Your next tender was probably written by a machine. Evaluate accordingly.
Somewhere in your current procurement pipeline is a bid drafted substantially by a large language model. You will not be able to tell which one from the prose, because the prose is exactly the problem: fluent, confident, structured, and potentially wrong in ways the author never noticed. PPN 017 is the government's answer to this moment, and its most useful quality is its refusal to panic. AI is not banned from bid writing. It is treated like any other tool a supplier might use, which means the burden shifts to the buyer to evaluate what is actually true.
The policy position most people misread
The instinctive institutional response to AI-written bids is prohibition, and PPN 017 explicitly declines to take it. Suppliers' use of AI is not prohibited during the commercial process. The comparison the PPN reaches for is the professional bid writer: nobody ever demanded suppliers disclose that an agency polished their tender, and the risks are of the same family. That framing matters because it kills the fantasy of an AI-free procurement. The tools are already in use, they lower the cost of bidding, and the PPN even names the upside: suppliers can bid for more public contracts.
What the PPN does instead is create visibility. Buyers can add disclosure questions asking whether AI was used and whether the output was checked for accuracy. The discipline is in the detail: the example questions are for information only and must not be scored. Marking a supplier down for honest disclosure would teach the market to stop disclosing, and discriminating on AI use invites challenge. The disclosure exists to shape your due diligence, not your scores.
Hallucination is an evaluation problem now
The PPN's background section says the quiet part clearly: large language models predict statistically plausible text, and plausible does not mean accurate. They cannot recognise or correct their own errors. In a tender context that means case studies with improved outcomes, references that read well but were never checked, certifications claimed in good faith by a model that pattern matched them from somewhere else.
The traditional evaluator heuristic, that a well-written bid signals a competent supplier, is dead. What replaces it is proportionate verification: clarification questions that ask for the evidence behind the claim, site visits, supplier presentations where the team that will deliver the work has to speak without the model. The PPN frames this as the same approach you would take to any uncertainty or ambiguity in a tender. The difference is frequency: uncertainty used to be the exception, and it is becoming the default.
The data question nobody asks until it hurts
Two quieter provisions of PPN 017 deserve more attention than they get. The first is about your information: buyers should have proportionate controls preventing suppliers from using confidential authority information as training data. Your tender documents, requirements and clarification answers are exactly the kind of text that improves a model trained to write future tender responses. Once absorbed, that information does not come back.
The second is about AI arriving inside ordinary services. The PPN's example is a video conferencing contract where the supplier offers generative AI transcription and translation. Suddenly a routine purchase involves meeting records that could be retained, classified data passing through a model, and the question of whether your conversations become somebody's training corpus. The recommended fix is contractual: data from the service is not used for training unless specifically agreed and approved in writing. Ask the delivery question in every procurement, because the supplier will not always volunteer it.
Plan for volume, not just risk
If AI makes bidding cheaper, you get more bids. The PPN tells buyers to plan for a general increase in activity: more tender responses, more clarification questions, suppliers with newly automated processes. The practical consequences are unglamorous: evaluation panels need more capacity, timetables need slack for due diligence, and moderation needs to be tight enough to stay consistent across a bigger field.
Teams that treat this as a workflow problem, with structured questions, consistent records and time budgeted for verification, will find the AI era manageable. Teams that keep evaluating prose quality will be outsourcing their scoring to whoever has the best subscription.
The takeaways
- AI-assisted bids are permitted; your job is verification, not prohibition.
- Use disclosure questions, but never score them and never discriminate on AI use.
- Treat fluent prose as unverified until due diligence says otherwise.
- Contractually bar your data from being used to train AI models without written agreement.
- Budget more time and evaluation capacity: cheaper bidding means more bids.
Want the full breakdown?
The complete explainer covers the key facts, the requirements in detail and a practical action list, free and printable in the Procurement Library.