AI Consulting Companies: How to Choose One That Actually Ships

Every company evaluating AI consulting companies right now is running the same broken process. Three vendors, three decks, three sets of case studies with the client names redacted, and a procurement spreadsheet that scores them on “AI expertise” — a column nobody knows how to fill in honestly. Six weeks later you pick one, and eight months after that you have a pilot that demos beautifully and touches nothing in production.

This is not bad luck. It is the predictable result of evaluating vendors on what they say instead of on what they are structurally able to deliver. The good news is that the structure is visible from outside, if you know which questions expose it.

The four kinds of AI consulting firm

There is no single market for AI consulting services. There are four, and they barely compete with each other despite pitching the same logos.

Type	Typical rate	Good at	Fails at
Large firm (Big Four, global SIs)	$300–600/hour	Board-level buy-in, regulated industries, very large programmes	Speed, and building anything themselves — the build is usually subcontracted
Technical boutique	$150–350/hour	Shipping a working system in weeks, deep expertise in one domain	Breadth. Ask them to cover HR, finance and supply chain at once and they will fail
Generic AI agency	$100–250/hour	Fast automation of well-understood workflows	Your specific business. They learn your domain on your budget
Independent consultant	$80–200/hour	Advice, architecture review, a second opinion that costs a day not a quarter	Continuity. One person cannot own a production system indefinitely

Notice what the table does not say: none of these is the “best” option. A regulated bank rolling AI across twelve countries genuinely needs the large firm. A brand that needs its ERP talking to an AI agent in a month does not, and hiring one is how a four-week job becomes a two-quarter programme.

The mistake is not picking the wrong type. The mistake is picking a type whose failure mode is the thing you actually needed.

Question 1: Who writes the code?

Ask it exactly like that, and listen for whether the answer contains the word “partner”.

The dominant model at large AI consulting companies is to sell the strategy engagement and subcontract delivery to an implementation partner, often offshore, often introduced after the contract is signed. There is nothing dishonest about it — it is how the economics work — but it has three consequences you are buying without being told.

The people who understood your business in the discovery phase are not the people building the thing. The timeline doubles, because every clarification crosses a company boundary. And when it does not work, accountability lands in the gap between two contracts.

A firm that builds what it sells will answer this question in one sentence, name the people, and offer to put them on the next call. Anything longer than one sentence is the answer.

Question 2: What exactly do I own when we finish?

The word to listen for is “platform”. As in “you’ll have access to our platform”.

Access is not ownership. If the models, the prompts, the integration code and the orchestration live in the vendor’s environment, then the thing you bought is a subscription with a consulting invoice attached, and your exit costs a rebuild. That may still be the right deal — plenty of good software works that way — but you should know you are buying it.

The stronger position is straightforward and easy to verify in a contract:

• Source code in your repository, not theirs.
• Deployment on your infrastructure, or infrastructure registered in your name.
• Documentation good enough that a different vendor could take over.
• Credentials and API keys held by you.

Ask for those four in writing. The reaction tells you more than the answer.

Question 3: What is the timeline, in dates?

“It depends on scope” is correct and useless. Any competent firm can convert scope into dates once scope is fixed, and the refusal to do so usually means the scope is not fixed either.

What good looks like: a fixed window, phases with dates attached, and an explicit statement of what is not in the window. Our own AI accelerators run on closed timelines for exactly this reason — two weeks for a connector, thirty days for a migration — because a date is the cheapest possible proof that the scope is real.

What bad looks like: a “discovery phase” of unspecified length, followed by a proposal. You are paying to be sold to.

Question 4: What happens if it does not work?

Almost nobody in this market answers this, which is why it is worth asking.

The honest versions range from “we keep working until the agreed metric moves” to “nothing, this is time and materials and you carry the risk”. Both are legitimate. What is not legitimate is a vendor who cannot say which one applies, because it means nobody has agreed what “working” means.

Before that question can even be answered, someone has to define the number. Not “improve efficiency” — the specific number, measured on your systems, with a baseline taken before the work starts. If your provider has not asked for that baseline, they are not planning to be measured.

Question 5: Who supervises the agents?

