How to Choose an AI Consultant in 2026: The Questions That Expose Pretenders

Choosing an AI consultant in 2026 is a filtering problem. The title is unregulated, the demand is enormous, and every agency, developer, and strategy firm has rebranded around it. Underneath the noise sits a market with measured failure rates worse than almost any other purchase your company makes. The good news: the pretenders are easy to expose if you know which five questions to ask. This article gives you the questions, the answers that should come back, and the current market rates so you can price what you hear.

I build AI systems for mid-market companies as the technical half of V3RSION, so I am a participant in this market, not a neutral observer. The evaluation below is the one I would run on us.

Know the Market Before You Shop It

Rates first, because they frame everything else. In 2026 the hourly market runs roughly: independent specialists at $150-$350, boutique AI consultancies at $250-$500, Big 4 AI practices at $300-$600, and the top strategy firms at $500-$1,000+ (AI consulting pricing guides, 2026). On a project basis: an AI readiness assessment runs $10,000-$25,000, an AI strategy and roadmap $25,000-$75,000, and a working proof of concept $50,000-$150,000 (ClearForge benchmarks, 2026).

Price tells you what the vendor costs. It tells you nothing about whether the project will work, and that is the actual risk you are pricing.

The Failure Data Nobody Puts in the Proposal

The numbers on AI project outcomes are stark enough that any honest consultant should raise them before you do:

• More than 80% of AI projects fail, roughly twice the failure rate of comparable non-AI IT projects (RAND Corporation, 2024).
• 95% of enterprise GenAI pilots produce no measurable P&L impact (MIT Project NANDA, The GenAI Divide, 2025).
• 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier, and the average organization scrapped 46% of its AI proofs of concept before production (S&P Global Market Intelligence, 2025).
• Gartner predicted 30% of GenAI projects would be abandoned after proof of concept by the end of 2025, and traces 85% of AI project failures to poor data quality or a lack of relevant data (Gartner, 2024 and 2025).

Read those again with a buyer's eye. The failures are not model failures. Every cause the research names is an operating failure: no business decision attached, data never made ready, no adoption plan, no measurement. Which means the failures are selectable. You choose them, or you filter them out, at the moment you choose the consultant.

One more data point that should shape your shortlist: MIT's research found externally partnered AI deployments succeed about 67% of the time, twice the rate of internal builds. Buying help is rational. Buying the wrong help is how you join the 80%.

AI executes. Strategy decides. Buy them in that order.

The Five Questions That Expose Pretenders

Ask these in order. Each one maps to a documented way AI projects die, and each has answers that cannot be faked with a slide.

1. "Which business decision does this system improve, and what is that worth?"

This is the strategy-before-tooling test. Pretenders answer with capabilities: chatbots, copilots, agents. Builders answer with a decision, an owner, and a dollar value: "quote follow-up currently takes 26 hours and loses you deals above $40K; the agent cuts it to minutes." If nobody can price the decision, the pilot will demo well and die quietly, which is the exact mechanism behind MIT's 95%. We wrote up the strategy half of this argument in AI Execution Without Strategy Is Just Faster Mediocrity.

2. "What is the adoption plan, with named people and named workflows?"

The shelfware test. A real answer names which roles use the system on a Tuesday, what training happens, and what old habit gets retired. If the plan is "we will train the team at handoff," you are buying licenses, not outcomes. Our own success criterion for the systems phase is 95% user adoption, and it is written into the engagement plan, not hoped for afterward.

3. "What gets measured from day one, against what baseline?"

The unprovable-ROI test. Half of all AI proofs of concept get scrapped before production, and the most common reason is that nobody can demonstrate the return, because nobody captured the before. A builder installs measurement in week one. A pretender promises a report at the end.

4. "Who owns the system when you leave?"

The dependency test. Ask who holds the admin keys, where the data lives, and what happens on the day the contract ends. You want the platform, the dashboards, and the automations to be yours. You own the platform. No consultant dependency. If the answer involves their proprietary black box and a mandatory retainer, the pilot was a lease.

5. "Where do humans stay in the loop?"

The governance test. Which decisions can the system take autonomously, which require sign-off, and how is that enforced technically rather than by policy memo? The right posture in 2026 is full automation of the repetitive layer with 100% human oversight retained on judgment calls. Anyone promising fully autonomous everything is selling you their liability.

What a Reference Implementation Looks Like

Calibrate the answers you collect against a real deployment standard. The systems layer we build at V3RSION runs on Savra.ai, and the published operating numbers across client deployments: 67% of manual tasks reduced on average, 3x faster decisions, 24/7 monitoring, 100% human oversight maintained. Those four numbers are what "it works" looks like when the strategy came first, the adoption was designed, and the measurement started on day one.

Order of operations is the entire trick, and it is why the systems build sits in weeks 5-10 of our 90-day engagement rather than week one: the strategy phase has to produce the decision framework the AI executes. I wrote the sequencing argument, with the implementation-order detail, in AI Strategy Before AI Tools.

Run the Evaluation Like You Mean It

Practical closing advice. Put the five questions in writing before the first call, and score the answers on specificity, because vagueness is the tell. Ask for two client results with numbers and dates, and call one. Ask who personally does the work; at large firms the partner sells and juniors build, which is fine as long as you priced it that way. And prefer a paid discovery with a written deliverable over an unpaid pitch, because a consultant who diagnoses before proposing is showing you how they work.

The AI consulting market in 2026 rewards buyers who filter hard. The failure statistics above are not a warning about AI. They are a census of companies that skipped the questions.