Blog/Evaluating AI delivery

The best AI consulting firms for an operating company, and how to tell them apart

The best AI consulting firms for an operating company are the ones that put a working system into production and stay accountable for it. The four kinds of firm, what each one is paid for, and the evidence that tells them apart.

The best AI consulting firms for an operating company are the ones that leave a system running in the company's books, and the rankings that come up when you search for the top AI consulting firms rarely measure that. They rank by revenue, headcount and brand. The gap between a brand and a result is wide. In McKinsey's 2026 State of AI survey, only thirty-seven percent of respondents said AI had added anything to their organizations' earnings, a share essentially unchanged from the year before even as more of them reported scaling the technology. So the useful question for a CFO, a COO or a PE operating partner is which kind of firm is paid to produce a result in production, and what evidence shows that a given firm has done it before. Four kinds of firm sell AI consulting, each is good at something different, and the way each one is paid explains most of what it delivers.

Four kinds of firm sell AI consulting

The first kind is the strategy house: McKinsey, BCG, Bain and their peers. A strategy house is paid for senior judgement and analysis. Its AI work produces an opportunity assessment, a business case for each use case, an operating model and a roadmap, and the strongest firms in this group are very good at choosing which problems a company should work on. The deliverable is a recommendation. We described why the engagement so often ends there in why your AI consultants left you a deck: the people on the team are trained to advise, and nobody on the team has the access or the remit to change a system of record.

The second kind is the Big Four and the large IT integrators: Deloitte, EY, KPMG, PwC, Accenture, IBM, Capgemini, Infosys and others. These firms are paid to staff large programs, and they are the right choice for a global ERP rollout or a data platform that has to work across forty countries. Accenture's results for fiscal 2025 show the scale of the model. The company reported generative AI new bookings of $5.9 billion for the year, against revenue of $69.67 billion and approximately 779,000 people serving clients. AI is a fast-growing slice of a business that sells people deployed at scale, and AI work is sold, staffed and governed the way the rest of that business is.

The third kind is the AI boutique: a small firm of machine learning engineers and data scientists who build custom models and systems. Boutiques are good at hard technical problems that a buyer can specify precisely, and they usually hand the finished system to the buyer's own team. The best of them tend to get bought. Faculty, the UK applied AI firm with more than 400 people, agreed to be acquired by Accenture in January 2026.

The fourth kind is the forward-deployed builder. Its engineers, and increasingly its agents, work inside the customer's environment, build on the systems the company already runs, and keep operating the result after it goes live. This kind of firm is usually paid on the result. We set out the customer each of these models was designed for in the customer all four delivery models leave behind, and the short version is that a mid-market company with no AI team fits the fourth model better than the other three.

What each kind of firm is paid for

The fee structure tells you what a firm is optimizing for, because a firm builds its staffing, its methods and its promotion criteria around how it earns money. Time and materials pays for hours. A firm on time and materials earns more when an engagement runs longer, and it carries no loss when the system it designed never runs. A fixed fee pays for completing a scope. It protects the buyer's budget, and it rewards the firm for reaching the end of the statement of work, which can happen with a working pilot and no production system. Outcome-based pricing pays for a measured change in the buyer's numbers, such as hours removed from accounts payable or days taken out of the close.

The large firms are moving toward the third model. Business Insider reported in November 2025 that about a quarter of McKinsey's global fees now come from outcome-based pricing. That share tells you the model works at scale. It also tells you that three quarters of the fees at the largest strategy house are still paid for something other than a measured outcome, so a buyer has to ask how a particular engagement is priced and can't assume it from the brand.

How to judge the best AI consulting firms on evidence

Every firm's proposal says it has deep industry knowledge, a proven method and experience in AI. Those claims are the same across firms, so they don't help a buyer choose. The evidence that does help is evidence about systems in production, and there are three pieces of it worth asking for.

The first is a production reference. Ask for a customer whose system has run for at least two quarters, and ask to speak with the person who handles that system's exceptions every week. That person knows how often the system stops, what it gets wrong and how long a fix takes. An executive sponsor knows how the kickoff went.

