Case study: the company we built the AI operator for first

Every AI agency has a case study. Most of them are a logo, a stock photo and a number nobody can check.

Here's ours. It's the only one we'll tell, because it's the only one we can stand behind line by line: Intelligent Living Solutions, the Canberra smart-home company the two of us own and run. The operator has been live there since early 2026, on real leads, real quotes and a real CRM. This is what we wired in, what a day looks like now, the numbers we can show, the numbers we won't claim, and what we got wrong.

If you run a building company, an electrical business or a clinic, the shape of the problem is the same as ours. Read it with your own inbox in mind.

The situation

Intelligent Living Solutions (ILS) designs and installs smart-home systems in Canberra. Lighting, blinds, audio, security, the lot. Two founders, Jonathan Kamel and Francois le Grange. No admin staff. No receptionist. No sales assistant. Two people who sell the work, design the work and then go and do the work.

Most of the leads come from Meta ads. Meta doesn't care what time it is. A lead form fills at 7am, at lunch, at 11pm on a Saturday, and each one lands as a notification while one of us is up a ladder or in a ceiling. It sits until someone gets to a laptop, then it needs a CRM entry, a first reply, a follow-up and eventually a priced proposal.

So the admin happened at night. Quotes at 10pm. CRM updates on Sunday. Follow-ups that should have gone Tuesday going Friday, or not at all. Same story as every two-person trade business in the country. You already have someone doing the office work for free. It's you, after dinner.

We build automation for a living, so we built the operator for ourselves before we offered it to anyone.

What we wired in first, and why in that order

We didn't try to automate everything on day one. We went after the jobs in the order that leaked the most money.

1. Lead intake into the CRM. First job, because a lead that isn't in the CRM doesn't exist. The operator now syncs Meta leads into our CRM every hour. Every enquiry gets a record, a source and a timestamp without either founder touching it. That alone killed the "did anyone see that lead?" conversation.

2. Follow-ups. Second, because that's where the money leaks. A lead that gets one reply and no chase is a lead you paid Meta for and then threw away. The operator schedules the follow-ups, flags the overdue ones, and keeps chasing until the lead books or goes cold. Leads that go quiet for 30 days are auto-archived so the pipeline stays honest instead of full of ghosts.

3. Proposals. Third, because writing quotes was the biggest chunk of night work. The operator runs a proposal pipeline: priced bill of materials in, branded PDF out. A founder still decides what goes in the system and signs off the price. What we stopped doing was formatting the document at 10pm.

4. Reviews. Fourth, because we kept forgetting to ask. Once a job is done, the review request goes out on schedule instead of when someone remembers.

Around those four sit the boring but essential bits: an unclaimed lead pool with claim and release so two founders don't double-handle the same enquiry, a weekly marketing review, nightly offsite backups, and Telegram approvals for anything client-facing.

What a day looks like now, from our side

No dramatised timestamps here. This is just the routine.

Morning. At 9am the operator sends the day's action list to both of us on Telegram, straight from the CRM. New leads sitting in the unclaimed pool. Follow-ups that are due or overdue. Proposals waiting on a decision. You read it with your coffee and claim what's yours.

During the day. When the operator wants to do something a client will see, a reply, a quote, a follow-up message, it asks. A yes or no on Telegram, from wherever we are. If it's a no, it doesn't send. New Meta leads keep landing in the pool every hour whether we're on site or not.

Evening. At 6pm the operator does the end-of-day CRM reconciliation. It looks at what actually happened during the day, the conversations, the proposals discussed, and proposes the CRM changes to match: this lead moved to this stage, for this reason. Clear ones it applies and tells us. Anything ambiguous, it asks. The CRM matches reality at close of business instead of drifting for a fortnight.

Overnight. Backups run offsite. Leads keep arriving into the pool. Nobody is up.

That's the whole rhythm. Two messages a day, approvals as they come, and the CRM stops being a chore.

