AI readiness checklist: 25 questions before you build an agent
Most agent projects don't fail on the model — they fail on a question nobody asked before week three. Run this checklist with the people who own the use case, the data and the risk sign-off, before you write a spec.
Is this worth building at all?
An agent that's technically impressive and commercially marginal is still a failed project. Answer these before scoping anything.
- 01What decision or task will the agent actually replace, end to end — not just assist with?
- 02What does doing it manually cost today, in hours, error rate or backlog?
- 03What's the single metric we'll use to say this worked, agreed before we build it?
- 04What happens when the agent is wrong 1 time in 20 — who catches it, and how fast?
Can the agent trust what it reads?
Agents amplify whatever the underlying data quality already is — good or bad.
- 05Where does the input data live today, and who owns it?
- 06Is the data clean enough to trust, or does it need a cleanup pass first?
- 07Can every field the agent reads be traced back to a system of record?
- 08What's the retention and deletion policy for anything the agent touches or stores?
What does it actually have to plug into?
The agent is often the easy part. The connectors to your CRM, ERP or helpdesk are where timelines slip.
- 09Which systems does the agent need to read from and write to?
- 10Do those systems expose an API, or would this need screen-scraping or manual export?
- 11Who owns access and credentials, and can we grant least-privilege scopes instead of admin?
- 12What happens when an integration goes down — does the agent queue, retry, or fail loud?
What would a regulator or a customer ask?
The EU AI Act (Reg. (EU) 2024/1689) already has obligations in force, and several US state rules apply today regardless of where your company is incorporated.
- 13Does the agent make or materially influence a decision about a person — credit, hiring, healthcare, insurance, pricing? That's the EU AI Act's high-risk test.
- 14Where are the people affected by its output located — does any of that reach the EU? Art. 2 of the AI Act reaches non-EU providers whose output is used there.
- 15Have the staff who operate or oversee it been given the AI-literacy training the AI Act's Art. 4 has required since 2 Feb 2025?
- 16If the system scores or profiles people, does it fall under a US state rule already active — California's CCPA ADMT rules (from 1 Jan 2026), Illinois HB 3773 (1 Jan 2026), or NYC Local Law 144 (already in force)?
- 17Who signs off before this goes live, and what's on the record if someone asks how it decided something?
Who's watching it after launch day?
Launch is the easy 10%. An agent nobody monitors degrades silently.
- 18Who watches it once it's live, and on what cadence?
- 19What's the rollback plan if a model or prompt update changes its behaviour?
- 20How do we catch drift or quality loss before a customer does?
- 21What's the cost ceiling per run, and who's alerted if it's breached?
Are the humans in the loop actually equipped?
An escalation path only works if the person at the end of it understands what they're looking at.
- 22Do the people supervising the agent understand its failure modes, not just its happy path?
- 23Is there a named person who can override or stop it, with the actual authority to do so?
- 24Has the team that will field complaints about it been briefed and trained?
- 25Who updates the runbook when the agent's scope changes?
If you answered "not sure" more than a few times, that's normal.
This checklist is the input to a proper scoping exercise, not a replacement for one. Here's where each gap gets closed:
AI Opportunity Sprint
A fixed-fee, 1–2 week engagement that ranks candidate use cases by value and feasibility, maps your data and systems, and classifies risk under the EU AI Act and relevant US state rules.
AI Governance Starter
An inventory of AI already in use, AI-literacy training for staff, a transparency check, and a policy pack your sign-off process can point to.
AgentOps
Monitoring, evaluation suites, model upgrades, cost control and compliance logs once the agent is live — with human-on-the-loop escalation built in.
Checklist done.Let's scope it.
Send us where you got stuck — we'll tell you honestly whether it's a two-week sprint or a bigger conversation.