An AI agent is a system where the model decides what to do next. You give it a goal and a set of tools it can use, and it works out the sequence itself, one step at a time, reacting to what it finds. That is the whole definition, and it is the only part of the word that means anything.
Everything else you have been told an agent is, autonomous, intelligent, a digital employee, is packaging. The useful question for a business owner is not "should we get an agent". It is "does this job need a system that figures out its own steps, or one that follows steps we already know". Most jobs, including most of the ones currently being pitched to you as agentic, are the second kind.
The distinction that decides the price
There are three things being sold under the same word, and they behave completely differently once they are running in your business.
A chatbot answers. Someone asks it something, it replies, nothing else in your business moves. Useful, limited, cheap.
An AI workflow does a job along a path you defined. An invoice arrives, the model reads it, the system matches it to a purchase order, flags anything that does not reconcile, and puts it in front of a person to approve. The model does the language-shaped part. The route was written by you. Every run costs roughly the same, takes roughly the same time, and can be tested, replayed and audited like any other piece of software.
An AI agent does a job by working out the path as it goes. You say "find out why this customer's order has not shipped", and it decides to check the order system, then the courier's tracking, then the email thread, then the stock records, in whatever order the previous answer suggests. Nobody wrote that sequence.
The reason this matters commercially is that the second and third have opposite cost profiles. A workflow has a known number of model calls, so it has a known bill and a known runtime. An agent runs until it decides it is finished, which means the bill and the runtime are variable, and the interesting failure mode is not an error message but a large invoice. We wrote the engineering version of this argument in agents versus workflows, and eighteen months of client work has not softened it: when the steps are knowable, encoding them is cheaper, faster and easier to operate than paying a model to rediscover them every time.
Most "AI agents" are not agents
Gartner gave this its own name in June 2025: agent washing, the rebranding of existing products such as AI assistants, robotic process automation and chatbots without substantial agentic capability. Their estimate at the time was that of the thousands of vendors marketing agentic AI, only about 130 were genuinely selling it. Treat that number as directional rather than precise, because Gartner did not publish a methodology for the denominator. Treat the direction as reliable, because it matches every vendor conversation we have sat in since.
This is not a purity argument. It matters because you are being quoted agent prices for workflow problems. If the thing you are buying runs the same four steps every time, it is a workflow, and a workflow should cost less, break less, and be easier to hand back to your own team.
The test is short. Ask the vendor to describe, precisely, what the system does on a run where something unexpected happens. If the answer is a fixed branch ("then it routes to a human"), you are looking at a workflow with good marketing. If the answer is genuinely "it decides", ask what stops it deciding for forty minutes.
The deployment numbers are worth knowing before you commit
Two published figures give a fair picture of where agents actually are in mid-2026.
McKinsey's most recent global survey found 62% of organisations at least experimenting with AI agents, but in any given business function no more than 10% said they were scaling them. Experimentation is near-universal; production is rare. And Gartner's June 2025 prediction, still the most-quoted number in the category, is that more than 40% of agentic AI projects will be cancelled by the end of 2027, on grounds of escalating costs, unclear business value or inadequate risk controls.
Read those together and the shape is clear. Enormous intent, thin deployment, and a forecast failure rate driven entirely by non-technical causes. Not model quality. Not capability. Cost, value and control. Those are exactly the three things a smaller business is worst placed to absorb when they go wrong, and exactly the three things a good scoping conversation settles before anything gets built.
When an agent is genuinely the right answer
There is a real category of work where agents earn their keep, and being precise about it is more useful than either enthusiasm or scepticism.
- Investigation. Where the next step depends on what the last step found. Tracing why an order is late across four systems. Working out why a reconciliation does not balance. Triaging a support ticket that could be five different problems.
- Genuinely variable inputs. Where the same job arrives in materially different shapes each time, and a fixed path would need so many branches that the branching becomes the bug.
- Long-running research. Tasks that need twenty or thirty lookups where each one shapes the next.
Notice what is not on that list. Customer support routing, document classification, extracting fields from invoices or delivery notes, drafting quotes from a template, summarising a call and updating the CRM. Every one of those is a workflow, every one of them is a good first AI project for a mid-sized business, and none of them needs an agent. The projects that quietly pay for themselves are almost all in that second list.
