Guide
An AI readiness assessment for a construction firm checks one task at a time, such as logging RFIs, checking a submittal against a specification or reading a tender pack. The same firm can be ready for one of those tasks and not for another, because each depends on different documents, systems and approvers.
Below is a filled-in sheet for one workflow, then the seven gates it uses, each with the sign that the workflow is not ready yet. A single failed gate, such as no owner or no shareable examples, outweighs a good score everywhere else.
Hugo Sequier runs State Method, which sells the assessments, pilots and builds these guides discuss. Apply the same tests to it as to any other provider.
Definition
An AI readiness assessment is a structured check of whether a specific process, its inputs and the people responsible for it can support an AI-assisted step under control. It asks whether this workflow is valuable, measurable and controllable enough to test, and what must change first, rather than whether the company as a whole is ready for AI.
A company-wide maturity scorecard can still help with policy, training and security planning. It cannot tell a project team whether submittal intake will survive scanned attachments, a re-submission or a reviewer who is on leave.
Worked sheet
This sheet applies the seven gates below to the illustrative Diagnostic excerpt published on the Diagnostic page: a product submittal checked against an approved specification section. The requirement and product data are synthetic and no cases were run, so it shows how a sheet records a failed gate, not a client result.
Trigger: a product submittal received against an approved specification section. Output: a compliance matrix row with a cited status per requirement. Owner and approver: the Technical Manager. Out of scope: professional, regulatory or design sign-off.
The current approved specification revision is not confirmed, and certificate completeness cannot be established from the sampled submittals.
The excerpt holds no time per case or weekly volume. On a real assessment this gate stays open until a sample of cases has been timed.
The workflow reads from the document store. The compliance register only receives a proposed update, never a direct write.
The model extracts requirement candidates and maps product claims as a cited proposal. Software checks the specification version, certificate presence and identifier match. The model may not set a compliance status; the Technical Manager determines and approves each one.
Five failure modes are written down with a signal and a response. If the approved specification version is unavailable, the row stops, goes to document control and no status is issued. Submitted text that contains instructions is treated as data, with no tool access, and the denial is logged.
No agreed rule exists for mapping a regional performance classification to the specified class, so pass and fail criteria for those rows cannot be set before a test.
Confirm the approved revision, assemble the certificate set and agree the mapping rule, then reassess. The sheet does not recommend a pilot yet.
Four of the seven gates pass, which a percentage score would report as 57 percent ready. The sheet records two blocking reasons instead, each with the fix that clears it.
01 / Workflow
Not ready yet if the request is “use AI for document control” or “AI across every project”. That is a programme. Pick one path through it first.
02 / Documents and data
Not ready yet if the only examples available are the clean ones, or if nobody can say which revision of a drawing was current on a given date.
03 / Baseline
Readiness includes knowing what the workflow costs today. Without a baseline, any later result is an opinion.
Not ready yet if the case for AI rests on an estimate of saved hours that nobody has measured.
04 / Systems and access
Not ready yet if the plan assumes write access to a live project register before anything has been tested.
05 / Authority
Construction decisions carry contractual, design and safety consequences. Readiness means those decisions already have named owners, so the workflow can route to them rather than replace them.
Classify a message, extract references, summarize a question, flag a possible conflict or draft a clarification.
Register rules, revision order, required fields, due dates calculated from the contract and who is allowed to see or change a record.
The RFI answer, the submittal status, the drawing review record, the bid decision and any action with commercial or safety effect.
Not ready yet if “a human checks it” is the whole answer. Name the role, the decision and what they see when they make it.
06 / Failure
Not ready yet if the only plan for a wrong answer is that someone will probably notice.
07 / Decision
Not ready yet if the decision to build has already been made and the assessment is expected to confirm it.
Scoring
Averaging these groups into a readiness score hides the gate that matters. A workflow with excellent data and no owner is not 85 percent ready; it is blocked. Record one of three outcomes for each workflow.
Every group passes. The next step is testing representative cases against agreed criteria.
One or two groups fail for a reason with an owner and a date, such as measuring the baseline or getting a sandbox.
The workflow, owner or inputs are missing. Narrow the scope, choose another workflow or keep the current process.
