AI agent development in London
AI agents are the most over-sold idea in business software right now, and also genuinely useful for a narrow set of jobs. This page explains which is which, so you can decide whether one is worth building for your business before anyone quotes you for it.
What an AI agent actually is
An AI agent is software that takes a goal, works out the steps, and carries them out using your other tools, instead of waiting for someone to click through each step by hand.
A useful way to picture it: ordinary automation is a set of instructions, and an agent is a very fast junior assistant who can read, make a judgement call, and use your systems. The assistant is quick and never gets bored. It also occasionally gets things confidently wrong, which is the whole reason the design around it matters more than the AI itself.
Where AI agents genuinely work
Agents are strong when the input is messy and the cost of a mistake is low or easy to catch.
- Reading things that arrive in a hundred different formats. Supplier invoices, client emails, forms, PDFs, scanned documents. This is the single best use, because it is exactly what rules-based software has always been terrible at.
- Sorting and routing. Working out what an incoming enquiry is about and putting it in front of the right person with the context attached.
- Drafting the routine reply. Not sending it. Drafting it, so a person approves and sends in seconds instead of writing from scratch.
- Pulling scattered information together. Gathering everything about one client or job from four systems into one summary.
- Flagging the exceptions. Reading through everything and surfacing only the handful of items that need a human to look.
Where they fail, and what that costs
This is the part most sales conversations skip.
- Anything where one silent mistake is expensive. Paying an invoice, sending a client a figure, filing something with a deadline. An agent can prepare these. A person should approve them.
- Long chains of steps. Reliability compounds downwards. An agent that is right nine times out of ten at each step is not right nine times out of ten across eight steps.
- Work that already follows clear rules. If your process is genuinely predictable, ordinary automation will do it cheaper, faster, and without the running cost.
- Anything needing real accountability. If a regulator or a client will ask who decided something, that needs to be a person, with the agent's working shown.
None of this means agents are not worth building. It means the value is in the boring engineering around them: what the agent is allowed to do on its own, where a human signs off, what happens when it is unsure, and how you can see what it did.
How I build them
- Start with the reading, not the deciding. The first version handles the messy input and hands a clean result to your team. Value arrives early and nothing can go badly wrong.
- Put a human in the loop where it counts. Anything that costs money, goes to a client, or carries a deadline gets approved by a person until it has earned trust.
- Make everything visible. Your team can see what the agent did and why, and can override it. A system nobody can inspect is a system nobody trusts.
- Use rules wherever rules work. The cheapest, most reliable part of any system is the part with no AI in it.
- Build it into the tools you already have. Your team should not have to learn a new place to work.
Why work with me on this
I build and run real software rather than only advising on it, which matters here more than in most areas, because the difference between an agent demo and an agent your team relies on is almost entirely in the engineering around it.
- DSPOps, an operations platform for Amazon delivery partners, including an AI check that scans van photos and automatically flags damage. Live and in daily use.
- Estate Revenue Manager, a staff rota and revenue management system for high-end residential property firms.
- A case management system for a UK immigration law firm, replacing paper files and shared inboxes with one place to work. In build.
To be straight with you: of those, the DSPOps damage check is the one that uses AI, and it is a focused, narrow job of exactly the kind described above. It reads images that vary every time and flags the ones a human should look at. That is the pattern that works, and it is the pattern I would build for you rather than promising an agent that runs your business on its own.
Find out whether an agent is even the right answer
A free 30-minute call. Tell me what your team is drowning in and I will tell you honestly whether an AI agent, ordinary automation, or better-built software is the right fix. Sometimes the answer is that you do not need AI at all.
Book a free workflow auditCommon questions
What is an AI agent?
An AI agent is software that can take a goal, work out the steps, and carry them out using other tools, rather than waiting for a person to click through each step. In a business setting that usually means reading incoming information, deciding what it is, doing something with it in your existing systems, and asking a human when it is unsure.
What is the difference between an AI agent and automation?
Traditional automation follows fixed rules you define in advance: if this happens, do that. An AI agent can handle work where the rules are fuzzy, like reading a supplier email that arrives in a slightly different format every time. Most real systems use both, with rules doing the predictable parts because they are cheaper and more reliable, and AI handling the messy parts.
Are AI agents reliable enough to use in a real business?
For some jobs, yes. For others, not yet. Agents are strong at reading unstructured information, classifying it and drafting a response. They are weaker where a single mistake is expensive and hard to spot. The practical answer is to give the agent the messy reading and sorting work, keep a human approving anything that costs money or goes to a client, and expand the agent's remit only once it has earned it.
How much does it cost to build an AI agent?
It depends entirely on how many systems it needs to touch and how much can go wrong if it gets something wrong. I quote a fixed price per project after the free workflow audit, once I understand the work involved. There is also a small running cost for the AI itself, which I will estimate honestly up front rather than leaving it as a surprise.
Do I need an AI agent, or would simple automation do?
A lot of the time, simple automation does the job for less money and breaks less often. If the work follows predictable rules, rules are the better tool. AI earns its place when the input is genuinely messy and varies every time. I will tell you which one you actually need, including when the answer is that you do not need AI at all.