Do you actually need AI? A checklist that says no a lot

Every readiness assessment we sell can end with a written recommendation not to build the AI system you came in asking about. That surprises people, mostly because the firms they talked to before us never put a no-build outcome on the menu at all.

Here’s the checklist behind those verdicts. You can run it as a 60-second interactive assessment or read on — either way you’ll know exactly what we look for before you spend a dollar.

Question 1: Can you write the task down?

Describe the task you want AI to do in one paragraph, including what “done correctly” means. If the paragraph keeps saying “it depends” — it depends on which customer, which document, who’s asking — that’s not a disqualifier; LLMs are good at “it depends.” But if you can’t say what a correct output looks like even for one concrete example, stop. You don’t have an AI problem yet; you have a process-definition problem, and a workshop is cheaper than a pilot.

The boring alternative that often wins: writing the process down and giving it to the team. A documented escalation flowchart is boring, costs almost nothing, and routinely beats a proposed “triage agent” on the metric that matters — because the agent would have been trained on a process nobody had defined.

Question 2: Does the knowledge exist anywhere your systems can reach?

RAG systems retrieve; they don’t know. If the answers your users need live in the heads of two senior employees, a retrieval system has nothing to retrieve. If they live in a wiki nobody has updated since 2023, the system will retrieve confidently wrong answers — which is worse than nothing.

The test: pick ten real questions your users would ask. Can you, with unlimited time and full access to your systems, find authoritative answers to at least eight? If yes, RAG can probably do it in seconds. If no, your project is a documentation project wearing an AI costume.

The boring alternative: a quarter spent fixing the knowledge base. Unsexy, transformative, and it makes the eventual AI system dramatically better if you still want it.

Question 3: What happens when it’s wrong?

Every LLM system is sometimes wrong. The question that decides whether you should ship one is: what does a wrong answer cost, and who catches it?

• Wrong answer gets caught by a human who was doing the task anyway → great use case (drafting, summarization, first-pass triage).
• Wrong answer goes straight to a customer with your logo on it → you need evaluation infrastructure, guardrails, and a human-in-the-loop design before you need a model.
• Wrong answer touches health, money, or law without review → the review step is the design. Systems in these domains work when they make the reviewer faster, not when they replace the reviewer.

None of this rules out regulated use cases — our most-cited build is a patient-intake system running in a HIPAA-eligible environment. It works because the design assumed fallibility from day one: structured handoffs to humans, audit trails, and clear boundaries on what the AI is allowed to decide (nothing clinical).

Question 4: Will anyone change how they work?

The graveyard of GenAI pilots is full of technically successful systems nobody uses. A pilot that “works” but requires people to open a second tool, trust a new workflow, and abandon habits that currently pay their bonuses — that pilot was dead before the kickoff.

The test is organizational, not technical: name the person who will own this system in six months. Not sponsor — own. Answers questions about it, watches its metrics, fights for its budget. If that person doesn’t exist, the pilot’s real deliverable is a slide, and you can buy slides cheaper.

Scoring it

Four questions: task defined, knowledge reachable, error cost survivable, owner named. Here’s how we read the score:

• 4 of 4 — build. Skip the ceremony, scope a pilot with acceptance criteria, ship in weeks.
• 3 of 4 — build the missing piece first. Usually it’s Question 2 (fix the docs) or 4 (find the owner). Then build.
• 2 or fewer — you don’t need AI yet. You need the boring thing. We’ll put that in writing, with the roadmap to change the answer — because “not yet” is a verdict with a date on it, not a rejection.

If you want the scored version with your numbers in it, that’s literally what the readiness assessment is. And if you score yourself at two, the honest move is the one nobody sells: don’t buy anything yet — including from us.