The numbers we can show

CRM data as of 17 August 2026, from the ILS system:

Measure	Number
Leads since December 2025	337
Of which from Meta ads	267
Won	10
Lost	89
Archived (went cold, auto-archived after 30 days)	203
Remaining in other stages (by subtraction)	35 (8 new, 8 contacted, 1 emailed no reply, 11 no answer, 5 proposal drafting, 2 proposal sent)

That's 337 enquiries in about eight and a half months, roughly 40 a month, 267 of them from Meta. Every one of them landed in the CRM without a founder typing it in. Before the operator, that entry work was manual and it was done late or not at all.

Yes, 203 archived is a big number. That's the truth of Meta lead forms for a premium install business: a lot of tyre-kickers, a lot of people who never answer. The point isn't that the operator turns bad leads good. The point is that the good ones now get chased on time and the dead ones get cleared out instead of clogging the list.

Ten won at smart-home install values is real work for a two-person outfit. We won't dress it up further than that.

The numbers we won't claim

We could put a "hours saved per week" figure here. Every AI agency does. We haven't measured it properly yet, so it stays blank until we have.

Claim	Status
Hours of admin saved per week	Not measured to a standard we would defend, so not claimed
Average first-response time to a new lead, before vs after	Not measured to a standard we would defend, so not claimed
Lead-to-won conversion, before vs after	Not measured to a standard we would defend, so not claimed
Revenue attributed to the operator	Not claiming it

If someone else states all four of those confidently after a three-month deployment, ask them how they measured. The answer is usually interesting.

What we can say plainly: two founders with no admin staff now run 40-odd enquiries a month through a CRM that stays current, with quotes going out as PDFs instead of late-night documents, and neither of us does data entry at 10pm.

What went wrong, and what we changed

This is the part most case studies leave out.

Leads were going to one founder instead of the pool. For a while, every new Meta lead was auto-assigned to Jonathan the moment it synced. Nobody decided that; the intake code set the assignee to whoever set the ads up. One founder had a stack of leads he hadn't asked for and the other couldn't see them. We changed the intake so new leads land unassigned in the pool, then either founder claims them. Source attribution stayed, ownership got fixed. Small bug, real consequence: the whole point of a two-person business is that either person can pick up the phone.

The operator needed rules about what it may not do. A capable assistant with no boundaries is a liability, so we wrote the boundaries down. Anything a client might see requires a Telegram approval before it goes. Pricing follows a locked price sheet the operator does not get to improvise on. Ambiguous CRM changes get asked about, not applied. It runs supervised because we designed it to, not because we forgot to take the training wheels off.

A template bug went out on real proposals. A label meant for one specific pricing scenario was hardcoded into the shared proposal template and printed on every proposal until we caught it. Fixed by making it conditional. Lesson: anything the operator sends to a client gets checked by a founder, and we build the checks so a wrong line gets caught before the PDF leaves.

None of these are exotic. They're what happens when a system meets a real business, and the reason we insist on the two-week supervised rollout for every client. We hit them on ourselves first.

What this means if you're a builder, a sparkie or a clinic

Strip the smart-home detail out and look at the shape:

• Enquiries arrive around the clock, from a form or a phone, and the owner is the only person who can deal with them.
• Each one needs to be recorded, replied to, followed up, quoted and eventually chased.
• All of that happens after hours because the day is spent doing the actual work.

That's an electrical contractor. That's a physio clinic with two practitioners and no front desk. That's a small builder whose partner does the invoicing at the kitchen table.

The wiring is the same as ours: intake into your CRM first, follow-ups second, quotes or bookings third, reviews and invoicing after that. Your accounts, your CRM, your Gmail or Outlook, your Xero. Telegram or WhatsApp as the approval front door. Two weeks supervised, then you decide how much rope to give it. If you leave, you export everything and take it with you.

And the math is the same as ours. An admin hire is about $75K a year once they're trained. The Managed plan is $2,000 a month, $24,000 a year. That's $51K back every year, and it covers the nights and weekends you can't roster anyone for anyway. You can do the sums yourself. It's not close.

Related reading

• What an AI operator actually does for a small business
• AI receptionist vs human vs voicemail: the real cost
• 9 questions before hiring an AI automation agency
• Plans and pricing
• Book the call

We built this for our own company first and we still run it there every day. If you want to see whether the same shape fits yours, book the 20-minute call and we'll walk you through it with your numbers, not ours.