Two agents walk into an RFP. It went well. But who’s accountable?

I have a love-hate relationship with RFPs. I love the competition, discovering new suppliers, challenging assumptions, and a good negotiation. At their best, RFPs get buyers and suppliers talking, challenge what both sides think they know, and lead to better decisions. Then there’s the version where everyone spends six weeks filling in a spreadsheet before having the conversation they needed in the first place. I’m less attached to that one.

But RFPs sit at the intersection of competing financial incentives, sensitive information, imperfect trust, and serious money. A bad decision can haunt a procurement team for just as long. Which makes the RFP the perfect testing ground for how far we can trust AI in commercial decisions. Now we’re putting AI on both sides of the deal. One agent works for someone trying to spend less. The other works for someone trying to sell more. They can both be doing an excellent job while working toward very different outcomes.

AI will soon be part of every RFP. If it isn’t already.

Follow an RFP from start to finish, and you can give almost every step its own agent:

• AI to identify business needs
• AI to structure business requirements and formalize the RFP
• AI to discover and qualify suppliers
• AI on the supplier's side to interpret the RFP and fill it in
• AI on the buyer's side to evaluate RFP responses against set criteria
• AI on both the supplier and buyer side to negotiate

Each step makes sense on its own. One agent writes 80 questions, another answers them, and a third summarizes the exchange so a human doesn’t have to read it. “Have your agent call my agent” used to be a Hollywood thing. Now it’s procurement. At some point, we should ask what this conversation is actually helping anyone decide.

When buyers can launch RFPs with a fraction of the effort, and suppliers can respond just as easily, what does a response actually tell you? A serious bid used to cost enough time and effort to suggest someone wanted the business. Now it might suggest their agent was available.

And then there's human psychology. RFPs are also about reading between the lines, challenging assumptions, and building trust. Biases can cloud that judgment, sure. But leave everything to AI, and you risk choosing the best proposal over the best partner.

What if some friction was there for a reason?

Procurement has spent years trying to eliminate friction, for good reason. Nobody needs another approval loop or three weeks of email chasing. But there's a difference between friction that wastes time and friction that forces people to think. Writing requirements makes a buyer articulate what it needs. Preparing a bid makes a supplier decide whether it can deliver. Negotiating exceptions makes both sides confront what they're committing to. Some of that friction is the sound of someone thinking.

An agent can launch 50 RFPs before lunch. That doesn’t mean anyone has 50 things to buy. A supplier’s agent can return 50 polished proposals just as quickly. That doesn’t mean the delivery team has agreed to any of them. Before we congratulate both sides on their productivity, someone should check whether the buyer has a budget and the supplier can keep the promises.

Three ways this could get weird

These agents represent different commercial interests. The buyer wants a sound decision. The supplier wants to win. Giving both sides AI gives them new ways to pursue those interests, including a few the other side might prefer they didn’t.

1. Prompt injection: A supplier could hide instructions inside its proposal designed to influence the buyer's evaluation agent. A proposal that quietly tells the evaluator to give it top marks is taking ‘persuasive selling’ fairly literally. That distinction becomes important when nobody is reading every page.
2. Algorithmic bias: An agent trained on historical sourcing decisions might look impressively consistent while quietly favoring the suppliers procurement already knows. We could take the least defensible part of a human decision, wrap it in a decimal score and call it objective.
3. Inflated demand: Buyers could use agents to test pricing across dozens of suppliers without seriously intending to award a contract. Responses might be cheap to generate, but suppliers still have to decide which opportunities deserve attention. If every RFP looks real, how do you know which ones are?

None of this is an argument for keeping humans buried in spreadsheets. It's an argument for making sure AI doesn't turn sourcing into an arms race of automated activity where nobody can tell what's genuine. More autonomy should demand more accountability, not less.

Who's accountable when AI evaluates AI?

I’d happily hand AI a substantial amount of RFP administration. Nobody needs to preserve the experience of copying the same answer into cell G147. But authority needs boundaries. What can the agent promise? What evidence supports its assessment? Who signs off, and who answers when it gets something wrong?

The money is still real. So are the contractual commitments, the implementation and the consequences of choosing badly. Let the agents talk. Before anyone signs, someone needs to stand behind what they’ve said.