What you can hand to an AI agent in 2026, task by task, and the four jobs to keep

The most common mistake I see when a team gets its first capable agent is that they give it the interesting work. Write our strategy. Draft the pitch. Decide which leads to call. Then they are disappointed, because the output is plausible and average, and they conclude agents are not ready.

Meanwhile the invoice queue, the meeting notes, the weekly report and the forty emails that need a two-line reply are still being done by a person, at 6pm, badly.

Agents in mid-2026 are very good at a specific kind of work: frequent, boring, and easy to check. Gemini Spark running in the background, Claude Cowork working through a folder on your desktop, an OpenClaw agent on a schedule, Copilot inside a Microsoft tenant: the tools differ less than the task selection does. Below is the list we use with clients, sorted by how much a person needs to check.

Hand over completely

These are tasks where the output is either right or obviously wrong, the cost of a wrong answer is small, and they happen often enough that the setup pays back in a fortnight.

Inbox triage. Sorting rather than replying. Label, prioritise, draft a one-line summary of each thread, flag the three that need you today. Every agent product now does this and the error mode is a mis-filed email, which you would have mis-filed yourself.

Meeting notes into the systems of record. A transcript becomes a summary, a list of decisions, a list of actions with owners, and an update to the CRM record or the project tracker. The human step is reading the summary, which takes a minute instead of the twenty minutes the write-up used to take. We wrote up the CRM sync version separately.

The weekly report. Pull the numbers from the four places they live, put them in the template, write two sentences about what moved. If the numbers are wrong the agent did not make them wrong.

Scheduling and rescheduling. Find a slot across five calendars, propose it, book it when everyone agrees, move it when someone cancels.

First-pass document extraction. Invoices, receipts, purchase orders, delivery notes, forms: pull the fields into a table. The check is a total that reconciles, which is a check a machine can also do.

Monitoring. Watch a regulator's page, a competitor's pricing, a supplier's status page, a set of search terms, and tell you when something changes. This is what Google's information agents and Spark's scheduled tasks are built for, and it was tedious for a person precisely because nothing changed most days.

Research briefs before a call. Company background, recent news, who you are meeting, what they have said publicly, what your CRM says about the last three interactions. Ten minutes of a person's time replaced by a page they read in two.

Hand over, then check

These tasks produce something a person should read before it goes anywhere, because the cost of a mistake is real and the agent's confidence is not a signal.

Replies to routine customer questions, where the answer is in the documentation. The agent drafts, a person sends. After a month of checking, the categories where the draft was never edited can move up to the first list.

Code that changes something. Coding agents are the most capable agents available and also the most confident when wrong. Review the diff. Run the tests. Every serious team runs its coding agent inside a pull-request flow for this reason.

Contract first reads. Pull out the parties, dates, payment terms, liability caps and anything that differs from your standard paper. A lawyer still reads the contract; they now read it with the map in hand. The contract review use case describes the full pipeline.

Anything that spends money below a threshold. Reorder stock, renew a subscription, book travel. Set the cap, set the approved vendors, and review the log weekly rather than each transaction.

Data cleaning and migration. An agent will deduplicate a customer list or map one schema to another far faster than a person and will make one systematic mistake across ten thousand rows. Sample the output before you load it.

Ask for the preparation, keep the decision

These are tasks where the agent's job is to make a person faster, and the person's judgement is the product.

Hiring. Screening against stated criteria, summarising a CV against a job description, scheduling: fine. Deciding who to interview and who to hire: a person, every time, and in several jurisdictions the law says so.

Pricing and quotes. The agent assembles the quote from the price book and the customer's history. A person decides whether to discount.

Strategy and planning. Ask the agent for the options, the evidence for each, and the strongest case against your preferred one. Then decide. The agent is a fast, tireless, slightly credulous analyst.

Writing that carries your name. First drafts, edits, alternatives, fact-checks. The final version, when it matters, comes from a person, because the reader can tell and because you are the one accountable for it.

The four jobs to keep

Some tasks should not go to an agent even with checking, and the reason is the same in each case: the action is hard to reverse and the accountability cannot be delegated.

Signing anything with legal or regulatory weight. Filings, attestations, contracts, consents. An agent can prepare every word. A named person signs.

Terminating people, contracts or accounts. Even where the criteria are clear, the decision to end a relationship should be a person's, and the person should have read the file.

Moving money above your threshold, or to a new payee. The threshold will be different for a startup and a bank. It should exist, it should be written down, and it should be enforced in code rather than in the prompt.

Speaking to a customer, a regulator or the press about a dispute or an incident. The agent drafts; a person sends. The first time an agent apologises on your behalf for something it does not understand will be the last time you let it.

How to start

Pick two tasks from the first list that someone on your team does at least weekly. Set the agent up with its own credentials and access to only what those tasks need. Run it for two weeks with the person still checking, and count the corrections.

Then move one task from the second list to the first, based on what the count told you. A team that does this every month has, within a year, handed over most of the work nobody wanted and kept the work that needed them, which is the whole point. It is also the shape of every agent engagement we run.

Frequently asked

What tasks can AI agents do reliably in 2026? Frequent tasks with checkable outputs and a small cost of error: inbox triage, meeting notes into CRM and trackers, weekly reports from existing data, scheduling, first-pass extraction from invoices and forms, monitoring pages for changes, and research briefs before calls. Tasks that change code, answer customers or spend money work well with a person checking the output before it goes out.

What should you not delegate to an AI agent? Four things: signing anything with legal or regulatory weight, terminating people, contracts or accounts, moving money above a written threshold or to a new payee, and communicating with customers, regulators or press about a dispute or incident. An agent can prepare all of these. A named person does them.

How do you decide what to automate with agents first? Score candidate tasks on three things: how often they happen, how easy the output is to check, and how bad a wrong answer is. Start with tasks that are frequent, easy to check and cheap to get wrong, run them with a person checking for two weeks, and promote or demote tasks based on how many corrections were needed.

Related reading

• AI agents vs automation: when to use which
• Gemini Spark works while your laptop is closed
• Use case library

The teams getting the most out of agents this year are the ones that handed over the dull work first and treated judgement as the thing worth keeping. Start with the invoice queue.