R.I.D.G.E. How we make delegation real.
Every RidgeHQ agent ships with five letters filled in: Role, Inputs, Decisions, Guardrails, Escalations. We review them with you on intake, again at week one, and as needed during the weekly review. Agents don't expand scope on their own.
- 01RRole
What the agent owns.
Every agent we deploy has defined ownership. Clear job, clear inputs, clear outputs. Not "AI for ops" — "drafts every Front ticket about order status, refunds, and registration." If you could write a job description for it, it belongs on the card. If you can't, we split the work until the boundaries are clear.
Examples · live agents- Helpdesk drafter (Front · order status, refunds, registration)
- Daily deposit reporter (Slack · 9am push, banking dashboard)
- KB question answerer (Slack · grounded in Notion + help center)
- 02IInputs
What the agent reads.
Named tools and data sources. Front, Stripe, the Postgres database, the help-center KB — listed explicitly, with credentials in our vault, never in the model. If a new input is required, it's a change request, not a silent expansion. The customer always knows what the agent can see.
Examples · live agents- Front (read inbound, draft replies)
- Stripe (lookup orders, refund history)
- Postgres (program + customer lookups, read-only role)
- KB articles (141 KB ingested, refreshed weekly)
- 03DDecisions
What the agent can decide on its own.
The narrowest letter. Listed as positive permissions, not vague capabilities. "Tone choices" is a decision. "Refunds under $50" is a decision. "What policy applies to this ticket" is a decision. "Whether to send" is a decision the human owns until the agent earns it — and even then, we check in.
Examples · live agents- Tone, phrasing, signature
- Policy lookup and citation
- Refunds under $50 (auto-draft; flag if reason ambiguous)
- Tag selection (1-3 tags per Front ticket)
- 04GGuardrails
Where the agent can't go.
Listed as negative permissions. Things the agent will not touch even if asked. "No legal advice." "No PII echo." "No medical guidance." "No price-quote without sign-off." Guardrails are where most production AI failures happen — usually because someone forgot to write one down. We write them down.
Examples · live agents- No legal or medical advice
- No echoing of customer PII (full name + DOB + address combinations)
- No price quotes outside the published rate card
- No off-policy refund offers
- 05EEscalations
When the agent has to ask.
The hand-off rules. Not "if confidence is low" — that's vibes. Specific triggers: "if refund > $200, route to manager." "If customer message contains the words 'attorney' or 'BBB', escalate immediately." "If agent's reasoning includes a policy lookup that returns no match, draft + flag." Every escalation has a destination and a SLA.
Examples · live agents- Refunds > $200 → manager review, 4-hr SLA
- Anything mentioning legal/regulator → human, immediate
- Policy KB miss → draft + flag for the team to write a new article
- Customer flagged angry by sentiment heuristic → human reply, agent backs off
One page per agent. No surprises.
Every agent we deploy ships with a one-page R.I.D.G.E. card. You can see it in the console, ask for it in a Slack message, or download the PDF for your auditor. The card is the contract: it tells everyone — you, your team, your vendor, your auditor — what the agent will do, what it can read, what it can decide, where it won't go, and when it has to ask.
When something needs to change, we change the card together, log the change, and the new card supersedes the old one. The agent's behavior follows the card. The card is the source of truth.
Most "AI delegation" frameworks miss the last three.
"Role" and "Inputs" are easy — almost every AI agent product has those, sometimes called something else. Decisions, Guardrails, and Escalations are where AI delegation usually breaks. Without explicit decision permissions, the agent either over-acts (refunds it shouldn't have approved) or under-acts (asks before drafting a reply about a tracking number). Without guardrails, edge cases turn into incidents. Without escalations, the human review loop never closes.
R.I.D.G.E. forces the conversation to happen on intake, in writing, before the agent goes live. Once the card exists, weekly review becomes mechanical: did the agent stay inside D and G? Did it escalate when it should have? What needs to move from G to D as trust earns up?