AI agents vs workflow automation vs chatbots: how to pick

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

Workflow automation follows steps you wrote in advance. A chatbot answers questions in a conversation. An AI agent is a language model that picks its own next step, calls tools such as your CRM or calendar, and keeps going until the task is done or it hits a limit you set.

The line that matters is who decides the path. Anthropic's engineering team draws it the same way in Building effective agents (opens in a new tab) (December 2024): workflows are "systems where LLMs and tools are orchestrated through predefined code paths", while agents direct their own process and tool use. The same post recommends the simplest option that works, adding complexity only when it pays. We agree, and most of this article is about how to tell when it pays.

The second column is the one most buyers miss. It is also the one that fits most business requests.

When does workflow automation win?

Workflow automation wins whenever you can write every step and every rule before the work arrives. If the input is structured and the decision is a rule ("deals over $50,000 go to a senior rep"), no model is needed.

The advantages are boring and large. A rule-based run costs a fraction of a cent on most platforms, produces the same output every time, and when it breaks it breaks loudly: a field is missing, an API returns an error, the run stops and someone gets an alert. You can read the logic on a canvas, and a new hire can maintain it.

Good fits include moving records between systems, scheduled reports, approvals with clear thresholds, onboarding checklists, payment reminders on a fixed cadence and syncing a web form to a CRM and a spreadsheet. We compare the 3 common platforms in n8n vs Make vs Zapier.

Automation loses the moment the input is language. A rule that looks for the word "cancel" in an email misfires on "please don't cancel my order". Teams then stack keyword rules until nobody trusts the routing. That is the signal to add one AI step, and a poor reason to throw the workflow away.

It is also our standing rule: if rule-based automation does the job, we build that instead, because it is cheaper to run and easier to hand over.

Where does an AI step inside a workflow fit?

Use an AI step when the path is fixed but one step needs to read or write language. The workflow still decides what happens next. The model only classifies, extracts or drafts inside a box you defined.

Typical AI steps:

• Classify an inbound email into one of a fixed list of queues, with a confidence score.
• Pull 12 named fields from a supplier invoice PDF and return them as structured data.
• Summarize a 40-minute call transcript into a CRM note with next steps.
• Draft a reply that a person approves before it is sent.

Because the output is constrained (one label from a list, fields that must validate, a draft that waits for approval), you can test it before launch. Run 200 past emails through the classifier, compare the labels with what your team actually did, and you know the error rate. Low-confidence results go to a person. That is human-in-the-loop design, and it is the cheapest form of safety you can buy.

Cost sits close to plain automation: one model call per run, usually a fraction of a cent to a few cents. In our pricing, a single workflow with an AI step costs $4,000 to $10,000 to build. The AI automation page has the detail.

When is a chatbot or assistant the right choice?

A chatbot is right when people need answers and the answers live in documents or a knowledge base. The person drives the conversation. The bot's job is to answer correctly or say it does not know.

The useful version is an assistant grounded in your own content through retrieval-augmented generation: it searches approved documents and answers with links to the passages it used. Staff questions about policy, product specs, past proposals and support macros are the common fits. Our guide on how to build a knowledge assistant that cites its sources covers the build step by step.

The failure mode is a confident wrong answer, and the business owns it. In Moffatt v. Air Canada (opens in a new tab) (February 2024), a British Columbia tribunal held the airline responsible for its website chatbot's incorrect statement about bereavement fares. For any customer-facing bot, the lesson is practical: ground it in current policy, show sources, and send anything about money, refunds or eligibility to a rule or a person.

A chatbot that only answers does not complete work. If the next sentence from the user is "now change my delivery address", you need a workflow behind the chat or an agent.

When does an AI agent win, and how do agents fail?

An agent wins when the steps cannot be listed in advance, the task needs several tool calls with judgment between them, and the value per case justifies the extra build and monitoring. Investigation, multi-system lookups and messy back-and-forth are where agents earn their cost.

Examples that fit: working out why an order is late across a carrier portal, the ERP and an email thread, then drafting the customer reply; preparing a meeting brief from the CRM, the last 3 call notes and the prospect's website; reconciling a vendor statement where each mismatch needs a different lookup.

The model receives a goal and a list of tools it may call (function calling), picks one, reads the result and decides again. That loop is the power and the risk.

Four failure modes to plan for:

1. They take a plausible but wrong path and complete it confidently. Nothing errors. The output is just wrong.
2. They loop, calling the same tool again and again, and the model bill grows with every turn.
3. They follow instructions hidden in content they read, such as an email that asks for a thread to be forwarded elsewhere. This is prompt injection, and any agent that reads outside content is exposed to it.
4. Costs drift as prompts grow and edge cases add steps.

