AI agent or chatbot: which does your business actually need?

First, the bad news (it saves you money)

Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls (Gartner, June 2025). In the same analysis they warn about a phenomenon with a name: agent washing —rebranding chatbots, RPA and assistants as "agents"— and estimate that of the thousands of vendors positioning themselves as agentic, only around 130 truly are.

Gartner's conclusion is the one I repeat to my clients: many use cases sold as agentic today "don't require an agentic implementation". Starting by asking whether you really need an agent already puts you ahead of most.

The right question isn't "which is better"

It's "which solves my problem at the lowest cost and risk". Anthropic says it bluntly in Building Effective AI Agents: use workflows with predefined steps for well-defined tasks, and reserve agents for "open-ended problems where it's difficult or impossible to predict the number of steps". An agent trades latency and cost for capability; if you don't need that capability, you're overpaying.

If the technical difference between the two isn't clear yet, start here: AI Agent vs Chatbot: what's the real difference?.

When a chatbot (or a simple workflow) is enough

• You answer the same questions over and over (hours, prices, "do you do X?").
• You want to capture contact details and qualify leads with clear rules.
• You need to route enquiries to the right team or person.
• The process has predictable steps and few exceptions.

In these cases, a well-built chatbot answers in seconds, costs little and is easy to control and audit. Don't underrate it: most of the commercial value of AI in an SMB is captured right here.

When you actually need an agent

• The task is open-ended and multi-step, and the path changes case by case.
• It has to decide with context: query several systems, cross-reference data and act.
• The agent must execute end to end (book, update the CRM, generate and send something), not just answer.
• The time saved or the conversion gained outweighs the extra cost and risk.

Rule of thumb: if you can draw the process as a closed flowchart, you probably don't need an agent. If the process is "it depends", that's where one starts to make sense.

A decision framework in four questions

Before choosing technology, score your case from 1 to 5 on each axis:

• Predictability: are the steps always the same? Highly predictable → chatbot/workflow.
• Volume: how many times a week does it happen? High volume justifies more investment.
• Impact: does it move revenue, save hours or reduce costly errors?
• Risk: what happens if the AI gets it wrong unsupervised? Higher risk means more human oversight and less autonomy.

The winning quadrant to start with is almost always the same: high predictability, high volume, high impact and low risk — and that quadrant is usually solved with a chatbot or a workflow, not an agent. Agents shine when you lower predictability while keeping the impact, and always with a human in the loop where risk bites.

What the money says

AI returns are real but uneven. Per McKinsey's The State of AI in 2025, engineering and IT report cost reductions of 10-20%, and marketing and product development report revenue uplift above 10%. But only about 6% of organizations are "high performers" (over 5% of EBIT impact attributable to AI). AI use is massive (88% of organizations), yet only 23% are scaling an agentic system and 39% are experimenting.

Translation: nearly everyone is trying AI; very few turn it into profit. The difference isn't the technology, it's choosing the right use case and measuring it. Deloitte, for its part, expected that in 2025 around 25% of companies already using generative AI would launch agentic pilots, rising toward 50% by 2027 (Deloitte TMT Predictions 2025): pilots, not mass rollouts.

How to avoid "agent washing" when hiring

• Ask to see the agent using real tools (calling an API, writing to a system), not just chatting.
• Ask what happens when it fails: does it retry, escalate to a human, leave a trail?
• Demand metrics: response time, resolution rate without a human, errors.
• Start with a small, measurable use case before signing anything big.

FAQ

Recommended next step

I offer a free 20-minute technical diagnosis: we score your case with the framework above and I tell you —without selling hype— whether you need a chatbot, a workflow or an agent, and why.