Guide · AI
How to choose an AI consulting firm.
Every firm now sells AI. This guide is for the buyer who has to pick one, and it gives you the five questions that separate the firms that build from the firms that present, the red flags, a selection checklist, and the cases where the right answer is to hire nobody.
This guide, in 6 parts
The demo is the easy part.
Foundation models made impressive demos cheap. Any competent firm can now show you a working prototype in two weeks, which means the demo no longer tells you anything about the firm. The numbers above describe what happens next: most pilots never touch the P&L, and a large share of projects are abandoned after proof of concept. The gap between the two halves of that story is delivery.
The firm you choose is the biggest variable you control. Firms that build are recognizable if you know what to ask, and this guide is the list of what to ask, written by consultants who operate. These are the questions we would put to any firm, ourselves included, from your side of the table.
« Choose the firm for the year after the pilot, not the month of the demo. »
Five questions that sort the field.
Ask all five in the first serious meeting, and take notes on the answers. What follows each question is what a builder's answer sounds like. The second question has a full guide of its own: from pilot to production.
Who owns production?
A firm that builds will name the engineer who carries the pager on day one and the date your team takes over, with a runbook. Listen for a named person and a transfer plan. If the answer is a governance slide and a steering committee, the running of the system has no owner yet, and you will discover that at 2am.
What happens after the pilot?
Ask to see the production line in the proposal itself: integration with your real systems, monitoring, a fallback for when the model is wrong, and a budget number for all of it. The pilot-to-production gap is where most AI investment dies, and we wrote a separate guide on exactly that: from pilot to production. A firm with no written answer is planning to stop at the demo.
Who actually does the work?
The people in the sales meetings are rarely the people in your systems. Ask to meet the delivery team before signing, ask the senior-to-junior ratio on your account, and ask what the seniors do weekly. In our own shop the people who scope the work do the work, backed by 100+ AI agents we run to build faster. Whatever a firm's answer is, it should be that specific.
How will success be measured?
The success criterion should be a business number fixed before any code: a cost, a delay, a defect rate, a conversion. Workshops run and documents delivered are activity, and activity is what gets invoiced when nobody agreed on a result. A firm confident in its delivery will accept a measurable target. Hesitation here tells you everything.
What governance ships with the system?
Controls, traceability, an audit trail and a named owner, designed in from the first workflow. Without them the system cannot pass internal review, the auditor or the regulator, and it will stall at the gate no matter how well it performs. Ask to see governance in the architecture, and in the price.
Red flags, from the buying side.
None of these is fatal alone. Two together should slow you down. Three end the conversation.
- A proposal that is mostly slides about your industry and thin on what gets built in month one.
- No engineer in the room by the second meeting.
- Success defined as activity: workshops run, documents delivered, hours logged.
- A pilot with no production budget, date or owner attached to it.
- Case studies with logos but no numbers, or numbers but nobody you can call.
- Pressure to sign a long engagement before a small one has proven anything.
- Certainty about your results before anyone has looked at your data.
The selection checklist.
Five steps, in order. Step one happens before any firm is in the room.
Fix the problem and the number before the first meeting
Choose one measurable problem: a cost, a backlog, a delay, an error rate. Write down the number today and the number that would make the project worth it. Firms are then evaluated against your problem instead of against each other's slides, which is a contest the best presenter would otherwise win.
Shortlist on shipped production work
Ask each candidate for one system that has been running in production for over a year, and for the client-side person who runs it now. Then call that person and ask about the second year, the maintenance, the surprises. References about launches are easy. References about the after are the ones that sort the field. Our own record is retail rollouts at scale: 150+ stores live across 10 countries, and 29 stores switched in one national date with zero revenue loss.
Meet the delivery team and put the five questions to them
Not the partners, the people who will be in your systems. Put the five questions from section two to them directly and listen for named owners, written after-pilot plans and acceptance of a measurable target. How a firm handles being questioned by a prepared buyer is a preview of how it handles your steering committee.
Demand the after-pilot plan in the proposal
Production, governance, monitoring, handover and exit, priced and dated, in the same document as the pilot. This is the single strongest filter we know, because it costs a builder nothing and a presenter everything. Our own way of running engagements is public at how we work, and you should expect the same transparency from anyone bidding.
Start scoped, then scale on evidence
A short, paid, bounded first engagement with its own success criteria tells you more than three months of pitch meetings. Our version is the FOCAL diagnostic, a fixed-scope assessment that ends in a decision, and most serious firms offer something comparable. If the first small thing lands on its number, scale. If it does not, you have lost weeks, not a year.
The transparency we ask you to demand in step four is published for our own engagements at how we work, and the scoped start from step five is our FOCAL diagnostic.
When you should not hire one.
An honest buyer's guide includes the cases where the right amount of consulting is none. We keep a longer version of this list at when not to hire a consulting firm. The short version has four cases.
Your data and workflows are not ready
AI performs on clean processes and structured data. If the process is broken, fixing it comes first, and much of that work is internal. Paying an AI firm to sit on top of a broken workflow buys a demo.
You need one team to adopt one tool
Rolling out an off-the-shelf assistant to a single department is a training and change job, not a consulting engagement. Buy training.
You have the team and need a decision, not delivery
If capable engineers are already in the building and the blocker is a choice, a scoped assessment or a few advisory sessions settle it. A delivery engagement would duplicate your own people.
The budget covers a pilot and nothing after
Given how many pilots die between demo and P&L, a budget with no production line is a budget for a demo. Wait until it covers the after, or shrink the scope until it does.
Frequently asked.
What is an AI implementation consultant?
An AI implementation consultant takes AI from decision to running system: wiring models into your existing software and data, building the workflow around them, setting up governance and monitoring, and staying through rollout and handover. The contrast is with an AI strategy consultant, whose work ends at the recommendation. Titles vary between firms, so ask what the person shipped last and who runs it now.
How much does an AI consultant cost?
It depends on the pricing model: firms bill time and materials, fixed scope, monthly retainers, or fees tied to a measured outcome. What drives the number is the seniority of the people actually delivering, how many systems the work touches, how ready your data is, and whether your industry is regulated. We do not publish figures because ranges without scope mislead. The practical move is to have each shortlisted firm price the same scoped first phase, and compare that.
What does an AI consulting firm actually do?
Day to day: pick use cases against a business number, do the data and integration work, build and test the system, set up governance, run the rollout and train the teams, then hand over. The mix between advice and building varies enormously between firms. Ask for the split between advisory hours and build hours in the proposal, because that split is the firm's real identity.
How long does an AI consulting engagement take?
A scoped diagnostic runs a few weeks. A first production deployment usually takes a few months, driven mostly by data readiness and the number of systems touched, and multi-site rollouts run longer still. Be wary of both extremes: a firm promising production in two weeks and a firm that cannot say when you would see a result.
You can run this checklist without us. Put the five questions to every firm on your shortlist, ourselves included. We are consultants who operate: once the advice is given, we build it with you and stay until it holds, and that standard is what this whole guide asks you to demand from anyone. What that looks like in practice is at AI in production.
Building a shortlist?
No deck, no gate. A working session on your use case, with the people who would build it.