Monday, June 29, 2026

Eyes and Doors: Why Telling the Truth About AI Is Opening Both


This morning we published a piece and opened enrollment for a program that teaches people to build their own AI agents. By the afternoon, the reactions had told a bigger story than the launch itself. So I want to write about the reactions, because they say something about why this approach is working when so much AI marketing is not.

Two things happened today, side by side. People's eyes opened. And doors opened. Here is what I mean.

The eyes

The idea we keep coming back to is simple and a little uncomfortable: AI now does the volume, and it does some of that volume confidently wrong. The skill that matters is no longer doing the thousand, it is catching the five that look clean but are not. When you say that plainly, something happens. People stop arguing about whether AI is good or bad and start recognizing their own work in it.

A partner who resells our program put it better than I did in a public comment today. A loan booked as revenue does not look like an error, he wrote, it looks clean until someone with business context catches it. Then he pushed the idea somewhere I had not fully said out loud: if catching the five becomes the job, it can quietly become the heaviest, least rewarded version of the work, all responsibility and none of the creation, with the strain invisible until your best person resigns. His point was not that human judgment is overrated. It was that we have to actually resource it. Spread the review load, cap the volume, protect and reward the expert doing the catching.

That is the kind of comment you only get when the underlying idea is true. People do not deepen marketing. They deepen a real idea. And a tax-transformation leader at a global firm said the same thing from his seat this morning: the review layer is what ultimately drives trust in AI for tax. Different chair, same recognition.

The doors

Here is the part I did not expect. Honesty about the hard parts did not slow conversations down. It opened them.

A bookkeeper named Anne had described a task that ate her afternoons, converting client emails to PDFs one at a time. We built her a tool. She tested it today and converted a hundred emails to clean PDFs in about two minutes. Her words: the PDFs are excellent, and I am very happy. That is one person, one task, one afternoon handed back. But it is also the whole thesis in miniature: let the machine do the volume, keep the human pointed at the judgment.

And it kept going all day. Firm owners, audit partners, tax pros, software people, even a few peers who build in this space reached out to compare notes or to ask how they could build alongside us. A fintech partner opened a referral conversation. People booked calls. None of that happens when your message is "the robot will replace you." It happens when your message is "the robot will do the typing, and here is how you become more valuable, not less." One promise closes doors. The other opens them.

Why this works

I think the reason is almost boring. We are not selling fear, and we are not selling magic. We are telling accountants and bookkeepers the truth: AI is powerful and it is fallible, your expertise is the safeguard, and the smartest move is to stop renting tools and learn to build your own, with the review step built in and the human in the driver's seat.

That message respects the person on the other end. It opens their eyes because it matches what they are actually seeing in their own files. And it opens doors because respect is the thing people walk toward. The eyes and the doors are the same move. When you tell people the truth about where the value is going, they lean in, and then they want to come build with you.

If today's conversations resonate

AI does the typing. You make the judgment calls. Say that honestly, resource the people doing the catching, and the eyes and the doors open on their own.

Yvonne

Become an Elite Operator, or Be Replaced by One


For the last few weeks I have written about a single idea: AI now does the typing, and your edge is the judgment to catch the entries it gets confidently wrong. That is the defensive skill, and every firm needs it now.

This post is about the offensive one. Because catching AI's mistakes is where you protect your value. Building your own AI agents is where you multiply it. The people who pull ahead this decade will not be the ones who rent a few clever tools. They will be the ones who can look at any task that eats their week and build the machine that does it, with the review step baked in. We decided to teach that, end to end, and to put real weight behind it.

So here is what we built, and who is teaching it.

One enrollment, two programs

The offer is simple. One enrollment. One thousand dollars, one time. You get two things, and they are built to work together.

  1. The AI Agent Sprint, a live, hands-on program where you build working agents alongside an instructor, with lifetime access.
  2. The Black Belt course (AI Workflow Engineering), a $1,000 value, included. Our complete self-paced methodology, nine belt levels, yours to keep. It is the foundation that makes the live build click.

Here is the value math, plainly. The Black Belt course is worth the full one thousand dollars on its own. Your one thousand dollar enrollment covers it, and the lifetime live Sprint, three weeks every month for as long as you want, is included on top. Two programs, one price.

Who is teaching it

This is the part I am most excited to announce. The Sprint is led by Steve Cunningham.

