Build vs. buy AI for small firms

Open your firm’s expense report and count the AI line items. Most founders I talk to find four or five: a transcription tool someone championed in January, a writing assistant with two active seats out of ten, a chatbot pilot that quietly auto-renewed. Each one was bought to answer the question “what should we do about AI?” None of them answered it. And now there is a new decision on the table, whether to buy another product or pay someone to build something, with no CTO in the room to referee it.

What follows is the framework I use inside client firms. It is not a pitch for building.

Buy when the tool is the system of record

Some software is the place where a category of truth lives. Your accounting package, your CRM, your document management system: these are systems of record, and the problems they solve are generic. Every firm reconciles books, tracks contacts, files documents. The vendors have absorbed decades of edge cases you have never met. Never rebuild these, and be suspicious of anyone who proposes to. Nothing about your firm is distinctive enough to justify owning ledger code.

The same logic covers vertical tools. If a product built for firms like yours matches the way you actually work 80 percent or better, and you can live with its shape on the remaining 20, buy it and move on. A subscription is cheap against the true cost of maintaining software, which is a forever cost, not a launch cost.

Build when the expensive part is the seams

“Build” is the wrong word for what a 15-person firm should ever do, so let me replace it: assemble. Thin automation and AI glue laid over the tools you already own, not a software product with your name on it. That distinction matters, because the case for assembling is never “we can make a better CRM.” It is one of three narrower cases:

• The cost lives between systems. The client’s answers are in the call transcript, the pricing history in a spreadsheet, the template in the document system, and a senior person spends Thursday moving it all by hand. No vendor sells your particular seam. Glue is the only thing that fits it.
• The workflow is your advantage. If the way you scope, review, or deliver work is the reason clients pay your rates, renting a generic version of that workflow flattens exactly the thing you charge a premium for.
• Off-the-shelf forces your process into someone else’s shape. When adopting a tool means retraining the firm to work the way the tool prefers, you pay twice: the subscription, and the process you gave up to fit it.

On one line, the ledger reads like this.

The real failure mode is neither

In firms your size, the common outcome is not a bad build or a bad purchase. It is subscriptions bought as a substitute for a decision. A tool gets adopted because deciding felt urgent and buying felt like deciding: no one mapped the workflow it was meant to change, no one owned adoption, and six months later it is a line item with two log-ins. The graveyard on your expense report is not evidence that AI does not work for firms like yours. It is evidence that nothing was ever implemented.

Buying a subscription feels like making a decision. Usually it is a way of postponing one.”

What implementation actually means

Here is the part the vendors skip: AI implementation, for a small business, is mostly not a technology problem. Model selection is perhaps a tenth of the work. The rest is workflow mapping, deciding who checks the machine’s first draft, what happens when it is wrong, and getting busy fee-earners to change how they work on a Tuesday. Adoption is where most small-firm AI actually dies, whether it was bought or built; I wrote about that gap in why most small-firm AI dies before production.

A tool nobody uses and a system nobody finished fail identically. Quietly.

Who should own the decision

In a founder-led firm of 5–30 people, somewhere between $1M and $10M in revenue, with no CTO or Head of AI, the build vs. buy call defaults to the founder. It should. The decision is about where your firm’s advantage lives, and nobody else can answer that. What you should not carry alone is the technical translation underneath it: which seams are cheap to automate, which are deceptively expensive, and what is safe to do with client data.

That translation is most of what I do as a Fractional AI Officer. It starts with a $5,000 two-week diagnostic: workflow interviews, an audit of what you already pay for, and an ROI-ranked roadmap you keep whichever way you decide. Where the roadmap says buy, you buy, and you cancel what the roadmap says is dead weight. Where it says assemble, a 90-day engagement ships roughly three production systems on your stack, each with monitoring, an evaluation suite, a kill switch you control, and a one-page runbook, ending in full IP transfer and your team trained to run them. What that costs sits in what a Fractional AI Officer costs; what the role covers, in what a Fractional AI Officer does.

However you resolve it, resolve it on purpose. The firms that get value from AI are not the ones that guessed right between building and buying. They are the ones where somebody actually decided, mapped the workflow, and stayed accountable for the sixty days after the invoice. The rest bought another subscription.