MillerKnoll AI
← All case studies
The practice · AI leverage· Jul 2026 · 4 min read

Not an agent. Not a skill. A practice.

The case study behind all the case studies: pointing AI at a volume of content a lean team could never read by hand — and getting structure back, not a summary.

Every story in this collection is a different agent, a different team, a different deadline — but look at what actually happened in each one, underneath the specific build.

3,300+ articles
Audited (AutoBid)
~9,349 quote lines
Mined as ground truth (AutoBid)
2,191 → 1,491
Pages → commands (SAM)
90 days
Of tickets analyzed (IT Support)
~3,115 pages
Cross-checked, 31 gaps caught (Scott AFB)
8 mailboxes → 224
Draft HR articles, PII stripped first (Knowledge Harvesting)
Undocumented
Format reverse-engineered (HR Studio)

None of that is a feature of any one agent. It’s the same underlying move every time: pointing AI at a volume of content — articles, tickets, quotes, price-book pages, an admin guide, an undocumented file format — that a team this size could never get through by hand on the timeline available, and having it come back with structure instead of a summary: a gap report, a control-question set, a parsed command index, a doc-coverage map, a re-architected orchestrator.

Lean development

A small team can ship agentic systems at a large team’s pace because the reading, parsing, and mining isn’t competing with the humans’ time — it’s what frees the few people on the team to spend their time on architecture decisions, stakeholder calls, and the judgment work AI can’t do. TracE was built and is still run by non-developers; HR Studio’s undocumented-format problem was charted in days, not weeks, by cloning and diffing rather than waiting on a vendor doc.

Quality assurance

The same pace that reads fast also checks fast — and checks honestly. AutoBid’s control-question set is graded against real dealer quotes, not synthetic cases. The Scott AFB bid didn’t silently skip 31 unmatched part numbers — it flagged every one, sorted by impact, with a likely fix already worked out. The AI Intake Coordinator’s design flaw was caught by evals before a single user ever saw it. In every case, the AI’s own output got verified against ground truth before anyone trusted it.

None of this replaces judgment. A person still designs the architecture, decides what “compliant” or “approved” means, and signs off on what ships. What changes is how much ground a small team can cover before it has to make those calls — and how rigorously it can check its own work once it has.