AI customer support agent, trained on your documentation
A first-line support agent answers from your own documentation, policies and past tickets — not from the internet and not from what a model assumes about your product. It resolves the repetitive questions with sources attached, checks its own confidence before replying, and hands everything else to a person with the context already gathered.
What problem this solves
“Sixty per cent of our tickets are the same eight questions, and our team answers them one at a time all day.”
Sixty per cent of tickets tend to be the same eight questions, answered one at a time, all day, by people you hired for the other forty per cent. The repetition is not just costly, it is what makes support roles hard to keep.
The trap is that everybody now knows this and the market is full of bots that answer the easy questions badly. The value was never in having a bot. It is in the percentage of tickets that actually close without a human and in how cleanly the rest are handed over — and a build that ignores the second number is a demo the client cancels in month three.
That is why this build is scoped around measurement from the start: deflection rate and handoff quality, reported monthly, so the thing either justifies itself with numbers or gets changed.
Typically 30–60% of repetitive queries resolved without a human, and the rest arriving pre-qualified.
Market contextChatbot development grew 71% year over year on Upwork. The value is no longer in having a bot — it is in the percentage of tickets closed and in a clean handoff to a person. Without that it is a demo the client cancels in month three.Source: Upwork, In-Demand Skills 2026
What runs, end to end
- 01Question arrives
- 02Search the documentation
- 03Answer with sources
- 04Confidence check
- 05Clean handoff to a person
Who this is for
E-commerce, SaaS and service businesses with a support inbox that has an obvious repetitive tail, and enough written documentation for an agent to answer from. If the knowledge lives only in your team's heads, writing it down is the first project, not this one.
It fits from a few dozen tickets a week upward. The threshold is less about volume than about repetition: two hundred varied tickets deflect worse than eighty near-identical ones.
It is not a replacement for a support team and we will not sell it as one. It is the machinery that stops your support team spending its day on the eight questions your documentation already answers.
- Sectors
- E-commerce · SaaS · Services
- Strongest fit
- United StatesEuropeLatAm
What it costs
| Size | Build | Monthly |
|---|---|---|
| Micro | Build$600–$1,200 | Monthly$150–$300/mo |
| Small business | Build$1,500–$3,000 | Monthly$400–$800/mo |
| Mid-market | Build$3,500–$7,000 | Monthly$900–$1,800/mo |
All prices are in US dollars. Out-of-scope work is $35–$60/hour. AI usage is billed at cost plus 20%, or you bring your own API key.
Frequently asked questions
- What proportion of tickets will it close?
- Typically thirty to sixty per cent of repetitive queries, with the rest arriving pre-qualified. The range is wide because it depends almost entirely on how repetitive your tickets are and how good your documentation is — both of which we can look at during the audit rather than guessing.
- How do we stop it making things up?
- It answers from retrieved passages of your own documentation and cites which ones. When nothing relevant is retrieved, it does not compose an answer — it hands over. That behaviour is the build; a model left to fill gaps on its own is precisely the thing that gets a support bot switched off.
- What happens to tickets it cannot handle?
- They reach a human with the conversation, the customer record and what the agent already established attached, so the person picking it up is not starting from the beginning. The handoff quality is the part clients notice most and it is what the monthly report tracks.
- Does it sit on our website or in our helpdesk?
- Either, and usually both. Chatwoot and Intercom are the common helpdesk integrations; a site widget covers the pre-sales questions that never reach a ticket. They share the same knowledge base, so there is one thing to keep current rather than two.
- Who keeps the knowledge base up to date?
- You do, and it is your documentation, so keeping it current is work you should be doing anyway. What the maintenance plan covers is the retrieval quality — spotting the questions that get answered badly and fixing the gap, which is the part that is hard to see from the inside.
We aren’t selling you a bot that says hello. We are selling you the share of tickets that stop reaching your team, measured every month.
Figures on this page are public industry benchmarks, not results measured on Amagenon client accounts. We will size the numbers against your own data on the audit call.
Find out what your process is actually costing you
A 30-minute call and a one-page report on the three highest-value automations in your business. Free while we build our first case studies.