AI Consultant vs. AI Agency vs. Build Shop: An Honest Taxonomy

If you run a business between five and fifty million in revenue and you have decided AI is no longer optional, the market will offer you at least four kinds of help under nearly identical language. An "AI agency," an "AI build shop," an "AI consultant," and — the newer entrant — an "applied-AI research lab" all promise to make your business better with AI. The words on the homepages rhyme. The work does not.

This is a category explainer written by one of those entrants, which is a reason to read it skeptically. So here is the discipline we will hold: every category gets its genuine best case, stated the way a satisfied client would state it, before we say anything about where it falls short. If you finish this and conclude you need a build shop, we have done our job. Knowing what you are buying is worth more than which door you walk through.

The marketing-side AI agency

Its best case is real and specific. You have a demand problem, not an operations problem. Your pipeline is thin, your content cadence is inconsistent, your paid channels are under-instrumented, and you know that AI can now produce and test creative faster than any team you could hire. An agency that has industrialized this — prompt libraries, generation pipelines, an analytics loop that classifies what wins and reweights toward it — will move your metrics faster than you can, because they run the same play across dozens of accounts and carry the pattern library you would spend a year building.

Hire the agency when the constraint is the top of the funnel and the deliverable is a stream of tested assets. The failure mode is not the agency being bad at its job; it is asking it to fix something downstream of demand. If leads arrive and then die in a broken handoff between sales and fulfillment, more leads make the fire brighter. That is a structure problem wearing a marketing costume, and no volume of creative resolves it.

The build shop

Its best case is the cleanest of the four, because it is the most bounded. You already know exactly what you want built. The spec is written or nearly so: an internal tool, a customer-facing feature, a retrieval system over your documents, an automation that eliminates a named manual step. A good build shop is a disciplined engineering contractor. Hand it a clear specification and it will return working software, usually faster and more cheaply than standing up your own team for a one-time build.

The dependency is the specification. A build shop optimizes execution against a target someone else has set; it is not structured to tell you the target is wrong. If you arrive with a spec that solves the second-most-important problem, you will get an excellent implementation of the wrong thing, on time and on budget. When the requirements are firm and the value is in shipping them well, the build shop is the correct and often the cheapest answer.

The consultant

The independent AI consultant — or the strategy firm's AI practice — sells judgment. Its best case is the situation where the decision matters more than the deliverable: a make-or-buy call on a platform, a read on whether a vendor's roadmap is credible, a diagnosis of why the last two initiatives stalled. A strong consultant compresses your uncertainty. They have seen your situation before, they will tell you what usually happens next, and a week of their attention can save a quarter of your wandering.

The structural limit is the handoff. Advice is delivered and then it leaves; whether anything changes depends on an organization that was, by definition, not changing on its own — which is why it hired advice. The recommendation can be correct and still not survive contact with the operation. Retain the consultant when the binding constraint is a decision and you have the capacity to execute it once it is made.

Where a research lab sits — which is none of these cleanly

An applied-AI research lab is not a fourth flavor of the same thing. It sits across the seams of all three, and it is honest to say that makes it the wrong tool more often than any single one of them.

It does what a consultant does — it diagnoses before it prescribes — but it refuses to leave at the recommendation. Our engagements open with a two-to-four-week performance diagnostic that produces a score, not a pitch, and most of them end there, with a client executing a clear set of findings themselves. It does what a build shop does — it ships working software — but it will not accept your specification unexamined, because the diagnostic exists precisely to test whether the thing you were about to build is the thing worth building. And it borrows the agency's instinct to measure everything, then points that instinct inward: the lab runs on the same operating system it installs, and it publishes what it learns, including the findings that undercut its own premise, as it did in its account of eleven weeks watching the regional AI market.

The engagement model is the tell. The lab installs an AI-native operating system inside a real business and then hands over the keys — the target is roughly a twelve-week handback, after which the operation runs without us. A build shop wants a longer backlog. A consultant has already gone. An agency wants a standing retainer against your ad spend. A lab is trying to make itself unnecessary and treats each engagement as an experiment it will report on. Those incentives do not overlap cleanly with any of the three, which is the point, and also the cost: you are paying for a diagnosis you might not like and a departure date you cannot postpone.

When you do not need the lab

If your problem is demand, hire the agency. If your spec is firm and correct, hire the build shop. If you need one hard decision made well and you can execute it yourself, hire the consultant. A lab earns its place only in the case none of those fit: when you suspect the problem you can name is not the problem you have, when a wrong answer is expensive, and when you want the fix to still be running after the people who built it have left. That is a narrower door than most homepages admit — ours included. The fairness is the whole point: the right first question is not "which vendor," but "which problem," and the honest ones will help you answer that before they quote you anything.