Buy or configure software first
The task is known and narrow, a product already does it, and the current project platform may cover it with a module or a setting. Run a trial on representative cases and measure the result yourself.
Guide
How to choose an AI consultant for a construction firm.
Construction experience, case studies, clear pricing and a phased plan all matter when hiring an AI consultant, but a firm can only judge them after a sales call, and every candidate will claim them. Written answers to the same brief are easier to compare.
The sections below cover when a consultant is the wrong first purchase, the brief to send every candidate, what a first engagement should leave behind, five pass/fail tests for comparing proposals and the signs that a proposal is selling a programme rather than answering a question.
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.
- Author
- Hugo Sequier
- Published
- 27 October 2026
- Reading time
- 6 min
- Updated
- 3 November 2026
01 / First decision
Decide whether a consultant is the right first purchase.
A consultant is one of three reasonable starting points. Choosing between them first saves a fee that answers the wrong question.
Hire a consultant first
The workflow crosses several systems or roles, nobody knows whether it is worth automating, approvals and liability need to be designed, or the firm wants evidence before committing a budget.
Neither yet
There is no named workflow, no owner and no examples. Any provider will be paid to find the question. Use the AI readiness checklist to pick one workflow first.
02 / Brief
Send every candidate the same one-page brief.
Proposals written against different assumptions cannot be compared. A short brief fixes the assumptions and shows quickly who engages with the actual work.
- The workflow: its trigger, its end state and the systems it passes through.
- The owner: the person who can say what a correct result looks like.
- Volume and baseline: cases per week or per project, time per case, known errors or late responses, measured or estimated.
- Examples: how many past cases can be sanitized and shared, including the awkward ones.
- Authority: which outputs must stay with a named person, such as an RFI answer, a submittal status or a bid decision.
- The decision that follows: build, buy, change the process, keep it manual or stop.
A candidate who answers the brief with a standard deck, without questions about the workflow, has answered a different brief.
03 / Scope
Keep the first engagement small enough to judge.
The first engagement should end in a decision the firm can act on, not in a plan for the next engagement. That is easier to check when the scope is bounded in writing.
- One workflow, named, with what is inside and outside its boundary.
- A fixed fee and a fixed duration for that boundary.
- Inspection of the firm’s own cases, with the number of cases stated.
- Named written outputs, delivered whether or not the work continues.
- Explicit exclusions: production deployment, security audits, legal advice and professional sign-off are usually separate.
- No obligation to commission the next stage.
Be cautious when the first proposal is a six to twelve month programme, an open retainer or a platform licence before anyone has looked at a single case.
04 / Deliverables
Know what the firm will hold at the end.
The outputs are the part of a consulting fee that stays with the firm. Ask for them by name, and ask whether they can be used internally or handed to another provider.
A bounded workflow map
The current process with its trigger, steps, handovers, exceptions and the point where work waits, drawn from real cases rather than an interview alone.
An evaluation on real cases
Results on the firm’s own examples against pass and fail criteria agreed before the results were seen, including the cases that failed.
An authority split
What a model may propose, what software must enforce and which named role approves each consequential output.
A failure register
The plausible errors, how each would be detected, and the stop condition and manual fallback for each.
A recommendation that can say no
Proceed, change the design, gather better data, buy a product, keep the workflow manual or stop, with the evidence behind the choice.
05 / Questions
Questions to ask before signing.
Ask them in writing and compare the answers side by side. Vague answers are information too.
- Which cases will be inspected, how many, and who selects them?
- How will the baseline be measured, and from whose records?
- What will the first version be allowed to do: draft, recommend or act? What will it never do?
- Which role approves each consequential output, and what does that person see when deciding?
- What happens when a document is missing, a check fails or a service is unavailable?
- Where are project documents processed and stored, who can access them, and how is access removed afterwards?
- What did the consultant decline to build, or narrow, on a previous engagement, and why?
- Does the consultant receive a commission, referral fee or other benefit from any product it may recommend?
- Who owns the code, prompts, evaluation cases and configuration produced during the engagement?
06 / Comparison
Score proposals with five pass/fail tests.
A weighted score lets a strong presentation hide a missing boundary. Record pass or fail for each test instead. A proposal that fails the bounded or exit test fails overall, whatever its price.
Bounded
One workflow is named, with an owner, inputs and an end state. Fails if the scope is “AI opportunities across the business”.
Evidence
Results come from the firm’s own cases against criteria fixed in advance. Fails if success is described as a workshop, a roadmap or a demo on vendor data.
Authority
Every consequential output has a named approving role. Fails if the answer is “a human stays in the loop” with no role or decision attached.
Exit
Stopping or staying manual is a stated possible outcome, the outputs are usable elsewhere and there is no obligation to continue. Fails if the next phase is assumed.
Price
The fee is fixed for the stated scope, or hourly with a cap and a reason the scope cannot yet be bounded. Fails if neither the fee nor its limit is clear.
07 / Red flags
Signs the proposal is selling, not assessing.
- Tools or platforms are recommended before anyone has seen the workflow.
- A savings percentage appears before a baseline has been measured.
- The first version is described as autonomous for RFI answers, submittal statuses or bid decisions.
- Case studies cannot state the starting position, the change and the measured result, even in anonymized form.
- There is no example of a project that was stopped, narrowed or kept manual.
- Questions about data handling, ownership or commissions get a general answer instead of a specific one.
Fees
Fixed fee or hourly for a first engagement.
Fees vary widely with scope, so compare what a fee buys rather than the headline number: the workflow, the cases inspected and the written outputs.
A fixed fee suits a first engagement on one workflow, because the boundary is known and the firm can compare offers like for like. Hourly or retainer billing suits open-ended strategy, change support or work whose scope genuinely cannot be bounded yet, ideally with a cap.
The AI readiness assessment guide lists what a paid assessment of one workflow should include, with a published example fee.
FAQ
Before the next decision.
Should a construction firm hire an AI consultant or buy AI software first?
Buy or configure software first when the task is narrow and a product already does it; trial it on representative cases. Hire a consultant first when the workflow crosses systems or roles, its value is uncertain, or approvals and liability need designing. If no workflow, owner or examples exist yet, neither will help until one workflow is chosen.
Continue with the evidence
Review the next relevant boundary.
Assessment / one workflow
Test the method on one workflow.
Describe the workflow, its owner, its inputs and how it is measured today. A person reviews every assessment.