The main myth about AI automation goes like this: you need a technical team of five developers and six months. In reality the first agent goes live in two weeks with no programmer on staff — on one condition that usually goes unmentioned.
The condition: those two weeks start once the process has been chosen and the documents for it have been gathered. Without that, three weeks is the honest number. Below is the plan day by day, marked out by who does what.
In short: days 1–4 are decisions only the client can make. Days 5–7 are data preparation, also on the client's side. Days 8–14 are the contractor's technical work. Delays almost always happen in the second stretch, not the third.
Days 1–2: process audit
Before automating, you need to know what. Three questions produce a shortlist:
- Which tasks do your staff do every day and openly hate?
- Where does the most time — and money — leak out?
- Which processes can be described as a clear "if A then B" algorithm?
The third question eliminates the most candidates, and that is fine. If you cannot describe a process in words to a colleague, you will not be able to explain it to an agent either.
Output of this stage: a list of three to five processes with a rough estimate of how many hours a week each one consumes.
Days 3–4: picking one process
One is taken from the list. Not two, and not "let's start with three and see."
High impact + low complexity is the ideal first project. High impact + high complexity comes second or third, once the team has worked with an agent. Low impact should not be touched at all, however technically interesting it looks.
The quick wins most often chosen first:
- answering routine customer questions in a messenger;
- processing and qualifying inbound enquiries;
- an internal knowledge-base bot for employees.
A detailed breakdown with numbers is in 5 processes to automate first.
And here, the success criterion
Before work starts, fix one number that will make it obvious in a month whether this worked. "It feels more convenient" is not a number. Workable ones: the share of enquiries closed without a human, first response time, hours freed up for a named employee. Without such a number a pilot can neither be defended nor closed. How to count it is in measuring AI ROI.
Days 5–7: preparing the knowledge base
The most important stage, and the only one that cannot be fully delegated. The agent will be exactly as good as the data it works from.
What to collect:
- routine questions and answers — 50 minimum, 100 is better;
- the product or service catalogue with prices and descriptions;
- sales scripts: how your managers actually reply to clients;
- internal procedures and instructions for the process.
Format: ordinary text documents or spreadsheets. No special markup is needed and formatting does not matter. What matters is that the content is current.
Write down separately what the agent must not do: quote prices, promise deadlines, discuss contract terms. A list of prohibitions saves more grief than a list of capabilities.
A realistic estimate of your own time at this stage is five to six hours. This is where schedules slip most often: documents turn out to be outdated, mutually contradictory, or to exist only inside one person's head. Technology has nothing to do with it — this is management work, and it has to be done regardless.
Days 8–12: build and configuration
This is the contractor's part:
- setting up a RAG system on your knowledge base;
- connecting the language model and working through the prompts;
- integrating with the channel — website, Telegram, Instagram Direct, CRM;
- escalation logic: when the agent is obliged to hand the conversation to a human;
- testing on real scenarios, including deliberately awkward ones.
What is needed from you: access to the channels and CRM, feedback on test answers, two or three half-hour calls. The feedback matters more than the access: only you can tell that an answer is formally correct but does not sound like your company.
Days 13–14: launch and observation
The launch is soft: the agent runs alongside people, not instead of them.
- Day 13: switched on for part of the traffic, every conversation reviewed by hand.
- Day 14: prompts corrected based on what was found, then rolled out to the full flow.
For the first week after launch, someone on the team reads the conversations every day. This is not a formality: nearly every real problem surfaces with live users rather than in testing. People phrase questions in ways nobody in the office would have thought of.
What you do NOT need
Not needed: programmers on staff — the contractor handles the technical side. Servers and infrastructure — it all runs in the cloud. Months of development — two weeks from start to a working agent. A large budget — a pilot $1,500.
What you do need
- an understanding of your own processes — you know them better than any contractor;
- five to six hours to prepare the data;
- two or three calls;
- one person on the company side who owns the project;
- a budget of $1,500 for the pilot.
The fourth point matters as much as the rest. Projects without an owner sink not because of technology but because the contractor's questions sit unanswered for a week.
Frequently asked questions
Is it really two weeks and not two months?
Two weeks is the timeline for one process with documents already prepared. Three parallel processes, an integration with a bespoke in-house system, or a missing knowledge base will stretch it. We quote that timeline because we take on exactly one process.
What if the agent answers badly after launch?
The first week is deliberately set aside for refinement — that is part of the pilot, not a rework. Bad answers almost always mean something is missing from the base or an instruction is vaguely worded. Both are fixed in hours.
Do we have to tell customers they are talking to a bot?
Yes. It is both a platform requirement and a matter of trust: the jolt of "that was a person — suddenly it is a different person" costs more than an honest note in the first message.
What happens after the two weeks?
You look at the number fixed on day three. If there is a result, you expand — a second process, or going deeper on the first. If there is no result, you work out honestly why, and that is also a legitimate outcome for a pilot. That is exactly why we start with a small sum rather than an annual contract.
Who owns the resulting solution?
You do. The knowledge base is yours, the credentials are yours, the agent runs in your channels. This is worth agreeing at the start rather than after the fact.
Where to start today
Not with choosing a contractor, and not with choosing a model. Start with day one of the plan: write out the processes that eat time and count the weekly hours for each. An hour later you will have a list concrete enough to have a real conversation about.
Want to walk through the first four days together? We run a free process assessment and will say plainly if there is nothing worth automating yet.