An AI automation agency in London that runs its own agents.

An AI agent is software that reads, decides and acts inside the tools you already use. It reads the email, the CV, the invoice or the form; decides what it is and what should happen next; and then does it: replies, books, files, updates the CRM, or hands it to a person. The difference from ordinary automation is that the input does not have to be tidy. It can be a scanned PDF, a rambling voicemail or a message in the wrong language.

What an AI automation agency does is build those agents around your process, connect them to your systems, decide where a person has to stay in the loop, and keep them running. Most of the work is the connecting and the deciding, not the AI.

Every one of these is running somewhere right now, either in a client system or inside our own products. The examples are real.

The market is wide and mostly unpublished. Where UK agencies do publish, a first workflow runs from roughly £800 to £5,000, a departmental build £5,000 to £20,000, and multi-system programmes £20,000 to £50,000 and up. Retainers for running the system afterwards sit between £350 and £2,500 a month. Some agencies sell a modular ladder: an audit under £2,000, a single agent deployment around £5,000, a monthly partner fee. That is the honest shape of the market and it is the shape of our pricing too.

One AI agent, built and integrated, costs us to deliver somewhere between £4,000 and £15,000 depending on how many systems it touches and how messy the inputs are. It costs pennies per task to run, and we tell you the number before go-live and cap it in production.

Market figures are published UK ranges gathered in September 2026.

Every vendor demo shows a call being answered and a natural conversation following. Our production data on outbound AI calls says something different: about half reach voicemail, a quarter never connect, and roughly one in eight reaches a person. When a person does answer, the conversation holds for a little over three minutes and nineteen in twenty produce a complete, scored assessment.

That is why the fallback matters more than the voice. Ours sends a text within the same minute with a link to finish by chat, records that voicemail was reached so nobody redials the same person an hour later, and spaces retries by hours. A voice agent without that is a very expensive way to reach voicemail. We publish the numbers because a buyer who knows them buys the right thing.

It cannot fix a process nobody can describe. If the exceptions live in one person's head, the first two weeks are spent writing them down, and that is discovery, not automation.

It cannot be trusted with money or compliance unattended. Anywhere a decision costs money, affects a person's rights or touches regulated data, the agent proposes and a person approves. We build the approval step in from the start rather than bolting it on after the first mistake.

It cannot make a bad channel good. An agent that emails people who never read email is fast, cheap and useless. The audit checks where your customers actually respond before we build anything that talks to them.

And it is never free to run. Every agent has a cost per task, and the channels around it, messages and calls especially, usually cost more than the model. We instrument both and cap both.

A 30-minute call, then the audit if you want one. A working prototype on your real inputs inside the first week, because the messy tenth example teaches more than the ten clean ones. Integration with your systems next, since it is the long pole. Then testing against a sample of real traffic with the person-in-the-loop rule switched on, a written acceptance threshold, and go-live with cost ceilings already triggered once on purpose so you know they fire.

At handover you receive the repository, the provider accounts, the credentials, a runbook and a recorded walkthrough, in your name. If you want us to run it, that is a separate line in the same document, not a negotiation after go-live.

Low-code tools are the right answer for a lot of automation and we use them when they are: connecting two well-supported apps, moving data on a trigger, a workflow your own team should be able to edit. They are the wrong answer when the logic is complex, the volume is high, the inputs are messy, or the failure mode is expensive. Then we write code, because code can be tested, versioned, instrumented and retried properly.

We tell you which we are using and why, and you own it either way. What we will not do is put a business-critical process on a tool that bills per operation without telling you what a busy month will cost.

Agents that contact people are regulated. Automated calls and messages need consent or a lawful basis, an opt-out that is honoured on every channel, and a clear statement that the person is dealing with an AI. Our agents state it, gather consent, and never dial or message anyone on the do-not-contact list. Augustova is ICO registered (ZC152144) and a data processing agreement is available on request.

If you trade with EU customers, the EU AI Act adds transparency duties for AI that interacts with people and stricter rules for anything that makes decisions about them. We map each agent against it in the audit so the design is compliant rather than retrofitted.

If you need calls answered, appointments booked and messages taken, buy an off-the-shelf AI receptionist. IONOS and Fasthosts sell one at £39 to £99 a month, Hey Jodie has unlimited minutes from £49, ARROW is built for trades, BookedSolid for clinics. They work, and building one bespoke would be a waste of your money. We have surveyed thirteen of them, with prices, in the guide below, and we will tell you which one to buy.

Build bespoke when the conversation has to do something specific to your business: screen an applicant against a role, qualify a lead against your criteria, take a booking that depends on your rota or your own systems, or call out as well as answer. That is the line, and the guide draws it with the running costs worked out to the penny.