The agent recommends, a person signs.
AI dealer sales for an industrial client.
7 months on the account. 3 approvals before any order goes out. 9 days from kickoff to the executive demo.
The client wanted agents that could sell on their own. I gave every agent a person to answer to.
A global industrial technology company sells through independent dealers. The equipment goes out on leases that run for years, and the software rides on top of it. They asked NextGen AI for AI agents inside that sales process: one to quote new solutions, and one to catch leases before they run out. I led the product design, on the same platform team as the NextGen AI case.
The people who would run these agents are dealer reps and the order desk, not engineers. They work from a list of accounts they know by heart. Real money, and no room for a wrong email.
An agent that prices a deal or emails a customer can lose money and trust in one action. Dealers keep their cost and margin hidden from the manufacturer, and from customers, by contract. Nobody could say yet what an agent was allowed to do on its own, or who would answer for it when it got something wrong.
Two products were also tangled together: ordering new solutions, and renewing leases near their end date. Same dealer, same list of accounts, different pace. A quote is a conversation. A lease renewal is a deadline.
Nine days from kickoff to an executive demo, with the client confirming the process on day three. Two developers, both in Albania, in the middle of a release. Stakeholders in five cities across North America, and me in Toronto. The confidentiality rules were in the contract: if cost leaks, that is a breach, not a bug.
The demo also had to run on the real build. Slides would not do it. A rep had to be able to click through.
A new layout for the leasing screens, built around the lease instead of the account. It demoed well and it taught badly: reps now had two ways of thinking about one job. I threw it out and rebuilt leasing on the account list they already knew.
I also sketched levels of trust, where an agent that had earned it could send low-risk email on its own. It died on one question: who eats the wrong email? Nobody in the room would own it, so that level could not exist.
The agent recommends. The platform writes it down. A person signs.
Every recommendation comes through one door, with the agent's reasoning, how sure it is, and which version of the agent made it. Three people have to sign before an order goes out: internal approval, customer approval, and the dealer submitting it. No setting turns those off. Picking an option only highlights it; a second, labeled click preps the email.
I mapped the order process before any screen was drawn: ten steps from discovery to licensing, and every step written to a log nobody can edit. The customer approves from an emailed link, or by signing in to their own portal. Either way it lands in the same log.
A round light shows where the order is. A square light means something went wrong.
Anything past recommending. Agents flag problems; they never resolve them.
Hiding dealer cost in the browser. Who sees what is decided on the server, by role. Opening the cost view is logged. The customer never receives cost or margin, because the server never sends it.
A build that only works as a demo. We logged fifteen decisions as we went, each with what we turned down and why, so the production team gets the reasons and not just the screens.
There is no separate rep role. Cost visibility is tied to the dealer role, so a wrong flag shows too much, and the fix is a process rule, not the product. The order also moves forward when the dealer submits, not when fulfillment confirms, because the demo needed a visible step. Both are logged as risks with a trigger to revisit them. Both are still open.
Seven months, start to finish.
The first version: the order process, the approvals, and who sees what, on a real screen.
The client took the prototype in house to build it. Ten weeks later that stalled and they came back.
The nine-day push to the executive demo. Selling and leasing on one screen pattern, so a rep learns the job once.
The real process, rebuilt with made-up data.
What was delivered, and where it stands.
Delivered: the whole order process, quote to license, with three approvals and a log of every step.
Delivered with it: the lease renewal list, on the same pattern as the selling list.
The dealer portal was mine end to end. The leasing agent was a teammate's; I designed the list it reports into.
It is a demo build. The client's order and ERP systems are stand-ins, and they say so.
The screens are rebuilt with made-up data. The process and the decisions are real.