Harness explained: the concierge AI that uses real tools

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

Every support team eventually discovers that its hardest information problem is not customer volume — it is the organization's own memory. What did we promise this account in March? Which discount did the handbook allow for repeat buyers, and did the last agent actually apply it? What is our current line on the customs-delay email? The answers exist in the workspace — in CRM records, the handbook, three months of conversations — and reaching them means a search, a scroll and usually a colleague's memory. Harness is the concierge layer built on top of that data: you ask a question in plain language, it answers from your workspace, and every answer carries visible sources — the CRM record, the handbook section, the exact thread. Ask costs 15 credits; add tools (so the answer may check account records as it reasons) and it is 24. Roughly a quarter, for an accurate memory.

The sources line changes what AI answers are for

Most AI answers must be checked; a Harness answer carries its evidence with it. That single design choice rebuilds the trust chain for a support desk: an agent quoting the AI quotes the CRM line visible beneath it; a manager reading a summary can click through to the thread; a new hire's answer shows the handbook section behind it. The sources are also the audit trail's friendliest face — "who decided this?" is a click rather than an interrogation — and they make the boundary of safety self-evident: anything without a source is the machine's speculation, and gets treated as such. This is why Harness is priced like a decision (15–24 credits ≈ $0.15–0.24) rather than like a draft: it is the expensive, grounded layer to use when being right matters more than being fast.

What a team actually asks it

Real Harness questions fall into families, and they are worth naming. Account archaeology: "what did we promise Acme about the March requote," "how many times has this buyer escalated," "who owns account 302." Policy lookups mid-conversation: "what does the handbook say about partial refunds," "our line on customs delays," "can we extend this return window." History search: "find the thread where we discussed the 40ft quote problem," "how did we word the August price increase." Onboarding: "how do we handle mixed-container shipments" — the new hire's question answered from the handbook instead of from hallway memory. And the meta-move that separates teams that love this layer: when an agent reaches for Harness and finds the handbook lacks an answer, that gap goes into the handbook feed — the tool's failure modes are the table of contents for next month's documentation.

Scope and permission: what it can and cannot see

Harness answers from the workspace it runs in — CRM records, handbook, conversation history, documents. It is bounded by your seat-based access: an agent's questions return what that seat may see, and the audit trail records who asked what. Two scope habits keep it clean. The permissions review: when a new seat joins, check what the seat can reach before their first Harness session, because the tool is only as tight as the seat behind it. The sensitivity line: workspace data is one thing (the desk is the record), but nothing about Harness changes what should not be written into customer threads at all — the security posture (TLS in transit, encryption at rest, audit trails on both deliveries) covers the container, not the judgment of what belongs inside it. On self-hosted, add the final dimension: the model that reads your graph is an endpoint you chose, so the data question ("where does this conversation live?") and the model question ("where does this endpoint live?") are answered in the same document — see the residency guide.

A worked week: what the ask layer costs in practice

Take a five-person team running Harness for one ordinary week: forty plain look-ups (15cr each) — account histories, policy checks, thread hunts — and six tool runs (24cr) for the cross-record jobs: a weekly "unanswered high-value orders" sweep and the occasional cross-account comparison. That is 600 + 144 = 744 credits, inside one $20 pack, for a layer that would otherwise consume a colleague's afternoon once a week. The comparison that matters is against the human cost of the same questions asked the old way: hallway memory, scrolling, the second agent who guessed and had to be corrected. At published credit prices, ask-and-tools is one of the cheapest upgrades a desk can buy per hour of senior attention returned — which was the point of pricing it above drafts in the first place.

Ask mode, then tool mode: choosing the depth

The two modes price different jobs. Ask (15cr) is look-up: question, answer, sources — one pass over the workspace. With tools (24cr) is operate: the answer may pull the account record, cross-reference the handbook, check a document's state, then compose — the shape of "find every open order above $5,000 with no reply in five days and list the customers." The judgment rule is the same as the copilot's: text-out jobs ride on Ask; anything where the answer should touch or change records is a tools run. Weekly cost for the ask layer is small even when daily — a hundred Harness questions a month runs under $20 — and the honest adoption advice is the reverse of most AI features: start broad here, because the questions people ask in week three ("how do I do X") are the handbook gaps your next hire would trip over.

Why this belongs to the desk and not to a chat tab

You can approximate Harness by pasting history into a general chatbot — the prompt guide even shows how — but the approximation leaks: the context you paste is the only context it has, the record cannot be reached, and nothing it "knows" can be cited or trusted into an action. Inside the desk, the linkage is the knowledge: CRM records, deals and accounts, threads, handbook, quotes and invoices in one graph — and on self-hosted, the model answering from your graph is an endpoint you chose. A workspace that talks back, with sources, is what "memory" was always supposed to mean here: not a better search, but an answerable record — yours, at a quarter per question.