AI chat agents
Chat that answers from your documentation, not from a guess.
The 2018 chatbot followed a script and annoyed everyone. This is a retrieval-grounded agent that reads your own material, answers from it, shows where the answer came from, and hands over cleanly when the question is not one it should be taking.
What it replaces
- A contact form that promises a reply within one business day, when the answer is already in your documentation.
- Ticket volume made entirely of questions your help centre answers.
- A decision-tree widget that offers four buttons, none of which is what the visitor wanted.
- Support staff answering the same question for the two-hundredth time.
We operate this ourselves
Vartalap
The same conversational agent we run on WhatsApp, deployed on the web. Building on something we operate ourselves means the failure modes were found on our own traffic first.
How it works
What we actually build.
- Retrieval over your documentation, policies and past tickets — answers grounded in your material, with the source attached so the visitor can check it.
- Explicit refusal when the answer is not in your material, instead of a confident invention. This is a configuration decision we make deliberately and can show you.
- Handover to your team with the transcript, so the conversation continues rather than restarting.
- Evaluation against a real question set before launch, so “it seems better” becomes a number.
- Deployed on your site or inside your product, styled to belong there.
The limits
What it does not do.
Every one of these has cost us a deal at some point. They have also stopped us selling something that would have failed six weeks in, which is the cheaper of the two outcomes for everybody.
- It can only be as good as the material it reads. If your documentation is wrong or missing, this surfaces that rather than fixing it — which is useful, but it is not the outcome people expect on day one.
- It will not invent policy. On anything commercially sensitive — refunds, pricing exceptions, contractual terms — it routes to a person by design.
- It is not a replacement for a support team. It removes the repetitive tier of the work; the hard tier still needs the humans, who now have time for it.
In practice
Where this shows up.
On a product with real documentation
Deflecting the answerable questions is where this pays for itself fastest, because the material already exists and only needs to be reachable in the moment someone is stuck.
On a considered purchase
Specification and eligibility questions answered at the moment of interest, rather than in a follow-up email the next morning.
Also
The other channels.
Different constraints, same job. Most clients start with one and add the next once it is running.
WhatsApp automation
Customer conversations on the channel they already use — handled automatically, inside Meta's rules.
AI email support
Shared-inbox triage, drafting and routing — so first response stops taking a day.
Tell us what ai chat agents would have to handle.
Pick a time that suits you. Tell us the problem, and we'll tell you honestly whether we're the right people to solve it.