Density Labs / Blog /Method

The AI readiness checklist: 20 questions

Twenty yes or no questions across workflows, data, team, and risk. Answer them in one meeting and you will know whether to build, and what to fix first.

July 21, 2026 · Updated July 21, 2026 · Federico Ramallo, founder of Density Labs · 6 min read

An AI readiness checklist is a set of yes or no questions that tests whether a business can turn an AI pilot into a production system. It covers four areas: the workflows AI would change, the data those workflows produce, the team that will own the result, and the risks that need guardrails before anything runs on its own.

That is the whole idea. No maturity matrix, no vendor scorecard, no forty-page survey. MIT’s NANDA project found in 2025 that 95% of AI pilots never reach production, and the failures trace back to questions nobody asked before the build started. Every question below can be answered in a single meeting with the people who do the work. Print this page, bring it to that meeting, and count the yes answers honestly.

Part 1: Workflows and value

AI creates value in workflows, not in mission statements. Before any tool talk, confirm you know where the hours actually go.

  1. Can you name three repeatable workflows that eat paid hours every week? Quoting, scheduling, document handling, support replies, reporting, data entry. If you cannot list three, you are not ready to shortlist one.
  2. Has anyone watched the work being done, step by step? Thirty minutes sitting next to the person, writing down every step, every exception, every system touched. A meeting about the work does not count.
  3. Do you know how many paid hours per week your top candidate consumes? A real number, not a feeling. Hours times loaded hourly cost is the only honest starting point for ROI.
  4. Does the decision inside that workflow repeat often enough for automation to pay off? A task done twice a quarter never earns back its build cost. A task done fifty times a day might.
  5. Would the projected return be at least three times the total cost? Count the build, the model usage, the human review time that stays, and maintenance. If the math does not clear 3x, pick a different workflow, not a cheaper vendor.

Part 2: Data

Every AI project is secretly a data project. This section is the gate: a no here blocks everything downstream.

  1. Do digital examples of the work exist? Past tickets with the replies that were sent, past documents with the data that was extracted. If the knowledge lives only in someone’s head, there is nothing to build on.
  2. Can you actually find and export those examples? Existing in a system you cannot query is the same as not existing. A no here is fixable, but fix it before you build.
  3. Could you assemble 20 to 50 real past examples with known correct outputs? This is the golden set, the single highest-leverage artifact in any AI project. It turns every vendor demo and every internal debate into a measurable pass rate.
  4. Is there a person who can label what “correct” looks like for those examples? Someone has to be the answer key. If nobody can say which output was right, no model can learn it and no test can prove it.
  5. Does fresh data keep flowing in the same shape? A one-time export gets you a prototype. Production needs the pipeline to stay alive after launch week.

Part 3: Team and ownership

“Nobody owned it” is the most common cause of quiet death after a promising pilot. Ownership questions feel soft until the first bad output lands in front of a customer.

  1. Is there one person who owns the output of this workflow today? Not a committee. One name.
  2. Can that person describe what “good” looks like in writing? If quality lives as folklore, the AI will learn the folklore, errors included.
  3. Is someone assigned to review the AI’s output daily during rollout? Review time is a real cost. Budget it or the reviews silently stop, and then the errors ship.
  4. Have you decided who owns the system when it misbehaves? Decide before launch. The first embarrassing output is the wrong moment to invent an escalation chart.
  5. Will the people who do the work today help build the replacement? Automation done to a team gets sabotaged or ignored. Automation done with a team gets adopted, because they wrote the exceptions into the spec.

Part 4: Risk and guardrails

Autonomy is earned in stages, never granted on day one. These questions test whether you can afford to be wrong while the system learns.

  1. If the AI is wrong one time in ten, does a human catch it cheaply? Some workflows tolerate error and correction. Some, like pricing or compliance, do not. Know which kind yours is.
  2. Can you run the system in stages? Sandbox first, where humans see every output and nothing leaves the building. Then assist, where a human approves every send. Then limited autonomy only for the categories with proven pass rates.
  3. Can you log every input, output, and human override? The overrides are next month’s training data and your audit trail. No logging, no learning, no defense.
  4. Is there an escalation path for the cases the system cannot handle? Every workflow has exceptions. The question is whether they route to a person or pile up unseen.
  5. Will you demand a measured pass rate before trusting any demo? Demos are the reason for the 95% statistic. A vendor who will not run against your golden set has answered your question about the vendor.

How to score it

Count the yes answers per section. Mostly no in any section means you have found your real first project, and it is not an AI build: it is mapping the workflow, assembling the golden set, or naming an owner, work that costs days rather than months. Mostly no on data specifically means stop, because nothing downstream survives a data gap. Mostly yes across all four sections means you are ahead of most companies that already signed a contract: pick the single highest-scoring workflow, build the golden set, and run a measured pilot before spending real money.

Get the full playbook

The full playbook, with the scoring card, the ROI math, and the staged autonomy ladder, is free: download the AI Readiness Playbook (PDF).

When yes answers are not enough

If the checklist surfaces gaps you cannot close in-house, that is exactly the line where our paid engagement starts. The AI readiness assessment is $2,500 and takes two weeks: a senior engineer runs this entire loop on your operation, watching the real work, building the golden set, measuring pass rates, and handing you a written plan your team can execute with or without us. We keep a 96% client retention rate, Clutch verified, in part because that plan sometimes says “build nothing yet.” See what the deliverable looks like in the redacted sample report, and if you run a smaller operation, there is a version built for small businesses.