Support triage
Answers the repeatable questions from your own documents and passes the rest on with a summary.
- Retrieval
- Guardrails
- Handoff
Service 04 of 05 — AI agents
An assistant that knows when to stop.
Support and sales assistants built on your documents, FAQs and product data — answering what they can and handing the rest to a person.
One builder
You talk to the person writing the code, not an account manager.
No handoffs mid-project.
Scope first
Written scope and price before work starts. Changes are priced before they are built.
No surprise invoices.
Your accounts
Domain, hosting, repositories and model keys stay in your name.
Everything is handed over.
After launch
A fix window is agreed in writing, with a maintenance option if you want one.
I do not disappear at go-live.
Scope
What this covers.
The useful version of an AI agent is not the one that answers everything. It is the one that answers from your material, says when it does not know, and passes the conversation on with context.
Lead qualification
Asks the questions a good salesperson would, scores the answer and books time with the right person.
- Qualify
- Score
- Book
Internal search
A private assistant over your SOPs, contracts and notes, with sources shown for every answer.
- Private
- Sources
- Roles
Evaluation
A test set of real questions, run before launch and after every change, so quality is measured rather than assumed.
- Test set
- Scoring
- Logs
How it runs
Four steps, no mystery.
The same sequence on every project. You always know what is being built, what it costs and what happens next.
01
Scope
We pick one job, list the questions it must handle and the answers that must never be guessed.
02
Sources
Your documents, pages and data are cleaned, chunked and indexed. Bad source material is fixed first.
03
Build
Retrieval, prompts, guardrails and a handoff path. Tested against real questions, not demo ones.
04
Measure
A scored test set, conversation logs and a review after the first month of live traffic.
Deliverables
What you get, and what I need.
No vague handover. The list on the left is what leaves my hands; the list on the right is what I need from yours to make it happen.
What you get
- A working assistant on your site, WhatsApp or internal tool
- Retrieval over your own documents, with sources cited
- A refusal and handoff path for anything outside scope
- A test set and a score you can re-run after changes
- Conversation logs and a monthly cost estimate
- Model keys and infrastructure in your account
What I need from you
- Your real documents, FAQs and product data
- A list of the questions people actually ask
- One person who can confirm what the correct answer is
- A decision on where the assistant lives: site, WhatsApp or internal
- Acceptance that some questions must go to a human
Straight answers
What I won't promise.
Most studios sell certainty they cannot deliver. These are the claims I will not make, and what I do instead.
If any of these matter to you, say so on the call.
- 01100% accurate answers.Language models make mistakes. I add retrieval, guardrails and a handoff so errors are contained.
- 02No running costs.Model calls are metered. I estimate the monthly volume before you commit.
- 03That it replaces your support team.It handles the repeatable volume and escalates the rest. Headcount is your call, not my promise.
- 04Magic from messy documents.If the source material is contradictory or out of date, the agent will be too. Cleaning it is part of the work.
Keep looking
The other four.
Start here
Tell me what you are building.
A short description is enough to start. I will reply within a day with whether I can help, what it would take and a straight answer on budget.