The market is learning this at some expense. Gartner predicted in June 2025 (opens in a new tab) that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. The same release estimated that only about 130 of the thousands of vendors marketing agentic products offer real agentic capability.

Our rules for agents in production: read-only tools first; write actions limited to ones that can be undone or that wait for approval; a hard cap on steps and spend per run; every run logged with its tool calls; and an evaluation set run before every change. The AI guardrails entry explains the controls. We build these as custom AI agents, from $15,000 to $90,000, with model usage billed at cost.

Five questions that pick the right build

Answer these in order. The answers usually point to one option before you reach question 5.

1. Can you write down the steps before the work arrives? If yes, start from a workflow. If the steps depend on what each lookup returns, keep going.
2. Is the input structured? Form fields and API records suit rules. Email, PDFs, calls and chat need a model somewhere.
3. Does the output go to a person or straight into a system? A draft for a person tolerates more error than a write to the ledger, the CRM or a customer's inbox.
4. What does a wrong result cost, and can it be undone? A mislabeled ticket costs a minute. A wrong refund or a sent email does not come back.
5. How many cases a month, and how different are they? 30 similar cases a month rarely justify any build. 3,000 similar cases suit a workflow. 3,000 cases that each need different lookups are where an agent pays.

Reading the answers:

How do cost and risk compare?

Plain automation is the cheapest to build and run, and agents are the most expensive on both counts. The risk profile flips too: automation fails loudly, while agents and AI steps fail without raising an error. Price the monitoring along with the build.

Running a system after launch is its own line item. Our automation care plans run $1,500 to $6,000 a month, and AI ops monitoring for model-based systems runs $1,500 to $5,000 a month with tooling at cost.

Three worked examples

These are illustrative examples, not client case studies. The volumes are assumptions chosen to show the reasoning.

Example: sales call follow-up

A 40-person B2B firm runs about 60 sales calls a week and wants the CRM updated and a follow-up email drafted after each one.

The steps never change: the call ends, the transcript arrives, a CRM note is written, a follow-up is drafted, the rep approves it, the email goes out. Only 2 boxes need language, the summary and the draft. That is a workflow with 2 AI steps and an approval. No agent required. The sales call follow-up blueprint shows the build.

Example: invoice intake

A finance team receives 1,500 supplier invoices a month as PDF attachments.

Extraction is a language problem. Everything after it is rules. A model or a document-processing service pulls the fields, validation checks the totals and the purchase order match, clean invoices post to the accounting system, and exceptions go to a person with the reason attached. The path never changes, so there is no case for an agent. See the invoice processing pipeline and intelligent document processing.

Example: support triage

An online retailer gets 4,000 tickets a month. Assume 70% are order status, returns or address changes.

Classification and routing are a workflow with an AI step. Order status is a lookup and a template, which is also a workflow. The rest ("my order arrived damaged and I was charged twice") needs several lookups and judgment, and that is where a narrow agent can draft a resolution for a support rep to approve. Refund rules stay in deterministic logic, never in model judgment. The support knowledge assistant covers the answering side.

When should you combine them?

Most systems worth building combine at least 2 of the 4: a workflow as the spine, AI steps where language appears, and an agent only at the open-ended edge. The workflow keeps the parts that must be predictable predictable.

Three shapes come up again and again:

• Chat front end, workflow back end. An assistant answers questions. When someone asks for an action, it calls a workflow that does the action the same way every time.
• Workflow that calls an agent. The workflow handles intake, logging and approvals, and hands the one messy step to an agent with a step limit.
• Agent that calls workflows as tools. Instead of raw write access to your CRM, the agent gets a "create follow-up task" tool that is itself a tested workflow with validation.

The last pattern is the one we reach for when an agent is truly needed. It shrinks what the model can get wrong to the choice of tool, and the tool enforces the rules. Model Context Protocol servers are a common way to expose those tools, and n8n, Make and Zapier all offer MCP support in some form as of October 2026.

When is none of these the answer?

Sometimes the right recommendation is to build nothing yet. We say so on scoping calls, even when it costs us the project.

• The process is not defined. If 3 people do the task 3 ways, automating it locks in one of them at random. Write the process down first. This is usually the first week of an AI readiness audit.
• Volume is too low. A task done 20 times a month at 5 minutes each is under 2 hours of work. A $4,000 build will take years to pay back.
• Nobody will own it. Every system on this page needs an owner who reads the alerts. Without one, it decays and people stop trusting it.
• The decision is regulated or high-stakes. Credit decisions, medical advice, legal conclusions and hiring decisions need human judgment and often legal review. A model can prepare the file. It should not make the call.
• A vendor feature already does it. Your help desk, CRM or accounting system may already have the routing or extraction built in. Turn that on before paying anyone to build.

If you want a second opinion on a specific process, a 30-minute scoping call is usually enough to tell which of the four it is.