Build, Buy, or Wait:
A Decision Tree for AI in Small Businesses

Most owners think they’re choosing between building AI and buying it. They’re usually doing a third thing: waiting. From the inside, waiting looks responsible. It feels like diligence, it avoids a vendor call, and it doesn’t show up on any budget. The build vs buy AI question is slippery, however, at a 50-person company because the default path is the one nobody prices. So let’s price all three.

The Cost of Buying AI

Buying means a seat license on a tool somebody else maintains. It’s the right call when the work is generic. Microsoft 365 Copilot Business is currently listed at $18 per user per month on an annual commitment, down from $21 as part of a promotional offer through December 31, 2026. The standard annual price is $21, and monthly billing runs $25.20 per user. At 40 seats, $18 per user per month is $8,640 a year. Against payroll, that’s a rounding error.

The real cost shows up in adoption.If you buy 40 seats and 12 people use them, your effective cost per working user triples. You also face a renewal conversation in eleven months with no evidence to bring to it. Buy when the job is generic: drafting, summarizing, cleaning up meeting notes, first-pass research. Buy when the data already lives inside that vendor’s walls, because a Copilot seat earns its keep from your Outlook and SharePoint content without anyone wiring anything.

Buying works poorly when your bottleneck is specific to how your shop runs.

The Cost of AI Build

Building means a workflow wired into your systems. It runs against your data and your rules. It earns its cost when the bottleneck is specific, repeated, and attached to a metric. For example, AI drafts quote follow-ups from deal notes in your CRM. Invoice coding is checked against three years of vendor history in QuickBooks. Inbound tickets get triaged against the way your team has resolved the same issue two hundred times before.None of those come in a box.

Price a build in three parts: discovery, build, and run. Discovery is where you realize your data is messier than anybody admitted. It’s way cheaper than finding out in week six. Build when the process is yours, the volume is worthwhile, and you can define a good month by saved dollars or hours. The Run cost is the line most forget: model usage, monitoring, and a human who owns the workflow when it starts drifting. Budget for review steps too. A workflow that drafts is a different risk profile than one that sends.

The Cost of Waiting

Waiting has a price. It’s smaller than most vendors claim and larger than most owners assume.  In the 2025 Small Business Credit Survey, 46% of employer firms said their business or employees used AI tools. Among users, only 31% reported increased sales, while 77% reported no change in labor costs.

Broad adoption has not automatically produced measurable payoff. If you’ve been holding off because the results look soft, the data backs you up. But the bill that comes due anyway in the first project. It’s always expensive – that’s when you discover which systems lack usable APIs, where your data is (really) dirty, and which team will champion change. The second project costs a fraction of the first. Firms that started two years ago are buying their third project at a discount you haven’t earned yet. The “we should automate that” list continually grows while it waits. The eventual catch-up is bigger and more rushed.

The real competitor here is inertia. It wins by default.

The Build vs Buy AI Decision Tree

Ask four questions, in order. Stop at the first one that answers you.

1. Can you name the bottleneck in a sentence, with a number attached?

“Our team spends about ten hours a week re-keying order data” qualifies. “We should use AI more” doesn’t. If you can’t finish the sentence, you’re not choosing between build and buy yet. You need an assessment first.

2. Is there a tool your team already pays for to do this job?

If yes, buy the seats and spend your energy on adoption: training, a named owner, and a check-in at 30 days. Most companies have unused capability sitting inside software they already renew.

3. Does the job depend on your data, your rules, or two systems talking to each other?

If yes, build. No seat license will learn how you price a job.

4. Is the metric from question one worth more than a year of run cost?

If no, find a different bottleneck. There’s always a better one.

Start Your AI Plan on Monday

Pick one bottleneck, write down what it costs you today, and give yourself 90 days to move it. That baseline is the whole game. It turns the build vs buy AI decision from a philosophy argument into a metric you can compare in a quarter.

Waiting is a valid option. It’s a defensible one when you do it on purpose with a date to review Taking it by accident is how three years go by.

If you want help running those four questions against your own operation, start a conversation with us.