If you do not know Steve, here is why it matters. He has spent fifteen years as the founder and CEO of Readitfor.me, turning the world's best business and personal development books into beautiful short-form videos and workshops for organizations around the world. He is the founder of Simple Academy, the number one AI productivity company, and the founder of the AI ROI Association. In other words, Steve has spent his entire career on one question: how do you take something complex and teach it so people can actually use it on Monday morning. The AI Workflow Engineering method is the product of that career.

Most "learn AI" programs give you one half of the equation. They hand you a technologist who has never closed a month, or a CPA who has never built a workflow. The Sprint gives you both. Steve leads the build. I am in the room as the accounting brain, so that anyone working on bookkeeping, tax, or accounting agents gets them built right for the actual work, not just clever-sounding prompts.

Included: the Black Belt (AI Workflow Engineering)

This is the complete methodology, from first principles to mastery, and it comes with your enrollment. Nine modules take you from using AI casually to thinking like a workflow engineer, Grey belt through Black:

  1. The AI Productivity Toolkit (Grey): https://aiaccounting.legacysbc.com/ai-workflow-engineering/1-foundation-concepts-grey-belt
  2. Understanding Generative AI (White): https://aiaccounting.legacysbc.com/ai-workflow-engineering/2-core-principles-white-belt
  3. Goals and Action Plans (Yellow): https://aiaccounting.legacysbc.com/ai-workflow-engineering/3-workflow-design-yellow-belt
  4. Adaptive Learning (Orange): https://aiaccounting.legacysbc.com/ai-workflow-engineering/4-advanced-workflows-orange-belt
  5. Critical Thinking (Green): https://aiaccounting.legacysbc.com/ai-workflow-engineering/5-system-design-green-belt
  6. From Role Ownership to Process Ownership (Blue): https://aiaccounting.legacysbc.com/ai-workflow-engineering/6-system-integration-blue-belt
  7. Creating AI Workflows (Purple): https://aiaccounting.legacysbc.com/ai-workflow-engineering/7-advanced-orchestration-purple-belt
  8. Multi-Role Processes (Brown): https://aiaccounting.legacysbc.com/ai-workflow-engineering/8-enterprise-patterns-brown-belt
  9. Reverse Engineering Tools and Deliverables (Black): https://aiaccounting.legacysbc.com/ai-workflow-engineering/9-complete-mastery-black-belt-1

Work through it and you will understand AI and workflows at a level most people in your field do not. That is the thinking. The live Sprint is where you turn it into agents you build yourself.

The AI Agent Sprint: lifetime access, live with Steve

Here is the live program, plainly. For your one-time enrollment you get lifetime access to a class that runs three weeks of every month, every month, for as long as you want to come back.

  • Live and instructor-led on Zoom with Steve.
  • Monday through Thursday, one hour a day, with daily homework so you actually ship.
  • Three weeks every month, forever. Enroll once, attend as many times as you like.
  • Each Monday-to-Thursday week is a full build cycle, so you can join any week and still walk out with a working agent:
    • Monday: pick the task that eats your week.
    • Tuesday: write the intake and the prompt that drives the agent.
    • Wednesday: build it and test it on real data.
    • Thursday: add your review step, ship it, and demo it.

You do not watch someone else build. You build, with help, until it works on your own files. And because lifetime access means you can run the live class three weeks a month for years, the value of a one-time price keeps compounding long after you enroll.

The next cohort starts Monday, July 6. July sessions run July 6 to 9, July 13 to 16, and July 20 to 23.

Who it is for

Anyone who wants to become an operator. You do not need to be an accountant, a coder, or technical. If you have ever thought "a machine should be doing this," this is where you learn to build that machine. Accountants, bookkeepers, tax pros, and firm owners get an extra edge, because I am in the room to make sure the accounting is sound. But the Sprint is for anyone in any role who is ready to stop doing the busywork by hand.

How to join

  1. Enroll in the AI Agent Sprint and join the July 6 cohort with Steve. Your enrollment includes the full self-paced Black Belt course (a $1,000 value): https://aiaccounting.legacysbc.com/sprint
  2. Want to see what finished agents feel like first? Claim one free and point it at a real task: https://aiaccounting.legacysbc.com/claim-agent
  3. Trade notes with people already doing this in our free community: https://aiaccounting.legacysbc.com/community

AI does the typing. You make the judgment calls. And now, if you want it, you can learn to build the whole machine. Become an elite operator, or be replaced by one. We would rather you became one, and we will teach you how.

