AI agents for business: where they make sense, and where they don't

An artificial intelligence (AI) agent is a software system that can reason over information, look up the knowledge it needs, use the tools you have authorized, and take controlled actions toward a goal, bringing in a person whenever judgment is required. A chatbot answers questions. Traditional automation follows rules written in advance. An agent sits between them: it works toward an outcome, decides which step comes next, and acts inside limits you set.

That definition also rules things out. An agent is not magic, and it is not an employee you can hand a job to and forget. It does well with some kinds of work and poorly with others. This article covers both, plus the controls that separate a useful agent from an expensive experiment.

Agent, chatbot, or automation: what is the difference?

Vendors use these words loosely, so the distinction is worth getting right before you spend money.

A chatbot holds a conversation. It answers from a script or a knowledge base, and stops there. Ask it to reschedule an appointment and, at best, it tells you how.

Traditional automation follows rules: when a form arrives, create a record; when an invoice goes past due, send a reminder. It is fast, cheap and predictable, and it breaks when the input does not match what the rule expected.

An agent gets a goal, a set of tools and a set of limits. It reads the situation, decides what to do next, uses a tool (your calendar, your document storage, or your customer relationship management (CRM) system, the software where you track customers and deals), checks the result, and either continues or hands off to a person. It copes with varied inputs, which is exactly where rules struggle.

"Agentic AI" is the broader label for systems built this way. When you see it on a product page, ask what the system does on its own and who approves what. Gartner has warned about "agent washing": existing assistants, chatbots and automation tools rebranded as agents without substantial agentic capabilities.

1. 01ReadThe request, in whatever format it arrives
2. 02RetrieveYour policies, records and history
3. 03Use toolsOnly the ones it is authorized to use
4. 04ActWithin defined limits
   Hand offWhen judgment is required, a person decides

Log and reviewEvery action recorded and checked

Where agents make sense

Agents earn their place in work with three traits: several steps, more than one system, and inputs that change from case to case. Typical examples:

• Intake and triage. Inquiries arrive by email, web form and phone notes, each written differently. An agent can read them, pull out what matters, check whether the contact already exists, rate the urgency, and route each one to the right person with a summary.
• Scheduling. Matching a client, a staff calendar and a room takes back and forth. An agent can propose options, confirm, and update the calendar, escalating requests that fall outside your policy.
• Quote follow-up. Open quotes are easy to lose track of. An agent can track open quotes, draft a follow-up that reflects the original request, and flag the ones where a salesperson should call.
• Document processing. Invoices, applications and contracts come in different formats. An agent can extract the fields you need, compare them against your records, and send exceptions to a person.
• Internal knowledge questions. Staff lose time asking where a policy lives or how a process works. An agent connected to your own documents can answer and cite the source, so people can check it.

In every case, the agent works with clear tools, clear limits, and a defined point where a human takes over. That handoff is part of the design, not a sign that the agent failed.

Where agents don't make sense (yet)

High-stakes, irreversible decisions without review. Approving credit, moving money, ending a contract, giving legal or medical advice. An agent can prepare the file and recommend. A person decides.

Tasks a simple rule handles better. If the logic fits in one sentence ("when X happens, do Y"), a rule is cheaper, faster and easier to audit. Adding reasoning to a task that doesn't need it adds cost and new ways to fail.

Processes nobody has written down. If the steps live in one person's head and change depending on who you ask, an agent will inherit the confusion. Document the process first. Often that exercise alone shows you the fix.

Work with no reliable data. An agent is only as good as what it can read. If customer records are duplicated, prices live in three spreadsheets, or the knowledge base is outdated, the agent will act confidently on bad information.

What the market is showing

Adoption is real, and uneven. McKinsey's State of AI survey, published in August 2026, found that 40 percent of respondents from organizations with more than $1 billion in annual revenue reported scaling AI agents, up from 27 percent the year before. Among smaller organizations, the share held flat at 22 percent. About one in five respondents also said AI operating costs were limiting their use of AI.

Caution is warranted too. In June 2025, Gartner predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Its advice matches the lines drawn above: use agents where decisions are needed, automation for routine workflows, and assistants for simple retrieval.

The controls that matter

A well-built agent is a controlled system, not an autonomous one. These are the controls we treat as non-negotiable.

Permissions and least privilege. Give the agent only the access the task requires. If it drafts follow-ups, it needs to read quotes and write drafts, not delete customer records. The Open Worldwide Application Security Project (OWASP), a nonprofit that publishes software security guidance, lists "excessive agency" among its top 10 risks for applications built on large language models (the technology behind most agents) and traces it to three causes: too much functionality, too many permissions, and too much autonomy.

Limits on actions. Enforce boundaries in the connected systems, not only in the agent's instructions: which records it can touch, how many messages it can send, which amounts it can quote. Prefer "draft for approval" over "send" until the agent has earned trust.

Human in the loop. Decide in advance which actions need a person's approval, and whose. The agent should hand off with context, so nobody starts over.

Logging and review. Record what the agent saw, what it decided and what it did. Review a sample regularly, especially in the first weeks.

Evaluation before rollout. Test the agent against real past cases, including messy ones, before it touches live work. Define what a correct result looks like and measure against it. Then start with a small share of the volume.

Cost monitoring. Agents consume paid model usage at every step, and a badly designed loop can run up a bill quietly. Track cost per completed task and set alerts.

A checklist for your agent idea

Before you commit budget, run the idea through these questions:

1. Can you state the goal, and what "done" looks like, in one sentence?
2. Does the work involve several steps, more than one system, and inputs that vary? If not, consider a rule.
3. Is the process written down, including the exceptions?
4. Is the data the agent needs accurate and accessible?
5. Which tools does it need, and what is the minimum access for each?
6. Which actions require human approval, and who gives it?
7. What is the worst thing it could do, and can that be undone?
8. How will you test it before launch and measure it after?
9. What will each completed task cost, and what is it worth in revenue, savings or reduced risk?

If several of these have no clear answer, the idea isn't ready yet. That is worth knowing before you build, not after.

What this means for your business

An agent is worth considering when it helps you grow revenue, reduce costs or reduce risk in a specific workflow you can name. The best first candidates are usually unglamorous: the inquiry that sits unanswered over a weekend, the quote that never gets a follow-up, the document someone retypes by hand.

Start with one bounded process, tight permissions, and a person who owns the handoff. Measure the result honestly. Expand only when the numbers justify it.

If you want to test an agent idea against these questions, book a meeting with us.