When an off-the-shelf AI tool isn't enough — and what a custom agent costs to build and run

If you run a small business, there is a fair chance you have asked yourself this: "We've got Copilot (or Gemini, or ChatGPT). Is that enough, or do we need something built?"

It's the right question. The tools you already pay for have become good at a lot of everyday work, and in Layer 4 of the getting-started series I argued that most businesses should switch on what they already have before buying anything new. Often that's the whole answer.

But not always. Some jobs sit just outside what an off-the-shelf assistant can do, and no amount of clever prompting will get it there. This guide is about telling the two apart, and what it costs when you do need something built.

Start with what you already have

An off-the-shelf assistant (Copilot in Microsoft 365, Gemini in Google Workspace, ChatGPT or Claude with their connectors) is usually enough when the job looks like this:

  • a person asks, the AI answers. Someone opens the assistant, asks a question, and reads or edits the reply;
  • the information lives in one place it can already reach, such as your shared drive, your mailbox or your calendar;
  • the output is a draft for a person to use: a summary, a reply, a first version of a document;
  • an occasional wrong answer is cheap, because a person reads it before anything happens.

If your job fits that shape, a well set-up assistant will probably serve you better than anything custom. It's cheaper, it improves without you paying for it, and your team already knows where to find it. This is what the Setup Day is for: £1,500 fixed for a business of up to ten people (£1,000 if it follows an audit) to get the tools you already pay for properly pointed at how your business works.

Five signs the job needs something built

The picture changes when the job has one or more of these features.

1. Nobody should have to remember to ask

An assistant waits for someone to open it. Plenty of useful jobs don't start with a person typing a question. They start with an email arriving, a form being submitted, a delivery date passing or a spreadsheet changing. If the work should happen because something happened, rather than because someone thought to ask, you need something that runs on its own.

2. It runs across several of your own systems

Off-the-shelf assistants are best inside their own suite. A job that reads from your booking system, checks a supplier's price list, looks up the customer in your accounts software and then drafts something in your inbox crosses four systems that were never designed to talk to each other. Connecting those reliably is a build, not a setting.

3. It needs to write back, not just read

Reading your documents is the easy part. Updating a record, creating an invoice, moving a job to the next stage or booking an appointment is different: a mistake changes something real. That needs proper permissions, checks before anything is written, and a clear record of what was done and why.

4. The data needs more care than a general tool should have

Some information shouldn't be handed to a general-purpose assistant with broad access: client case files, health information, financial records, anything covered by a contract that limits where it can go. A custom build can be given access to exactly what the job needs and nothing else.

5. It has to be right in a specific, checkable way

"Usually helpful" is fine for a draft email. It isn't fine for a quote that has to use this month's prices, or a reply that must follow your terms exactly. When the result has to meet a known standard, the agent needs to be built and tested against real examples from your business, and watched once it's live.

If none of these apply, stick with what you have. If one or two do, it's worth a conversation.

What a custom AI agent actually is

The phrase sounds grander than it is. A useful custom agent is a focused piece of software built around one recurring job. In practice it has four parts:

  1. A trigger: what starts it, such as a new enquiry, a daily time or a change in a system.
  2. Access to the right information: connections to the specific systems and documents the job needs, with permissions limited to those.
  3. The work itself: the AI step that reads, decides or drafts, wrapped in ordinary code that checks its output.
  4. A hand-off: where the result goes, and where a person approves it before anything important happens.

It is not a chatbot you have to remember to open. When it works well, your team mostly notices that a job they used to do has started getting done.

Sometimes the right answer turns out not to be an agent at all. A plain piece of software or a tidier process does the job more cheaply and more reliably. The job decides, not the technology.

What it costs to build

At Simon Studios, a custom business build starts from £2,500.

That starting point covers one recurring job across your own systems, for one user or a small team. The price goes up with the number of systems involved, how many people use it, how much it writes back, and whether regulated data is involved. Every build is scoped to the job before work starts, so you agree a figure rather than watching an hourly meter.

Most builds start with an AI Opportunity Audit (£950 fixed). The audit works out which job is worth building first, and whether a tool you already pay for would do it instead. If you then go ahead with the build, the full £950 comes off the price.

What it costs to keep running

This is the part most people don't budget for, and it's where custom AI projects usually go wrong.

An agent that's built and handed over starts decaying almost straight away. The AI models it relies on are updated or retired. The documents and systems it reads change shape. Edge cases nobody thought of start turning up. Usage costs creep upward. Six months later it's either quietly wrong or quietly switched off.

So I don't hand agents over and walk away. Running costs from £150 a month, scaled to the complexity of the job and how much it's used. That covers:

  • hosting, so there are no servers for you to manage;
  • the AI usage itself, metered and managed inside the subscription, so there's no surprise bill at the end of the month;
  • tuning against what actually happens once it's live;
  • keeping up with model changes, so an update upstream doesn't break your job;
  • fixing whatever breaks, and support when something looks odd.

You're paying for a job that stays done, not for a piece of code you'd then have to look after yourself.

A rough way to decide

Before talking to anyone, try writing the job down in one sentence:

"When [this happens], use [this information] to [do this], and [this person] checks it before [this result]."

If you can fill that in, you have a buildable job. Then ask:

  1. How often does it happen? Daily or weekly jobs pay back. Twice-a-year jobs rarely do.
  2. What does it cost you now? In hours, in delays, or in mistakes that slip through.
  3. Could your existing tools do it? If the answer is "yes, if someone remembered to ask", try that first.
  4. What happens if it gets it wrong? The bigger the consequence, the more checks it needs, and the more it costs to build properly.

A job that happens often, costs real time, sits across several systems and can be checked by a person before anything final happens is a good first build. A job that's rare, vague or high-stakes with no room for review usually isn't, at least not yet.

The simplest next step

If you have a job in mind, book a free intro call and bring that one sentence. We can usually tell quickly whether it's a Setup Day, a custom build, or something that's better left alone.

If you're not sure which job to start with, the AI Opportunity Audit is designed to answer exactly that.

I'm Simon. I run Simon Studios, an AI-first development studio in Teddington. I help small businesses find the useful, manageable AI project first, then build and run it. No jargon, no hype.