Yvonne

Sunday, June 28, 2026

Confidently Wrong: The One Skill Every Firm Needs Now That AI Codes the Books


Here is a scene that is already playing out in firms everywhere. An AI tool codes a thousand transactions in about a minute. It is fast, it is tireless, and it is mostly right. Then five of those entries are wrong. Not obviously wrong. Confidently wrong. They look exactly as clean and reasonable as the 995 it nailed.

That gap, between fast-and-mostly-right and quietly-wrong, is the most important thing happening in accounting right now. And it points to a new skill that every firm needs to build immediately.

Why AI gets these wrong

The model is not dumb. It lacks business context. An AI tool sees a cash inflow and defaults to revenue, because in the absence of any other information, money coming in usually is revenue. It has no idea that this particular deposit was a loan draw, an owner contribution, or a transfer between two of your own accounts. That knowledge does not live in the transaction. It lives in the business.

So the machine fills the gap with the statistically obvious answer, states it with total confidence, and moves on. Multiply that across a thousand rows and you get books that are 99.5 percent right and 100 percent untrustworthy until someone checks.

The five it gets wrong

These are the errors I see again and again when AI codes a set of books without a human in the loop:

  1. Owner draws coded as income or expense, when they belong in equity.
  2. Account transfers double counted, inflating both sides of the activity.
  3. Loan proceeds booked as revenue, turning a liability into phantom income.
  4. Personal spend swept into deductions, which is a problem the moment it touches a return.
  5. Big purchases expensed instead of capitalized, which quietly distorts the balance sheet and the depreciation schedule.

None of these are rounding errors. Each one flows straight into the financials and, eventually, onto a tax return.

The skill that actually matters now

Here is the shift, and it is a big one. For decades the entry-level accounting skill was data entry: key it in, key it accurately, key it fast. AI just took that job. What it cannot take is the judgment to look at a confident, clean-looking entry and know it is wrong.

So the new skill, the one to train your team on starting today, is spotting confident but incorrect output. It is less "can you enter a thousand transactions" and more "can you catch the five that look right but are not." That is a higher-value skill, it is harder to automate, and it is exactly where your people become more valuable, not less.

The firms that struggle will be the ones that paste AI output straight into the books. The firms that pull ahead will be the ones that let AI do the thousand and train their people to catch the five.

How we built this thinking into our tools

This is the whole philosophy behind what we launched on Friday, and it is why every agent we build keeps you in the driver's seat rather than pretending to replace you.

Our two new tax agents are built around exactly this principle. The Depreciation Schedule Builder will compute straight line, declining balance, and full MACRS across your entire asset list in seconds, and carry forward prior depreciation, but it leaves the judgment calls (the conventions, the elections, what counts as a capital asset in the first place) to you. The Cost Segregation Estimator flags where building components could be reclassified to accelerate depreciation, then hands you a defensible estimate to review, not a black-box answer to trust blindly. Both do the typing. You make the calls.

We also launched the Practice Growth Bundle for firms that want to turn this efficiency into more and better clients, and the AI Agent Sprint, a program where you learn to build your own AI agents, with the review step baked in, for your own workflows. Because the goal was never to hand your books to a robot. It was to let the robot do the thousand so your team can master the five.

Try it on your own books

The fastest way to understand this is to watch it happen on real data.

AI does the typing. You make the judgment calls. The firms that internalize that one sentence will spend the next decade getting stronger, not nervous.

Yvonne

Friday, June 26, 2026

Everything We Are Releasing Today: Two New Agents, a New Bundle, and the Sprint

Today is a big one. We are shipping four things at once: two new tax agents, a new bundle for growing your firm, and the training a lot of you have been asking for. Here is the three minute tour, and then the details.

Watch on YouTube: https://youtu.be/W-PGhJ4jsCY

If you would rather read than watch, everything in the video is below.

Agent #121: Depreciation Schedule Builder

Hand it your fixed assets and it builds a clean depreciation schedule in seconds, asset by asset, with accumulated depreciation and net book value calculated for you. There are three ways to get your data in: upload or paste a CSV, import a PDF of a prior schedule and let it read the assets, or enter assets one at a time in the form.

