An AI agent is a job description, not a strategy

Octopus Energy put an AI agent called Arlo on its customer emails for three months, and published how it went. By Octopus's own figures, Arlo scored 76% for customer satisfaction, against 72% for human advisers. It did that by doing one thing. It handled about 8,000 emails a week, around 4% of the total, and it "never handles vulnerable customers, sensitive cases or complex complaints". The narrowness is why it worked. On 8 October, Google pitched the opposite: "your new single, universal agent for work". For a 50-350 person company, an agent is worth buying when it is scoped like a hire for one job. Bought as a strategy, it rarely is.

What an agent actually is

An agent is software that decides its own next step and uses your systems to finish a task. Anthropic defines agents as "systems where LLMs dynamically direct their own processes and tool usage". The UK government is plainer: agents are "small, specialised pieces of software that can make decisions". Read either definition and you find a task, some tools and the freedom to choose between them. You do not find a colleague.

The vendors prefer the colleague. OpenAI said its Frontier platform was "designed to work the same way companies manage human employees". That language makes an agent harder to scope, not easier. A new member of staff given a vague role can work out what the job is. An agent given a vague role works out something, and acts on it in your systems.

An agent is a process with permissions, not a member of staff.

Why agent projects get cancelled

In June 2025, Gartner forecast that over 40% of agentic AI projects will be cancelled by the end of 2027, citing "escalating costs, unclear business value or inadequate risk controls". Its analyst Anushree Verma added that "many use cases being marketed as agent-based today do not, in fact, require agent-based implementations". In Confluent's own survey of 4,625 IT leaders, 77% of those running agentic AI in production reported stalled projects.

Unclear value is rarely a technology problem. It usually means the project started with no named job, no number taken before it began, and no one person who owns that number. When the review comes, the cost is the only figure on the table.

Nobody cancels a project that is beating its number.

Vague jobs are also where agents are weakest, because their errors compound. On CRMArena-Pro, a benchmark of business tasks in a CRM, leading agents succeeded about 58% of the time on single-turn tasks, and about 35% once the task ran over several turns. Every extra step an agent decides for itself is another chance to go wrong. A narrow job keeps the steps few.

A good first agent at 150 people

The previous piece in this series argued that the money sits in specific work between your systems, found and costed before anything is bought. That costing is where an agent's job description comes from.

A good first agent does one process. That means a supplier invoice matched against a purchase order, or an order email turned into a line in the ERP, not "operations". The process is measured before go-live and after it, on the same terms: hours, errors, turnaround. One person owns the number, and answers for it. When the agent gets something wrong, the mistake is cheap and somebody sees it: a draft held for approval, not a payment sent. Anything outside the job goes to a human, as at Octopus.

Anthropic says customer support suits agents partly because "success can be clearly measured". METR, which measures what AI can do, warns that some tasks "require 98%+ success probabilities to be worth automating".

Almost all of this evidence comes from large firms and benchmarks, not from a company of 150. The only number that will tell you whether your agent works is the one you take before it starts.

If you cannot write the job description, you cannot hire the agent.

Where agents work today

Agents do work, in the right place. Marc Benioff said Salesforce cut its support staff from 9,000 heads to about 5,000, with half of customer interactions now handled by agents, by his own account. Tata Steel, according to Google, which sells it the technology, deployed more than 300 specialised agents in nine months and cut complaint turnaround by 50%.

Look at where those wins sit: support queues and complaints, where there is a measure. Even Tata Steel's success, in Google's own telling, is more than 300 specialised agents, not one universal one. Each of them has a job.

Klarna's own story moved. In 2024 it said its assistant handled two-thirds of customer service chats in its first month. In 2025 its chief executive said "investing in the quality of human support is the way of the future for us".

And not every job needs an agent. Anthropic's own advice is to find the simplest solution, and "this might mean not building agentic systems at all". In one study of customer support staff, an AI assistant that only suggested replies raised productivity by 14%, and by 34% for novices. It was not an agent.

Is it worth doing? Yes, one named job at a time. Take the baseline before go-live, and give one person the number. Not as a programme, not as a platform, and not as a headcount plan.

Hire the agent the way you would hire anyone: for one job, against one number, and let it earn the second.