> ## Content Index
> Fetch the complete content index at: https://www.doitwithfriday.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI Agent vs Workflow: Which One Your Business Task Needs
- URL: https://www.doitwithfriday.com/blog/ai-agent-vs-workflow/
- Published: 2026-11-15T07:35:57.000Z
- Updated: 2026-11-20T14:00:44.000Z
- Description: A workflow follows steps you set in advance, and an agent picks its own. Here is how to tell which one a task needs and what each costs in money and in risk.
- Author: Kevin Anderson
- Tags: AI agents, Costs & ROI

In a workflow, you fix the order of steps and AI handles single steps, such as reading an email. In an agent, the model picks the steps, the tools and the end.

Anthropic's guide to [building effective agents](https://www.anthropic.com/research/building-effective-agents?ref=doitwithfriday.com) draws the same line and says to start from a simple solution, adding agent autonomy only once a task needs it.

[![](https://i.ytimg.com/vi/0DB4bOJGfO8/hqdefault.jpg)](https://www.youtube.com/shorts/0DB4bOJGfO8?ref=doitwithfriday.com)

## Who decides the next step

Workflows and agents can use the same AI model and the same tools. What separates them is who picks the next action: code written in advance, or the model while it works.

### What a workflow does

Anthropic defines workflows as systems where models and tools are "orchestrated through predefined code paths." Someone sets the order of steps before the first run. AI can still do real work inside a step, such as pulling the amount from an invoice or drafting a reply, but it cannot add a step, skip one or change the order.

### What an agent does

In the same guide, agents are systems where the model directs its own process and tool use. The agent works in a loop: it reads the request, picks a tool, looks at the result and decides what to do next, until it judges the job done or reaches a limit you set. Anthropic's tool-use documentation describes the moment of choice: with the default setting, Claude decides on each turn whether to call a tool or answer directly.

The model's choice arrives as a structured request, and your application runs the tool and sends the result back. So the business that builds the tool keeps the final say over what it may do.

### The difference at a glance

Five practical points separate the two in daily use.

__How workflows and agents compare in daily use__
| Point               | Workflow               | Agent                                     |
| ------------------- | ---------------------- | ----------------------------------------- |
| Steps               | Defined in advance     | Chosen by the model                       |
| Input that fits     | Structured, consistent | Messy, varied, open-ended                 |
| Running cost        | Lower and predictable  | Higher, varies per task                   |
| Testing             | Check each step        | Check outcomes across a set of real cases |
| Explaining a result | Follow the path        | Read the agent's trace                    |

## When a fixed workflow is the right choice

Anthropic says workflows offer "predictability and consistency for well-defined tasks." Four signs show that a small-business task is one of those well-defined tasks.

### You can draw the process on a whiteboard

If every step can be mapped before the first run, fixed steps do it the same way each time. Anthropic's prompt chaining pattern is meant for tasks that "can be easily and cleanly decomposed into fixed subtasks."

### The same input should take the same path

Invoice routing, lead assignment and appointment reminders follow rules. A customer who books on Tuesday should get the same reminder as one who booked on Monday, and a fixed path makes sure of it.

### The task runs at high volume

Every agent run pays for the model to decide the steps. When those steps never change, that spend buys nothing, and it grows with each run. A workflow pays for AI only in the steps that need it, and it can send work that is not urgent to discounted batch processing.

### Someone must explain each decision

When a client, an auditor or a staff member asks why a record went where it did, a workflow answers with the rule it followed. With an agent, the answer sits in its trace, and someone has to read it.

## When an agent earns its place

Anthropic suggests agents "for open-ended problems where it's difficult or impossible to predict the required number of steps." Four signs point to a task like that one.

### The next step depends on what the last one found

Researching a new supplier means following leads that differ each time: a website, then a certification, then a reference. A fixed path cannot know in advance which of those it will need.

### The input does not follow a format

Free-text requests, mixed document packets and open questions do not fit one path. An agent can read what arrived and decide which tool or source fits it.

### Several routes lead to a good result

A customer question might need the knowledge base, the order system or a person. If listing every branch would take longer than the task itself, the choice can go to the model, inside limits you set.

### The judgment is worth the extra cost

Anthropic's engineering team puts it in money terms for multi-agent systems: they need tasks where the value is "high enough to pay for the increased performance." The same logic applies to a single agent: if a person would spend an hour thinking through a case, an agent's higher cost per run can still be small next to that hour.

## The five patterns between one AI call and an agent

Anthropic's guide names five patterns that sit between a single model call and a full agent. Naming the one a task fits makes it easier to scope, test and price before anything is built.

### Prompt chaining

The task is split into fixed steps, each working on the output of the one before, with checks in between. For example: pull the details from a purchase order, compare them with the price list, then draft the confirmation.

### Routing, and what it saves

Routing sorts each input and sends it to the right follow-up, such as billing, scheduling or sales. The guide's example sends easy questions to a smaller model and hard ones to a more capable model. On Anthropic's [API pricing page](https://platform.claude.com/docs/en/about-claude/pricing?ref=doitwithfriday.com), Claude Haiku 5.5 costs $0.50 per million output tokens for prompts up to 100,000 tokens, against $10 for Claude Sonnet 5.5 and $20 for Claude Opus 5.5.