It handles the methods that actually matter: straight line, declining balance at 150 and 200 percent, and full MACRS using the IRS tables. It also carries forward prior accumulated depreciation, so a mid-life asset picks up exactly where it left off instead of starting over.

Agent #122: Cost Segregation Estimator

This is the companion to the depreciation tool, and it is where the real tax dollars live. It helps you spot the building components that can be reclassified into shorter recovery lives, five, seven, and fifteen years, instead of sitting on a 39-year schedule. The result is accelerated depreciation, which front-loads deductions and keeps your client's cash in their pocket sooner.

It gives you a fast, structured, defensible estimate, not a chatbot guess, and it flags what a full cost segregation study would confirm. It is an estimate, not a study, and not tax advice. It is the tool you use to decide whether a study is worth commissioning.

The honest part: five places AI gets depreciation wrong

Both of these tools are fast. But depreciation is one of those areas where fast and wrong live very close together, so we want to be straight with you. There are five places an AI tool will burn you if you trust it without knowing the rules:

  1. Book is not tax. MACRS is a separate system from your book schedule, with its own recovery periods and tables.
  2. The convention. MACRS defaults to half-year, but more than 40 percent of purchases in Q4 forces mid-quarter, which changes every number.
  3. Section 179 and bonus. They reduce basis before MACRS even starts, and the limits change yearly.
  4. Salvage value. Book subtracts it. MACRS ignores it entirely.
  5. Real estate. Buildings are 27.5 or 39 years on a mid-month convention, different again from everything above.

The tools do the building. You make these five calls. We wrote the full breakdown in a separate post today, "AI Did the Depreciation in 30 Seconds. Here Is the One Place It Will Burn You."

New: the Practice Growth Bundle ($249)

Your agent library already covers doing the work. This bundle covers getting and keeping the work: attract, convert, and expand. It is 14 agents in all, and the flagship is the LinkedIn Connection Engine, which productizes the exact outreach system we run here every single day. It segments your connections, drafts tailored intros, sets up a follow-up tracker, and gives you reply playbooks. The rest of the bundle fills the top of the funnel (lead magnets, niche positioning) and the bottom (referrals, reviews, expansion) so the whole "grow your firm" system is covered in one place.

The big one: the AI Agent Sprint ($1,000)

By the end of this program, you will be able to build your own AI agents. Not borrow ours. Not poke at a chatbot. Build the real thing, for your own firm and your own clients. Become an elite operator, or be replaced by one.

Here is what one enrollment gets you, for $1,000 one time:

  • The live Sprint, with lifetime access. It runs on Zoom, Monday through Thursday, one hour a day, three weeks of every month, forever. Each Monday-to-Thursday week is a complete build cycle, so you can jump in any week and walk out with a working agent. Buy once, attend as many times as you want.
  • The Black Belt course, included. Our full self-paced AI Workflow Engineering program, nine belt levels from Grey through Black, a $1,000 value, included with your enrollment. Black Belt is not sold separately. The Sprint is how you get it, and the lifetime live Sprint comes on top.

You also get both halves of the equation most AI training is missing. Steve Cunningham leads the build and teaches the method for turning a real task into a working agent. Yvonne Razo is the accounting brain in the room, so your agents are built right for actual bookkeeping, tax, and accounting work, not just clever prompts.

It is for anyone who wants to become an operator. You do not need to be an accountant, a coder, or technical. If you have ever thought "a machine should be doing this," this is where you learn to build that machine.

The next cohort starts Monday, July 6. Live weeks are July 6 to 9, 13 to 16, and 20 to 23.

How to start today

Whatever you pick, the idea is the same. Let AI do the typing. You keep the judgment.

Yvonne


 

Three Reasons Good Firms Stay Stuck — and What We Built for Each

 

Most accounting and bookkeeping firms are not stuck because the owner isn't good at the work. They're stuck for the opposite reason: the owner is so good at the work, and so buried in it, that there's no room left to grow. The skill is there. The hours are not.

When you look closely at why a capable firm plateaus, it almost always comes down to three specific gaps. They're different problems with different fixes, and confusing them is why so much effort goes nowhere — you can't market your way out of a workflow problem, and you can't automate your way into new clients. So before anything goes live today, I want to name the three gaps plainly, because once you see which one is actually yours, the path forward gets a lot simpler. Everything we're releasing today maps to one of them.

