From a $200K/yr buy decision to a bid engine the business owns
A VP was ready to license an external pricing tool at $200K/yr. Instead, AISE and the Specials team built one together — a multi-agent engine grounded in MillerKnoll’s own documented rules that does days of pricing research in moments. It went live on August 22, 2026, and in its first 30 days it agreed with the quoters’ actual decisions 91.7% of the time.
The opportunity
At the October 2025 VP Summit, a vendor was lined up to give the Specials team AI pricing support — roughly $200–240K a year, plus implementation. The underlying problem was real: pricing a special is a research task. A complex request has to be checked against MillerKnoll’s own documented rules — what’s approved, under what conditions, at what upcharge — and a question that requires deep research and review could take days to come back. On October 27, the spend was reframed as a possible internal “medium bet,” and the Lean AI team took the challenge, standing up a Confluence-grounded proof of concept in days.
What AISE built
The proof of concept was deliberately narrow: a single agent with nothing behind it but its own instructions and the Specials Knowledge space — pure retrieval-augmented generation — to prove that MillerKnoll’s own documented rules could carry a defensible decision. They could. From there the build compounded into a multi-agent decisioning platform: an orchestrator that extracts the deviations from a request and delegates to dedicated children that classify each one against the knowledge, score complexity, and apply the pricing math. Retrieval moved onto Atlassian’s Rovo and Teamwork Graph; a deterministic Dataverse lookup backs the pricing; and the latest generation, AutoBid 2.1, is replatformed onto Copilot Studio’s skills-based experience on Claude Sonnet 5, with the Dynamic Pricing API resolving a real quoted list price.
The discipline underneath all of it is determinism: the engine cites the documented rule it used and never invents an approval or a price. When the knowledge isn’t there, it abstains to Engineering Review rather than guess. That is what makes its answer something a human can trust and act on.
Built with the business, not for it
The single biggest reason AutoBid worked is collaboration — the project would have been impossible without it, and that collaboration is what unlocked its true potential. It took people recognizing that the AI wasn’t there to replace a job, but to enable them to do more of the human-centric work only they can do.
Nate Zylstra played a massive role as the technical analyst sitting between the AISE engineers and the Specials business partners — translating their needs and their early-testing feedback into actionable stories and sprint items, and sitting in on the workshops where some of the most critical innovations happened. Early testers Jack Clark, Jaime, and Anna put the engine through its paces and, crucially, made edits to the documentation itself — which turned out to matter more than anyone expected.
The engine made the documentation better
Grounding a decision engine in the Specials rules made the quality of those rules visible. The early builds surfaced exactly where the documentation was thin, stale, or ambiguous — and gave the team a concrete reason to clean and harden it. The result is a knowledge base that got materially better because an agent was reading it: thousands of edits across hundreds of pages, maintained by the people who own the rules.
Distinct Specials Knowledge pages created and updated per month (space OK, AutoBid era). Building the engine gave the team a reason to tighten the rules behind their work.
The benefit to the business
The headline number is the $200K/yr vendor license the internal build avoided — but the operational win is bigger than the spend. Research that used to take days now happens in moments: the engine does the same rule-checking at a scale a person couldn’t match, via Pub Layer advanced search, product validation, and the Confluence knowledge. Faster RFQ turnaround changes what the Specials team can do with its time.
That reclaimed capacity is what lets the team stand up a bid desk — pulling people to prioritize customer-facing dealer work, help with more complex bid questions, and, when a request comes back as a no-bid or a not-approved deviation, work with the dealer on an alternative: a swap into a different product line, or another solution entirely. The engine handles the research so the humans can handle the relationship.
Into production
AutoBid went live on Saturday, August 22, 2026 at noon, on Claude Managed Agents. That build had started a month earlier as a hedge: the same instructions and skills carried over from the Copilot Studio build and graded on the same bids. It came out about 30% cheaper to run, so it took production, and the Copilot Studio build is kept in step as the Azure digital twin. Nothing changed for the people using Product Portal. The contract they had tested against stayed exactly the same.
The first weekend had no failures, no lost requests, and no manual intervention. In the first 30 days the engine evaluated 5,700 real dealer requests and agreed with the quoter’s actual decision 91.7% of the time, above what the pre-launch control testing had predicted. Quote response time came down from about 72 hours to 26.1, and the goal from here is under 24.
How the first month went is its own story: a daily hypercare collaboration with the Specials and Product Portal teams that shipped more than 40 fixes. It’s told in the hypercare case study.
The Solutions Desk
The most important change isn’t in the software. With the engine doing the first pass of research on every special, the team has launched a Solutions Desk that takes appointments. Dealers can book time to talk a special through with a person, and when a request can’t be approved as asked, the desk works with them to find something that can be built instead. It’s a deliberate move toward being more approachable and personal with dealers, not just a quote coming back from behind a keyboard. The engine does the research so people can spend their time with dealers.
What’s next
Two things. First, the knowledge base: most escalations come down to a missing article rather than a hard engineering question, and the engine has already identified which ones. The top recurring gaps have been drafted as ready-to-publish articles for the Specials team. Second, selective autonomy: letting the engine handle more requests without human review, starting with precedent-based approvals where MillerKnoll has recently built exactly this and a person has already signed off. The long-term goal is a full auto-bid platform that brings in a person only for new, undocumented deviations, and a Solutions Desk that has more time for the dealers who need it.