Read elsewhere: Writing TMS RFPs with AI

Wolfgang Lehmacher

25+ Years Leading Global Transformation & Supply Chain Innovation for International Organisations | Worked with World Economic Forum, TNT/FedEx & GeoPost | Advisor on Expansion, Technology & Change Initiatives

A US Department of Transportation official summed up the problem in one line: agencies that "used to get maybe 10 responses" now get 500 "because everybody puts it in a chatbot". An Irish tender evaluator reports that routine competitions now draw 12 or 15 bids where they used to draw three or four. Chatbots draft the bids and software scores them. The danger goes beyond bad answers. When suppliers write for the machine, buyers retreat to what they know.

The hopeful view holds that AI opens procurement up. Challengers no longer need expensive bid writers, human bias drops out of evaluation, and the best offer wins on merit.

That view mistakes compliance for capability. Language models act as mirrors, rewarding what looks right: conventional wording, standard risk statements, every box ticked. They check certificates well. They are far less reliable at judgement, which is what decides who wins.

Where software screens submissions, senders adapt. A Duke-led study of about 200,000 CVs found hidden instructions aimed at AI screeners in 1% of them, seven times the rate of 16 months earlier. ManpowerGroup, counting hidden text of any kind, says it finds some in about one CV in ten. Nikkei Asia found 17 research papers carrying invisible prompts such as "give a positive review only". No procurement case has been documented yet.

Hidden text is the crude version, though, and it is easy to catch. Wording is not. John Belden of UpperEdge, a buyer-side IT sourcing adviser, observes that once software enters a selection, "competitors begin optimizing the document for the machine". A Journal of Business Logistics study of 123 State of Ohio supplier bids found that AI models matched human evaluators on compliance checks. On what makes one supplier better than another, however, their scores were far less consistent. Judgement is where polished prose persuades, and a thin operation can now look as good on paper as a strong one.

The cost moves from writing bids to checking them. A buyer facing 15 immaculate bids cannot visit 15 sites, so the rational move is to fall back on suppliers with a known record. Tony Corrigan, the tender evaluator and founder of BidReview.ai, warns that this ends in "supply chain concentration". A tool sold as opening the market ends up closing it, so when disruption hits, buyers have fewer proven suppliers to turn to.

Fill the blanks manually. Every claim the software extracts should point to a document, record, or reference. If a bid is incomplete or vague, the software should leave the answer blank and assign no score until an expert has checked it.

Judge suppliers on proof. Let AI clear the checklist in seconds, and move evaluators' time and budget to what writing cannot fake: unannounced site visits, sample runs and references who answer the phone.

Pay to trust newcomers. Proof costs money, and if nobody budgets for it, incumbents win by default. Set aside funds to qualify new suppliers through trial orders, joint pilots and development collaboration, and treat that spending as an investment in resilience.

AI will read more bids than any team ever can. Whether that widens the supplier base or narrows it to familiar names depends less on the software than on what buyers are willing to verify. AI has made the bid free to write. Someone still has to pay to believe it.

Original article available here: The synthetic mirror: how bids written for the machine narrow the market