Chatbot, automation or AI agent: which one does your business need?

Which one do you need? The short answer

Founders ask me this often, so here is the short answer.

A chatbot answers questions. An automation runs the same fixed steps every time. An AI agent does a job that needs judgment, and it asks you before the step that costs money.

Most businesses need two of the three. The rest of this page shows you which two.

One job, three ways: a customer asks for a quote

The same email arrives on a Tuesday afternoon: “Can you quote me for twenty chairs like the ones I bought last year, delivered by the end of the month?”

Chatbotanswers

1. The email: quote 20 chairs
2. Answers from the price page
3. Stops. A person writes the quote.

Automationfollows rules

1. Quote form filled in
2. Template filled
3. PDF sent
4. “Like last year’s?” Not on the form: stops.

AI agentdoes the job

1. Reads the email
2. Finds last year’s order
3. Drafts the quote
4. Checks the discount rules
5. OK to send at this price? You tap OK.
6. Sends it and logs it

The chatbot reads it and answers from your price page: here are the chair prices. Then it stops. Someone still has to find last year’s order, work out delivery and write the quote.

The automation only starts if the customer used your quote form. The form fills a template, the template becomes a PDF, the PDF goes out. Fast and cheap, until a request arrives that the form did not expect, like “like the ones I bought last year”. Then it does nothing, or the wrong thing.

The agent reads the email, opens the order history, finds last year’s chairs, drafts the quote with delivery for the end of the month, checks the total against your discount rules, and then stops: “Quote ready. OK to send at this price?” You tap OK. It sends the quote and logs it.

Anthropic’s own guide names customer support as a good fit for agents for exactly this reason: it mixes conversation with tools that read customer data and order history (Anthropic, December 2024). OpenAI’s guide lists the jobs where an agent earns its place: decisions that need judgment, rule books that have grown too messy to maintain, and work that lives in emails and documents (OpenAI, April 2025).

Chatbot vs automation vs AI agent: the one-line definitions

One line each, and where each one stops.

	What it does	Where it stops	What it needs from you	The risk to watch
Chatbot	Answers questions from what you gave it	At the answer. It does not act.	Good pages and documents	A confident wrong answer
Automation	Runs fixed steps when a trigger fires	At any input the steps did not expect	A clear path, written once	Silent failure on the odd case
AI agent	Works towards a goal: looks, acts, checks, finishes	At the gates you set	A goal, tools, rules, and a person at the gate	Doing the wrong thing confidently and fast

OpenAI’s guide says it in plain words: apps that use a language model but do not let it run the work, “think simple chatbots, single-turn LLMs, or sentiment classifiers”, “are not agents” (OpenAI, April 2025). Google draws a ladder: a model alone, a model with tools, a model that plans many steps, then a team of agents, and at the top agents that build new tools for themselves (Google’s “Introduction to Agents”, November 2025, via PPC Land). Zapier, which sells both, puts it in one line: automation is “When this happens, do that”, and an agent is for “several steps and some need for adaptability” (Zapier, 2026). Anthropic draws the line between workflows, where the steps are fixed in code, and agents, where the model decides the path as it goes.

The decision chart

Four questions. Answer them in order.

1. 1Is the path the same every time?

   YesAutomation. No AI needed.

2. 2Does it only need to answer from things you already wrote down?

   YesChatbot

3. 3Does it need judgment, messy inputs or many small decisions, and can you check the result?

   YesAI agent

4For each step the agent takes: is a mistake expensive or hard to undo?Yes: a person approves that step first.

Anthropic’s advice to people who build agents is the same: start by “finding the simplest solution possible, and only increasing complexity when needed” (Anthropic, December 2024). The reason the chart matters is in Gartner’s June 2025 note: “only about 130 of the thousands of agentic AI vendors are real”; the rest were chatbots and automations with a new name, which Gartner calls “agent washing” (Gartner, via W.Media, June 2025). The chart tells you which name you actually need.

Where you stay in the loop: the approval gates

You do not approve every step. You approve the ones that matter.

OpenAI’s guide says: “Human intervention is a critical safeguard”. Its examples of high-risk actions that should wait for a person are “canceling user orders, authorizing large refunds, or making payments” (OpenAI, April 2025). Anthropic studied how people really supervise its agents and found the same pattern: “effective oversight doesn’t require approving every action but being in a position to intervene when it matters”. New users let the agent run on its own in about 20% of sessions; experienced users in over 40%; and even the experienced ones still step in on about 9% of turns (Anthropic, February 2026).

So the gates go where a mistake costs money or trust: paying or refunding, sending anything outside the company, and deleting. Everything else the agent does and you check afterwards. Real products work this way. Qonto, which runs finance admin for more than 600,000 small businesses, says of its agents: “Any action with a financial or customer relationship impact requires user confirmation” (Qonto). Supermetrics creates ad campaigns “in a paused state, not live, so nothing goes to market without a human actively choosing” (Supermetrics).

Keep the gates few. Anthropic found that people approve 97% of permission prompts in its coding tool, and in a controlled test with 1,053 paid professional testers, human review caught just 13.6% of dangerous commands (Anthropic, August 2026). A gate works when it is rare and about something that matters.

Time, effort and risk by type

In words, because the shape matters more than any number.

