How to Choose an AI Consulting Firm: A Buyer's Guide for 2026

Step 1: Start with one workflow and one number

The most common reason AI projects stall is that they start from the technology ("we should use LLMs") instead of a process. Before you talk to any firm, write down:

• The workflow. One process, described in a sentence: "match supplier invoices to purchase orders and flag exceptions."
• The metric. What should change, and how you measure it today: hours per week, error rate, cycle time, cost per transaction.
• The data. Where the inputs live (ERP, email, shared drives, a database) and who can grant access.
• The constraint. Anything non-negotiable: data must stay in your cloud, the system must integrate with SAP, a deadline.

A firm that cannot work with this brief, or immediately widens it into an "AI strategy," is telling you something about how it sells.

Step 2: Know which kind of firm you are talking to

"AI consulting firm" covers three different businesses, and each is the right answer for a different project.

Large consultancies increasingly do both strategy and build, and many build firms offer a short audit before they write code. The useful question is what the firm's people spend most of their time doing. That is usually what you will get.

Step 3: Understand the engagement models

• Fixed price after an audit. The firm studies your process and systems first, then commits to a scope and a price. You know the budget before the build. The audit itself is short and usually paid separately or credited against the project.
• Time and materials. You pay for hours worked. Flexible when the scope is genuinely unknown, but the budget risk sits with you.
• Retainer or dedicated team. A team works for you month to month. Good for an ongoing roadmap, expensive for a single system.
• Outcome-based pricing. Fees tied to a measured result. Attractive in theory, but it only works when the metric is clean and both sides agree how to measure it.

For a single well-defined workflow, a fixed price after a short audit usually gives the best balance of speed and budget certainty.

Step 4: Twelve questions to ask on the first call

1. Which of your AI systems are running in production today, and what do they do?
2. Who will write the code on our project? Can we meet them before signing?
3. How soon can we see a working prototype on our own data?
4. Is the price fixed, and what exactly would change it?
5. Who owns the code, the prompts, any fine-tuned models and the infrastructure when the project ends?
6. Where does our data go? Can the system run inside our own cloud account?
7. Will our data be used to train any model, yours or a third party's?
8. Which models do you use, and how hard is it to switch to another provider later?
9. How do you measure whether the system is working, and how do you catch it when it is wrong?
10. What happens when the AI is not confident: which cases go to a human, and how?
11. What does handover look like: documentation, training, a support period?
12. Which project would you turn down, and why?

The last question is the most revealing. A firm that says it is the right choice for every project is not listening to yours.

Step 5: Red flags

• No working software before a large commitment. A firm that needs months of paid discovery before showing anything is pricing in its own uncertainty.
• Only demos on the vendor's data. Every model looks good on curated examples. Insist on your documents and your edge cases.
• A proprietary platform you cannot leave. If the system only runs on the firm's platform, you are signing a subscription, not buying software.
• Seniors on the call, juniors on the project. Ask by name who works on the code each week.
• Vague answers about data. "It's secure" is not an answer. Ask where data is stored, who can access it and what is logged.
• Accuracy promises with no evaluation plan. Nobody can guarantee a precise accuracy figure before seeing your data. A credible firm proposes how to measure accuracy and agrees on targets after the prototype.

Step 6: What the contract should say

Before you sign, make sure the agreement covers:

• Intellectual property. Assignment to you of the code, configuration, prompts and any model weights trained on your data.
• Data use. A written statement that your data is used only for your project and not to train models for anyone else.
• Hosting and access. Where the system runs, whose accounts hold it and how access is revoked at the end.
• Scope and change control. What is included in the fixed price and how changes are priced.
• Acceptance criteria. How you will decide that the delivered system works, ideally on a test set of your own cases.
• Handover and exit. Documentation, a knowledge-transfer session and a defined support period after launch.

Step 7: Compare your shortlist with a scorecard

Ask every firm the same questions and score the answers side by side. A simple version:

Typical timeline for a single workflow

A focused AI project usually runs in three phases. An audit of a week or two maps the process, the data and the integration points. A prototype runs on your real data, typically within two to four weeks of starting. Production hardening adds integrations, monitoring, access control and handover, and usually takes one to three months depending on how many systems are involved.

FAQ

Next step

Compare the firms in our ranking of the top AI consulting firms, and use the scorecard above on your shortlist. If you want to put the same questions to us, Building Agentic starts every project with an architecture audit.