Turning a team’s inbox into its documentation

The problem

HR Studio can only answer questions it has documentation for. Much of how HR actually works — how a payroll correction gets handled, what to tell someone going on leave, where the holiday calendar lives — wasn’t written down. It lived in shared mailboxes like Payroll, Benefits and Talent, as thousands of individual replies. That’s common for any team that does service work over email: the knowledge is real and used every day, but it only exists as correspondence.

Asking the team to stop and write it all up wasn’t realistic, and guessing which documents “should” exist would miss what people actually ask about. So AISE went to the source.

What we built

Knowledge Harvesting reads a designated shared mailbox and turns it into structured documentation. It only runs on a mailbox after the owning stakeholder has approved it and the privacy and retention constraints are written down. Then it works in stages:

Extract pulls the messages for an agreed time window through Microsoft Graph, counting them first so the run has a known size. Redact runs a set of scripts that strip names, emails, IDs and other personal data before any model sees the text. There’s a stricter variant for HR and a standard one for other teams, plus vocabulary tuned per mailbox. Cluster groups the redacted threads into recurring topics. Draft writes a how-to article for employees or an internal SOP for each topic. Certify has a separate, more capable model check each draft against the source threads. Publish renders the approved drafts into MillerKnoll’s document templates and files them in the team’s SharePoint library.

Each stage uses the right-sized model: Opus for the judgment-heavy pre-flight and the independent certification, Sonnet for clustering and drafting, Haiku for rendering. Runs are held to a dollar budget. Because many of these inboxes receive mail from outside the company, every stage treats email content as data, never as instructions.

“Mailbox mining isn’t a one-off project. It’s a capability we built once and lend to other teams.”

— AI Solutions Engineering

What it produced

Across eight HR mailboxes — Benefits, Payroll, HR Shared Services, Disability & Leave, Talent, Compensation, Retirement and Careers Help — it produced 97 employee-facing how-to articles and 127 internal SOPs. Each mailbox also got a summary report and a documentation-gap report showing which recurring questions had no published answer. Every article is marked as an AI-written draft that needs approval, and the HR Documentation team reviews and refines them. That review feeds back into the process: one lesson so far is to split articles that combine several policies into one article per topic, so each serves a single audience.

The drafts are live as HR Studio knowledge sources behind a routing rule. HR Studio cites the articles, but it never presents a mailbox summary or gap report as guidance, because those are the map used to write the answers, not answers themselves.

Built to hand off

The first runs needed the engineer who built it. In September the process was packaged as a set of Claude Code skills: one orchestrator plus five stage skills that a teammate installs and runs on their own machine against a mailbox they have access to. Hardening it for that handoff turned up problems that had only worked by coincidence on the original machine, like hard-coded paths, templates missing from the package, and a redaction choice that was silently ignored. All were fixed and covered by tests before Trevor Cline ran it on his own. More HR mailboxes are queued: Exec Comp, Belonging, Global Mobility, Associate Relations and Recognition.

Where else it fits

Any team that does service work over email but hasn’t documented how it works can use the same capability. Customer Care, IT support queues, Specials, and product enquiries all qualify. Each run produces a first draft of the team’s own playbook, grounded in what customers and colleagues actually asked, which the team can correct and approve instead of writing from scratch. It also produces the knowledge an agent needs to answer those questions.