Service
Bots and AI — designed around being wrong.
A bot or AI assistant closes part of your inbound without a person: FAQs, order status, intake, document lookup. A scripted bot typically takes 1–2 weeks; an LLM assistant grounded in your own documents takes 2–4 weeks. After a short brief we return within 3–5 days with a scenario map, stages and a firm quote, and that stage is not billed.
- Scripted bot
- 1–2 weeks
- LLM assistant
- 2–4 weeks
- Estimate
- 3–5 days
What happens when it is wrong
Every serious question about deploying an assistant is a question about its failure mode. A scripted bot fails visibly: it does not understand, it says so, it offers a button. An LLM fails invisibly — it produces a fluent, confident, wrong answer, and the customer acts on it.
The engineering is mostly about converting the second failure into the first.
Grounding is how. With retrieval-augmented generation the model does not answer from memory: it retrieves the relevant passage from your documents and answers only from that, citing the source. Where nothing relevant is found, the assistant says it does not know and hands the conversation to a person. We build that path first.
Escalation is the other half, and it is a business decision rather than a technical one. Which questions must always reach a human, what happens outside working hours, how much context travels with the handoff. We write those rules down during the brief and implement them as rules, so the behaviour is auditable rather than emergent.
Running cost belongs in the business case too. An assistant is billed per conversation by the model provider, and that bill scales with volume rather than with headcount. Usually a fraction of the equivalent staffing. Not zero. In our AI assistant project the measurable outcome was a 60% drop in operator load and replies inside 30 seconds instead of five minutes.
What the work covers
Included
- A map of conversation scenarios
- Channel setup: Telegram, WhatsApp, website widget
- Knowledge base preparation and indexing
- RAG retrieval with source citation
- Handoff to a human and escalation rules
- Integrations: CRM, stock, orders
- Conversation logging and quality metrics
- Testing against real historical questions
- Admin: editing answers and the knowledge base
Not included
- LLM provider usage fees
- Training your own model from scratch
- Writing the knowledge base content
- Hiring the operator team
- Registering messenger business accounts
What you receive
- A working bot on the chosen channels
- Repository and prompts in your account
- The knowledge base and a procedure for updating it
- Conversation logs and a metrics dashboard
- Escalation rules, written down
- Monitoring and a 3-month warranty
What drives the price
We do not publish a price list: a scripted bot and a document-grounded assistant are different pieces of work. These move a quote.
- Type: a button-driven script, or an assistant that reads free text
- Channel count: Telegram, WhatsApp, a site widget — each is connected separately
- Knowledge base size and condition: tidy documents or scattered files
- Integrations: stock, orders, CRM — what the bot needs to be able to look at
- Escalation: operator shifts, working hours, handoff rules
The scenario map, the stages and a firm quote reach you within 3–5 days of the brief. Free, and it commits you to nothing.
Stack
Chosen per task. A simple script does not need an LLM, and we do not put one there.
AI
- LLM API
- RAG
- LangChain
Backend
- FastAPI
- PostgreSQL
- Celery
Channels
- Telegram
- WhatsApp API
- Webhooks
Operations
- Docker
- CI/CD
- Monitoring
Stages
-
Brief and scenario map
3–5 daysThe real question list, the channels, and which cases must always reach a person. Output: a map and a quote.
-
Knowledge base
2–4 daysCollecting, cleaning and indexing documents. This is where it emerges that some answers were never written down.
-
Prototype conversation
2–4 daysA first working version on one channel. We test it against real questions and read the answers with you.
-
Integrations and escalation
3–8 daysStock, orders, CRM. Handoff, working hours, queueing — and the "I do not know" path.
-
Testing and release
1–2 daysA run against a hundred real conversations, correction of weak answers, channels taken live.
-
Warranty and tuning
3 monthsWe read the logs and fix weak answers. In the early months this is where the bot improves, not by changing model.
When you do not need this
A bot pays back on repetition. We raise these on the brief ourselves.
Ten enquiries a day and every one different
No repetition, nothing to automate. A person is faster and cheaper.
The answers do not exist in writing anywhere
A bot can only retrieve knowledge that exists. If prices, terms and rules live in people’s heads, they have to be written down first, and that is separate work.
A wrong answer is expensive
For legal, medical and financial advice an automated answer is a liability. There the bot should collect the question and pass it to a person, which is how we build it.
The expectation is that it replaces the team
It will not. A bot takes a share of the inbound and the rest stays with people. If the plan is to close the support desk, the result will disappoint.
If one of these describes you, write anyway. We will look at your actual conversation history and tell you for free what share can be automated and what cannot.
Related work
Frequent questions
Will it make things up?
Answers are assembled from your documents rather than the model’s memory — that is what retrieval-augmented generation means. Where no relevant passage is found, the assistant says it does not know and hands over to a person. That behaviour is standard, not an upgrade.
Do we need to train our own model?
Almost never. An off-the-shelf model plus your knowledge base is cheaper and more accurate in practice. Training your own model is a large separate project and we propose it only where there is a real reason.
What does it do with a question it cannot answer?
It does not guess: the question and the conversation context go to an operator, the customer is told how long the wait is, and the case itself is logged — the next version of the knowledge base is built from those.
Which channels does it work on?
Telegram, WhatsApp and a site widget are the common ones. Each channel is connected separately and appears as its own line in the quote, but the logic is shared rather than rewritten per channel.
Who maintains the knowledge base?
You do, through the admin panel: add a document and the index updates itself. No developer needed for it. After large changes the answer quality is worth re-checking. We show you how at handover and leave a written procedure for maintaining the knowledge base.
Which questions should the bot take
A short brief is enough: which channels, how many enquiries a day, and which questions repeat. We come back within 3–5 days with a scenario map and a quote.