AI agents for business: where they fit, where they do not, and how to start
← INSIGHTS

AI agents for business: where they fit, where they do not, and how to start

Interest in AI agents has surged ahead of experience. Here is where they fit, the controls they need, and a six-step way to start.

BY STEFAN MARITZ··8 MIN READ

TL;DR

  • An AI agent is a system where a model directs its own process and tool use, rather than following a fixed path. That autonomy is the source of both the value and the risk.
  • Interest has surged. Google searches for "ai agents for business" rose from about 1,900 a month in September 2025 to 18,100 in August 2026.
  • Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls.
  • Agents suit open-ended work with clear success criteria, feedback loops, and human oversight. Stable rule-based steps are better served by plain automation.
  • Give any agent the least access it needs, a human checkpoint before consequential actions, and a log of what it did.
  • Start with a workflow, which follows fixed steps, and move to an agent only when the steps cannot be predicted.

Every software vendor now sells an agent, and the word has stretched to cover almost anything with a chat box. For a leader deciding where to spend money, a precise definition and a few filters are worth more than another product list.

This guide covers what an agent is, where the evidence says it fits, the controls it needs, and how to start.

What an AI agent is

Anthropic's engineering guide to building effective agents draws the line most clearly. It calls both variations agentic systems, and then separates workflows, where models and tools are orchestrated through predefined code paths, from agents, where the model dynamically directs its own process and tool use. An agent starts from a command or conversation, plans and operates independently, uses feedback from its environment at each step, and can pause for human input at checkpoints or when it hits a blocker. In practice, the guide notes, agents are typically models using tools in a loop.

That definition excludes plenty of products sold as agents. A chatbot that answers questions is not one. A fixed sequence of AI steps is a workflow, which is often the better choice, as the guide itself argues. We covered the underlying distinction in our guide to automation versus AI.

Where interest and evidence stand

Searches for "ai agents for business" rose from about 1,900 a month in September 2025 to 18,100 in August 2026, according to Google Keyword Planner data. Demand for information is clearly ahead of experience.

McKinsey's 2026 State of AI survey reports that nearly nine in ten respondents use AI regularly in at least one function, and that 44% say AI is scaling across their enterprise. It also says the share reporting meaningful EBIT impact is essentially unchanged. I could read only McKinsey's summary of the survey, not the full report.

Gartner's prediction from June 2025 is that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value, or inadequate risk controls. That is a forecast, and I could read only a summary of Gartner's release. It still names the three questions to answer before starting: what will it cost, what is it worth, and how will we control it.

Where agents fit

Anthropic's guide says agents add the most value for tasks that need both conversation and action, have clear success criteria, allow feedback loops, and include meaningful human oversight. Its two examples are customer support, where tools can pull customer data and issue refunds and success can be measured by resolution, and coding, where automated tests verify the result.

OpenAI's practical guide to building agents describes agents as suited to workflows where traditional deterministic, rule-based approaches fall short. Its examples include systems with rule sets so extensive that updates become costly or error-prone, and scenarios that rely heavily on unstructured data such as documents and conversation. I could read only a search summary of that page.

Put those together and a pattern emerges. Agents fit work that is open-ended, language-heavy, and checkable, where you can tell afterwards whether the result was right.

Where they do not fit

Three situations call for something simpler.

  • The rule is stable and writable. Plain automation is cheaper, faster, and easier to audit.
  • A wrong action is expensive and nobody reviews it. Anthropic warns that autonomy brings higher costs and the potential for compounding errors, and recommends extensive testing in sandboxed environments with appropriate guardrails.
  • The process is undocumented. An agent cannot be checked against a standard nobody has written down.

Anthropic's own advice is to find the simplest solution possible, which might mean not building an agentic system at all.

The controls an agent needs

An agent can act, so access is the first control. IBM's 2026 breach report page says security teams must secure agentic identities with identity-based access controls, tightly scoped permissions enforced at runtime, human attribution, and auditability. Kiteworks' summary of the same research reports that 92% of organisations with an AI-related security incident lacked proper AI access controls. That is a vendor's summary, so treat it as indicative.

Transparency is the second. The Cloud Security Alliance's analysis of the EU AI Act notes that the Article 50 duty to disclose when a person is interacting with an AI system applied from 2 August 2026, whatever the risk tier. A customer-facing agent in the EU needs to say what it is. Ask legal counsel how this applies to you, and see our guide to AI governance for the wider framework.

Starting with a first agent

  1. Pick a process with a clear success measure and record the baseline: how long it takes, what it costs, and how often it goes wrong.
  2. Build it as a workflow first. Fixed steps are easier to test. Let the model direct more of the process only where the steps cannot be predicted.
  3. Give it least privilege. Read access before write access, and small limits on anything that spends money or contacts customers.
  4. Add a human checkpoint before consequential actions, and keep it until the error rate justifies removing it.
  5. Log everything it does, so you can see what happened and why.
  6. Set a cost cap and a stop date. Compare the result with the baseline and decide whether to continue.

Questions to put to a vendor

  • What does the agent do on its own, and what needs approval?
  • What can it read and change, and who set those limits?
  • How do I see what it did, step by step?
  • How is success measured, and who measures it?
  • What happens when it is wrong?
  • What does each completed task cost at my volume?

If you want a scored view of whether your organisation is ready for agents, the AI Readiness Diagnostic covers strategy, data, people, process, and tooling, and is free to take.

Frequently asked questions

What is an AI agent?

An AI agent is a system where a model plans and carries out a task using tools, directing its own process and using feedback from its environment. It differs from a workflow, where the steps are fixed in advance.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions. An agent can also take actions, such as updating a record, sending a message, or issuing a refund, across several steps.

Are AI agents safe for business use?

They can be, with controls: minimum access, a human checkpoint before consequential actions, and a full log of activity. The risk comes from autonomy combined with broad access and no review.

How much do AI agents cost?

There is no single price. Anthropic notes that agentic systems often trade latency and cost for better task performance, so measure cost per completed task against your baseline before scaling.

Should I use an agent or plain automation?

If the rule is stable and you can write it down, use automation. Use an agent where the steps cannot be predicted, the input is mostly language, and you can check the result afterwards.

WORK WITH PRAXES

Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.

Take the AI Readiness Diagnostic