A managed IT firm cut inbound admin time by 87% for £140 a month

For this week’s AI in practice, we spoke to the founder of a regional managed IT services provider that had grown faster than its admin processes could keep up with. Two engineering teams, 250+ support and onboarding requests a month, and a phone line that never stopped ringing.

Out-of-hours web form requests piled up overnight with no one to pick them up. Calls, emails, and portal submissions were all landing in different places with no consistent way to process them. Clients with urgent issues were often waiting hours for a response.

Coordinators were also losing around 15 hours a week to phone handling and rescheduling. Many of these needed 10-15 minutes of work before any engineering even started. Checking which engineer was free, looking up the client’s contract, confirming response commitments, then logging it all manually. And with engineers split across two teams, keeping track of who was available and when was a constant challenge – clients got promised slots that didn’t exist.

The owner wasn’t just looking to save coordinator time. He’d watched the business grow and knew managing every request by hand wasn’t going to hold. He wanted a front door that worked at any hour without adding headcount.

The build: off-the-shelf tools, £140 a month

They built a first working version from off-the-shelf tools in around 8 hours, then iterated over the following weeks.

Twilio provides the phone layer with a local geographic number. When someone calls, OpenAI’s API powers a scripted intake flow. It captures site location, issue type, urgency level, and preferred time window. The firm runs scheduled triage for onsite and complex work. Calendar-based booking where the agent checks availability through Google Calendar suits how they already work.

Zapier Pro connects the tools, pushing confirmed bookings into the firm’s existing job management system. Once booked, the agent sends a confirmation message with ticket reference and next steps. Reschedules follow a similar flow.

Anything outside the agent’s scripted scope gets handed to a human coordinator during working hours. Loss of service, contract queries, multi-site problems, or any caller who asks for a person. Out of hours, the agent captures the details and flags it for morning review.

The whole stack runs at around £140 a month – £47 for Zapier Pro, £35-40 for Twilio depending on volume, and £20-30 for OpenAI.

What went wrong and how they fixed it

The first version tried to handle full triage on the phone, categorising severity, matching response timeframes, and suggesting next steps. Callers gave incomplete information and the model made bad guesses. Coordinators ended up re-doing the work anyway. Switching to a “capture and book” model – where the agent just collects the details and books a slot, nothing else – cut the repeat work.

The early build used a voice model that sounded tinny and misheard jargon. It tripped up on technical terms and site references. A better model cost an extra £10 a month, and the confirmation message helped with the rest. Clients saw what the agent had captured and replied to correct anything before someone got dispatched.

Even after those fixes, the system needed ongoing fallback controls. Zapier occasionally failed to pass data between tools when call volumes spiked. They ran a daily reconciliation, comparing calendar entries against the job management system and flagging anything that slipped through. If a booking didn’t land, the caller still had their reference number and a coordinator followed up the next morning.

And while the agent nailed down engineer availability more efficiently, it couldn’t account for jobs that ran longer than planned. That remained a coordination problem between field teams and the office, handled the same way it always had been – engineers flagging delays and coordinators reshuffling the afternoon.

Six months on

The majority of requests are straightforward booking or rescheduling. Instead of spending 15 hours a week on the phone chasing details, coordinators now spend under 2 hours reviewing what the agent captures and correcting anything it gets wrong. That’s about 87% less time on inbound handling.

With the agent covering intake, more enquiries convert into booked jobs. Roughly a third more based on the firm’s own tracking, largely because calls stop going unanswered. The owner reckons it paid for itself inside the first month through coordinator hours saved and fewer missed calls.

They also learned that not every client wanted to deal with an AI. Early on they added a simple bypass, “press 1 for a person”, and started the rollout with new enquiry calls only. Existing clients with open tickets stayed on the human line until the team was confident the agent handled routine intake cleanly.

On the compliance side, they initially treated GDPR like a documentation exercise. A quick privacy notice update, keep the recordings, and figure out the rest if anyone asked.

A few weeks into live calls, it was obvious that wouldn’t hold. Voice AI creates recordings by default, pushes data through multiple tools, and sends automated messages. So they tightened up, setting recording retention to 30 days with automatic deletion, restricting access to recordings and customer data, updating privacy notices to explicitly cover AI-assisted call handling, and adding a clear disclosure at the start of each call with an opt-out route to a person. They also started documenting how they’d handle data subject access requests and objections to automated processing as volumes grew.

What they’d do differently

When we asked the founder what he’d change if he was starting again, the first thing he said was scope. Start with one job for the agent: capture the details and book the slot. They tried to automate triage on day one and had to strip it back. The narrower the scope, the faster you learn whether it works.

He also felt strongly that you need to budget as much time for rollout as you do for the build. The bypass, the staged introduction, the weekly call reviews – that’s what kept clients comfortable. The technology was the easy part.

And finally, perhaps most importantly, don’t leave compliance until after launch. Recording retention, disclosure, opt-out routes, processor terms. It’s all manageable if you build it in from the start. Less so if you’re retrofitting it at scale.