How businesses use AI agents

If you are reading this you are probably past the question of whether to use AI agents and onto the harder one of how, specifically, and what is realistic. That is the gap worth naming at the outset.

More than half of UK firms now use AI in some form, but only around one in six have actually deployed an agent with a defined business purpose.

Most of the rest are dabbling with tools rather than putting an agent to work on a real task, and the distance between those two states is exactly what this guide is about.

Using an agent in a business does not mean buying a subscription; it means handing a bounded, repeatable workflow to a system that observes, decides and acts, and then checking that it holds up.

Written for an owner or operations lead in a UK small or midsized company who wants a concrete picture of how agents are actually deployed, department by department, with realistic budgets and an honest account of what gets in the way. It is not written for engineers.

It sits under our pillar guide, the complete guide to AI agents, and pairs with AI agent use cases if you want the tasks themselves catalogued rather than the deployment.

Our position, so you can weigh what follows: we build and then maintain these systems for clients, which means we see the difference between a demo and a deployment every week.

So this page is candid about where a platform tool is the right starting point and where a bespoke build genuinely earns its place.

What does it mean to use an AI agent in your business?

To use an AI agent is to give a system a workflow it can run end to end, observing an input, deciding what to do, acting on it, and reporting back, without a person directing each step. That end to end quality is what separates an agent from a chatbot.

A chatbot answers when spoken to and then stops; an agent takes a task and carries it through, calling your other software as it goes.

In practice that means an agent is only useful where there is a defined workflow to give it, with a clear trigger, a clear finish, and a way to tell whether it did the job.

Businesses that get value from agents almost always start by picking one such workflow and scoping it tightly, rather than adopting AI as a general capability and hoping a use emerges.

The government’s AI Security Institute has published an analysis of how agents are used in practice, drawn from a very large sample of real agent interactions, in its work on how AI agents are used, which is a rare grounded look at real deployment rather than vendor claims.

How businesses use AI agents, department by department

The clearest way to picture deployment is to walk through where it lands in an organisation. Each department below gets one realistic entry workflow and a note on the route to standing it up.

Customer service and support is usually the first place an agent lands, because the work is high volume and repetitive.

The entry workflow is handling routine incoming queries and escalating the rest, and the realistic route is to connect the agent to your existing helpdesk and knowledge base so it answers from your own documented material.

Sales and lead qualification is the next common landing spot. The entry workflow is triaging inbound enquiries, enriching them with public context and scoring them against your criteria so a person only spends time on the ones worth a call.

The route is to wire the agent into your CRM so its output lands where the sales team already works.

Finance and operations is where agents earn their place on document heavy work. The entry workflow is reading invoices, statements or orders that arrive in varied formats, extracting the figures, and posting them for a person to approve.

The route runs through your accounting system, with a value threshold above which everything is held for human sign off.

HR and internal knowledge is a quieter but reliable deployment. The entry workflow is answering staff questions about policy and process from your internal documents, which deflects a stream of repetitive questions away from a small HR team.

The route is to point the agent at a controlled set of your own current documents and nothing else.

Legal and compliance is the most oversight heavy deployment and the one to approach last. The entry workflow is surfacing relevant clauses or precedents and producing a first draft for a qualified person to review.

The route keeps a human firmly in the loop on every output, because the cost of an unchecked error here is high.

The pattern across all five is the same: a bounded workflow, connected to the system the work already lives in, with human oversight scaled to the stakes.

Platform built or bespoke, which route fits your business?

Once you have picked a workflow, the real decision is how to build it, and here the honest answer depends on the shape of the work rather than on what any vendor sells.

For a bounded, low complexity workflow sitting on tidy data, a no code or low code platform is the fastest and cheapest way to start, and often all you need; there is no virtue in a custom build when an off the shelf tool does the job.

A bespoke build earns its place in the opposite case: when the process is genuinely complex, when it depends on proprietary data or logic a generic tool cannot hold, or when the integration required across your systems exceeds what a platform can reach.

That last case is our own wedge, and we will say plainly when it does not apply to you.

The mistake to avoid is reaching for a custom build out of ambition when a platform would do, and the opposite mistake of forcing a complex, deeply integrated process onto a platform that cannot really handle it. Match the route to the workflow, not to the marketing.

The broader discipline of designing these processes is covered in our complete workflow automation guide.

What stops businesses from getting results with AI agents?

The reason so many firms have adopted agents but so few run them in production comes down to a small number of barriers, and naming them is more useful than pretending they do not exist.

The first and largest is the skills gap: a majority of UK businesses cite a shortage of the in house capability needed to build, deploy and maintain agents as their main blocker, which is why so many pilots stall after the demo.

The second is what gets called dark data, the fragmented, unstructured, inconsistent internal data that agents choke on; an agent reasoning over a mess produces confident nonsense, and most firms discover the state of their data only when they try to automate on top of it.

The third, quieter barrier is scoping: workflows defined too broadly never quite work, because the agent is asked to handle too many exceptions to be reliable.

The path through all three is the same and it is unglamorous: start narrow, get the data behind one workflow into decent shape first, and either build the skills or bring in help to stand it up and maintain it.

The firms that clear these barriers are not the ones with the biggest budgets; they are the ones that scoped smallest.

How to measure whether your AI agent is working

Deployment is not the finish line; knowing whether the thing works is, and measurement is the part almost every guide omits. The right metric depends on the workflow. For a support agent, track the resolution rate and how much of the queue it clears without a person.

For a sales agent, track pipeline velocity and the quality of the leads it passes through. For an operations agent, track processing time against the manual baseline and the error rate on what it posts.

For an internal helpdesk agent, track the deflection rate, the share of questions answered without a human.

In every case the discipline is to fix the before state in numbers first, so you have something honest to compare against, and to keep watching after launch, because an agent’s inputs drift and its performance with them.

An agent that is not measured is an act of faith; one that is measured is a decision you can defend and improve.

A realistic plan for a first AI agent

The route that works for a smaller business is short and deliberately narrow. Scope one bounded, repetitive workflow, the kind with a clear trigger and a clear finish.

Audit the data that workflow depends on and get it into decent shape before anything else, because that is where most deployments quietly fail.

Choose the route honestly, a platform tool if the workflow is simple and self contained, a bespoke build if it is complex or spans several of your systems. Define the one metric that will tell you it is working, and keep a person reviewing consequential output while trust builds.

Prove that single agent against its baseline before you extend to a second.

Where the process is genuinely complex or integration heavy, a build shaped around your actual stack pays back faster than bending a generic tool to fit, which is the point at which our kind of work becomes relevant.