For two years, the enterprise conversation about artificial intelligence was dominated by demos: a model summarising a document, drafting an email, answering a question. The new conversation is about completion. Can the system take a task from request to done, with a log of every step and a human able to intervene?

That shift is smaller than the hype suggests and more consequential than the demos. Agents are not general-purpose workers. They are narrow pieces of software wrapped around a language model, a set of tools and a policy layer that decides what they are allowed to touch.

Key facts

  • Most large enterprise deployments are in support, claims, procurement and reconciliation.
  • Typical pilots automate 15% to 35% of a single process, not an entire department.
  • Cost per completed task varies by more than a factor of ten between vendors.
  • Human exception handling remains the largest hidden cost.
  • Audit logging is now a standard procurement requirement.
  • Few deployments have reduced headcount; most have absorbed rising volumes.

What actually changed

The technical change is unglamorous. Models became better at using tools, following multi-step instructions and recovering from errors. Context windows grew. Costs per token fell. None of that makes an agent reliable on its own, but together they make a narrow workflow automatable in a way it was not eighteen months ago.

The commercial change matters more. Vendors stopped selling a model and started selling an outcome: claims processed per day, tickets resolved without escalation, invoices matched to purchase orders.

Server racks in a data centre with blue indicator lights
Most enterprise agents run in the cloud, where costs scale with every step they take.

Where it works

The successful deployments share three features. The process is repetitive. The inputs are structured or semi-structured. And the cost of a mistake is recoverable.

Insurance claims triage is a good example. An agent can read a claim, check it against policy data, request missing documents and route it to the right adjuster. A wrong decision is caught by the adjuster before payment. Customer support is another: agents handle password resets and order status queries, then hand off anything involving money or complaints.

"The model is maybe 20% of the work. The other 80% is deciding what the system is not allowed to do."Head of automation at a mid-sized insurer

Where it fails

Failures cluster around the same edges. Long-tail cases, ambiguous instructions, systems without clean APIs, and anything that requires accountability rather than throughput.

The most expensive failure is silent. An agent that confidently completes a task incorrectly is worse than one that stops. Teams that deploy without exception logging discover the problem weeks later, when a customer or an auditor finds it for them.

Customer service team working with headsets in a modern office
Early deployments cluster in support, claims and back-office reconciliation.

The cost model is still unsettled

Agent pricing is a mess. Some vendors charge per seat, some per task, some per token, and some take a share of the savings. The result is that two seemingly similar proposals can differ by an order of magnitude once volume scales.

Buyers are learning to ask for cost per completed task, including retries and human review. That number is usually three to five times higher than the pilot estimate, because pilots run on clean data and motivated teams.

Close-up of hands typing on a laptop beside a notebook with process diagrams
The hard part is not the model. It is mapping the process precisely enough to automate.

The labour question

Despite the rhetoric, few deployments have produced layoffs. Most have absorbed rising volumes with the same headcount, or redeployed staff to exception handling and quality control. That is a real productivity gain, but it is not the same as replacement.

The roles that change fastest are middle-skill processing jobs. The roles that grow are oversight, process design and data quality. Both trends favour workers who understand the underlying workflow, not just the tool.

That has implications for the shorter-week experiments now spreading through service industries, where automation is often cited as the productivity source that makes fewer hours feasible.

Governance is the bottleneck

Procurement teams now ask for audit logs, human override, data retention rules and a clear statement of model limitations. Regulators are moving in the same direction. The NIST AI Risk Management Framework has become a common reference point even outside the United States.

None of this is exotic. It is the ordinary work of making software accountable. The difference is that language models introduce variability that traditional software did not, and organisations are still learning how to test for it.

Manager reviewing a printed exception report with a colleague
Exception handling remains the largest hidden cost in agent deployments.

The next stage

The most interesting deployments are not single agents but small teams of them, each with a narrow role and a shared log. One reads, one checks, one drafts, one escalates. The architecture looks less like a chatbot and more like a back-office process that happens to be written in software.

That is also where the risk concentrates. When several systems act in sequence, a small error can compound before a human sees it. The firms doing this well are the ones treating the whole chain as a regulated process, not a clever demo.