By KYN AI Advisory Team — AI implementation specialists, Singapore
Before you hire an AI agent consultancy, ask twelve questions: what they will build, who has run it in production, who owns the output, how it is secured, who maintains it and what it costs over three years. A good consultancy answers each with a document, a setting or a test. A weak one answers with a deck. Use the questions below in the first call, and score them afterwards. If you are comparing products rather than consultancies, see how to evaluate an AI agent vendor: 15 questions.
Why ask about production, not prototypes?
Anyone can show a prototype. A prototype handles the cases its builder imagined. Production handles the cases customers send: a voice note instead of text, a reply in two languages, a duplicate record, a model that was updated last week.
Relevance AI, a platform vendor, makes a similar point in its writing about production-grade agents. The gap between demo and daily use is where projects fail. Your questions should push there.
Scope: what will they build?
1. Which three tasks will the agent do, and which will it not do? A clear answer names tasks and limits. "A workforce of agents for your whole business" is not an answer.
2. How will we know it works? Ask for the acceptance test before work starts. For example: 100 real cases marked by your team, with a target for correct outcomes and for missed handovers.
3. What will you need from us? Access to systems, a named reviewer, 100 real examples. If the consultancy needs nothing from you, ask what it assumes.
Experience: who has done this?
4. Which systems have you run in production, for how long, in what industries? Ask for industries and durations, and a reference you can call. We name clients by industry only, and so should they if the clients have not agreed to be named.
5. What went wrong on the last project, and what changed? Every real project has a failure. A consultancy that cannot name one has either not shipped or is not candid.
6. Who will do the work? Names, roles and how many hours. A senior person in the sales call and a junior on the build is common. Ask directly.
Ownership: what do we keep?
7. Who owns the code, prompts, configuration and data? Rights should pass to you on full payment, in the contract. See who owns the data and prompts in a custom agent project.
8. Where does the system run, and in whose account? The strongest answer for an SME is a cloud account in your name, with the consultancy as a revocable role. Then, if you end the contract, the system keeps running.
Safety: how is it controlled?
9. What can the agent do without a person? Outbound messages, money, filings and deletions should wait for a named person. Ask to see the rule, not only hear it.
10. How are people and agents authenticated, and how is data isolated? Look for named accounts, no shared logins, restricted agent accounts and row-level security. Our security note shows the standard we apply, and our list of questions to ask an AI vendor about security goes further.
Cost: what will it cost?
11. What are the build fee, the monthly cost and the usage cost? Three lines, in writing. Ask for the three-year total. See what a custom AI agent project costs in Singapore.
12. What happens in month 13? Models change, tools change, your process changes. Ask who handles it and what it costs. A project with no plan for month 13 is a project with a short life.
What red flags should you watch for?
- A fixed price quoted before anyone has looked at your systems.
- No test on your data.
- "Proprietary platform" with no export.
- Claims of full automation with no approval steps.
- A refusal to say which model providers they use.
- A promise of fixed results for a system that depends on your data.
None of these proves bad faith. Each is a reason to ask a follow-up question.
How should you run the selection?
- Write the three tasks and the acceptance test before you speak to anyone.
- Speak to three consultancies. Ask the same twelve questions.
- Score each answer 0, 1 or 2: 0 for no answer, 1 for a general answer, 2 for a specific answer backed by a document or a test.
- Ask the top two for a paid review of your stack, with a written output.
- Choose on the review, not on the pitch.
The paid review is the cheapest protection you have. It shows how the consultancy thinks about your problem, at a cost far below a failed build. See also how to pick an AI agent development company.
What should the first call cover?
Keep the first call to forty-five minutes. Spend ten on your problem, twenty on the questions above and fifteen on next steps. Ask the consultancy to restate your problem in its own words. If it cannot do that, it has not listened. Ask what it would build first and what it would refuse to build. Ask for a sample of a review output from an earlier project, with client details removed. After the call, write down what you heard before you speak to the next firm, because the pitches blur together. Compare notes against the scoring table, not against your impression of the presenter.
After the pilot, score the consultancy against your question list: which answers held up in practice, which did not, and what you would ask differently next time. Keep the scorecard with the contract and review it at month six before agreeing any change.
Where KYN fits
We encourage you to ask us these questions. KYN builds custom systems in a cloud account in your name for SMEs in Singapore, Hong Kong, Indonesia and Malaysia. The first 14 days are a review: an inventory of your stack and interviews with your team. A full build ships in under 90 days. Pricing is a one-off build fee and a monthly retainer of at least twelve months, quoted after the review.
Rights in the custom code and configuration pass to you on full payment, and you can revoke KYN's role while everything keeps running. Every action is logged with the actor. See case studies and security.
Frequently asked questions
Should we hire a consultancy or build in-house?
If you have an engineer who has shipped an agent to production, build in-house. If not, a consultancy shortens the time and removes the learning cost. Either way, insist on ownership and documentation so you are not locked in.
How long should a first project take?
Scope it to weeks, not quarters. At KYN, the review takes 14 days and a full build ships in under 90 days. A narrow first agent can go live earlier.
Do we need a technical person on our side?
You need a named owner who understands the business process, and someone who can approve access to systems. You do not need an engineer, but you do need a reviewer for the agent's output.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
