Should My Company Build or Buy Its AI Agents?

Buy the AI for every process that runs the same way in every company, like receipt capture, meeting notes, support triage and invoice reading, and build your own agents only for the 1 to 3 processes where your rules, your data or your customers are the reason you win. Build means your own agent on a model you rent from Anthropic or OpenAI, owned by a named person inside the company, and for most 30 to 500 person companies it ends up as about 4 processes bought for every 1 built.

Why companies get this wrong

Most companies make the call once, for the whole company, in a board meeting. Then they get one of two bad outcomes.

The first is buying a tool per department. Sales buys one, finance buys one, support buys one, HR buys two. A year later I usually find 8 to 12 AI subscriptions, each holding a slice of the company's data, none of them talking to each other. The CEO still can't get a straight answer to which customers are profitable, because that question crosses four of those tools.

The second is building everything. A CTO hires two engineers and spends 9 months rebuilding meeting notes and invoice reading, which you could rent for less per month than one engineer's day rate. The process that would have moved the numbers, say pricing approvals or the cash forecast, is still on the list for next quarter.

I make the call one process at a time. Do that and most of the argument goes away.

The four questions I ask about each process

I list every process a client wants to hand to AI, usually 15 to 30 of them, and put each one through the same four questions. It takes about a day. Here's where the answers usually land.

1. Would a competitor run this process the same way?

If yes, buy. Your expense receipts aren't why customers pick you. A supplier who serves 10,000 companies will keep that tool better than you will, and you get every improvement for free. I don't spend a week of build time on anything a competitor would do identically.

2. Does it run on rules or data only you have?

If yes, build. Your discount rules, how you score a deal, which customers get credit and on what terms: that's where a company makes or loses its margin. A bought tool makes you fit its idea of the process. Your own agent follows yours, and you can change the rule on a Tuesday without raising a ticket with a supplier.

3. Does it cross 3 or more systems?

Then build the agent that does the work and buy the parts it uses. The close is the clearest case. It touches the bank, the billing system, the ledger and the payroll export. No single bought tool sees all four, so you end up with a person copying between them. I build one agent that reads all four and buy the pieces that are the same everywhere, like bank feeds and invoice reading.

4. Who owns it on day 300?

Every agent you build needs a named person who spends 2 to 4 hours a week keeping it right. If nobody inside can take that, buy the process or leave it alone for now. A built agent with no owner drifts within a quarter, and it's worse than a bought tool with no owner, because nobody outside is maintaining it either.

5. Look at the list again every 6 months

What you built last year might be a product this year. When a supplier does a process well enough, I switch the client over and move the build time to the next process that matters. Nothing on the built list is permanent. The only things that stay built are the ones a competitor couldn't copy.

What build actually means in 2026

Nobody I work with trains their own model. Build means writing your own agent: its instructions, its rules, its tests, its connections into your systems and the log of everything it decides. The model underneath is rented from Anthropic or OpenAI and can be swapped. I write every agent so the model can change in a day, because prices and quality move every few months.

This matters for valuation. A buyer doing diligence doesn't give you credit for 12 AI subscriptions. Anyone can buy those the week after completion. They will look hard at a pricing agent that has run on your rules for 18 months, with a log of every decision and a person who runs it. That's an asset, and it's yours.

From my operating seat

Inside one client I run, the cash runway is live every morning. I built that agent, because the forecast depends on how they bill, when their biggest customers pay and which costs move with headcount. No bought tool knew those rules. The receipt capture and bank feeds underneath it are bought, and I'd never build them. Building the first version took about 6 weeks, and the finance lead owns it now.

I've spent twenty years in the C-suite, with three exits and three FT100 listings. In every sale the buyer asked the same thing about our systems: what do you own, and what could we just buy? The companies that sold well had a short, clear answer to the first half of that question. With AI agents, I'd want that answer ready before anyone asks.

What does it cost to build an AI agent in house?

In my builds, one agent for one process takes 4 to 8 weeks to reach production, including 2 to 4 weeks running in shadow next to the people who do the work today. After that it needs 2 to 4 hours a week from the person who owns it. The model bill is the small line. The owner's time is the big one, so price that before you price the software.

Build vs buy AI talent: should I hire an AI team or bring in an operator?

Hire a team once you have 3 or more agents live and a list of the next 5. Before that, an in house team has nothing to run and spends its first 6 months choosing tools. I'd bring in an operator, fractional, to build the first 2 or 3 agents and hire the person who'll own them, then hand over with a runbook for each one.

Do I own an AI agent if it runs on someone else's model?

You own what matters: the instructions, the rules, the tests, the connections to your systems and the record of every decision it made. The model underneath is rented and swappable. I write every agent so the model can be changed in a day. If a supplier holds your rules and your logs, you've bought a tool, whatever the contract calls it.

Where I come in

This is what I do inside companies. I sit down with your leadership team, sort every process into buy or build in a day, build the 1 to 3 agents that run on your own rules, and hand each one to a named owner with a runbook. I do it with the exit in mind, so what you build shows up as something a buyer pays for. Which of your processes would a competitor run exactly the way you do? See how I work as a fractional CFO and AI operator, or start at meethayat.com.