Four weeks from scoping call to a live agent: what actually happens
Our deployment process, week by week — what you do, what we do, where it usually gets stuck.
When we tell people Coluum can put a live AI agent in their business in four weeks, the most common reaction is polite scepticism. Fair enough — most software projects in this space slip by months and then ship something underwhelming. So here's exactly what the four weeks look like, and where they tend to get stuck.
Week 1 — Listen
We sit with your team for two days. Not a kickoff slide deck — a real two days at the desk, watching the work happen.
We tag every inbound message, ticket, and call for a representative period. We mark which ones got resolved on the first reply, which bounced around, which got escalated and to whom. By the end of the week we know — with numbers — what slice of your team's time is repetitive and what isn't. That slice is the agent's job description.
Where it gets stuck: sometimes the work isn't actually that repetitive once you measure it. We'd rather find out in week one than week eight. If the agent doesn't have enough volume to justify the project, we tell you and refund the deposit.
Week 2 — Design and train
We translate the job description into an agent specification:
- What can the agent decide on its own?
- What gets escalated, and to whom?
- Which tools does it need access to (CRM, email, WhatsApp, calendars)?
- What's the tone of voice — and whose past messages do we train on?
In parallel we ingest your structured data. For a real estate agency that's the property portfolio. For a clinic it would be services and protocols. For a law firm it would be matter types and document templates.
Where it gets stuck: access. The data is almost always somewhere — but it's in a custom CRM, or a spreadsheet someone updates manually, or three different tools that don't talk to each other. Week 2 is mostly plumbing.
Week 3 — Shadow mode
The agent goes live, but only as a draft. Every reply it would send appears in a queue for your team to approve, edit, or reject before it goes out.
You see exactly what the agent would have said in the wild. We see exactly where it's wrong. The model gets corrected, the property/protocol data gets sharpened, the escalation rules get tuned. By the end of week three the approval rate is usually above 90%.
Where it gets stuck: not stuck so much as humbling. Week 3 is when agencies discover edge cases they didn't know they had. Always property #14 with the broken hob. Always a guest asking about the bathroom in three different ways.
Week 4 — Cutover and handoff
The agent starts replying directly. The team goes from approving every message to approving only the borderline ones. We run daily check-ins for the first week and weekly thereafter.
You're now in steady state — which for Coluum means you're not on your own. We operate the agent. Weekly reviews of every conversation. Continuous prompt tuning. A human at our end you can call.
Where it gets stuck: the handoff itself, occasionally. Some agencies get nervous about cutover and want a fifth week of shadow mode. That's fine — we extend. The four-week target is the median; we'd rather ship right than ship fast.
After week 4
The agent gets sharper over time. New properties get added, new edge cases get learned, new channels get plugged in. The longer it runs, the better it gets — which is the opposite of how most software ages.
If you want to see what week one looks like for your specific business, book a scoping call. It's free and takes 30 minutes.