How to use AI to write government proposals without inventing facts
AI drafting tools are useful for structure and wording, but they produce statements that read as true and are not. In a government bid, one wrong contract value, lapsed certificate or invented client can cost you the award, and in US federal work it can also create legal exposure. The fix is simple: let AI draft, but make every checkable fact trace back to a record you hold before you submit.
Why AI gets facts wrong in proposals
Large language models predict likely wording; they do not look facts up. The UK Cabinet Office's guidance to public buyers, PPN 02/24, warns that AI output can contain "statements, facts or references" that "appear plausible, but are in fact false". In proposals the errors cluster in predictable places:
| Claim type | Typical AI error | What to check it against |
|---|---|---|
| Contract values | Rounding or transposing ($1.18 million becomes $1.8 million) | The contract or completion certificate |
| Dates and durations | Shifting a period, or merging two contracts' dates | Start and end dates on the same record |
| Certifications | Stating a certificate is current after it expired | Certificate number and expiry date |
| Named clients | Attributing work to the wrong agency or inventing one | Your past performance register |
| Staff and percentages | Inflating headcount or performance rates | Payroll summaries and performance reports |
| Superlatives | "Leading", "only", "100%" with no support | Evidence you could show an evaluator, or delete it |
Why it matters more in government work
Buyers increasingly ask. PPN 02/24 gives UK central government buyers optional questions asking whether a bidder used AI to prepare its response and whether AI-assisted content was checked for accuracy. Answer them truthfully, and be able to show how you checked.
In the United States, the stakes are higher. Statements made to win a federal contract can be pursued under the False Claims Act. In one example, the Department of Justice announced in 2017 that a contractor and its chief executive paid $300,000 to settle allegations that its proposal misstated prior experience. Those were allegations only, with no finding of liability, but the lesson is plain: proposal facts must be provable.
A five-step check before you submit
1. Build an evidence register first. One row per fact you can prove: each contract, certificate, staff figure or commitment, with the document that proves it. Our free evidence sheet template has the columns you need.
2. Draft from the register, not from memory. Give the AI tool only the facts in the register and tell it not to add figures, dates, clients or certifications of its own.
3. List every checkable claim in the draft. Pull out every amount, percentage, date, certification and named organisation, sentence by sentence. Paste your draft into the free Claim Scanner to do this instantly.
4. Match each claim to one record. Every fact in a sentence should come from the same record. A value from one contract and a date from another is the most common hidden error. Check precision too: if the record says $1,180,000, "$1.2 million" is a fair rounding and "$1.8 million" is wrong.
5. Have a named person sign off. Someone accountable should confirm every flagged claim and every superlative before submission, and keep the record of what was checked. That record also answers a buyer's AI-use question.