AI agents: where they help your operation — and where they do not

Revised September 4, 2026.

An AI agent is useful when completing a task requires interpreting information and deciding what to do next. That does not mean every business process needs one. A fixed rule, an integration, or a better screen in your existing system may solve the problem more reliably.

Start with the operation, not the technology: where does work wait, where do people re-enter the same information, and which exceptions consume experienced staff time?

Automation, assistant, or agent?

Consider an incoming service request.

• Automation: copy structured fields into a ticket and assign it using an agreed rule.
• Assistant: summarize the request and draft a response for a person to review.
• Agent: investigate the request, choose relevant tools, and decide the next step within defined permissions.

These are different design choices, not maturity levels. A well-designed assistant may deliver the value you need without permission to contact customers or change records.

Choose a problem you can measure

A useful first project has a clear owner, recurring work, accessible information, and an observable result. “Improve productivity with AI” is not a testable brief. “Prepare a service-request summary that staff can verify without reopening three systems” is.

Before building, collect a representative sample of the current work. Include incomplete requests, conflicting records, and cases that require escalation. Record handling time, correction effort, and the point at which an experienced person makes a judgment.

Agree on the outcome before evaluating a demo. Useful measures include accepted outputs, total handling time including review, missed exceptions, and cost per completed case. A fast draft that needs extensive correction is not a faster operation.

A real case, with the right limits

In a November 2024 customer story, Microsoft reported that Markerstudy used generative AI to summarize claims calls. The company handled roughly 840,000 calls per year, with reported savings of about four minutes per call — the basis for an annual estimate of 56,000 hours. Staff checked the summaries for accuracy. This is a third-party case about assisted summarization, not evidence that an autonomous agent will produce the same savings in your business. Read Microsoft's case study.

The useful lesson is the shape of the project: a specific task inside an existing process, with people still responsible for the result.

Fix the workflow before adding autonomy

If nobody agrees which system contains the authoritative customer record, an agent will inherit that ambiguity. Establish data ownership, access rules, and escalation paths first.

Separate reading from acting. Finding an order is not the same permission as changing it. Drafting an email is not permission to send it. Start with the access needed for the pilot and add actions only when their benefit and failure handling are clear.

Ask what happens when the agent is uncertain, a dependency is unavailable, or an action partly succeeds. Someone needs to own those cases after launch.

When purpose-built software makes more sense

Expensive legacy software can force teams to work around the product: duplicate entry, spreadsheet handoffs, and manual reconciliation. An agent on top may help, but it may also hide a workflow that needs redesign.

Custom software is worth considering when the way you operate is genuinely different and a standard tool repeatedly gets in the way. Keep commodity capabilities that work. Build around the parts where better coordination, faster decisions, or fewer errors could create an advantage.

Compare the full cost: implementation, migration, integrations, support, training, and operating the new system alongside the old one. Replacing a subscription creates no advantage if its maintenance burden costs more.

What to ask before approving a pilot

1. Which task will change, and who owns its result?
2. What is the baseline, and what would count as an improvement?
3. What data and actions are in scope?
4. Which decisions must remain with a person?
5. How will we stop, recover, or return to the previous process?
6. Who maintains the system once the pilot ends?

You do not need a company-wide transformation to answer these questions. You need one bounded workflow and evidence that the new approach is better.

For implementation choices, read our agent architecture guide. If legacy software is the constraint, start with modernizing an existing system.

Talk to us about the workflow you want to improve.