	Chatbot	Automation	AI agent
Time to set up	Quickest	A little longer: map the steps	Longest: the tools, the rules and the gates are the work
Effort to keep it running	Easiest: keep your pages up to date	Almost nothing, until the case nobody drew	You check its results and answer at the gates
Risk if it goes wrong	A wrong answer said with confidence	Fails quietly on the odd case	Biggest, because it acts. The gates keep it small.

A chatbot is the quickest to set up and the easiest to run; its risk is a wrong answer said with confidence, and that risk is real: in February 2024 a Canadian tribunal held Air Canada to a refund its chatbot had wrongly promised, after the airline argued the chatbot was “a separate legal entity” (Moffatt v. Air Canada, 2024). An automation takes a little longer to map out, runs for almost nothing, and fails quietly on the case nobody drew. An agent takes the longest to set up well, because the tools, the rules and the gates are the work, and its risk is the biggest, because it acts. The gates make that risk small.

AI agents in business: 2025, 2026 and a 2027 guess

More companies run agents every year, and most of them are still small, bounded agents.

Companies scaling AI agentsMcKinsey surveys, 2025 and 2026. Share of people who say their company is scaling agents.

• 2025
• 2026
• My guess

Enterprise apps with task-specific AI agentsGartner, August 2025. 2026 is Gartner’s forecast, so it is hatched.

• 2025
• Forecast

23%of people said their company was scaling an agent somewhere in the businessMcKinsey, November 2025

15xmore active agents in Microsoft 365, year on yearMicrosoft, May 2026

3 in 4business API conversations with Claude are automationAnthropic, January 2026

40%+of agentic AI projects cancelled by the end of 2027Gartner forecast, June 2025

See the numbers

What	Number	Source
Scaling an AI agent somewhere in the company, all respondents	23%	McKinsey, 5 Nov 2025
Experimenting with AI agents, all respondents	39%	McKinsey, 5 Nov 2025
Scaling AI agents, companies with US$1 billion+ revenue, last year	27%	McKinsey, 25 Aug 2026
Scaling AI agents, companies with US$1 billion+ revenue, 2026	40%	McKinsey, 25 Aug 2026
Scaling AI agents, smaller companies, both years	22%	McKinsey, 25 Aug 2026
Enterprise apps with task-specific agents, 2025	under 5%	Gartner, 26 Aug 2025
Enterprise apps with task-specific agents, end of 2026 (forecast)	40%	Gartner, 26 Aug 2025
Agentic AI projects cancelled by the end of 2027 (forecast)	more than 40%	Gartner, 25 Jun 2025
Growth in active agents in Microsoft 365, year on year	15x	Microsoft, 5 May 2026
Claude API conversations classed as automation	about 3 in 4	Anthropic, 15 Jan 2026

McKinsey’s November 2025 survey found that 23% of respondents said their organisation was scaling an agent somewhere in the business, and another 39% were experimenting (McKinsey, November 2025). Its 2026 survey, run from 4 May to 8 June 2026 with 1,719 respondents, shows the split opening: among companies with more than a billion dollars of revenue, the share scaling agents rose from 27% to 40% in a year; among smaller companies it stayed flat at 22% (McKinsey, August 2026). Microsoft counts the number of active agents in its Microsoft 365 ecosystem as up fifteen times year on year (Microsoft, May 2026). Gartner reported that under 5% of enterprise applications had task-specific agents in 2025 and expects 40% by the end of 2026 (Gartner, via UC Today, August 2025). Anthropic’s own usage data from January 2026 says about three quarters of what businesses send through its API is automation, where the model does the task rather than helping a person with it (Anthropic, January 2026).

For 2027, Gartner expects more than 40% of agentic AI projects to be cancelled by the end of that year, mostly from cost, unclear value or weak risk controls (Gartner, via W.Media, June 2025). My guess is the same in plainer words: the agents that survive are the small ones with one clear job and a person at the money step.

What an AI agent can do in a small business: quotes, orders, invoices, reports

The jobs that fit first are the ones that repeat every day and need a little judgment each time.

• Quotes and leadsRead the request, find the customer, draft the quote.Autotorino: 2,394 hours returned to its CRM team in 2025
• OrdersRead the order, check stock, update the record.Same shape: read, look up, act, ask at the gate
• InvoicesDraft them from the job, you approve them.Qonto: invoice creation in a third of the time
• ReportsPull the numbers, write the summary.Supermetrics: a ten-hour report in twenty minutes
• BookingsFind a time, hold it, confirm it.Same shape: read, look up, act, ask at the gate
• SupportAnswer, look up the order, act within the rules.Zendesk: a million agent runs in seven weeks

Quotes and leads from emails and calls: the Autotorino dealer group routes lead intake into its CRM and counted 2,394 hours returned to the CRM team in 2025 (Zapier). Invoices: Qonto’s agents cut client invoice creation to a third of the time, and one customer generated more than 500 monthly invoices in a single batch (Qonto). Reports: Supermetrics customers get a ten-hour marketing report done in twenty minutes (Supermetrics). Documents: EvenUp turns legal drafting that took eight to fifteen hours into a draft in about thirty minutes, and “an attorney reads, checks, and signs it” (EvenUp). Support: Zendesk’s custom agents ran a million times in the first seven weeks of early access, with each customer deciding which actions an agent may take (Zendesk). Bookings and orders follow the same shape: read the request, look up the record, act, ask at the gate.

This is what I build as custom AI apps for daily work: one job, the tools it needs, and the gates where you want them.

Watch: what an AI agent is, in seventy seconds

This is how a build like this runs with my team.