AI Agents for Business: What They Can Do Today

Do you need an agent or a workflow?

Use a workflow if you can write the steps on one page. Use an agent only if the steps depend on what the software finds along the way and you can check the result afterwards. Most small business tasks fall on the workflow side, and workflows are cheaper to run and easier to trust.

Ask three questions about the job:

1. Can you list the steps in advance? If yes, build a workflow and use AI only inside one or two steps.
2. Can you check the result in under a minute? A drafted reply or a booked slot can be checked. A research report with invented sources cannot be, not quickly.
3. What happens when it is wrong? If the answer is a refund, a legal problem or an unhappy customer, keep a person in the loop.

Anthropic's engineering guidance makes the same point from the builder's side. It defines agents as systems where the model directs its own process and tool use, and says to start with the simplest solution because agentic systems trade cost and speed for flexibility and risk compounding errors. If the people who sell the models say to start simple, a small business should too.

What agents do well today

Agents work today on narrow jobs where the input is text, the answer is in a document you control and a mistake is cheap or caught. The table sets out the main ones.

What they do badly

Agents do badly at open-ended work and at anything where being confidently wrong is expensive. Three patterns show up again and again.

• Making things up. A model asked a question it has no answer to may produce a plausible one. Restrict it to your documents and tell it to say when it does not know.
• Long chains of steps. Each step can go wrong, and mistakes build on each other. An agent that does eight things in a row is far less reliable than one that does two.
• Judgement calls on money or people. Refunds, discounts, hiring screens and credit decisions need a person. Article 22 of the GDPR also restricts decisions based solely on automated processing where they significantly affect someone.

Gartner forecast in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs and unclear value. It also said that many vendors rebrand chatbots and assistants as agents, and estimated that only about 130 of the thousands of agentic AI vendors are genuine. Nobody can say how that forecast will turn out. It is a good reason to ask any vendor what the agent decides on its own.

How agents are priced

Most vendors charge per conversation or per outcome, so the bill grows with your volume. Prices below come from each vendor's own pricing page, checked in November 2026.

Read Intercom's definition twice. An outcome includes the customer not asking for more help after the agent replies. A customer who gives up counts the same as one who is satisfied, so you are paying for silence as well as success. Track real resolution yourself, for example with a one-click thumbs-up on each reply.

A worked example

This is illustrative arithmetic, with assumptions you should replace. A business handles 500 customer conversations a month. A person takes about 6 minutes each, and staff time costs EUR 20 an hour.

• All by a person: 500 x 6 minutes = 3,000 minutes, or 50 hours, which is EUR 1,000.
• With an agent that resolves 60% at USD 0.99 an outcome: 300 outcomes cost about USD 297. The other 200 still go to staff, 200 x 6 minutes = 20 hours, or EUR 400.
• Total: about 700 a month against 1,000, a saving of about 300 before setup, seat fees and the time to review the agent's answers.

If the agent resolves only 30%, the saving falls to about EUR 150 a month (150 outcomes cost about USD 149; the other 350 conversations take 35 hours, EUR 700), before setup, seat fees and review time, which can erase it. The number that decides the case is the resolution rate, and you only learn it by running the thing.

Guardrails to insist on

Set these before an agent touches a real customer or a real system:

• Least privilege. Read access to the calendar, not the whole mailbox. Draft access, not send access.
• Spending and volume caps. A limit on messages per hour and on monthly spend.
• A visible log. Every conversation and every action stored where you can read it.
• A human handover path. One tap or one word to reach a person, always.
• A written scope. What it may answer and what it must refuse. Test the refusals.
• A way to turn it off. One switch, known to more than one person.

Questions to ask a vendor

• What exactly does the agent decide without a person? Ask for an example, not a slogan.
• How is a resolved conversation defined, and who measures it?
• Where is conversation data stored, and will you sign a data processing agreement under GDPR Article 28?
• What happens when it does not know? Ask to see it fail.
• What do I pay in a month with twice the expected volume?

Where to go from here

Pick one narrow job from the table, run it for a month with a person approving the output, and measure the resolution rate before you pay for more volume. Our guide to AI agents for small business shows five real jobs with the hours worked out, and the AI agents examples page lists build-or-buy options by department. For the customer-facing side, read AI customer service agents: what works and what does not. If you want help scoping a pilot, see our AI automation service. We are a new studio, and we quote in writing.