Rethinking procurement criteria in an era of AI-generated submissions

When every proposal arrives sounding equally assured, procurement stops getting useful information from the page.

That is the deeper problem lurking behind the latest wave of AI-assisted vendor submissions. A shortlist that once would have split cleanly between polished contenders and obviously weaker ones can now come back almost flat: each response neatly structured, technically fluent, responsive to the brief and hard to separate on first read. The result is not that supplie...

rs have suddenly become indistinguishable in quality. It is that presentation has become far cheaper to manufacture.

For years, polish served as an imperfect but still valuable proxy for competence. A proposal that anticipated objections, handled detail properly and read cleanly usually reflected time, judgement and subject-matter knowledge. It was never proof of skill, but it was a costly signal that tended to correlate with it. That relationship has now weakened. With generative tools able to produce persuasive, well-organised prose at negligible cost, the old association between elegant writing and real capability has broken down.

That matters because human readers are not well equipped to compensate. Research published in Communications of the ACM found that people are close to chance when trying to tell AI-generated material from human-created content. Other studies have shown the same broad pattern in text, images, audio and video, with detection especially poor once material has been lightly revised. In one 2024 paper in Scientific Reports, identical information appeared more credible when presented in a fluent conversational style, and readers were less likely to spot factual errors in the smoother version. In other words, fluency does not merely hide weakness; it can actively lower scrutiny.

That is a dangerous combination in vendor selection. If the document sounds convincing, evaluators tend to rate it more favourably. If the proposal is machine-assisted, that fluency may have little to do with the vendor’s ability to deliver. And if the evaluation process is still weighted heavily towards written answers, it may now be measuring the one thing AI can cheaply imitate best.

The issue is not limited to procurement. In hiring, LinkedIn has reported a sharp rise in application volume, with a significant share of the increase attributed to AI-assisted mass applying. Gartner has also said many B2B buyers now turn to sales representatives specifically to validate AI-generated insights. That instinct is telling: buyers are no longer treating the written submission as the final word, but as something that must be tested against evidence.

The practical lesson for procurement teams is simple. Do less judging by prose, and more judging by proof.

That means setting the scoring rubric before any submission is read, so the criteria are not quietly reshaped by the first impressive answer. It means checking a handful of claim-heavy statements against their source material rather than accepting every cited outcome at face value. It means giving more weight to references you choose yourself, live problem-solving, and artefacts from previous work that can be inspected independently. It also means asking candidates to explain a failure, a limitation or a project that did not go well. Generated proposals often glide past that terrain. Real operators usually do not.

The best test is often the least polished. A short unscripted conversation can reveal more than pages of well-formed text, because it asks for judgement, memory and specificity rather than generic confidence. The vendors who are genuinely stronger at the work may no longer stand out on the document alone. If a process cannot see them any more, the problem is no longer AI. It is the process itself.

Source: Noah Wire Services