

Why announcing it makes it less likely

There is a decent body of work on this, and the uncomfortable version goes like this: telling someone your intention gives you some of the satisfaction of having acted on it. The social credit arrives early. You have already been seen as the sort of person who is going to learn this, fix this, start this. The part of you that wanted to be seen that way is now fed.

And the actual task, which is still sitting there untouched, has quietly become less urgent.

I am not above this. I have a list of things I have told people I was going to do this year, out loud, with enthusiasm, and a shorter list of things I have actually done. The two lists barely overlap. The things I finished are almost all things I never mentioned to anyone until they were finished.

If that describes you too, then the useful thing is not to feel bad about it. The useful thing is to notice that the moment you feel like replying "I'll sign up" is precisely the moment to sign up instead, because the reply is going to cost you the doing.

Now the part where I have to eat my own argument

On Monday I published a piece arguing that a thing with no cost has no deadline, and that free offers collect people who are never going to turn up because nothing forces the issue.

That same evening, we cancelled the October week.

Which meant that a few hours after publishing that argument, I had sent nearly four thousand people an email telling them there were two free weeks left and that they had until the second of October to register for the second one. That is now wrong. It was wrong before the day was out.

I am sending a correction today, and I want to be straight about why, because the easy thing would have been to let it go and hope nobody had picked October.

Some of you did pick October. You read that email, decided the later date suited you better, and put it out of your mind. On the twenty-fifth of September you would have gone looking for a week that does not exist, and you would have found a closed door and no explanation.

I made the exact mistake I had spent that morning writing about. I gave people a date far enough away to be comfortable, and comfortable dates are the ones nobody acts on.

So the correction is not just housekeeping. It is the argument, demonstrated on myself, in public, four hours after I made it.

What is actually true now

There is one free week left.

Monday 28 September to Friday 2 October. Registration closes Friday 25 September.

There is nothing behind it. No October week, no November week, no quiet relaunch in the new year under a different name. We are ending them because they have done their job, and because a free thing that never ends stops being an offer and becomes furniture.

An hour a day, five days, live at 3 PM Central. Every session recorded, so if you are in India or the Philippines or the UK and 3 PM Central is the middle of your night, register anyway — registering is what gets you the replay. That is the actual reason to register, not attendance. I have watched too many people in other time zones quietly decide it was not for them because they could not make the live hour.

No card, nothing renews, and nothing is sold to you during it.

Who it is for, and it is broader than it looks

I specialise in accounting, so most people assume this is a bookkeepers' week. It is not.

It is for the business owner who is the finance department after everyone goes home. The CEO who wants to understand what these tools actually do before someone sells her a platform. The operations person who has sixteen spreadsheets and a recurring headache. The staff member whose job involves a great deal of typing that a machine could do while she checks it.

If your work contains a repetitive layer that you have privately accepted as the cost of the job, the week is for you.

The one thing I would ask

Do not reply to this to tell me you are coming.

I mean that kindly. I like hearing from people and I read everything. But the reply is the trap. It will feel like progress, it will feel like a decision, and then the twenty-fifth will arrive and the decision will turn out to have been a feeling.

If you want the week, take the thirty seconds now and register. Then tell me, if you like, and I will be glad to hear it, because at that point the telling costs you nothing.

Register here: https://humansplusagents.ai/legacy-sbc-aia

Registration closes Friday 25 September. The week runs 28 September to 2 October. It is the last one.

Yvonne Razo is the founder of Legacy Small Business Consulting, CFO of Simple Academy, and co-founder of AI for Accounting. If a whole week is genuinely not realistic, claim a free agent instead — it is yours to keep either way. The community stays free and open long after the last free week has gone.