Retrieved, not remembered
The draft is written from articles pulled out of your own knowledge base for this specific question, not from what a model absorbed during training.
How it works
Answers from your sources, with the source attached.
The risk with AI in support is not that it writes badly. It is that it writes confidently from nothing. Rezolo Desk is built so an answer either rests on something you wrote and can point to, or it goes to a person.
What grounded actually means here
Grounded is a word a lot of products use loosely. In Rezolo Desk it has a specific mechanism behind it, and the mechanism is what makes the claim checkable.
Only what was actually cited
Several articles may be offered to the model. Only the ones it genuinely cited are attached to the draft, so the citation list reflects what the answer rests on.
Decided by the server
Whether a draft counts as grounded is computed from those citations on the server, not reported by the model about itself. A model cannot mark its own homework.
Stale sources are flagged
An article can carry a review date. If a cited article is overdue for review, the citation is flagged on the draft so the agent sees the answer rests on something unverified.
No source means no answer
When nothing in your knowledge covers the question, the request goes to an agent. That is the designed outcome, not a failure mode.
Degrades honestly
If retrieval fails outright, the draft is marked ungrounded rather than passed off as sourced.
What happens to one request
Every incoming request goes through the same four steps, whichever channel it arrived on.
- 01
Search your knowledge
The request is used to search your knowledge base. Matching articles are collected with their titles and the relevant passage, and the same ranked search backs your customer portal, so agents and customers work from one source of truth.
- 02
Redact before the boundary
Personal data is replaced before the request reaches an external model. Payment cards, Aadhaar numbers, phone numbers and any provider credentials are swapped out, then restored in the finished draft, so the answer reads normally without that data having left.
- 03
Draft with citations
The model writes a reply and names the articles it used. The server keeps only the citations that match articles it was actually given, marks any that are overdue for review, and derives from that whether the draft is grounded.
- 04
An agent approves
The agent sees the draft, its sources and whether any of them are stale, then edits and sends. Nothing reaches the customer without that step.
Why this matters more in support than anywhere else
A wrong answer in a chat demo is embarrassing. A wrong answer to a customer about their refund, their fee deadline or their policy entitlement is a commitment your business has to honour or retract. Both are expensive. Grounding is what makes an AI answer something you can audit after the fact, because the article it came from is recorded against the ticket.
This is also why Rezolo Desk does not offer unattended AI replies. The automation layer handles the mechanical work, tagging, prioritising, routing, assignment and status, and a person stays responsible for what is said to a customer. See everything Rezolo Desk does.
Questions about grounded AI
- What does grounded AI mean?
- A grounded answer is written from a specific source you own and can point to, rather than from whatever the model learned in training. In Rezolo Desk that source is an article in your own knowledge base, and it is attached to the draft so an agent can check it.
- Can the AI invent an answer?
- The draft is written from articles retrieved out of your knowledge base, and only the articles the model actually cited are attached to it. Whether a draft counts as grounded is decided by the server from those citations, not claimed by the model. If nothing relevant is found, the request goes to an agent instead of getting a guess.
- Does anything get sent to a customer automatically?
- No. Rezolo Desk drafts; an agent approves and sends. There is no action in the automation system that sends a customer-facing AI reply on its own. Automation can tag, prioritise, route, assign and change status, which is the mechanical work, not the answering.
- What happens when the knowledge base is out of date?
- Articles can carry a review date. When a cited article is past its review date, the citation is flagged as stale on the draft, so the agent knows the answer rests on something nobody has checked recently rather than discovering it after the customer does.
- Is customer data sent to the AI model?
- Personal data is redacted before anything crosses the boundary to an external model. Payment card numbers, Aadhaar numbers, phone numbers and provider credentials are replaced before the request leaves, then restored in the draft, so the reply still reads naturally without that data having been sent.
- What if the search fails?
- The draft degrades to an ungrounded one and is marked as such rather than pretending it had a source. An agent sees that state before deciding what to do with it.
Talk to us
Have a support queue you want to fix? Write to us.
Reach out at hello@rezolo.in. We will understand your requirements first, then guide you on the commercials.