This is the question that has quietly become the whole game, and the market is split down the middle on it.

Gartner’s research puts organisations running fully autonomous, unsupervised AI agents at 15–17%, with roughly three quarters running agents under human supervision. The same body of research expects over 40% of agentic AI projects to be cancelled before the end of 2027 — and the cited reason is governance, not model quality.

So when a vendor’s pitch is that you will not have to supervise anything, they are selling you into the category that gets cancelled. The question to ask is concrete: when the agent changes a price, adjusts a budget or writes to a customer record, who approved it, where is that logged, and how do I undo it?

A system that cannot answer those three is not ready for production regardless of how well it demos.

Question 6: Have you done this in my industry?

Not “do you have a case study”. Have you personally operated in this domain.

The difference is enormous in practice. A firm that has run commerce operations knows what a stockout costs, what a vendor chargeback is, and why the catalog team will not accept an agent that rewrites titles without review. A firm learning that on your project will get there — in month four, at your expense, and the education is not transferable to you.

Vertical experience is also the only defensible reason to choose a smaller firm over a larger one. If they cannot demonstrate it, the large firm’s brand risk mitigation is probably worth more.

Question 7: What will you tell me not to do?

The strongest signal in a sales process is a vendor arguing against their own revenue.

Most AI work that gets proposed should not be built. The process is fine as it is, the data is too messy to support the model, the volume is too low to repay the integration, or the thing that is actually broken is a process nobody wants to change. A provider who has never told a prospect “this is not worth doing” is either extraordinarily lucky or not paying attention.

Ask directly: what have you turned down in the last six months, and why? The specificity of the answer is the answer.

What this looks like when you get it right

The engagements that work share a shape, and it has nothing to do with the size of the firm.

Scope written down before money changes hands. A date. Read access before write access, so the system proves itself on real data before it can change anything. An approval gate on everything that touches money or customers. A handover that includes the source. And a number, agreed at the start, that everyone is measured against at the end.

That shape is not exotic. It is just rare, because it transfers risk from you to the vendor, and most vendors would rather not.

If you want to see what it looks like written down, our AI consulting page publishes the market rate ranges, the four rules we do not negotiate, and the guarantee that goes with the packaged projects. You can also read the detail on what AI consulting costs in 2026, or on why enterprise AI pilots never reach production, which is the failure this whole checklist exists to prevent.

FAQ

What do AI consulting companies actually do?

They diagnose which parts of a business can be automated or augmented with AI, then design, build and deploy those systems. In practice the market splits between firms that mainly produce strategy and roadmaps and firms that mainly produce working software, and the two are priced similarly while delivering very different things.

How much do AI consulting firms charge?

Rates in 2026 run roughly $80–200/hour for an independent consultant, $150–350/hour for a technical boutique and $300–600/hour for a large firm. Project fees typically span $25,000 for a readiness assessment to $500,000 or more for a multi-system build, with a first-year enterprise GenAI implementation commonly landing between $170,000 and $680,000.

Should I hire a large consulting firm or a boutique for AI?

Choose the large firm when the problem is organisational — board alignment, regulation, a programme across many countries. Choose the boutique when the problem is technical and bounded, and you need it working in weeks. The large firm’s weakness is speed and the fact that it usually subcontracts the build; the boutique’s weakness is breadth.

How long should an AI consulting project take?

A bounded integration or connector should take two to four weeks. A migration or a forecasting system, around thirty days. Anything quoted as “six to twelve months” before a discovery phase has even happened is a programme, not a project, and should be broken into pieces that each deliver something usable.

What are the warning signs of a bad AI consultancy?

Four recurring ones: they cannot say who writes the code, they describe deliverables as capabilities rather than artifacts, they will not commit to dates before an open-ended discovery phase, and they have never talked a client out of a project. A fifth is promising fully autonomous agents with no human approval — the category Gartner expects to be cancelled at scale over governance.

Do we need our own technical team to work with an AI consultancy?

No, but you do need one person who knows how the process genuinely works today rather than how the documentation says it works. That person is worth more to the project than an engineer, and their absence is the most common reason a well-built system goes unused.