Home/ Journal/ Memo

Build vs buy AI: should you commission a build or buy a tool?

Buy off-the-shelf for any commodity, standardized workflow (email help, transcription, generic chat, meeting notes) where a product already reaches 90 percent or more of what you need. Commission a custom build only for the one workflow that is both your actual bottleneck and specific to how your company runs, where no product can reach the part that matters. For most mid-market operators ($8M-$50M revenue), buying is the right call more often than building. The mistake is the reverse of what people fear: not over-building, but buying a generic tool that almost fits the workflow that was supposed to be the whole point. Build narrow, buy broad, and reserve the custom spend for the workflow that is genuinely yours.

"Should we build or buy AI" is the decision every mid-market operator faces once the budget is real and the vendors are circling. This memo gives the honest answer, which is buy more than you build, and a plain test for telling which side any given workflow belongs on. The commodity-versus-differentiated line, the trap of the tool that almost fits, the 90 percent test, and where a custom build actually earns its cost.

MemoJune 2026
Read time7 minutes
AudienceOwner-CEOs, COOs, Managing Partners

The honest answer, up front.

Buy more than you build. That is the answer most consulting houses will not give you, because they sell builds. For a $8M-$50M company, the large majority of the workflows you would point AI at are commodity workflows, and for commodity workflows a product almost always beats a commission on cost, speed, and risk. The custom build earns its place on exactly one kind of workflow: the one that is both your real bottleneck and specific to how your company operates. Spend custom money there, and buy everything else.

The fear in the room is usually backwards. Operators worry about over-building, about commissioning something expensive and bespoke when a tool would have done. That happens, but the more common and more expensive mistake is the opposite: buying a generic tool that almost fits the one workflow that was the whole reason you started looking. The tool handles the easy 90 percent that every company shares and misses the 10 percent that was actually your constraint. You pay a small subscription and the bottleneck is still there.

Commodity versus differentiated.

The entire decision turns on one distinction, so it is worth stating carefully. A commodity workflow looks roughly the same at every company in your size band. Drafting routine emails, transcribing calls, summarizing documents, answering common questions, taking meeting notes: these are not where your business is different from the firm down the street, and a product built for ten thousand companies will do them better than anything you would commission, because the vendor has amortized the work across all ten thousand.

A differentiated workflow is shaped by decisions, data, and sequencing that are specific to how you operate. It is usually the workflow that has resisted every tool you have tried, the one where someone on your team says "the software does not really work the way we do this." That resistance is the signal. A product never quite fit because the workflow was never standard, and the part that does not fit is, more often than not, the exact part that is your bottleneck.

Put plainly: you buy the commodity layer and you commission the differentiated one. The error is treating the whole company as differentiated (over-building) or treating the differentiated workflow as commodity (buying a tool that almost fits). Both are common. Both are avoidable with a short test.

The decision test, in three questions.

Before you decide build or buy on any workflow, answer three questions in order. They take a few minutes per workflow and they settle most cases without a vendor call.

One: is this workflow commodity or differentiated? Is it done roughly the same way at companies like yours, or is it shaped by something specific to how you run? If it is standard, lean buy and stop here. If it is specific to you, keep going.

Two: does an off-the-shelf tool reach 90 percent or more of it? Look at the real products, not the demo. Map your actual workflow against what the tool does. If a product covers nearly all of it and the rest is cosmetic, buy. If the tool covers the easy front of the workflow and stops exactly where your work gets specific, that gap is the tell.

Three: what does the last 10 percent cost you? This is the question that decides it. If the missing slice is a minor annoyance, buy the tool and absorb the gap. If the missing slice is the bottleneck, the part that consumes staff hours, produces the errors, or slows the response time you actually care about, then that 10 percent is not a rounding error. It is the reason to commission a build, because it is the only part worth building and it is the part no product will ever reach.

The tool that almost fits.

The most expensive purchase in mid-market AI is not the custom build that ran long. It is the SaaS tool that almost fits. Here is how it goes wrong. A product is sold on the 90 percent it does cleanly, demoed on sample data, and priced at a subscription that feels like a bargain next to a commission. You buy it. Then it meets your real workflow and the missing 10 percent turns out to be the specific part, the part that was your constraint in the first place.

From there, one of three things happens, and all three cost money the subscription line never shows. Your team bends its process to fit the tool, which means you have changed how you operate to suit software instead of the other way around. Or your team runs a manual workaround for the gap, which is the same labor you were trying to remove, now with a tool on top of it. Or people quietly stop using it, and you are paying for shelfware while the bottleneck sits exactly where it was. The cheap number was real. The solved problem was not.

This is the same failure pattern, viewed from the buy side, that shows up on the build side as the thin wrapper. We covered the build-side version in what a mid-market AI engagement costs: a generic chatbot pointed at a problem that needed a custom build works for the average company because it was built for the average company, and the mid-market operator with a non-standard workflow is, by definition, not the average company. Whether the not-quite-fit arrives as a tool you bought or a wrapper you commissioned, the diagnosis is the same: the workflow was differentiated and got treated as commodity.

When a custom build is actually right.

A custom build is the right call when three things are true at once. The workflow is differentiated, specific to how you operate rather than standard across your peers. An off-the-shelf tool cannot reach the part that matters, the 90 percent test fails on the slice that is your constraint. And the cost of that gap is real, measured in hours, errors, or lost speed that you can actually name. When all three hold, you are not choosing custom out of vanity; you are commissioning the only thing that can reach your bottleneck.

