Blog · 12 November 2026
A customer asks a question in plain words. The assistant searches the company's help documents and answers with citations. If it cannot answer, it opens a request that is already triaged, with a priority, a summary and a first reply drafted, and the team picks it up in a staff inbox. The customer follows the request and sees the answer.
The company in this example
Picture a services business of about fifty people. Three people share the support inbox on top of their real jobs. The same twenty questions arrive every week: how refunds work, what the notice period is, whether a document is needed, when a delivery is due. The answers exist, spread across a drive folder, a wiki nobody updates and two years of resolved tickets.
This is the shape of the demo we built to show the pattern, and the shape of most support inboxes we see. The names in a real project change, the problem does not.
What the customer sees
- They type the question in their own words, in a chat on the company's site or help centre.
- The assistant answers from the help documents and shows the source under each claim, so they can check it without asking anyone.
- If nothing matches, it does not guess. It offers to open a request, and pre-fills what it already knows from the conversation.
- They follow the request in their own portal, with a timeline, instead of wondering whether the email arrived.
What the support team sees
The inbox only receives what the assistant could not answer, and each item arrives already read.
- A priority and a short summary, so the urgent one is not behind ten password resets.
- A first reply drafted from the documents, which the agent can send or hold for a person.
- The customer's history, including earlier visits, so nobody asks for the order number twice.
- Internal notes, so the team can correct the assistant where it was wrong.
Why the citations change the conversation
People accept an answer they can check. A citation turns a suspicious chatbot into a colleague pointing at the policy, and it changes what happens when the assistant is wrong: the source is on screen, so the correction is specific instead of an argument about whether AI works.
It also changes the work. When the assistant answers, nobody re-reads the policy. When it escalates, the person starts with the summary and the draft rather than the blank page.
What we measure in the first month
- Deflection: the share of conversations answered without a person, and how many of those got a thumbs down.
- First reply time, split into answered by the assistant and answered by a human.
- Escalation quality: the share of requests where the priority and summary were right.
- Accuracy on thirty held-back questions with known answers, run again after every change.
- Cost per conversation, including the model calls.
If deflection rises and accuracy falls, the assistant is guessing and the citations will show it. That is the number to watch, not the one that looks best in a slide.
How it is built on Keel
- Knowledge holds the help documents, chunked and indexed, with the source returned on every search.
- An agent runs the conversation, searches first and calls tools for the rest: open a request, look up an order, escalate.
- Records store the requests, the messages and the internal notes, so nothing lives only in an inbox.
- A workflow triages each new request the moment it is created, which is why the inbox is never unsorted.
- Approvals keep a person in the loop wherever a refund or a promise is involved.
- Runs and traces record every answer with the passages it used, because a support answer is evidence.
What it takes
A working version on the company's own documents is the 48 hour proof of concept. Getting it into daily use, with the help centre connected, permissions set and the test set agreed, usually takes two to six weeks after that.
See how retrieval works underneath. Chunking, local embeddings and hybrid search, in detail.
FAQ
Can the assistant hand a conversation to a person?
Yes, and it should. Anything it cannot answer from the documents, anything a customer asks to escalate and anything that touches money goes to the staff inbox with the context attached.
Do we have to replace our help desk?
No. Keel can sit in front of the tools you already use, answering from your documents and creating records in your system. Replacing the help desk is a separate decision and usually a later one.
What if our documents are out of date?
That is the most common real blocker, and the assistant makes it visible instead of hiding it. We start with the documents people already trust, mark the rest as out of scope, and add an owner for the ones that get updated.
How long before it is answering real customers?
A working version on real documents takes 48 hours as a proof of concept. Running it in front of customers usually waits two to four weeks, until the test set and the escalation rules are agreed.
Infrenic builds AI agents on your data and tools, ships them to production and runs them, all on Keel. Send us a problem and see it working in 48 hours.
The Infrenic AI Dispatch
Practical architecture notes on AI agents, RAG, and production systems. No buzzwords, no spam. Sent when we have something real to share.
Read earlier editions at infrenic.beehiiv.com