An enquiry for consultancy services is published, bids due in ten days. AI can help produce the technical submission for a construction tender, but the time is usually lost before the first line: you answer out of habit, hesitate for three days, then write a generic submission under pressure. The method in AI for Architecture and Engineering (chapters 19 and 20) comes down to three steps: decide in writing whether to bid, build the submission on the scoring grid, and write without a single unsourced commitment.
The workshop · a worked case
Every criterion has to lead to evidence.
| Criterion in the scenario | Response to build | Evidence and action |
|---|---|---|
| Method · 30% | Construction sequencing on an occupied site | Programme + phasing diagramTo prepare |
| Team · 20% | Roles, competences, availability | CVs + organisation chartAvailability to confirm |
| Environment · 10% | Comparison of the options | Indicator matrixSource to attach |
| Price · 40% | Consistent workload, services and fees | Fee breakdownDirector’s decision |
These weightings illustrate a scenario; they are not a standard grid. The instructions to tenderers for your own enquiry govern the structure of your response.
0 / 4 points reviewed — a local aide-mémoire; it never decides GO on its own.
Fictional teaching example · Adapt it to the data and the scope of your own commission.
01 · Decide whether to bid, before you write
Go/No-Go is settled before any drafting. The book reduces it to six questions:
- Conditions of entry: capacity, guarantees, documents required. Are we eligible?
- Experience: do our references match the references demanded?
- Chances of winning: scored criteria, likely competition.
- Team: do the CVs required exist, and are those people available on the dates?
- Programme: do the bid deadline and the commission duration fit the workload plan?
- Profitability: do the budget and the scope allow a profitable commission?
The tool works through each question, citing the clause in the instructions to tenderers, from the company’s record of references, CVs and methods. In the anonymised extract the book gives, it marks one reference “COMPLIANT” on floor area and year. That status is only earned once the project sheets have been read: type of facility, full scope of services, year of handover. Chances of winning, workload and costs appear neither in the instructions nor in the company record: you supply those.
The verdict stays human and is written down with its reason: GO, NO-GO or TO BE NEGOTIATED, with the precise condition to be lifted. File it with the outcome. After a year, you know your real win rate. Plate 47 takes the exercise as far as a weighted score, on a fictional grid: a favourable score does not lift a blocking condition, and a partner’s promise is not a secured resource.
02 · Turn the scoring grid into a work plan
Once GO is decided, the submission is not written as free text. It is assembled from evidence, in the order of the criteria the panel will score. In the conversation for that enquiry, load the instructions to tenderers and the published scoring grid, then make this request:
Here are the instructions to tenderers and the published scoring grid [documents]. Produce the requirement-to-evidence matrix: for every scored criterion in the instructions: exact requirement (cite the clause) / evidence expected / document or reference in the company record that answers it / person responsible for drafting it / status. Mark every criterion with no evidence available in RED — those define this week’s drafting plan.
What you have to check: read the table in both directions. From the instructions to the table: every scored criterion and sub-criterion has its line, with the exact clause, because a criterion missed here will be missed in the submission too. From the table to the company record: every piece of evidence cited exists, and says what the table claims it says. The RED lines become the work plan, each with an owner and a date.
The contents page of the submission then follows exactly the structure of the sub-criteria in the instructions. The assessor finds their own grid without searching for it. You then write criterion by criterion: the tool proposes text supported by the evidence indicated, and you check that the evidence is cited exactly. The panel scores what it can verify.
03 · Write without inventing commitments
A language model left to itself will happily produce plausible figures. In a bid, a figure written down is a commitment. So the book imposes a rule, which it calls the no-embroidery rule:
“You will produce no operational figure (duration, percentage, threshold, volume) without an enforceable source identified in the project documents.”
In practice, when no document fixes a duration, the tool writes “Concept design issued at [week __ — to be confirmed with the director]” instead of inventing a value that sounds right. Before submission, search the assembled document for every remaining bracket.
The other drafting rules from chapter 20:
- talk about the client’s project, not about “our solid experience”;
- prove through a named reference, placed at the point in the method it illustrates;
- commit to dated milestones, approved by the director;
- never round up a CV: if the specialist has eight years and the threshold is ten, you change the person, not the figure.
Plate 48 applies these rules to a submission for a commission: every requirement cited with its source, then method, owner, deliverable and check, and an eight-week programme that separates confirmed commitments from resources still to be secured. Missing data and programme conflicts are flagged. No reference and no availability is invented.
04 · Move to an agent when the volume justifies it
For a practice answering several enquiries a month, chapter 20 describes a drafting agent, using a fictional company as the example. Its instructions work like a job description. It proceeds in nine steps, six of which end with the director’s approval. The contents page comes after the team has been confirmed, so as not to promise a competence you will not have. Every deliverable ends with a self-assessment against seven criteria; below 18/21, the agent reworks it. The book is explicit that a model tends to over-score itself: the grid forces a second read, it does not guarantee quality.
After every outcome, won or lost, a debrief starts from the tender assessment report and turns each observed cause into a written rule in the next version of the instructions. This level takes several rounds of tuning. The simple mode — one project and one conversation per enquiry — works with an ordinary professional assistant account.
05 · Prepare the submission without delegating the commitment
- The account first. Tender documents and CVs are sensitive data. Upload nothing to an account your organisation has not approved for project documents (chapter 3). Strip from the CVs, without AI, the personal data the work does not need, such as a date of birth.
- The “Tender responses” project holds the company only. The company record lives there. The instructions, the tender documents and the submission stay in that enquiry’s conversation, closed on submission.
- The financial offer. The tool checks the documents required, the totals and any duplication between stages, without correcting a single amount. Pricing strategy and amounts stay with the people who commit the practice.
- Submission. The day before, fifteen minutes of checking against the list of documents required: formats, consistency of names and amounts, signatures, anonymity, a test of the portal.
The agent neither signs nor submits the bid. It carries no commitment. For producing the documents on the design side, see AI for writing a specification. For other requests ready to copy, see engineering prompts for construction.
Author of “AI for Architecture and Engineering”.
One approach: document the inputs, produce, check, decide.