Note what this rules out. It rules out commissioning a build for a workflow a product already serves, which is paying bespoke prices for a commodity. It rules out "AI across the whole business" as a single custom engagement, which is over-building by definition, because most of the business is commodity workflows you should buy. The disciplined version is narrow: one differentiated workflow, scoped, built, and owned, with the commodity layer bought off the shelf around it. The reasons mid-market builds drift wide and stall are worth understanding before you commission anything, and we walk through them in why mid-market AI rollouts stall in month four.

There is also a sequencing point. You do not have to decide build-versus-buy in the abstract; you decide it per workflow, and the right order is usually to buy the commodity layer first, live with it, and let the genuinely differentiated bottleneck reveal itself by being the thing the tools still cannot touch. That is the workflow to commission, and it is the only one worth the custom spend.

How to run the decision without us.

You can do most of this yourself before any vendor is involved. List your workflows. Mark each one commodity or differentiated using the first test question. For the commodity ones, go buy the best product and move on; there is no engagement to scope and no reason to overthink it. For the differentiated ones, run the 90 percent test against the real tools, and for any that fail the test on the part that is your constraint, you have found a candidate for a custom build.

If you want the map made for you, the AI Maturity Index does exactly this sorting, separating the workflows you should buy from the one or two worth commissioning, and our diagnosis process exists to pressure-test that call before any money is spent. The framing to hold onto is simple. Build narrow, buy broad. Reserve the custom build for the workflow that is genuinely yours, and let a product handle everything that is not. Most of the time, that means buying, and that is the honest answer.

Field-note context

Where the build-versus-buy call sits.

Build-versus-buy is decided per workflow, not per company.

The mistake operators make is treating it as one decision for the whole business. It is not. Almost every company is a mix: a long list of commodity workflows that should be bought, and a short list of differentiated ones, often just one or two, that may be worth commissioning. The work is sorting the list, not picking a side. Once the list is sorted, most of it answers itself, and the only real decision left is the handful at the top.

The hire question sits next to this one.

Once you have decided to commission the differentiated workflow, the next question is who builds it: an internal hire or a commission. That is a different decision with its own logic, and getting it wrong is expensive in a different way. We work through timeline, ownership, and when each path is right in AI consultant vs in-house hire. The short version: for a single named workflow, the commission ships faster and transfers ownership cleanly, while a hire makes sense once the work is continuous and broad enough to fill a role.

Why we tell you to buy.

We commission custom builds for a living, so the advice to buy off-the-shelf wherever you can is against the obvious short-term interest. We give it anyway because the operators we want to work with are the ones who can tell the difference between a workflow that needs a build and one that needs a subscription. An operator who has already bought the commodity layer and isolated the one workflow a product cannot reach is the operator a commission actually helps. The honesty selects for the right engagements, which is the entire point.

Extended questions

The build-versus-buy questions buyers ask next.

Should a mid-market company build or buy AI?

Buy for the commodity workflows and build only for the differentiated one. The majority of what you would point AI at (routine email, transcription, document summaries, generic chat) is standard across companies your size, and a product will do it better and cheaper than a commission. Reserve the custom build for the single workflow that is both your real bottleneck and specific to how you operate, where no off-the-shelf tool reaches the part that matters. For most operators, that means buying more than building.

How do I tell a commodity workflow from a differentiated one?

Ask whether the workflow looks the same at every company in your size band. If it does, it is commodity, buy it. If it is shaped by your specific data, sequencing, or decisions, and if it is the workflow where every tool you tried "did not quite work the way we do this," it is differentiated. That phrase is the signal: the part that did not fit is usually the part that is specifically yours, and often the part that is your constraint.

What is the 90 percent test?

Map your real workflow against what an off-the-shelf tool actually does, not the demo. If the product reaches 90 percent or more and the remaining slice is cosmetic, buy it. If the product covers the easy front of the workflow and stops exactly where your work gets specific, ask what that missing slice costs you in hours, errors, or lost speed. A cheap last 10 percent means buy; a last 10 percent that is your bottleneck means build, because that slice is the only part worth commissioning and the only part no product will reach.

Why is a tool that almost fits worse than no tool?

Because it solves the part you did not need solved and misses the part you did. The 90 percent it handles is the commodity work; the 10 percent it misses is the specific work that was your constraint. Teams then bend their process to the tool, run manual workarounds for the gap, or stop using it, all while paying the subscription. The bottleneck stays in place and you have added a tool on top of it. A tool that fits 60 percent and is honest about it is easier to reason about than one that fits 90 percent and hides the gap that mattered.

Should we just buy everything and skip custom work entirely?

For the commodity layer, yes. For the one differentiated workflow that is your actual bottleneck, no, because that is precisely the workflow products cannot reach, and buying a tool that almost fits it leaves the constraint in place. The disciplined posture is buy broad and build narrow: purchase the best products for everything standard, and commission a single scoped build for the workflow that is genuinely yours. Skipping custom work entirely only works if you have no differentiated bottleneck, which is rare for a company that grew to this size by doing something specific well.

In what order should we decide build versus buy?

Per workflow, and usually buy-first. List your workflows, mark each commodity or differentiated, and buy the commodity ones immediately; there is nothing to scope. Live with the commodity layer for a while and let the genuinely differentiated bottleneck reveal itself by being the thing the tools still cannot touch. That surviving workflow is the candidate for a custom build, and by then you will be able to name exactly what the missing slice costs you, which is the number that justifies commissioning it.

Not sure which of your workflows to build?

Start with the AI Maturity Index. Ten minutes, no call, and it sorts your workflows into buy-off-the-shelf and worth-commissioning, so you spend custom money only where a product cannot reach.