How we build ·

Build it, own it, license it

Software used to be too expensive for a small business to build, so it rented. AI changed what code costs to write. The question now is whether what you build lasts long enough to be worth owning.

Who should hold the budget ↓

01 · The statement

A new line on the card

AI TOOLS MOVES EACH MONTH
A new line on the card: the same steady charges each month, and one new one whose amount moves.

Many company card statements now have a new line: what the AI coding tools cost this month. Sometimes that’s a few seats. Sometimes it’s a bill for tokens, the units these models charge by, and it moves from month to month with how much code got written.

Sooner or later someone asks what it bought. That is the old question, build or buy, about a new kind of spending. For most small businesses the answer used to be easy: buy. Building meant hiring engineers for months, so you rented what someone else had built, one subscription at a time, for as long as you used it.

02 · The accounts

What the accountants now say

THE BOOKSbooked asTokens: building the orders appASSETAI subscription, whole teamEXPENSETokens: keeping old app runningEXPENSEAN EXAMPLEone build, spread over its life$60,000Year 112,000Year 212,000Year 312,000Year 412,000Year 512,000ASU 2025-06: the asset starts once the project iscommitted and probable. Required for years beginningafter December 15, 2027.
What the accountants now say: the cost of a build, spread over the years it stands.

This year Deloitte and PwC both wrote down how to book AI spending on software, and they agree on the shape of it. When a company builds software for its own use, the cost of building it can be recorded as an asset and spread over the years the software is used, instead of counted as an expense the month it’s paid. AI tokens can be part of that cost, but Deloitte is plain that a token charge doesn’t qualify just because it can be measured. It has to be directly tied to building a specific piece of software.

So a general AI subscription the whole team uses for everything is an expense, and so are tokens spent keeping existing software running. PwC draws the same line: tools dedicated to building something specific may count; tools for general use usually don’t. In practice that means tracking which project every token went to, and what kind of work it was.

The rule about when that starts is changing too. Under a new standard, ASU 2025-06, a company starts recording the asset once it has committed to the project and it’s probable the software will be finished and used. It’s required for years beginning after December 15, 2027, and companies may adopt it sooner.

Taxes run on separate rules. Since the tax law signed July 4, 2025, US companies can deduct domestic software development in the year they spend it, or choose to spread it over at least five years (PwC, Crowe). The books and the tax return can treat the same spending differently. Your accountant decides both.

03 · Rent or own

A subscription is rent

LICENSE2027fee4,8002028fee4,8002029fee4,8002030fee4,8002031fee4,800total24,000at the end you holdnothingBUILD2027build12,0002028upkeep2,0002029upkeep2,0002030upkeep2,0002031upkeep2,000total20,000at the end you holdthe softwarean example, not a quote
Rent or own: a license fee every year, or a build once and its upkeep, and what you hold at the end of each. An example.

Software works like housing. A subscription is rent: the same payment every month, nothing to show for it when you stop, and a landlord who decides when the price goes up and what the rooms look like. Software you build is a house you own. It costs more up front, it needs looking after, and it’s yours.

Recording it as an asset and spreading the cost over its useful life is the accountant’s way of saying the same thing. A house is worth owning only for as long as it stands, and so is software. That’s the part that decides whether token spending was an investment or just a bigger bill.

04 · Useful life

How long it lasts depends on quality

THE REPOdocs/figures.mdvoice.mdsrc/lib/figure-style.tsfigure-style.test.tsthe rule, its card,and a test that can failCHANGEair at least 30% emptymeant tokeep figures plainwrong ifa wireframe passesproven byfigure-style.test.tswritten before the change
A change with its record: what it is meant to do, what would prove it wrong, and the test that checks it.

AI can write a great deal of code quickly. Code nobody on the team understands, with no tests and no record of why it does what it does, is a house built without a foundation. It stands for a while. Then the first real change breaks something nobody can find, and it’s rebuilt. On paper that spending was an asset. In fact it was rent, paid up front.

What makes software last hasn’t changed because a model typed it. We write the test before the code, so what “working” means is settled before anyone argues about it. We write down what each change is meant to do, and what would prove it wrong, before building it; we give away a kit for that. Build a small piece, show it working, and only then build the next one, which is how we run our own work.