What an agent needs underneath it before it can be trusted
If you do have a genuine agent-shaped problem, the constraint is rarely the model. McKinsey's data is blunt about this: eight in ten companies cite data limitations as a roadblock to scaling agentic AI. An agent is only as good as its ability to reach and interpret your records, and in most mid-sized businesses those records disagree with each other before any AI touches them. That is a data foundations problem, not an agent problem, and no amount of autonomy fixes it.
Beyond the data, four things are non-negotiable in anything we would put in front of a client:
- A human gate on anything irreversible. Spending money, sending a message to a customer, changing a record in a system of record. The agent proposes; a person approves.
- Hard limits. Maximum steps, maximum spend, maximum time, and a hard stop when exceeded. Without these you will eventually pay for a confused loop.
- A visible trace. Every step it took, logged and readable, so that when it gets something wrong you can see where the reasoning went sideways rather than guessing.
- A POPIA-aware view of what it can see. An agent with broad access to your systems is, by design, a system that reads personal information across contexts. That needs deciding deliberately, not discovering later.
The short version of all four: the more the system decides, the more you need to be able to check it afterwards.
The better question
"Do we need an AI agent" is a technology question, and technology questions produce technology answers, which is how businesses end up with capable systems nobody uses. The question that actually gets you somewhere is the one we open every engagement with: which decision, or which repetitive job, is costing you time or money right now, and what is the smallest thing that would fix it? Nine times out of ten the honest answer to that is a workflow. Occasionally it is an agent. Once in a while it is neither, and the right advice is to fix the process and buy nothing.
Working out which of those three you are looking at, before anyone quotes you for a build, is most of what good AI consulting is for. What it then takes to wire any of them into the systems you already run is the five things you are actually connecting.
FAQ
What is an AI agent? An AI agent is a system where the model decides its own next steps to achieve a goal you set, using tools it has been given access to. That is what separates it from an AI workflow, where the sequence of steps is written in advance by a developer and the model only performs specific tasks inside it.
What is the difference between an AI agent and an AI workflow? A workflow follows a path you defined, so its cost, runtime and behaviour are predictable and testable. An agent works out the path at runtime, so it is more flexible but its cost, runtime and behaviour vary from run to run. Workflows are the right answer for most business tasks; agents suit work where the next step genuinely depends on what the last one found.
Does my small business need an AI agent? Probably not as a first project. Most high-value tasks in a mid-sized business, such as classifying documents, extracting data from invoices, drafting quotes or updating records after a call, are repetitive and predictable, which makes them workflow problems. Agents suit investigation-shaped work where the steps cannot be written down in advance.
What is agent washing? Agent washing is Gartner's term for rebranding existing products, such as chatbots, AI assistants and robotic process automation, as agentic AI without substantial agentic capability. It matters to buyers because agent-branded pricing is often being applied to workflow-shaped software.
Why do so many AI agent projects fail? Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, and attributes it to escalating costs, unclear business value and inadequate risk controls rather than to model capability. McKinsey separately finds that eight in ten companies cite data limitations as a roadblock to scaling agents.
References
- Anthropic: Building effective agents (the architectural distinction between workflows and agents). https://www.anthropic.com/news/building-effective-agents
- Gartner: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025; includes the agent-washing definition and the ~130 vendor estimate). https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- McKinsey: The state of AI: how organisations are rewiring to capture value (62% experimenting with agents; no more than 10% scaling in any function). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- McKinsey: Scaling agentic AI with data transformations (eight in ten companies cite data limitations as a roadblock). https://www.mckinsey.com/capabilities/mckinsey-technology/our-insights/building-the-foundations-for-agentic-ai-at-scale
- Forbes: Why 40% Of Agentic AI Projects May Be Canceled By 2027 (July 2026 analysis of the governance and ownership causes). https://www.forbes.com/sites/robertszczerba/2026/07/07/why-40-of-agentic-ai-projects-may-be-canceled-by-2027/
Written by JP, Sixees Labs. Last reviewed July 2026.