AI Readiness Assessment: How to Check a Workflow Before You Build AI on It

An AI agent will not fix what is already broken in a workflow. These are the checks a construction, logistics or manufacturing process has to pass before anyone builds on it, and the evidence that shows whether it does.

Ground Floor Audit · Updated

What an AI readiness assessment should answer

Most AI readiness assessments score a whole company: strategy, data platform, talent, culture, governance, each rated on a scale and averaged into a maturity level. That number is useful to a board. It is close to useless to an operations director, because it does not say which piece of work to change first.

For an operations team, a readiness assessment is worth doing when it answers three questions:

  1. Which workflow is ready to be rebuilt with AI first, and which ones are not ready yet?
  2. What has to be fixed in that workflow before an agent touches it: a missing owner, an undefined exception path, data that never gets reconciled?
  3. How you will know it worked: the baseline the workflow runs at today and the result that counts as a pass.

Everything below is built to answer those three questions for one workflow at a time. Run it on three to five candidate workflows and compare the results; the prioritization frameworks piece covers how to pick the candidates.

Why assess the workflow, not the company

A company can have a modern ERP, a data warehouse and an AI policy and still run its change orders over email, with a different approval path on every project. It can also have ageing systems and one dispatch process that is documented, followed on every shift and measured weekly. The enterprise score treats the first company as more ready. The second is the one that can put an agent into real work.

Readiness is a property of a specific process: its steps, its inputs, the people who own it and the decisions made inside it. Assessing at that level also keeps the scope honest. An assessment that ends in a platform recommendation before it has named a workflow has answered a different question from the one you asked.

The six gates to check

Treat each area as a gate the workflow passes or fails, not a score to average. A workflow that is strong on data but has no accountable owner is not two-thirds ready; it is not ready, and the assessment should say what fixes it.

1. Process stability

Is there one repeatable process to build on, or several versions that share a name?

  • Does the work follow the same steps across crews, shifts, sites or desks?
  • Does the written procedure match what people actually do?
  • Is the volume steady enough that a change can be measured?

Evidence: walk-throughs with the people doing the work, a sample of recent cases followed end to end, the written procedure side by side with what you observed. Not ready if:every case depends on one person's judgment, or teams run materially different versions of the work. The stability threshold test goes deeper on this gate, and reading the gap between procedure and field behavior covers how to observe it.

2. Inputs and data

Can the information the workflow runs on be reached, trusted and kept current?

  • Which systems, spreadsheets, inboxes and paper records feed each step?
  • Who corrects a record when two of them disagree, and how often does that happen?
  • Is there enough history of past cases and outcomes to test against?

Evidence: an inventory of every input for this one workflow, where it lives and who owns it, plus a sample checked for gaps and conflicts. Not ready if:a step depends on information that only exists in someone's head or a personal inbox, or records are never reconciled. See baseline data requirements: you need the data this workflow uses in order, not a company-wide data program.

3. Handoffs and exceptions

Is it clear where work moves between people or systems, and what happens when it goes wrong?

  • Where does work pass between teams, companies or systems, and what is passed?
  • What are the common exceptions, and who decides each one?
  • Which workarounds keep the process running today?

Evidence: a current-state map that shows handoffs and exception paths, and a list of workarounds with the reason each exists. Not ready if: exceptions are handled ad hoc, or a handoff only works because someone happens to notice. The guides to handoff failures, documenting workarounds and shadow workflow mapping cover how to find them.

4. Ownership and decision rights

Is one person accountable for the outcome, and are the decisions inside the workflow defined?

  • Who owns the result, not just a step, and can they change how the work is done?
  • Which decisions follow a rule, and which need judgment?
  • Who approves what, and where do approvals queue?

Evidence: a named owner who has agreed to the work, and a list of the decisions in the workflow marked rule-based or judgment. Not ready if: several people can change the workflow but nobody owns the outcome. On the judgment calls, see auditing human judgment calls and approval bottlenecks.

5. A measured baseline

Do you know how the workflow performs today, in numbers you can measure again later?

  • Cycle time from start to finish, and where the waiting happens.
  • Volume per week or month, and how much of it is exceptions.
  • Error and rework rate, and what each costs downstream.

Evidence: current figures taken from real cases, with how they were measured written down. Not ready if:nobody can say what “better” would look like in a number. Rework rate is often the most honest of these in field operations.

6. Controls and people

Is it clear what an agent may do on its own, and are the people who do the work part of the change?

