AI is good at writing the descriptive, repetitive parts of project documents (the project description, the practice’s usual construction systems, a finishes schedule in plain language) as long as it works from a project data sheet and the practice’s own approved text. It’s bad, and risky, at justifying regulatory compliance, quoting regulations from memory and making up data it doesn’t have. The method that works: one data sheet per project, a library of the practice’s own text blocks, drafting section by section, a cross-check against the data sheet, and a final review by the professional.

This guide looks at each type of document, what AI does well and badly in each, the five-step method and the most common mistakes.

Names differ by country: memoria descriptiva and memoria de calidades in Spain, relazione tecnica and capitolato in Italy, projectbeschrijving and technische omschrijving in the Netherlands, design and access statement and specification in the UK. The principles are the same.

What AI does well in each type of document

Type of document What it does well What it mustn’t do Who decides
Design statement Write the description from the project data, in the practice’s style Make up areas, uses or planning data The author
Construction specification Adapt descriptions of systems the practice already uses Propose new construction solutions without professional judgement The author
Compliance statement Organise sections and draft the descriptive parts Justify compliance, calculate, quote clauses from memory The professional, with calculation software
Finishes schedule for buyers Turn the technical specification into plain language Add sales language or materials that won’t be installed The developer and the professional
Tender quality submission Organise content by award criteria and reuse the practice’s own text Fill pages with generic text The bid team

Design statements and construction specifications

This is where it saves the most time. Much of a construction specification describes solutions the practice uses again and again: foundations, structure, envelope, windows, services. If AI starts from that text and the new project’s data, the draft is ready in minutes and in the practice’s style.

The risk is the opposite of what people expect: not that AI writes badly, but that it carries over a figure from the previous project that sounds perfectly plausible. That’s why every number must come from the project data sheet, never from the starting text.

Compliance statements

Here AI can organise the structure and draft the descriptive parts, but it mustn’t justify compliance. Justifications come from the professional’s calculations and checks, usually with dedicated software. AI models also quote regulations from memory and sometimes mix versions or superseded clauses. If you want help with reviewing, paste the official text and ask for a checklist, as we explain in 15 AI prompts for architects and engineers.

Finishes schedules for buyers

AI does a good job of turning a technical text into one a buyer understands, organised room by room. But remember that in most countries the finishes schedule is part of what the buyer is told and may form part of the contract: every material, brand or finish listed must be the one that will be installed. No adjectives the AI adds on its own (“premium”, “exclusive”) unless they can be backed up.

Tender quality submissions

Where tenders are scored on quality, the written submission decides many awards. AI helps organise the content exactly according to the award criteria, avoid missing any section and reuse well-written text from the practice. What doesn’t work is generic text: evaluators score what’s specific to that project (organisation, programme, resources, specific risks) and quickly spot paragraphs that would fit any tender, whether a person or an AI wrote them. More pages don’t mean more points.

A five-step method

1. One data sheet per project

A table with all the objective data: areas, heights, number of units, uses, main materials, planning and land references. It’s the only source numbers come from. If a figure changes, it changes there.

2. A library of your own text blocks

The practice’s reviewed and approved text: descriptions of construction systems, standard sections, finishes paragraphs. Organised by topic, with the date of last review. It’s what gives the AI the practice’s style and judgement.

3. Draft section by section

Asking for the whole document at once produces long text that’s hard to check. Better section by section, with the relevant data and blocks, and a rule to mark anything not in the data sheet as [OUTSTANDING].

4. Cross-check against the data sheet

Before reading the text, ask the AI to compare the whole document with the data sheet and flag any figure that doesn’t match (it’s prompt 5 in our list). It’s a quick check that catches most carried-over figures.

5. Professional review and signature

The final read is the professional’s: consistency with the drawings, construction solutions, regulatory references and, for finishes schedules, that everything promised is what will be built.

Common mistakes

  • Starting from the previous project’s document instead of the new project’s data sheet. It’s the quickest way to carry over wrong areas, names and references.
  • Letting AI quote regulations from memory. References are copied from the current official text or left out.
  • Accepting the sales tone AI adds to finishes schedules.
  • Padding tender submissions with text that adds nothing specific.
  • Not updating the library. A block describing a solution the practice no longer uses ends up in a new project.
  • Uploading documents with personal data to the wrong account. See our guide on ChatGPT and data protection.

When it’s worth automating

If the practice writes these documents often, the next step is a system that generates the draft straight from the data sheet and the library, with each figure marked by its source, instead of everyone copying and pasting into an assistant. To see whether it pays off, do the sum as we explain in how much AI automation costs.

Frequently asked questions

Can I ask AI to write the compliance justification?

It can organise the structure and draft the descriptive parts. The justification itself comes from your calculations and checks; AI mustn’t state that something complies.

Do tender evaluators notice when AI wrote a text?

They notice when a text is generic, whoever wrote it. A submission with concrete details of the project, reviewed by the team, scores as well or better; padding scores worse.

Can AI read the whole project and write the document on its own?

It can read a lot of documentation, but it still misreads drawings and doesn’t know which decisions were made outside the documents. It works better with a clear data sheet than with the whole project.

How much time does it save?

For design statements and construction specifications with a good library of blocks, the draft goes from hours to minutes; checking still takes time, as it should. Measure how long a typical document takes you now and compare after a few weeks.