The second is operating data. A firm that runs its systems keeps the numbers: how many transactions the system handled last month, what share it completed without a person, how many errors reached the ledger and how they were caught. Ask for those numbers from a live deployment. A firm that hands systems over at go-live won't have them, because nobody at the firm watches the system after that day.

The third is a test of the label. Gartner, in a June 2025 forecast, named the practice of agent washing, the rebranding of chatbots, robotic process automation and ordinary assistants as agentic AI, and estimated that only about 130 of the thousands of vendors claiming the label offer something genuinely agentic. Ask what the system does with a case it was not built for. A real answer describes a designed path: the case is routed to a named person with the evidence attached, and the system records the decision for next time. We set out a fuller procedure in how to tell an AI delivery vendor that ships from one that demos.

Who stays after go-live

A back-office system on AI changes after launch whether anyone plans for it or not. A supplier changes its invoice layout, a new approval rule arrives with a new controller, an ERP upgrade moves a field. A system with no owner degrades, and the team falls back to doing the work by hand while the system still appears on the vendor's list of references. So the second question for any firm, after what it will build, is who will run the system a year later, by name and by team, and what that costs.

The answer also tells you whether the firm intends to change the process or only to automate it. McKinsey's survey found that among its high performers, nearly three-quarters reported fundamentally redesigning workflows because of their AI use, against a quarter of other respondents. Redesign takes months of adjustment after the first release, because the new process meets cases the old one handled by habit. A firm that leaves at go-live has left before most of that work is done.

Pricing tied to outcomes, and what to check in it

Outcome-based pricing is only as good as the outcome it names. A fair contract defines the baseline before work starts: the hours a process consumes, its cost per transaction, its error rate and its cycle time, measured on the company's own data. It names the measure that sets the fee, and it says who measures it and how often. A contract that ties the fee to "use cases identified" or "models deployed" has tied the fee to activity, and activity is what time and materials already paid for.

Two further terms protect the buyer. The first is a fixed price, or no fee at all, until the system runs in production on real transactions, so the firm carries the cost of a pilot that never ships. The second is ownership: the company should hold the right to keep running the system, with its configuration and documentation, after the engagement ends, so a change of firm doesn't mean starting again.

A reasonable counter, answered

A reasonable counter is that a large, well-known firm is the safer choice. A board, a lender or an acquirer recognizes the name, the firm will still exist in five years, and a CFO who hires a top AI consulting company is rarely blamed when the program disappoints. That safety is real, and it matters most to the person who signs the contract. It protects the buyer's reputation more than it protects the result, and for a mid-market company the larger risk is paying for a program that never runs in production. The continuity risk of a smaller firm is real as well, and a buyer can test it directly: ask for production references, operating data and the right to run the system without the firm. A firm that passes those tests is a safer choice for the process than a famous one that can't.

Common questions

What are the best AI consulting firms?
The answer depends on the job the firm is hired to do. For a board-level strategy or a market view, the large strategy houses are hard to beat, and for a multi-year program across a global enterprise, the Big Four and the large IT integrators have the people to staff it. For an operating company that wants one back-office process running on AI this year, the best firm is the one that builds the system on the company's existing software, runs it in production and stays accountable for the result. Published rankings mostly measure size and brand, which tells you a firm can staff a large engagement and very little about whether its systems keep running.
How do AI consulting firms charge?
Three pricing models are common. Time and materials bills the hours the firm's people spend, so the buyer carries the risk that nothing ships. A fixed fee caps the cost of a defined scope, and it rewards the firm for finishing the scope. Outcome-based pricing ties the fee to a measured result against a baseline agreed before the work starts, and McKinsey has said that about a quarter of its global fees now come from arrangements of that kind.
What should I ask an AI consulting firm before I hire it?
Ask for a reference customer whose system has run in production for at least two quarters, and ask to speak with the person who handles that system's exceptions rather than the executive who signed the contract. Ask who owns the running system after go-live, by name, and what happens when a supplier changes its invoice format or a field moves in the ERP. Ask what accuracy the system must reach before it touches a live transaction. Ask how the fee changes when the result falls short.

Book a demo

Start with one department

We map how one team’s work runs, show you where the hours go, and automate from there.