05 · The budget

Who should hold the budget?

$ npm run buildprebuild: npm run checks &&          node scripts/build-tools.mjs"checks":scripts/check-rhythm.mjs&&scripts/check-share.mjs&&scripts/check-confirm.mjs&&scripts/check-open.ts&&scripts/check-scenes.mts&&scripts/check-figures.mts&&scripts/check-releases.tsnext buildruns only if every script above exits clean
Measured, not eyeballed: the seven scripts this site runs, in order, before every build, and the build waits on all of them.

Your token spend is a foundation pour. Everyone now has the same concrete. What differs is who is standing at the trench with the level: someone who can take a business problem all the way to working software and knows what sound software looks like, or someone who has seen it done and roughly knows how. From the street the two pours look alike. The difference shows the first time the house takes weight.

So who holds the budget matters more than how big it is. Someone who can see all of it, from the problem to the running product, spends it on one piece of software you can name, tracked by project and kind of work. That’s also what the accounting asks of an asset. Spread across everyone for everything, it isn’t one.

Measure it every week, against the business, not against how much code came out. Each week we look at the product and the engagement together and ask whether what the business wanted has moved: orders handled, hours saved, revenue. Lines of code are easy to count and say nothing about that.

06 · License

Then rent out a room

DIGITAL STUDIOOne catalog behind everythingPrices from your own formulaOffers first, a cart for the shopA daily check on works in galleriesYour own mailing listShows, calls and deadlinesbuilt with one artist:the first testGood CitizensPaulaco-ownerpart of every license
Rent out a room: the Digital Studio as a license, and the share that runs on to Paula, its co-owner.

Once you own a house that stands, you can rent out a room. Software you built well can be licensed to other businesses in your trade, and they pay you for it, the way you used to pay someone else.

We’ve done this once. We built a set of tools for Sara Paula Hoffman, a visual artist: her catalog, her prices, the questions waiting on her, the shows she might enter. She called it her digital studio. Now we offer the Digital Studio to other artists. Paula co-owns it, so part of what each artist pays goes to her.

That’s the change worth noticing. The smallest businesses in a trade, an artist, a landscaper, a shop of three, have always rented. When building costs less, they can build what fits how they work, own it, and share in what it earns when others in their trade want the same thing.

07 · Limits

What would prove us wrong

WE LOSE IFreadingwritten October 2026AI-written code does not last,whoever directs itrenting stays betternone yetAuditors read ASU 2025-06 morenarrowly than the firms didless of it countsnone yetSubscriptions get cheaper fasterthan building doesbuying wins on pricenone yetLicensing: one product, built with one artist. Thefirst test, not proof it scales.
What would prove us wrong: the three tests, written down in October 2026 with the line where we lose. There are no readings yet.

If AI-written code turns out not to last no matter who directs it, the asset argument falls apart, and renting stays the better deal. If auditors read the new guidance more narrowly than the firms that wrote it, less of the spending will count. If subscriptions keep getting cheaper faster than building does, buying wins again on price alone.

And licensing is a business, not a side effect: someone has to support the people who license it. One licensed product, built with one artist, isn’t proof that it scales. It’s the first test, and we’ll report how it goes.

Build what lasts, keep it, and when it’s good enough, let others pay to use it.

Sources

Where every rule comes from

Read on October 28, 2026. This is how the rules read, not advice for your books: ask the person who signs them.

  1. Guidance

    Technology Spotlight: Accounting for AI Costs Associated With Internal-Use Software Development

    Deloitte ·

    DART, August 3, 2026

    dart.deloitte.com

  2. Guidance

    Accounting for AI tools used in software development

    PwC ·

    Viewpoint, In depth

    viewpoint.pwc.com

  3. Article

    FASB Revises Internal-Use Software Cost Guidance

    Crowe ·

    on ASU 2025-06

    crowe.com

  4. Article

    Optionality restored to tax treatment of US research activities

    PwC ·

    on Section 174A

    pwc.com

  5. Article

    Domestic R&E Deduction Returns Under Section 174A

    Crowe ·

    crowe.com