  • Which actions need a person to review or approve them before they take effect?
  • What could go wrong if an action is wrong, and how would you notice?
  • Have the people who do the work today been involved, and who trains them on the new way?

Evidence: a short list of what the agent may and may not do, the review points, and who signs off. Not ready if: the plan has no review point for consequential actions, or the team doing the work first hears about it at launch. See human-in-the-loop design.

The assessment on one page

Copy this into a spreadsheet and fill in one row per gate for each candidate workflow.

GateThe questionEvidence to collectFails when
StabilityIs it one repeatable process?Walk-throughs, sample cases, procedure vs. practiceVersions differ by team, or one person holds it together
Inputs and dataCan its inputs be reached and trusted?Input inventory with owners, a checked sampleKey information lives in heads or inboxes, or is never reconciled
Handoffs and exceptionsAre handoffs and exceptions defined?Current-state map, list of workaroundsExceptions handled ad hoc; handoffs rely on someone noticing
OwnershipDoes one person own the outcome?Named owner, decisions marked rule or judgmentMany editors, no owner
BaselineDo you know how it performs today?Cycle time, volume, rework from real casesNo number for “better”
Controls and peopleIs it clear what an agent may do, and is the team involved?Allowed actions, review points, sign-off, training ownerNo review on consequential actions; team not involved

How to run it

  1. List the candidates. Pick three to five workflows where the pain is known and the result can be measured: in construction, change orders, RFIs, submittals or keeping the schedule in step with procurement; in logistics, catching late loads, dispatch exceptions or matching proof of delivery to billing.
  2. Talk to the people who do the work. Interview and shadow the people who run each step, not only the managers who describe it. The interview techniques piece has the questions.
  3. Follow real cases end to end. Take a handful of recent cases and trace each one through every system, email and handoff it touched.
  4. Collect the evidence for each gate using the table above, and mark each gate pass or fail with the reason.
  5. Choose one workflow that passes, or the one whose failed gates are cheapest to fix, and write down what changes before any agent is built.

Keep the first pass narrow. Assessing one department's three to five workflows well is more useful than a survey of every process in the company.

What you should have at the end

  • A ranked shortlist of candidate workflows with the pass or fail result for each gate.
  • A current-state map of the chosen workflow, with handoffs, exceptions and workarounds.
  • The list of fixes to make before an agent touches it, each with an owner.
  • Baseline figures and the result that counts as a pass for the first rebuild.
  • A named owner and a clear statement of what the agent may and may not do.

If an assessment ends with a maturity score and a platform roadmap instead, it has not told you what to fix first. The pre-AI process audit engagement walk-through shows what this looks like in construction and logistics.

Doing it yourself or bringing someone in

A team with an operations lead who can spend time on the floor can run this assessment on its own. Outside help is worth it when nobody has that time, when the workflow crosses several companies or systems, or when the findings need to carry weight with leadership; see making the internal case.

Many software vendors offer a free AI readiness assessment. Before you accept one, ask:

  • Will the result name a workflow, or a product?
  • Will they observe the work and real cases, or rely on a survey?
  • Will they tell you a workflow is not ready, and what to fix first?
  • Is the scope and fee fixed in writing before the work starts?

The vendor shortlist criteria and what separates a real diagnostician from an integrator cover how to tell them apart. Terminal Use, an AI-first rebuilder for construction, logistics, manufacturing and retail operations, embeds engineers with the team to map a process in days and put the first rebuilt workflow live in weeks.

Questions operators ask

How long does an AI readiness assessment take?

It depends on how many workflows are in scope and how easy the people and records are to reach. Scoped to one department's few candidate workflows, it is a short piece of work; scoped to the whole company, it can run for months without naming a first workflow. Ask anyone quoting one to say which workflows are in scope and when you get the result.

What does one cost?

Prices vary with scope and with who runs it, and many vendors offer one free as part of a sale. Ask for a fixed scope and fee in writing, and compare what each one delivers against the list above.

Do we need a data platform before we are ready?

Not for a first workflow. You need the inputs that one workflow uses to be reachable, owned and checked. A company-wide data program can run alongside; it should not be the gate for starting.

What if every workflow fails a gate?

That is a useful result. Pick the workflow whose failed gates are cheapest to fix, fix them, and assess it again. Most of those fixes, such as naming an owner, defining the exception paths and writing down the real procedure, make the work better before any AI is involved.

Read next