AI Agents Stopped Demoing and Started Working

An AI agent is software that is given a goal instead of a script. You say “chase this invoice”, and it decides which steps to take, uses the tools it has, and reports back. That idea is old. Making it survive a Tuesday afternoon in a real company is new.

For most of 2024 and 2025, agents were a conference slide. They worked in the demo, then met a form that had changed, an API that timed out, or an instruction that was slightly ambiguous, and fell apart. In 2026 something shifted. Not a breakthrough moment — a slow, unglamorous one, visible in finance departments and support queues before it was visible anywhere else.

What actually changed

Three things, none of them cinematic.

Scope got smaller. The teams getting results stopped asking an agent to “run sales” and started asking it to do one repeatable task with a clear finish line. Match this receipt to this transaction. Draft this reply. Chase this document.

Checkpoints got added. The successful setups look less like a robot employee and more like an assembly line with inspection points. The agent does five steps, a human approves, it does five more. Errors get caught at step six instead of step sixty.

Someone owns the output. This is the part nobody puts in a keynote. Where agents work, a named human is responsible for what the agent produced. That single organisational change does more for reliability than any model upgrade.

The companies winning with agents did not buy better AI. They wrote better job descriptions for it.

Where they are genuinely working

Finance operations. A FloQast industry report published on 11 August 2026 found that AI-native accounting and finance teams cut manual work by nearly half and closed their books about two days faster. Month-end close is a perfect agent task: repetitive, rule-heavy, well-documented, and wrong answers are catchable.

Support and outbound. The software newsletter SaaStr has described running eight-figure revenue with single-digit headcount by handing agents outbound, support triage, content and revenue operations — reporting an AI sales development agent that sent roughly 15,000 messages in 100 days at a 5–7% response rate. Read that carefully: the response rate is normal for cold outbound. The change is that it took almost no human hours.

Phone calls. Voice was the surprise of the year. Agents that call a shop to check stock, or sit on hold with an airline, moved from novelty to product. Roughly $1.8 billion went into AI agent startups in July 2026 alone across a dozen-plus deals, with average valuations up about 40% quarter on quarter — a lot of it chasing voice.

Where they still fall over

Agents remain bad at anything where the right answer depends on context nobody wrote down. Judgement calls. Negotiations. Situations where the rule and the sensible thing disagree.

They also degrade over long chains. Every step has a small chance of going wrong, and those chances multiply. Twenty steps at 98% accuracy each is a coin flip. This is why checkpoints matter more than model choice.

And they are confidently wrong. A human who does not know usually hesitates. An agent produces a clean, well-formatted, completely incorrect answer in the same tone as a correct one. If your process has no verification step, you will not notice.

The India angle

India runs an enormous share of the world’s finance operations, support desks and back-office processing. Those are precisely the workflows agents are getting good at, which makes this less abstract here than almost anywhere else.

The realistic near-term shape is not mass replacement. It is that the entry-level rung — the job where you learned by doing the repetitive task — gets thinner, while the review, exception-handling and client-facing rungs get busier. That is good news for people who are already two years in and a genuine problem for people trying to get in.

How to tell hype from reality

When someone tells you their agent works, ask three questions:

• What exactly does it do, start to finish? A vague answer means a demo.
• Who checks it, and how often is it wrong? No one has measured this if they cannot answer.
• What happens when it fails? Real deployments have a fallback. Demos have an apology.