The playbook
- 1
1 — Show me 3 live UK production deployments
If they can't, walk away. Demos and pilots don't count.
- 2
2 — Who owns the IP and data?
You should. If they say 'we license it back to you,' walk away.
- 3
3 — Where is data processed and stored?
UK or EU. If US — interrogate the GDPR position.
- 4
4 — What happens if you go bust?
You should be able to take the code, prompts and configs and run them yourself or with another agency.
- 5
5 — What's the realistic accuracy on day 1 vs month 3?
Honest answer: 85–90% day 1, 95%+ month 3. If they say 99% day 1 — they're lying.
- 6
6 — How do you handle edge cases and escalations?
Should be a defined human-in-loop pattern, not 'the AI figures it out.'
- 7
7 — What's included in ongoing run cost?
Hosting, inference, monitoring, retraining cadence, support SLA. Get it in writing.
- 8
8 — Who's liable if the AI gets it wrong?
Contractually clear liability split. Most agencies cap at fees paid. Acceptable if you have your own insurance.
- 9
9 — How long to first production deployment?
4–6 weeks for one agent. Anything longer = over-scoped. Anything shorter = under-engineered.
- 10
10 — What happens at month 12?
Ongoing relationship terms, retraining cadence, exit terms. No long lock-ins.
- 11
11 — Can I speak to 2 reference clients?
If they refuse, walk away.
- 12
12 — What would make you turn down this project?
Tells you whether they think critically about fit. Best agencies say 'no' often.
What you walk away with
Frequently asked
What if the agency refuses to answer?+
That's your answer. Walk away.