Agencies
AI Knowledge Assistant: the deliverable your clients are already asking about
A private assistant over the client's own knowledge base — chat with source citations and real admin tools — delivered by your agency instead of a chatbot widget vendor.
The AI Knowledge Assistant is a productized package agencies deliver on Ciao: a private knowledge base built from a client's documents, a chat interface that answers with source citations, and admin tools for managing content and reviewing conversations. It ships as a real React, TypeScript and Supabase application the agency owns, with role-based access, audit trails and zero-retention model contracts — an AI deliverable serious enough for a paying client.
Published 2026-07-03 · Last updated 2026-07-03
Every client wants AI; almost none can buy it safely
Somewhere in the last year, every one of your clients asked the question — in a board meeting, at a conference, or straight to you: "should we be doing something with AI?" What they usually mean is specific and modest: their staff and customers ask the same questions over and over, the answers live in PDFs and policy docs nobody reads, and they would like a system that answers accurately from their own material. What they find on the market is either a toy chat widget or an enterprise platform sized for a bank.
Agencies have been rightly wary of filling this gap. A chatbot that invents answers about a client's pricing or policies is a reputational grenade with your agency's name on the pin. The wariness is about controls, not the idea: where does client data go, is it used to train someone's model, can anyone tell why the assistant said what it said, and who is accountable when content goes stale?
The Knowledge Assistant package answers those questions structurally. Answers cite their sources so users can verify rather than trust. The knowledge base is curated through admin tools, not scraped and forgotten. Inference runs under zero-retention model contracts, and customer code is not used to train models. It is an AI deliverable your agency can defend in the room where the client's lawyer asks the awkward questions.
What ships in the package
Private knowledge base
The client's manuals, policies, FAQs and product docs, ingested and organized with admin curation — a governed library, not a one-time scrape that rots quietly.
Chat with citations
A branded chat interface that answers from the knowledge base and cites its sources, so users can open the underlying document instead of taking the answer on faith.
Admin tools
The client's team adds and retires content, reviews conversations, flags weak answers and sees what people actually ask — which is market research they have never had.
Access control
Public-facing, staff-only, or tiered: role-based access decides who can ask what, and SSO via SAML and OIDC ties staff access to the client's identity provider.
Escalation paths
When the assistant lacks a grounded answer, it says so and hands off — to a contact form, ticket queue or human inbox — instead of improvising.
Usage analytics
Question volume, topic clusters, unanswered-question reports and content gaps, feeding a monthly improvement loop your agency runs as retainer work.
How the build runs
1. Brief
Define audience and boundaries in plain language: who asks, what corpus answers, what topics are out of bounds, where escalations go. The boundary list matters most.
2. Build
Generate the assistant in the Builder — Blocks wires the AI features — then load the corpus and shape tone, layout and escalation flow in the live preview.
3. Review with the client
The client's experts interrogate the assistant with real questions, including hostile ones. Weak answers are content fixes, made in the review session, before any user sees them.
4. Govern
Guardrails applies plain-English policies to sensitive surfaces — corpus changes, access rules, escalation logic — and records human review with an audit trail.
5. Ship
QA replays the critical paths: grounded answer with citation, out-of-scope refusal, escalation handoff. Deploy to the client's domain, public or behind staff login.
6. Retain
Run the monthly loop: review analytics, patch content gaps, retire stale documents. An assistant is a garden, and gardening is recurring revenue.
Packaging and economics
Price the assistant on the cost of the questions it absorbs — support tickets, staff interruptions, slow onboarding. Platform context: serious agency development programs on Ciao start at USD 10,000 per year.
| Package | Typical scope | Delivery rhythm | Revenue model |
|---|---|---|---|
| Assistant launch | Knowledge base, cited chat, admin tools, one audience | Two to three weeks | Fixed project fee |
| Knowledge care | Hosting, monitoring, monthly content and analytics loop | Ongoing, monthly review | Recurring monthly fee |
| Second audience | Staff-only tier, partner tier or new language on the same base | One to two weeks per tier | Fixed fee per expansion |
| Assistant plus | Integrations — ticketing, CRM handoff, in-product embedding | Scoped per request | Iteration billing on the care plan |
White-label and ownership notes
The assistant is your agency's deliverable end to end: the client's brand on the chat surface, your name on the engagement, and underneath it standard React, TypeScript and Tailwind over Supabase — 100% owned and exportable, corpus and all. You are not reselling a chatbot vendor's widget with a margin; you are delivering software, which is why the care retainer holds up.
Data posture is the part to put in writing in every proposal: customer code is not used to train models, inference runs under zero-retention model contracts, and a multi-provider model ladder with fallback means the assistant does not depend on any single model vendor's uptime or pricing. Deploy to Ciao cloud or the client's own cloud account. If the assistant is your first paying client build, the Agency Build Grant covers up to 2,000 credits.
Frequently asked questions
What stops the assistant from making things up about our client's business?
Scope and citations. The assistant answers from the curated knowledge base and cites the source behind each answer; when no grounded answer exists, it escalates instead of improvising. The client review step stress-tests exactly this before launch, and the monthly analytics loop catches drift after.
Where does the client's data go?
The knowledge base lives in the client project's own Supabase backend. Inference runs under zero-retention model contracts, and customer code is not used to train models. For stricter postures, deployment to the client's own cloud account, private VPC or on-prem is available.
Who owns the assistant and the knowledge base?
Per your contract, as with every Ciao build — the code is standard React and TypeScript, the corpus sits in a real Postgres database, and both export cleanly. Most agencies retain the build under a care agreement and hand over on exit.
Do we need AI expertise to sell and deliver this?
You need editorial judgment, not machine-learning staff. Blocks provides the AI wiring; your team's work is curating the corpus, shaping tone and boundaries, and running the review discipline — which is much closer to what agencies already do than it looks.
What does the client see versus what we see?
The client's users see the branded chat; the client's admins see the content and conversation tools. Your agency works in Ciao — building, governing, monitoring, and managing every client assistant from one Conductor view.
Can we resell hosting and the monthly content loop?
Yes — knowledge care is the recurring engine here, and unlike generic hosting it is visibly valuable: the client gets a monthly report of what users asked, what the assistant could not answer, and what you fixed. That report renews the retainer by itself.