Questions to ask before hiring an AI consultant
Ask us these too. A good consultant will enjoy answering them; the answers tell you more than any portfolio.
The AI market is crowded and loud, and most buyers are choosing a consultant for the first time. That asymmetry is not your fault, but it is your risk. These questions come from the failure modes that are actually documented in industry research and the Israeli business press, not from horror stories about anyone in particular. Ask them of everyone you talk to, including us.
1. Will the person who scopes my project also build it?
In many firms, a senior person designs the solution and a different team implements it. That handover is where context, assumptions, and accountability quietly leak. Ask who exactly will write the code, and whether that person sat in the scoping meetings. If the answer is “our delivery team”, ask to meet them before you sign.
2. How will this integrate with the systems we already run?
Israeli business press coverage of enterprise AI adoption (Calcalist/CTech) reports lack of integration with core systems as one of the most-cited barriers, and treats ERP and CRM connectivity as mandatory from day one, not a phase-two luxury. A model that lives outside your ERP produces a report someone has to re-type. Ask for the specific integration plan for your systems, by name, and who has done it before.
3. Are our data actually ready for this, and will you tell us if they are not?
Data readiness is the most common real blocker. Israeli tech press reporting (Geektime) attributes the large majority of failed AI initiatives to fragmented, unready data. A consultant who audits your data before quoting, and who is willing to conclude “you are not ready yet, fix these three things first”, is protecting your budget. A consultant who quotes without looking at your data is quoting a fantasy.
4. What is your plan for getting from a working demo to production, and what is your track record on that step?
An MIT study widely covered in the Hebrew business press found that the overwhelming majority of enterprise GenAI pilots deliver zero measurable ROI, with only about one in twenty capturing significant value. That is a statistic about pilots, not about consultants, but it tells you where projects die: after the demo. Ask what “done” means in the proposal. If “done” is a demo, the riskiest part of the project is unpriced.
5. When it is delivered, who owns the model, the code, and the pipelines?
Some engagements end with you owning an outcome; others end with you renting one. Neither is wrong, but you should know which you are buying. Ask directly: if we part ways a year from now, does the system keep running, can our own people or another vendor maintain it, and is anything metered or licensed per use?
6. Can you show me predictions graded against what actually happened?
Anyone can show accuracy on a slide. The strongest evidence a forecasting or prediction vendor can offer is a ledger: predictions made in advance, compared openly against actuals, over months. Ask whether they track this for their clients and whether they will commit to tracking it for you. A vendor confident in their models will welcome being graded.
7. What happens without internet? Can this run inside a closed network?
If you operate in defense, critical infrastructure, or anywhere with sensitive data, ask early whether the proposed solution can run fully self-hosted inside a closed network, with nothing leaving it. Many offerings are architected around external cloud APIs and cannot. Sovereign, disconnected AI deployments now exist at the highest security levels internationally, so “it is impossible offline” is no longer a given; it is a design choice.
8. What success metric will we agree on before you build, and what happens if you miss it?
Insist on a success metric (KPI) defined in business terms before the build starts, measured against how you work today, with an agreed decision point if the numbers fall short. Projects without a pre-agreed metric do not fail; they just never end.
9. What is this project NOT going to do?
An honest scope has edges. Ask what is explicitly out of scope, which problems this approach is wrong for, and what kind of client the consultant turns away. Someone who has never said “we are not the right fit for this” to a paying customer will not start with you.
10. What happens after delivery?
Models drift as reality changes. Ask who monitors performance after handover, what retraining looks like, what it costs, and whether support is a contract or a favor. A delivered system with no after-plan is a countdown.