Gap 1: The growth gap

This is the firm that does excellent work and still has a feast-or-famine pipeline. Clients come from referrals and luck, marketing happens in bursts when things get slow, and the moment the owner gets busy with delivery, the outreach stops — which guarantees the next slow patch. The work is great. The system for getting more of it doesn't exist.

The trap here is thinking the answer is "do more marketing." It isn't. The answer is a repeatable engine: a steady way to start conversations, a reason for happy clients to refer and review, a lead magnet that earns attention, and clear messaging about who you're actually for. Most owners know they need this and never build it, because building it competes with billable work every single day.

That's exactly why we built the Practice Growth Bundle ($249) — a full suite of agents that runs the growth work for you: LinkedIn connection and outreach, a referral and review engine, a lead magnet plus nurture sequence, niche positioning and messaging, and more. It turns "I should really market more" into a system that runs whether or not you have a free afternoon. → https://aiaccounting.legacysbc.com/pro-bundles

Gap 2: The operating gap

This is the firm that has tried AI. The owner has used ChatGPT a few times, maybe bought a tool or two, watched some videos. It helped a little, then fizzled — because dabbling and operating are not the same thing. Dabbling is asking AI a question now and then. Operating is having AI built into how the work actually moves through your firm: intake, cleanup, reconciliations, reporting, client communication, all of it, reliably, every week.

The gap between those two states is not more tools. It's knowing how to design the workflows, where AI belongs and where your judgment has to stay in control, and how to build it so it holds up under real client volume. That's a skill, and it's learnable — but not by collecting more tips.

That's why we built the July Sprint (Live) + Black Belt (Self-Paced) — $1,000. The Sprint is live and hands-on; the Black Belt is the self-paced program you keep. One enrollment, both tracks. It's the difference between "I've messed around with AI" and "AI runs in my firm and I know exactly how." → https://aiaccounting.legacysbc.com/sprint

Gap 3: The task gap

The third gap is narrower and very specific: the individual high-value tasks that eat disproportionate hours, or that you avoid and outsource because they're fiddly. You don't need a whole program or a bundle for these. You need the one tool that does the one job, today.

Two of those came up over and over from you, so we built them — and they're now available on their own, $25 each, a la carte:

  • Depreciation Schedule Builder (#121) — a clean fixed-asset schedule, book and MACRS, in seconds instead of an afternoon with a spreadsheet.
  • Cost Segregation Estimator (#122) — size the tax benefit of a cost segregation study before you commission one, so you can have an intelligent conversation with a client about whether it's worth it.

→ Both at https://aiaccounting.legacysbc.com/build-a-bundle

These exist because you asked for them. That's worth saying plainly: two of today's releases are direct answers to what this community kept requesting.

How to tell which gap is yours

Here's the quick diagnostic. If your work is great but your pipeline is unpredictable, that's the growth gap — start with the Bundle. If you've tried AI but it never became part of how you actually operate, that's the operating gap — that's what the Sprint and Black Belt are for. And if you're mostly running fine but there are one or two specific tasks bleeding hours off your week, that's the task gap — grab the single agent that fixes it and move on.

Most firms have one dominant gap at a time. The mistake is throwing effort at the wrong one — buying tools when you have a growth problem, or chasing clients when your delivery workflow is the bottleneck. Name the real gap first. Then close it.

Start where you are

Everything above is live as of today. And if you're not sure yet, the free door is always open:

You're not stuck because you're not good enough. You're stuck because the work fills every hour you have. The fix is never working harder at it — it's building the system, the skill, or the tool that gives the hours back.

Yvonne

Thursday, June 25, 2026

AI Coded 1,000 Transactions in a Minute. Here Are the Five It Will Get Wrong.

 


Hand a year of bank activity to an AI tool and it will categorize the whole thing before your coffee is cool. A thousand transactions, sorted into accounts, with descriptions cleaned up and vendors matched. The speed is real, and for a bookkeeper buried in cleanup work, it feels like magic.

It is also exactly where a lot of books are about to go quietly wrong.

Transaction coding is one of those areas where AI is fast and mostly right, and "mostly right" is the dangerous part, because the wrong ten percent does not announce itself. It sits in the ledger looking perfectly reasonable until the financials are off, the tax return overstates income, or a reviewer starts asking why owner distributions are buried in "office expense." So let me show you, plainly, what the machine codes beautifully and the handful of places it will walk your books straight off a cliff if you trust it without checking.

