We build the agent into a business process: task and data first, then integration, and finally a measurable result. A single-channel, single-task agent typically goes live in 2–4 weeks.
Timelines are for a single-channel, single-task agent. Agents that connect to several systems are built in stages.
Which work to hand to the agent, how long it takes now and how the result will be measured. Baseline KPIs come from the CRM and reports.
The catalogue, FAQs, scripts and SOPs are gathered and cleaned up. Outdated and conflicting documents are removed, and every source gets an owner.
What the agent does, when it hands off to an employee and what it never does are written down and approved.
The channel (Telegram, Instagram, website), CRM, ERP and calendar are connected via API. A separate technical user is created for the agent.
First on a test set, then with employees, then on a share of real customers. Errors are fixed along the way.
Weekly reviews in the first month: KPIs, error samples, unanswered questions. Monthly reports after that.
Every document stays with you, so another team can support the agent later.
| Stage | Document or deliverable |
|---|---|
| Task and KPI | Agent profile: task, boundaries, KPIs and baseline values |
| Data preparation | A cleaned-up knowledge base and a list of source owners |
| Scenario and guardrails | Scenario, handoff rules, list of off-limits topics |
| Integration | Integration diagram, permissions table, data flow map |
| Testing | Test set, results report, list of fixes |
| Monitoring | KPI report, staff guide, 30 days of support |
Specific targets are agreed during the diagnostic, based on the baseline.
The agent’s owner must be on the client side.
Updates the knowledge base, reviews handed-off conversations, tracks KPIs.
Take part in testing, flag wrong answers, handle the tasks handed off to them.
Architecture, integration, scenario, testing, staff guide and first-month monitoring.
For the first agent, we pick a repetitive, measurable, low-risk task, such as frequently asked questions on Telegram or creating drafts from documents. Once the results are visible, we move on to the next agent. The most widely used agents are the sales assistant, the booking agent and the knowledge base.
Not at first: it works on an internal test set and with employees. Then it is switched on for part of the real traffic, for example only outside business hours.
Yes. Your employee updates the knowledge base and reply templates. We provide a guide for changing the scenario and integrations.
You can, but we don’t recommend it. The first agent establishes how data, permissions and monitoring are handled. The next agents are built on that foundation faster and with fewer errors.
We review your processes, channels and systems, rank the tasks that could be handed to AI by benefit and risk, and propose a task, KPIs and a plan for the first agent.
If KPIs fall short of the agreed level during testing, we analyze why: the data, the scenario or the choice of task. Sometimes the right conclusion is not to hand this task to AI yet.
We review your processes and recommend a specific task, KPI and plan for your first agent.