Feed your AI: a handbook Copilot can quote

By the Onsites AI team · Last updated August 28, 2026 · 5-minute read

Every AI feature in a help desk is fed by the same thing — the workspace's written truth. A copilot drafts from context; Harness answers from records, handbook and history; tools inspect what the records contain. Which means the desk's AI quality has a single upstream variable, and it is embarrassingly analog: the handbook your team can actually find, read and keep current. Teams that skip it get AI that speaks fluently on their behalf with nothing to stand on; teams that keep a lean one get drafts that cite real policy and answers a manager can defend. This guide is the small book of keeping that handbook good: five sections, one page each, and a maintenance ritual that takes less time than the meetings it replaces.

The five sections, one page each

1 · What we sell, to whom, at what price. Products with their catalog names (the same names the quote-and-invoice tools use), the buyer types, and current pricing — because half of first replies are price-and-fit answers, and the AI can only quote what the record states. 2 · Money. Refund policy with the numbers (window, methods, who approves what), discount authority, credit rules — the section that turns every money conversation from improvisation into citation. 3 · Logistics and promises. Shipping timelines as stated, delay procedures, what we tell customers about customs and carriers — the delay playbook's home. 4 · Voice. The greeting, the closing, the apology shape, the formality line per language — so drafts inherit register instead of inventing it. 5 · The edge cases, in writing. Every weird case the team has met, written as it was handled — this section is the handbook's compounding asset, and it exists because of the update ritual below.

The update ritual: fifteen minutes, weekly-ish

Handbooks decay from good intentions — "we'll document it after the rush" — so anchor the ritual to the desk's own signals: monthly, read the month's escalation summaries (the handoff cards are literally a syllabus) and ask "what did we decide that the handbook doesn't say?"; every time Harness fails to find an answer, that question goes straight into the next week's edits — the tool's misses are the table of contents for documentation; and once a quarter, a full read-through with a deletion pass, because stale rules are worse than none. The fifteen-minute budget is honest: it covers one or two entries, not a rewrite — which is exactly the point, because a handbook maintained at the pace of the desk's reality is the only kind that stays true.

The onboarding dividend nobody prices

Every question a new hire asks in week one is a handbook question. "What's our requote policy when the buyer misses the container?" has an answer at the desk — the senior knows it — and the difference between teams that compound and teams that restart is whether that answer is written where hire #4 will find it. The handbook, fed to the AI, flips onboarding from weeks of hallway training to days of reading followed by supervised volume: the new agent's first week answers are drafted from the same truth the seniors use, and their first escalations are read from handoff summaries of real cases. The secondary effect compounds too: new readers are the sharpest reviewers of stale policy, because they question what habit has stopped noticing — a quarterly fresh-eyes pass from the newest teammate is documentation quality control disguised as training. Teams that measure this honestly report onboarding time cut in more than half, and — quieter, more valuable — the same answer given by every agent, which is what the customer experiences as "this company knows what it's doing."

The handbook and the audit trail are the same document

Two audiences read the handbook, and pretending otherwise is how audits go wrong. The desk's readers — the copilot quoting policy into a draft, Harness citing the section behind an answer, a tool run checking a record — consume it at machine speed. The human readers — the new hire, the manager reconstructing a decision, the auditor asking "on what basis did you tell customer 302 thus?" — consume the same pages at evidence speed. When the two read one document, the desk's answers are auditable by construction: every draft has its clause, every citation its section, every money promise its rule and its approver. When they read two documents — a wiki for the humans, nothing for the machines — the audit question reopens on every case: which truth did the agent quote? The access and audit records close the circle at the container level (TLS in transit, encryption at rest, seat-based access, trails on both deliveries); the handbook closes it at the substance level. One document, five pages, both audiences — that is the entire architecture.

Where teams go wrong, in order of frequency

First: the absent handbook — AI drafts anyway, fluently, from nothing; the first money answer the machine invents usually funds the next fifteen minutes of handbook work, retroactively. Second: the novel — sixty pages drafted in one heroic weekend, updated never, and stale by autumn; lean-five-beats-fat-sixty, every time. Third: the shadow version — the handbook in someone's private doc, invisible to the AI layer and to the new hire, so drafts cite the truth as of last winter. Fourth: the deleted edge cases — the quarterly deletion pass that removes the weird cases actually worth keeping; deletion is for stale rules, not for hard-won examples. The fixes are all ritual, not technology: anchor the fifteen minutes to the escalations, feed the misses straight in, and let the newest reader sharpen the pencil. A handbook is not a project; it is the desk's memory, maintained at working pace — which is precisely what the AI layer reads, and what the customer experience compounds upon.

Why grounded AI is a different product

The difference between grounded and ungrounded AI shows up in the two moments that cost teams money. The quote: a copilot that has read the handbook drafts "we can offer requote #1041 as agreed, or a 40% credit — the credit posts in 48 hours" — a manager reviews a sentence that cites policy; ungrounded, the same button produces a fluent paragraph inventing both the number and the promise. The compliance question: "did we tell this buyer the truth about the delay?" — Harness answers with the handbook section and the March thread visible, ungrounded it answers with something that sounds like policy and is not. That is also why the handbook belongs to the desk rather than to a wiki nobody feeds: the audit trail, the tool runs, the access records and the drafts all read the same document, and it is the document you chose. The honest framing for a skeptical team — "the handbook is how the AI tells the truth about us" — has converted more than one workflow from skepticism to habit.