A chat answers a question. An AI agent is supposed to take the next steps without you standing over it. Most small teams already have the first. McKinsey’s State of AI 2026 survey finds that 47 per cent of respondents say their organisations are scaling chatbots across the enterprise. Only about two in ten report scaling AI agents in the same way. That gap is the useful fact for anyone asking when to use an AI agent. The extra loop is worth it only when the extra cost and the extra ways it can go wrong are worth it too.
Anthropic’s engineering note on building effective agents, published 19 December 2024, says the same thing in working terms. Find the simplest solution that works. Increase complexity only when you must. For many jobs, a single model call with the right files attached is enough. An agentic system trades latency and cost for task performance. Know why you are paying that extra cost before you add the loop.
What an AI agent is, in working language
McKinsey’s March 2025 explainer defines an AI agent as software that has the agency to act on behalf of a user or a system to perform tasks. A chatbot that drafts a reply is a copilot. A system that pulls the order, checks stock, drafts the email, and waits for you to send it is already doing more than chat. Anthropic splits that middle ground. A workflow is a predefined path. An agent is a model that chooses its own next tool call. Both are agentic. Only the second is allowed to wander.
That distinction matters more than the marketing word. If you can write the steps on a whiteboard, you probably want a workflow, a checklist the model walks. If you cannot predict the number of steps, and you trust the model enough to let it choose, you are in agent territory. Anthropic is explicit that agents cost more, and that errors compound. Test in a sandbox. Put a stop on the number of turns.
Two other surveys, asked of different people, point the same way. Gartner’s C-suite AI survey of 1,303 respondents, published 1 September 2026, finds that only 22 per cent of organisations have successfully scaled AI across multiple business units. Deloitte’s Finance Trends 2026 press release, dated 8 October 2025, finds that 14 per cent of surveyed finance leaders are using fully integrated AI agents, even though 63 per cent of those teams have fully deployed and actively use AI. Chat is already common. A free-running agent is still the exception. Scaling AI agents beyond pilots is the later problem. This page is the earlier one. Should you build the loop at all?
When chat is enough
Stay with a conversation when the output is a draft a person will edit, the facts sit in one file you can attach, and nothing has to be posted, paid, or emailed without a human. Meeting notes, a first-pass proposal, a translation of copy you already approved, a list of questions for a supplier. Anthropic’s rule is the same. Optimising one model call with retrieval and examples is usually enough.
Chat is also the right tool when you do not yet know what “done” looks like. An agent that cannot tell whether it succeeded will keep calling tools. That is how token bills jump. McKinsey records that about 20 per cent of respondents say AI-related operating costs, including token costs, constrained their AI use. Cost is not only a finance issue. It is a signal that someone let a loop run without a stop.
NIST’s generative AI profile flags human-AI configuration, including over-reliance, as a live risk. If staff stop reading the draft because the chat “usually gets it”, you do not have a productivity gain. You have a quieter error. Keep the human on anything that leaves the firm. That is the same control as moving agentic AI from pilot to production, applied to a five-person office.
When a fixed workflow beats a free agent
Invoice matching, a month-end commentary pack, a weekly competitor scan with the same five sources. The path is known. The failure mode is skipping a step, not inventing a new one. Anthropic calls this a workflow. Prompt chaining, routing, a check after each stage. You write the path. The model fills the boxes. That is closer to digital transformation for SMEs than to a science project. You are documenting work you already do, then letting the model do the boring middle.
Customer support is the case both sources keep returning to. McKinsey’s research on gen AI’s economic potential reported that organisations using gen AI-enabled customer service agents increased issue resolution by 14 per cent per hour and reduced handling time by 9 per cent. Treat those as a measured case, not a target you can copy. Anthropic likes support because success is visible, tools can pull order history, and a refund can wait for a human. The pattern is a conversation plus a short list of allowed actions, not an unsupervised wander through the ledger.
When to use an AI agent
Use an agent when the job is open-ended, the success test is mechanical, and a person can still stop it. Anthropic’s own examples are a coding agent that can run tests, and computer-use demos where the model operates a machine with checkpoints. A small firm version looks like this. Research a supplier across several public filings, then produce a one-page brief with URLs the writer opened. Triage a backlog of support tickets that do not share a template. Draft a first pass of code against a test file you already own.
