AI agents are coming to your business software. Here's what to hand over first

Shubham SharmaShubham Sharma5 min read

Agents are arriving inside the tools you already pay for. Most of the value sits in small, boring tasks, and most of the risk sits in handing over too much too soon.

Open almost any business app this year and you will find something called an "agent". Your CRM has one, your help desk has one, and your accounting software will offer to "work on your behalf" if you let it. Some of this is real progress. Some of it is a chatbot with a new name.

The difference matters, because an agent does more than answer questions. It takes a goal, works out the steps, and then acts through the tools it has been connected to. It can send the email, update the record or book the meeting. That is useful when it gets things right, and expensive when it doesn't.

Agents are arriving whether you plan for them or not

You may never go shopping for an AI agent. They are coming to you through software you already use. In August 2025, Gartner predicted that 40% of enterprise applications would include task-specific AI agents by the end of 2026, up from less than 5% in 2025. For most small and mid-sized businesses, that means the next update to a familiar tool will quietly add a button that lets the software act on its own.

Gartner has also been blunt about the other side. In June 2025 it said more than 40% of agentic AI projects would be cancelled by the end of 2027, because of rising costs, unclear business value and weak risk controls. It also called out "agent washing", where vendors rebrand ordinary chatbots and automation as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI were the real thing.

Put those two forecasts together and the message is simple. Agents will be everywhere, and a lot of the early attempts will fail. The businesses that get value will be the ones that pick the right first job.

What makes a good first task

The best starting points share four traits:

  • It repeats. The same kind of work comes in every day or every week.
  • The rules are already written down, or could be written on a single page.
  • Mistakes are cheap and easy to spot. A wrong draft is fine. A wrong refund is not.
  • There is a clear finish line. You can tell when the job is done and whether it was done well.

A useful test is to ask whether a careful new hire could do the task from a checklist in their first week. If yes, it is a reasonable candidate. If the job depends on judgement, relationships or context that lives in someone's head, leave it with a person for now.

Good candidates

These are the kinds of jobs that tend to fit:

  • Sorting and routing incoming enquiries, so a sales lead goes to sales and a complaint goes to support.
  • Drafting first replies to common questions for a person to check and send.
  • Updating the CRM after a call or a meeting, from notes or a transcript.
  • Sending polite reminders for overdue invoices, using wording you have approved.
  • Pulling a weekly report together from three or four tools that don't talk to each other.

None of these will make headlines. All of them give back hours that people currently spend copying, pasting and chasing.

Keep these with people, at least for now

Some work should stay with a human until you have months of evidence that the agent is reliable:

  • Pricing, quotes and discounts.
  • Refunds, credits and anything that moves money.
  • Anything published under your brand name, including social posts and press replies.
  • Hiring decisions and anything involving personal data you would struggle to explain.
  • Advice that carries legal, medical or financial weight.

The agent can still help here by preparing a draft or gathering information. It just should not press the final button.

Run a small, honest trial

Pick one task and write down how a person does it today: the steps, the time it takes and the mistakes that usually happen. That becomes your baseline.

Then switch the agent on in its most cautious mode, where it prepares work and a person approves it. Run it for about four weeks and track three things: time saved, how often its output needed fixing, and whether anything went wrong that a person would have caught. Give it the smallest set of permissions that lets it do the job. If it only needs to read the inbox and write drafts, it should not be able to send or delete.

Set a spending limit too. Many agent products charge per action or per task, and a loop that keeps retrying can run up a bill surprisingly quickly.

Questions worth asking any vendor

Before you connect an agent to real customer data, get clear answers to these:

  1. 1What can it do without asking for approval, and can I change that?
  2. 2Can I see a log of everything it did and why?
  3. 3What data can it see, and where is that data stored and processed?
  4. 4What does it do when it is unsure?
  5. 5How is it priced, and what stops costs from running away?

If a vendor cannot answer these plainly, that tells you something.

Start boring

The companies getting real value from agents are rarely the ones with the most ambitious plans. They start with one dull, well-understood task, measure it properly and expand only when the numbers hold up. That approach is slower to announce and much faster to pay off.

Have a project in mind?

Tell us what you need. We will reply with a plan, a team, and a quote.