How we onboard a new AI employee in week 1

The most common pre-sales question after "is this a chatbot" is "how long does setup take, and what do you need from us?" The honest answer is five business days, sometimes less, with about three to four hours of your team's time spread across them. The rest is our work.

This post walks through what those five days actually look like. The defaults we ship. Why those defaults. Where we deviate. What you should expect to see at end-of-week-one, and what month two looks like once it's live.

The five-day default

The default shape, with a Monday kickoff:

• Mon: brief and scope
• Tue: knowledge ingest
• Wed: voice tuning and system prompt
• Thu: quality test
• Fri: go-live

If your content is clean and the role is one of our well-trodden ones (Shop for a webshop, Support for a SaaS, Receptionist for a clinic), we sometimes compress to three days. If your catalog is in a long-form spreadsheet that needs reshaping, or your service docs live in a Notion no one has touched in 18 months, we extend.

Day 1: brief and scope

One call, 60 to 90 minutes, with whoever owns the role on your side. Three people on our side: the operator who'll run the agent through month two, an engineer who'll handle the catalog or KB integration, and someone from product taking notes.

We come out of the call with five concrete things written down:

• The role. Shop, Support, or Receptionist. Specific enough that the system prompt writes itself.
• The top 20 questions your visitors actually ask. Not the questions you'd like them to ask. The ones in your inbox today.
• Tone. A handful of sample sentences from your own writing, or three to five adjectives ("calm, direct, no jargon").
• The escalation rule. When the agent doesn't know, where does it route. Email inbox, Slack channel, CRM stage, all of the above.
• What it must never say. Pricing it can't confirm. Promises about delivery windows. Anything regulated.

We send the writeup the same day. You have until Tuesday morning to push back on anything that doesn't match your read.

Day 2: knowledge ingest

We start the catalog or KB pipeline. Most of this runs unattended on our side; what we need from you is access and a confirmation of which sources are authoritative.

Three patterns we see most often:

• Webshops. WooCommerce or Shopify sync, then a pass to flag categories where the agent should escalate instead of quoting (custom builds, B2B accounts, anything regulated). This is the fastest of the three, usually under 2 hours of attended time on our side.
• Service businesses. Pages crawled, policies pulled from PDFs you send us, FAQ documents collated. Slower because the canonical source is often "the founder's head," and we have to write down what's only been said aloud.
• Reception roles. Schedule, service catalog, intake forms, plus the office's stated hours and the after-hours rule. The variance here is huge; we sometimes finish ingest in 45 minutes for a small law firm and take a full day for a larger clinic with multiple locations.

You'll get a link to the KB in your dashboard the same day. We ask you to skim it. Wrong sources, stale content, things that shouldn't be in there: flag them in a comment thread.

Day 3: voice tuning and system prompt

We write the system prompt against the brief. This is the file that defines who the agent is, what it does, what it won't do, and how it talks. It's longer than people expect; a well-tuned prompt for a Shop Employee is typically 1,200 to 2,000 words.

Then we run a dozen sample exchanges with you. You send a question the way a customer would; we paste the agent's reply back. After three or four iterations, the voice clicks. The signal that we're done with day 3 is when you read a reply and say "yes, that's how we'd say it."

This is the day with the most back-and-forth. Block 30 minutes for it in the afternoon. The rest of the day you don't see us.

Day 4: quality test

We run a structured test: 30 to 50 questions you and we wrote together on day 1, covering the cases that matter. Common pricing question, edge-case product, out-of-scope ask, escalation trigger, the angry customer, the spammy "test test test."

Each answer gets one of four labels:

• Pass. Answer is correct, voice matches, no fabrication, escalation handled if needed.
• Fail (knowledge gap). Answer is wrong because the KB is missing something. We add the source.
• Fail (prompt issue). Answer is wrong because the system prompt told it to do the wrong thing. We adjust.
• Fail (model limitation). Rare, but it happens. The agent's response is structurally fine but the underlying model misunderstood the question. We add a guardrail.

You get the full test transcript with our labels. Pushback is welcome; sometimes you have a "fail" we have as a "pass" because we don't know your business well enough yet. Day 4 is the last cheap day to catch mismatches.

Day 5: go-live

You paste one line of JavaScript on your site. We watch the first hour. We're in our own dashboard reading every conversation as it comes in, ready to escalate to your team manually if something we missed in testing surfaces.

By end of day Friday: every conversation from the first day has been read by a human on our side. If anything weird shows up, you hear from us before you'd notice yourself. If it's quiet, you don't.

Why these defaults

Three constraints shaped this shape.

A working day for the customer is one hour or less. People agreeing to a managed AI employee are SMB owners or ops leads. They cannot disappear for a week. Day 1 is the longest at 90 minutes; the rest are 30 to 45 minutes of attended time. The five-day timeline isn't five days of your time. It's five days of ours.

The test happens before go-live, not after. The default at the cheap-tier widgets is "ship it, see what happens, fix later." Our default is the opposite. Day 4 is the entire reason a Co-lega agent costs what it does. Failures caught in test cost nothing; failures caught in production cost a real customer.

Go-live is on a Friday, not a Monday. Counterintuitive, but: a Friday launch gives us two full days of low-traffic conversations to watch the agent in the wild before Monday's load arrives. If something's off in the prompt or KB, we catch it on Saturday morning and fix it before the customer notices on Monday.

When we deviate

Cases where the five-day default isn't right:

• Heavily-regulated industries. Medical, legal, financial. We extend day 3 to write tighter "must never say" rules with your compliance contact in the room. This typically adds 2 to 3 days.
• Multi-language sites. Each additional language is half a day of voice tuning on day 3 plus a doubled quality test on day 4. The five-day shape becomes six or seven.
• Content that doesn't exist yet. If your FAQ is verbal and your service docs are in screenshots, we'll pause onboarding and either (a) help you write them or (b) recommend a content shop. Either way we'd rather extend by a week than ship an agent on top of garbage.
• You're replacing an existing solution. If you're migrating off a chatbot or a human team that's currently handling these conversations, day 4 expands to include a parallel-running test with the legacy stack. We compare answers side-by-side before we cut over.

What month two looks like

The setup week is the visible part. The work that makes the agent worth what it costs is what happens next.

Every month, we read the transcripts. Not a sample. Every one. You get a report: what customers asked, what the agent answered well, what it missed, what it escalated, and what we changed. On Growth and Managed plans we walk through it on a call. On Starter, it's a written summary in the dashboard.

If you've read the comparisons post you've already seen this argument. The €200+ per month isn't paying for the chat widget. It's paying for the person reading the transcripts.

For the role-by-role detail on what an AI employee actually does on a live site, the role-guide covers it. For the pricing tier that matches your volume, the pricing page breaks it down. If you want to see the result before you book a setup call, the FlexiRENT demo is a real agent on a real customer site you can click through right now.

See it on your site

When you're ready to start your own week 1, request a demo and we'll walk you through what your shape of business looks like in our pipeline. We'll tell you honestly whether five days is the right window or whether yours is the case where we'd extend.