The paragraph on AI makes it into the proposal, either because the client asks for it or because the firm adds it unprompted. And nearly always it says the same thing. A made-up but familiar example:
"We use AI responsibly, transparently and in line with applicable laws and regulations, under human oversight."
Every word of that may be true. You just can't do anything with it. The client can't set it beside another firm's proposal, can't fall back on it during delivery and can't attach a question to it when something goes wrong. For the firm itself it's worse: the sentence promises something nobody has agreed.
Why the paragraph stays empty
Not because firms are hiding something. There's no fact in it because, at the moment of writing, there's no fact to hand. The paragraph describes an intention ("responsibly") while the client wants to know what happens. Whoever can't write that down writes an intention.
Clients also know by now why they want to know. In 2025 Deloitte Australia had to repay part of the fee for a report to the Australian employment department, after invented references and an invented court quotation turned up in it. The corrected version added a reference to generative AI in an appendix. That's the question a client can now ask: where did this come from, and who saw it?
If you write for a public body, the European Commission's model contractual clauses are worth reading too. On 5 March 2025 the Commission published an updated version of the EU model clauses for procuring AI solutions, with a full version for high-risk systems, a light version for the rest and a commentary. They're voluntary. They are meant to help public buyers procure AI solutions that are trustworthy, fair and secure. Anyone writing for such a buyer can hold their proposal up against them.
The paragraph in four sentences
SEAL describes where in knowledge work traceability gets lost. That makes it a ready-made frame for the paragraph.
SEAL describes the four phases knowledge work goes through: Source, Edit, Approve, Launch. In every phase the traceability of a source can be lost, and the weakest phase determines what the work is worth.
Write one sentence per phase, and let it answer the question of that phase.
Source (where does the information come from, and does that source really exist?): "For every factual claim in the report we record the source underneath it and who checked it. We supply the source file on request."
Edit (does the source still cover the text after editing?): "An approval is tied to the passage it refers to. If the passage is rewritten, we go through it again."
Approve (who reviewed the content, with what expertise, and against what?): "[Name, role] reviews the conclusions and recommendations, with the source passage alongside, and records the outcome."
Launch (who releases the work, and with what visibility?): "[Name] releases the report, and only once the three points above are complete."
Under that, add one sentence on what AI does and doesn't do at your firm. For example: "We use AI to search sources and to draft first versions of sections of text. We write conclusions and recommendations ourselves." That's a limit you can point to, and so a limit the client can hold you to.
Five sentences, around a hundred words. Short, and you can do a lot with it.
Only write what you do
This is the one hard rule. Every sentence has to be something you can show, a year later, actually happened. If your way of working doesn't yet record a source for each claim, don't write that sentence, or write it only after a trial run of two or three assignments. A sentence that isn't true is worse than a sentence that's missing: it sits in a proposal, and a proposal is a commitment.
It's tempting to make the sentences sound like a system. Don't. "[Name] does that here" is stronger than "a review takes place". A name is something a client can trace.
The same goes for a tool. If you use a system that shows the source for each claim and ties an approval to the passage, that can go in. In TRACES those are the Trail and Retrieval layers. But say what the system does, not its brand name.
Three questions for whoever reads the paragraph
Got a proposal with a paragraph like this on your desk? These are the three questions that get an answer:
- Who? Is there a name or a role against the review and the release? Without a name, nobody reviews.
- Against what? What exactly is checked: the source, the passage, or just "the content"?
- How do we see it? Can you get an example of the source file on delivery, with names left out where necessary?
A firm that answers the first two in its proposal usually has the third ready too.
What you can do this week
- Take the AI paragraph from your last proposal or tender.
- Strike every sentence where, a year on, you couldn't show it actually happened.
- For each SEAL phase, write one sentence that is true of how you work today. If a phase is missing, write nothing and put it on your list.
- Add names: who reviews, who releases.
- Have a colleague who doesn't know the project read the paragraph and say where they'd check something.
The phase you can't write a sentence for is probably also the phase where your work leaks.
Sources
- European Commission, Public Buyers Community (5 March 2025). Updated EU AI model contractual clauses now available. public-buyers-community.ec.europa.eu
- Cyber Daily (2025). Deloitte to refund government after using AI in $440,000 report. cyberdaily.au
- City AM (2025). Deloitte refunds Australian government after AI 'made up citations' in report. cityam.com