Do not use an agent as the first tool for a new process you cannot yet describe. Do not let it send mail, move money, or change a live record. McKinsey’s explainer still puts a human in the loop for extra input and feedback. That is not optional for a firm without a dedicated AI team. Name the reviewer. If nobody owns the stop, the loop runs without a check.
Larger organisations are moving faster on this than smaller ones. McKinsey’s 2026 survey shows respondents from firms over $1 billion in revenue scaling agents in one or more functions at 40 per cent, up from 27 per cent, while smaller organisations stayed flat at 22 per cent. Catching up does not mean copying their stack. It means picking one open-ended job with a hard stop, measuring hours returned and errors caught, then deciding whether a second loop is justified.
What to do this week
List the last ten jobs you put into a chatbot. Mark each as draft, fixed path, or open-ended. If nine are drafts, keep the chat. If two are the same path every week, write that path down and run it as a workflow with a human gate. If one is a hunt with no known steps, and you can say what “done” looks like, that is the candidate for an agent. Give it a turn limit. Read every outbound sentence.
A weekly competitor scan is almost always a workflow. The sources repeat. The output is a one-page note. The failure is skipping a source, not inventing a new one. A supplier diligence pack that has to follow new filings is closer to an agent, because you cannot list the clicks in advance. Still write the stop. Still open the URLs the brief cites. The classification is the work. The tool comes second.
When to use an AI agent is a cost and control decision. Chat is already the common tool. Agents remain the minority because they cost more and errors compound. Start with the conversation. Add a predefined path when the steps are known. Add a free loop only when the path cannot be written, the test of success is real, and a named person can still say stop.
Key takeaways
- McKinsey’s 2026 survey has 47 per cent scaling chatbots and about two in ten scaling agents. Plan from those two shares.
- Anthropic’s working split is a single model call, a predefined workflow, then an agent that chooses tools. Start on the left.
- Chat is enough when a person will edit the draft and nothing is posted unsupervised.
- A workflow beats a free agent when the steps are already known. Write the path. Keep a human gate.
- Use an agent for open-ended work with a mechanical test of done, a turn limit, and a named reviewer. Cost and compounding errors are the tax.
FAQ
When should a small team use an AI agent instead of a chatbot?
When the path cannot be written in advance, success can be checked, and a named person can still stop the loop. If a conversation produces a draft you will edit, stay with chat.
What is the difference between a workflow and an agent?
Anthropic treats a workflow as a predefined path the model walks, and an agent as a model that chooses its own next tool call. Both are agentic. Only the second is allowed to wander.
Why are agents less common than chatbots?
They cost more, they take longer, and errors compound. McKinsey’s 2026 survey shows chatbots already scaling in 47 per cent of organisations, and agents in about two in ten.
Can an agent send email or post a journal?
Not as a first step. Draft and flag. A human sends and posts. Unsupervised outbound action is a control hole a small firm cannot absorb.
How do I know the agent is done?
Write the test before you start the loop. A test file that passes, a brief with opened URLs, a ticket in a resolved state. If you cannot name done, you do not have an agent job. You have a chat.
Sources
- McKinsey, “The state of AI in 2026: On the road to ROI”, 25 August 2026. mckinsey.com/…/the-state-of-ai
- McKinsey, “What is an AI agent?”, 25 March 2025. mckinsey.com/…/what-is-an-ai-agent
- McKinsey, “The economic potential of generative AI: The next productivity frontier”, 14 June 2023. mckinsey.com/…/the-economic-potential-of-generative-ai-the-next-productivity-frontier
- Anthropic, “Building effective agents”, 19 December 2024. anthropic.com/engineering/building-effective-agents
- Gartner, “Gartner Survey Finds Only 22% of Organizations Have Successfully Scaled AI Across Multiple Business Units”, 1 September 2026. gartner.com/…/successfully-scaled-ai-across-multiple-business-units
- Deloitte, “Finance Trends 2026: Finance Leaders Take Helm in Strategic Decision-Making Amid Global Challenges”, 8 October 2025. deloitte.com/…/finance-trends-2026-survey-release.html
- NIST, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” (NIST AI 600-1), 26 July 2024. nist.gov/publications/…generative-artificial-intelligence
This article is general information, not investment advice. See section 4 of our Terms & Conditions.
This is the kind of decision we advise on. If you are weighing it for your own business, see how we work or start a conversation.
