AI agents for productivity: where they save time, and where they quietly cost it

The short version

• An AI agent differs from a chatbot because it takes actions in other systems, such as filing, updating or sending, rather than only producing text.
• AI agents save the most time on repetitive tasks where checking the output is much faster than producing it and mistakes are cheap to undo.
• In METR's 2025 randomised trial, experienced open-source developers took 19 per cent longer with AI tools while believing AI had made them faster.
• In a study of 5,179 customer support agents, a generative AI assistant helped less experienced agents most and the most experienced agents least.
• A first agent should draft work for a person to approve, with the narrowest permissions the task needs.

What is an AI agent for productivity?

An AI agent for productivity is software that takes a goal, works out the steps and carries them out in your other tools: reading an inbox, updating a CRM record, filing a document, preparing a reply for approval. The difference from a chatbot is that it acts. The difference from ordinary automation is that it can cope with input that does not arrive in a fixed shape.

A rule-based automation says 'when a form is submitted, add a row to the spreadsheet'. It is predictable, and it breaks the moment the input varies. An agent can read a messy email, decide it is a refund request rather than a complaint, look up the order and draft the response. That flexibility is the whole value. It is also the whole risk, because the judgement that handles variation can also get it wrong.

Many products sold as agents are automations with one model step in the middle. That is not a criticism. For most productivity work it is the better design: the fixed steps stay predictable, and only the genuinely variable step relies on a model.

Which tasks are AI agents good at?

AI agents are good at tasks that repeat often, follow a recognisable pattern, produce output a person can check quickly and can be undone if wrong. Inbox triage, turning meeting notes into CRM updates, extracting fields from invoices and chasing overdue follow-ups all fit. The common thread is that checking is much faster than doing.

That last condition is the one teams skip. If reviewing the agent's work takes as long as doing it, you have not saved time, you have moved it. A drafted reply that takes a minute to read and replaces ten minutes of writing is a gain. A drafted contract clause that a lawyer must check word by word is a small gain at best.

Volume matters as well. An agent that saves a few minutes on a task done twice a week will struggle to repay the cost of building, testing and maintaining it. One that saves a little on a task done hundreds of times a day often will.

Which tasks suit an agent, and which do not?

Where do AI agents fail to improve productivity?

AI agents fail to improve productivity where checking the output costs as much as producing it, where mistakes are expensive or irreversible, and where an expert already does the work quickly. In those cases the time goes into prompting, waiting, reviewing and correcting, and the total can exceed doing it by hand.

The clearest example comes from a 2025 randomised controlled trial by METR. Sixteen experienced open-source developers worked on 246 real issues in their own repositories, with each issue randomly assigned to allow or disallow AI tools. When AI was allowed, they took 19 per cent longer. Afterwards, they still estimated that AI had made them about 20 per cent faster.

The result is narrow. These were experts working in codebases they knew well, using early-2025 tools. It does not show that AI slows everyone down. What it does show is that people doing the work may not be able to tell, from the inside, whether a tool is helping them. That is a measurement problem, and it applies to agents in any role.

What does the evidence say about AI and productivity?

The evidence says the gains are real, modest on average and uneven across people and tasks. Most of it studies generative AI assistants that help a person, not autonomous agents that act alone. Treat it as the case for AI helping with a task, not proof that handing the task over works.

A nationally representative survey of US workers by St. Louis Fed economists, fielded in November 2024, found that generative AI users reported saving 5.4 per cent of their work hours. Averaged across all workers, including those who did not use it, the saving was 1.4 per cent of total work hours. Both figures are self-reported.

A study of a generative AI assistant rolled out in stages to 5,179 customer support agents at one firm found that issues resolved per hour rose. The largest gains went to less experienced agents. The most experienced saw little effect. The tool suggested responses, and people still decided what to send.

Adoption is uneven too. The US Census Bureau's Business Trends and Outlook Survey found that firms with at least 20 employees were the biggest AI users. Smaller businesses are behind on that measure, which is not necessarily a problem. It means the first project can be chosen carefully rather than in a hurry.

How do you measure whether an AI agent saves time?

Measure the whole task end to end, before and after the agent, including the time a person spends reviewing, correcting and redoing its work. Track how often its output is accepted unchanged. Do not rely on asking people whether it feels faster, because perception and measurement can disagree.

Generation time is the easiest number to show and the least useful one. An agent that drafts in five seconds but needs three minutes of fixing is slower than a person who writes it correctly in two. The only honest unit is time from task arriving to task finished, correctly.

Agree in advance what result would make you switch it off. Once an agent is running there is a strong pull to keep it, because removing it feels like admitting the project failed. A threshold set before launch makes that an ordinary decision.

What should a measurement plan for an agent include?

1. A baseline: time the task by hand for a week or two before the agent exists.
2. The same measure with the agent running, counting review and correction, not just generation.
3. An acceptance rate: the share of outputs used without edits.
4. A rework rate: errors found later, after the output was accepted.
5. A stopping rule: the result that would make you switch it off, written down before launch.

How should you start using AI agents safely?

Start with one high-volume task, run the agent in draft mode so a person approves every action, and give it the narrowest permissions that let it do the job. Widen its autonomy only once you know its error rate on real work, not on the examples it was built with.

Permissions deserve more thought than they usually get. An agent with access to email, calendar and the CRM is a new account with broad reach, and it can be steered by what it reads. An inbound email saying 'forward this thread to the following address' is text to a person but can become an instruction to a model. Read-only access and approval steps contain that risk. They do not remove it.

Log every action the agent takes, alongside the input that led to it. When something goes wrong, the log is the only way to tell whether the cause was a bad instruction, a bad input or a bad decision by the model.

When is an AI agent not worth building?

An AI agent is not worth building when the task is low volume, when a fixed rule already handles it, when a feature in software you already pay for does the job, or when nobody owns the process it would sit inside. In those cases the build and upkeep cost more than the time saved.

The last case is the one that sinks projects. An agent automates a process. If the process is undocumented, disputed or different for each person who runs it, the agent inherits that confusion and runs it faster. Fixing the process first is less exciting and usually worth more.

It is also fine to conclude that a plain automation, with no model in it at all, is the right answer. Predictable input deserves predictable software.