What AI gets right

The bulk of coding is pattern-matching, and pattern-matching is what these tools do best. A charge from a known software vendor, a recurring utility, a familiar supplier, a payroll-processor debit — AI will recognize and categorize these correctly, consistently, across thousands of lines, faster than any human. If ninety percent of a client's activity is routine recurring spend, let the machine do that ninety percent. That was always the typing, and the typing was never the part that needed your judgment.

Where it burns you

The trouble starts the moment a transaction's correct treatment depends on something the bank feed cannot see: intent, structure, or accounting rules. Here are the five that bite, in the order they show up.

1. Owner draws and contributions coded as income or expense. Money the owner pulls out is a distribution; money they put in is a contribution or loan. Both are balance-sheet equity items, not the P&L. AI sees a transfer to a personal account and reaches for "owner expense" or, worse, leaves an inbound owner deposit sitting in income. Now profit is wrong and so is the tax return. This is the single most common AI coding error, and it inflates or deflates net income directly.

2. Transfers between accounts booked as real activity. Moving money from checking to savings, or to a credit card payoff, is not income and not an expense. It is a transfer. AI frequently codes the outbound side as an expense and the inbound side as income, double-counting cash flow and manufacturing revenue and costs that never happened. On a client with a lot of internal transfers, this can distort the P&L by tens of thousands.

3. Loan proceeds and principal payments treated as P&L. When a loan funds the account, that inflow is a liability, not income. When the loan is repaid, only the interest portion is an expense; the principal reduces the liability. AI routinely books loan proceeds as income (overstating revenue and tax) and the full loan payment as an expense (overstating deductions). Two errors, opposite directions, both wrong.

4. Personal spending coded as a deductible business expense. The bank feed cannot tell you whether that restaurant charge was a client dinner or date night. AI guesses from the merchant, and it guesses "business." Every personal charge it sweeps into deductible expense is an overstated deduction the owner cannot defend in an audit. Intent is a human call, every time.

5. Capitalize versus expense. A $4,000 laptop or a piece of equipment is a fixed asset to be capitalized and depreciated, not a one-line expense. A new roof is a capital improvement; patching the old one is a repair. AI does not apply your capitalization policy and will happily expense a major purchase, understating assets and overstating current-year expense. (And once it is mis-expensed, it never makes it onto the depreciation schedule.)

One more that belongs on the list: sales tax collected and payroll withholdings are liabilities, not income or expense. AI often codes them straight to the P&L. That is money you are holding for someone else, and it does not belong in revenue or cost.

The point underneath all of this

None of this means AI is bad at bookkeeping. It means transaction coding is the perfect example of the rule that should govern how every firm uses AI: the machine is a phenomenal first-draft engine for the mechanical work, and a dangerous final authority on the judgment.

Coding the routine recurring spend, cleaning vendor names, matching the obvious charges across thousands of lines — that is the typing, and AI should absolutely do it. Knowing that this transfer is not income, that this deposit is an owner contribution, that this $4,000 charge gets capitalized, that this dinner was personal — that is the judgment, and that is you.

The bookkeeper who gets burned is the one who imports the AI-coded file and locks the period. The one who gets ahead lets AI code the books in a minute and then spends their time on the handful of lines that actually decide whether the financials are true.

How to use it the right way

Let AI do the first pass: code the recurring and obvious activity, normalize the descriptions, and flag anything it is unsure about. Then run your eyes down the exceptions and the big-ticket items. Specifically, before you close the period, check: every transfer, every owner draw or deposit, anything tied to a loan, any charge that could be personal, and any purchase large enough to capitalize. Reconcile to the bank, confirm the balance-sheet accounts actually moved the way they should, and keep a short list of coding rules the AI does not know about your client.

That is the workflow: AI for speed on the ninety percent, you for judgment on the ten percent that holds the financials together.

Try it on your own books

If month-end coding and cleanup is the work that eats your evenings, see what a purpose-built agent does with it.

AI will code your books in a minute. Your job is knowing the five places it should not be trusted to. That is not a weakness of the tool. That is the value of you.

Yvonne

Wednesday, June 24, 2026

AI Did the Depreciation in 30 Seconds. Here Is the One Place It Will Burn You.

