AI agent costs could rise fivefold by 2028 - what UK businesses should do now

AI agents have quickly moved from impressive demos to boardroom ambition. The pitch is simple: give an AI system a goal, connect it to tools and let it complete useful work with less human hand-holding.

But the economics are becoming harder to ignore. Gartner predicts that AI inference costs per agentic workflow will increase more than fivefold through 2028, as more powerful models and more sophisticated workflows consume more tokens per completed task. Gartner's forecast is not saying every company's total AI bill will automatically rise by five times. It is saying the cost of running an individual agentic workflow could become much higher.

That distinction matters. For UK businesses, this is not a reason to pause AI adoption. It is a reason to stop treating agentic AI as a cheap chatbot with a better job title.

What AI inference costs actually mean

Inference is the cost of running a trained AI model to produce an answer or take part in a workflow. In plain English, it is the meter running every time your system sends information to a model and receives output back.

With a simple chatbot, the cost is usually tied to one user prompt and one model response. With an AI agent, the same user request might involve planning, searching, calling business tools, reading documents, checking its own output, retrying a failed step and then producing the final answer.

That is why agent costs can rise even when model prices per token improve. The unit of work changes from a single answer to a chain of model calls and tool interactions.

Why AI agents can cost more than chatbots

A chatbot is often a question-and-answer interface. An agent is closer to a workflow engine powered by one or more AI models.

Research on agent infrastructure describes this shift from static, single-turn inference to multi-turn agentic workflows. It also notes that design choices such as reflection depth, few-shot prompting and parallel reasoning affect the trade-off between accuracy and cost. The infrastructure paper is useful because it frames agents as systems, not just smarter text boxes.

In practice, a single agent task might include:

• Breaking a task into sub-tasks.
• Calling a search, database, email, CRM or file tool.
• Sending retrieved content back into the model.
• Generating intermediate reasoning or draft outputs.
• Validating whether the work is good enough.
• Retrying when something fails.

Each of those steps can add input tokens, output tokens, reasoning tokens, tool costs and latency. The user may only see one final answer, but the system may have done several expensive things behind the scenes.

The hidden cost trap is cost per successful outcome

The most important metric is not cost per prompt. It is cost per successful outcome.

Suppose an agent handles supplier queries, drafts customer responses or reviews internal policies. A cheap run that produces unusable work is not cheap. It is waste with a nicer interface.

A practical cost model looks like this:

Cost per useful task = model calls + input tokens + output tokens + reasoning tokens + tool fees + retries, divided by successful outcomes.

This is where many pilots look better than production. In a demo, the happy path is controlled. In a live UK business, the agent meets messy documents, vague requests, permission issues, duplicated data, edge cases and human review requirements.

If the agent keeps looping, overusing tools or escalating everything to a more expensive model, the bill can rise without a matching productivity gain.

Why this matters for UK organisations

UK companies are often under pressure to do more with less: reduce admin, improve customer service, speed up internal processes and help teams work through document-heavy tasks. AI agents can genuinely help with that.

But businesses also have to manage procurement, data protection, internal governance, sector-specific compliance and operational risk. An uncontrolled AI agent is not just a cost problem. It can become a process problem.

If an agent touches customer data, employee records, commercial documents or regulated workflows, UK teams need clear decisions about what data is sent to which model, where logs are stored, who can access outputs and what level of human approval is required. That is before you get to the cost of repeated inference.

This links directly to a broader point I have made before: AI is not automatically cheaper than people once you include compute, management, review and failure costs. If you are weighing the economics, my earlier piece on Nvidia's compute cost reality for UK businesses is a useful companion read.

How to control AI agent costs before they scale

1. Route tasks to the cheapest model that is good enough

Not every task needs the strongest reasoning model. Simple classification, summarisation, extraction and rewriting should usually go to a cheaper model if quality is acceptable.

Save expensive models for tasks where better reasoning changes the business outcome. That might include complex policy interpretation, multi-document analysis or high-value sales and technical work.

2. Cache repeated context

Agents often resend the same system prompt, policy text, tool descriptions and schema information. If your provider supports caching, repeated context may be cheaper than sending everything from scratch each time.

Even when caching details differ by provider, the principle is the same: keep stable instructions stable, avoid rewriting boilerplate unnecessarily and track whether repeated context is actually being reused.

3. Make prompts shorter and more precise

Long prompts feel safer, but they are not free. Repeated instructions, excessive examples and irrelevant tool descriptions increase context size and can encourage the agent to overthink simple jobs.

Good agent design is not about stuffing every policy into every call. It is about exposing the right information at the right time.

4. Limit tools, retries and reasoning depth

An agent with access to every tool is not necessarily more capable. It may simply have more ways to spend money.

Set practical limits: maximum tool calls, maximum retries, maximum reasoning effort and clear stop conditions. If the agent cannot complete the task within those boundaries, route it to a human or a different workflow.

5. Monitor intermediate usage, not just the final answer

Pricing can include underlying token consumption and tool usage, including intermediate steps in agent loops. Google describes agent usage in terms of token and tool consumption within the underlying workflow, which is a useful reminder that the visible response is only part of the bill. Gemini API pricing sets out this kind of usage-based framing.

Your dashboards should show model calls, token use, tool calls, failed runs, retries and cost per successful workflow. A monthly invoice is too late to discover that a badly configured agent has been doing expensive busywork.

A practical 30-day AI agent cost audit

If your organisation is already piloting agents, run a short audit before expanding access. It does not need to be complicated.

1. Pick one agentic workflow with real business value.
2. Track every model call, tool call, retry and failed run for 30 days.
3. Measure cost per successful task, not just total API spend.
4. Identify which prompts, tools or documents create the most token volume.
5. Test cheaper model routing for low-risk steps.
6. Add hard limits for retries, tool calls and monthly spend.
7. Review whether human approval is being used at the right points.

This is the difference between AI experimentation and AI operations. The former proves something can work. The latter proves it can work reliably, affordably and safely.

If you are still separating hype from useful deployment, my article on agentic AI hype versus reality for UK businesses explores the adoption gap in more detail.

Treat agentic AI as a product, not a magic labour saving

The fivefold forecast should not scare UK businesses away from agents. It should make them more disciplined.

AI agents can be valuable when they are pointed at well-defined workflows, measured against useful outcomes and governed like any other production system. They become risky when companies assume autonomy means low cost, low oversight and unlimited scale.

The winners will not be the firms that connect every process to the most powerful model available. They will be the ones that design lean workflows, use the right model for the job, cap unnecessary agent behaviour and measure value in pounds saved or earned per successful task.

In other words: do not ask whether agents are expensive. Ask whether each agentic workflow is worth what it costs to complete.