The pilot that always gets approved
The pilot is approved because it demos well. A chat surface, a buyer asks for a thousand fittings of a specific type, the agent returns options and places the order. The board sees it and asks why the trade desk is not running on this already.
The trade desk knows why. The buyer in the demo is not a real trade buyer. The real trade buyer arrives with a part number, a job reference and a price they already have in their head from the last order. They are not in a conversation. They are in a transaction, and the conversational layer is in the way.
The five things the first pilot got wrong
A chat-first surface. The buyer types more than they would in a quick-order form and gets less information back. The interface optimises for the wrong axis.
No real ERP integration. The agent quotes a price the contract does not honour, or fails to enforce a credit limit that the ERP holds. The first incident makes the trade desk wary and the second one makes the buyer go back to the phone.
No approval awareness. The agent places an order that needs sign-off and the sign-off never arrives because nobody told the approver. The buyer hears about it from the site manager and the agent gets blamed for the process gap.
No catalogue specificity. The agent is grounded in the marketing catalogue rather than the customer's assortment, and starts offering products the customer cannot actually buy on their contract.
No path back to a human. When the conversation breaks, the buyer cannot hand the context to the trade desk and start again. They start again on the phone, with no record of what happened.
"B2B buyers do not want a conversation with the storefront. They want a faster way to place a complex order without phoning the trade desk."
What the second attempt should look like
Start with a narrower job. Reorder, similar-product, and search-by-description. Not "place a complex order through chat". The narrower job is the one the agent can do well and the one the buyer is willing to delegate.
Ground the agent in the customer's assortment, the contract pricing and the ERP credit position. If the agent cannot see those, it is not an agentic checkout. It is a marketing chatbot in a B2B coat.
Build the handoff before the conversation. The trade desk needs to be able to pick up any session, with full context, and finish the order. The buyer should not have to repeat themselves and should not have to ask for the handoff explicitly when the agent has stopped helping.
Measure on order completion without trade-desk intervention, not on conversation length. The first metric is what the business cares about. The second metric is what the demo cares about.
A modest expectation, deliberately
A second-attempt pilot that handles reorder and similar-product cleanly, against the real catalogue and the real credit position, will earn the trust required for the harder jobs. A pilot that tries to be the whole checkout will fail again, and the next time the business will be slower to fund it.
The point is not to be unambitious. The point is to ship something the trade buyer will use on a Tuesday morning. Everything else follows from there.