 

Hand a list of fixed assets to an AI tool and you will get a depreciation schedule back in about thirty seconds. Asset, cost, method, annual depreciation, accumulated depreciation, net book value, all of it, instantly. The speed is real, and it is wonderful.

It is also exactly where a lot of people are about to get burned.

Depreciation is one of those areas where AI is fast and mostly right, and "mostly right" quietly becomes "wrong on the tax return." So let me show you, as plainly as I can, what the machine handles beautifully and the handful of places where it will lead you straight off a cliff if you trust it without knowing the rules.

What AI gets right

Straight-line book depreciation is simple arithmetic. Cost minus salvage, divided by useful life, the same amount every year. An AI tool will do that perfectly, every time, across a thousand assets, faster than you can open the spreadsheet. Declining balance is a little more involved, but it is still a formula, and AI handles it.

If all you need is a clean book depreciation register for your financial statements, the machine is genuinely excellent. Let it do that. That is the typing, and the typing was never the part that needed your CPA brain.

Where it burns you

The trouble starts the moment "depreciation" means tax depreciation. Here are the traps, in the order they bite.

1. Book is not tax. This is the big one. Your financial statements use straight-line or declining balance over the asset's useful life. Your tax return uses MACRS, a completely separate system with its own recovery periods and its own IRS percentage tables. A vehicle is "5-year property" for tax no matter what useful life you put on the books. AI tools constantly blur these two, applying a book life to a tax method or quoting one when you needed the other. They are two schedules, not one.

2. The convention you did not think about. MACRS does not start depreciating on the day you bought the asset. It uses a convention. The default is the half-year convention, which assumes everything was placed in service at mid-year. But if more than forty percent of your asset purchases for the year landed in the fourth quarter, you are forced onto the mid-quarter convention instead, which changes every number. Almost no AI tool checks that forty percent test. It just assumes half-year and moves on. If your client bought heavy equipment in December, that assumption is wrong.

3. Section 179 and bonus depreciation. These let you deduct a large chunk, sometimes all, of an asset in year one. They are also where the dollars really live for your clients. Two things AI routinely gets wrong here: it does not reliably apply them, and when it does, it forgets that Section 179 and bonus reduce the asset's basis before MACRS even starts. The limits and the bonus percentage also change from year to year, and a model trained on last year's rules will confidently give you last year's answer.

4. Salvage value, in the wrong place. Book depreciation subtracts salvage value. MACRS ignores salvage entirely. Mix those up and your tax depreciation is wrong by exactly the salvage amount on every asset. It is a small rule that AI flips constantly.

5. Real estate plays by other rules. Buildings are not five-year property. Residential rental is depreciated over 27.5 years and nonresidential over 39, both on a mid-month convention that is different again from everything above. If an AI tool tries to run a building through the equipment tables, the result is not a little off. It is meaningless.

The point underneath all of this

None of this means AI is bad at depreciation. It means depreciation is the perfect example of the rule that should govern how every firm uses AI: the machine is a phenomenal first-draft engine for the mechanical work, and a dangerous final authority on the judgment.

Building the fixed-asset register, computing the book schedule, doing the arithmetic across hundreds of assets, that is the typing, and AI should absolutely do it. Knowing that this client crossed the forty percent threshold, that this asset qualifies for Section 179 but that one does not, that the building belongs on a 39-year mid-month schedule, that is the judgment, and that is you.

The accountant who gets burned is the one who pastes the AI output onto the return. The accountant who gets ahead is the one who lets AI build the register in thirty seconds and then spends their time on the five decisions that actually move the client's tax bill.

How to use it the right way

Use AI to build and maintain the fixed-asset register and the book depreciation schedule. Let it carry forward accumulated depreciation, flag fully depreciated assets, and produce the journal entries. That is hours back, every close.

Then apply the tax layer yourself, or with a tool that is honest about its limits. Run the Section 179 and bonus decisions. Check the convention. Keep book and tax depreciation reconciled, because the difference between them is a real number that flows into your deferred tax. And verify anything tax-specific against the current-year IRS rules, not last year's, and not the model's memory.

Try it on your own assets

If your fixed-asset register lives in a spreadsheet you dread updating, see what a purpose-built agent does with it.

AI will do your depreciation in thirty seconds. Your job is knowing the five places it should not be trusted to. That is not a weakness of the tool. That is the value of you.

Yvonne