First workflow
A firm usually has more candidates than it can test: RFI logging, submittal checks, tender reading, progress reports. On 24 and 25 October 2026, State Method asked the ChatGPT app, the Gemini app and Google AI Mode how architecture and engineering firms decide which AI use case to start with, and how to pick one workflow when every department has an idea. All six answers weighed value against ease or feasibility, and five ranked the candidates on a 1-to-5 scorecard or a value-against-effort matrix. Six answers show the advice a buyer is given, not what firms actually do.
In the three scorecards, a measured starting point was at most one criterion: 15 percent of the total in the one answer that gave weights, one of four in another, and absent from the third, which left the baseline to the pilot. A high value rating can therefore put first a workflow whose current effort nobody has timed. Scores order candidates that could all be tested. They do not reveal the ones that cannot, so apply gates 01, 03 and 05 as filters and rank only what passes.
A candidate without a named workflow and owner, a way to time the work today, or a role that approves each consequential output does not get a low score. It waits until that gap has an owner and a date.
A workflow that runs every week leaves enough past cases to test against known outcomes. One that runs twice a year leaves too few, however painful each instance is.
Prefer work whose output already passes a person before it leaves the firm, such as a submittal reviewer or a document controller. An error found there costs a correction rather than a dispute.
If the project platform already offers the step, configuring it may beat building one. The RFI and submittal guide compares what Procore and Autodesk Forma cover for intake.
Each workflow example on this site names one owner role, from document controller to handover manager: gate 05 written down before anything is built.
Firm-level frameworks
Ask an AI assistant how a design or construction firm can tell whether it is ready, and the answer usually describes the firm rather than a task. Between 24 October and 1 November 2026, State Method put three readiness questions to the ChatGPT app, the Gemini app and Google AI Mode, and 14 answers came back. Twelve structured the check as a firm-wide audit of areas such as data, technology, workflows, people and governance. The other two were ranked lists of assessment products. Four of the 14 named the AIA AI Firm Toolkit as a place to start. It is a small sample of AI answers, not a survey of firms.
That toolkit is free and built for architecture practices. Its maturity quiz has five questions, on AI policy and awareness, tool usage, documentation and training, verification and quality control, and culture and leadership. The answers place a firm on one of four levels, from ad hoc experiments to an AI-first firm, each with suggested next moves.
Gates 01 to 04 and gate 07 have no counterpart in the quiz: which task, triggered by what and owned by whom; which documents and revisions; how long the work takes today; which systems are read or written; who will decide on the result.
The quiz asks how the firm verifies AI output in general. Gates 05 and 06 ask the same thing of a single output: which role approves it, what the most damaging error would be, and what stops the workflow when it happens.
Written policy, how staff already use AI tools, training, and how leadership approaches adoption. A workflow can pass all seven gates in a firm with no AI policy, which is a firm-level gap the quiz will show and this sheet will not.
Use both, in order: the quiz to set policy and training across the firm, then the seven gates on the one workflow chosen for a test.
Next step by gate
Scope a proof of concept: one question, past cases with known outcomes and thresholds set before the run.
Write the workflow control sheet: approver, permission matrix, audit record and stop conditions.
Send a one-page brief and test each proposal before signing.
Run the three filters in the first-workflow section, then read the guide for each workflow that passes: what a model can prepare, which checks stay in software and who approves.
Cost and scope
A self-assessment like this one is free and useful for choosing which workflow to examine. A paid assessment should go further: inspect your own cases against the baseline and produce evidence you can use whether or not you continue with the same provider.
Ask for a fixed scope and a fixed fee when the subject is one workflow. Hourly billing suits open-ended strategy work, where the scope cannot be bounded in advance.
The State Method Diagnostic is priced this way: one workflow, 10 business days, a fixed €4,000 excluding VAT. It ends with five decision artifacts (workflow map, evaluation summary, permission matrix, failure register and go/no-go recommendation), and the excerpt behind the sheet above shows each of them in short form. A pilot is scoped separately, and only when the evidence supports one.
FAQ
The person who owns the workflow outcome, someone who does the work today, and whoever controls access to the systems involved. Approving roles, such as the design lead for RFI answers, should confirm where their decision sits.
Continue with the evidence
Assessment / one workflow
Describe the workflow, its owner, its inputs and how it is measured today. A person reviews every assessment.