![Bar chart of Claude API output prices per million tokens: Claude Haiku 5.5 $0.50 for prompts up to 100,000 tokens, highlighted, Claude Sonnet 5.5 $10 and Claude Opus 5.5 $20.](https://storage.ghost.io/c/67/23/6723d566-f4a4-4c03-8b45-4f8f0948342b/content/images/2026/11/claude-output-price-per-mtok.webp)

Sending routine inputs to the small model cuts the output price per million tokens from $10 or $20 to $0.50; Haiku 5.5's price applies to prompts up to 100,000 tokens. Source: Anthropic, Claude API pricing, checked Nov 16, 2026.

Input tokens follow the same order, at $0.10, $2 and $4 per million. A router that sends routine messages to the small model and keeps the larger ones for hard cases pays far less per message, and it is still a workflow, because code decides where each message goes.

### Parallelization

Several model calls work on the same task at once, and code combines the results. A contract can be checked for payment terms, liability and renewal clauses as three separate calls, each with a narrow question.

### Orchestrator-workers and evaluator-optimizer

In orchestrator-workers, a central model breaks a task into subtasks it could not predict in advance and hands them out. In evaluator-optimizer, one call drafts and another reviews the draft against clear criteria, in a loop, until it passes.

Only these two give the model real say over what happens next. If a process fits prompt chaining, routing or parallelization, the answer is a workflow with AI steps.

## What autonomy costs to run

An agent's bill grows with each step it takes, because each tool call sends the conversation so far back to the model. Tokens, tool overhead, lost discounts and waiting time make up the gap.

### Tokens per task

Anthropic's engineering team reported in June 2025 that, in its data, agents used about 4× more tokens than chat interactions, and multi-agent systems about 15× more. In its [research system write-up](https://www.anthropic.com/engineering/multi-agent-research-system?ref=doitwithfriday.com), token usage by itself explained 80% of the variance in performance on BrowseComp, a test of browsing agents.

![Bar chart of token use relative to a chat interaction, from Anthropic's data: chat 1×, an agent about 4× and a multi-agent system about 15×, highlighted.](https://storage.ghost.io/c/67/23/6723d566-f4a4-4c03-8b45-4f8f0948342b/content/images/2026/11/agent-tokens-vs-chat.webp)

In Anthropic's data, an agent used about 4× the tokens of a chat and a multi-agent system about 15×. Source: Anthropic, "How we built our multi-agent research system" (June 13, 2025), checked Nov 16, 2026.

In Anthropic's research system, then, better results came partly from spending more tokens. For a task that runs a few times a week, that spend can stay small. For one that runs thousands of times a month, the multiplier lands on every run.

### What a run is billed for

On the Claude API, a tool call is billed as ordinary tokens, and the bill covers more than the question and the answer. The tool definitions sent with each request, the tool\_use and tool\_result blocks and a tool-use system prompt all count. Anthropic's tool-use page lists that system prompt at 286 tokens for Claude Sonnet 5.5, added to each request that uses tools.

### Discounts that only fixed steps can use

The Batch API processes requests asynchronously at a 50 percent discount on input and output tokens, and Anthropic's cost advice is to use it for tasks that are not time-sensitive. A nightly workflow step that summarizes the day's support tickets qualifies. Anthropic's hosted agent service, Claude Managed Agents, has no batch mode, because its sessions are "stateful and interactive," and it adds $0.08 per session-hour of running time on top of tokens.

### Time the customer waits

Anthropic's guide warns that agentic systems can trade latency and cost for better task performance. An agent makes a model request for each round trip to a tool, so the wait grows with each tool it uses. That matters when a customer is on the phone or a chat window is open, and matters less for a report that runs overnight.

## Where an agent's freedom turns into risk

Letting the model pick the steps also lets it pick wrong ones. The guide names "the potential for compounding errors," and the risk shows up in three places that fixed steps keep narrow.

### When the model fills in a gap

Anthropic's tool-use documentation says that when a request lacks a required detail, Claude Opus is much more likely to ask for it, while Claude Sonnet might ask but "might also infer a reasonable value." Its example shows the model calling a weather tool with a city and a unit the user never gave.

![The Claude API documentation's example of a tool_use block in a terminal window: type tool_use, an id, the name get_weather and the input line with location "New York, NY" and unit "fahrenheit" highlighted and tagged "values Claude guessed".](https://storage.ghost.io/c/67/23/6723d566-f4a4-4c03-8b45-4f8f0948342b/content/images/2026/11/tool-use-inferred-values.webp)

Asked "What's the weather?" with no place, the model can call the tool with values it inferred. This behavior is not guaranteed, especially for more ambiguous prompts and for less capable models. Source: Anthropic, Claude API docs, Tool use with Claude, checked Nov 16, 2026.

The documentation adds that this behavior is not guaranteed, especially for ambiguous prompts and less capable models. In a workflow, a missing field stops the step or sends it to a person. In an agent, that check has to be built in: make the field required, or tell the agent to ask.

### Errors that feed the next step

In a fixed path, a bad output from one step is caught by the next check or fails in a known place. In an agent, a wrong tool result can become the basis for the next choice. Anthropic's guide says agents need "ground truth" from the environment at each step, such as a tool result, to judge their progress, and it recommends extensive testing in sandboxed environments with guardrails.

### Excessive agency, in OWASP's terms

The OWASP Top 10 for LLM Applications traces excessive agency to three causes: excessive functionality, excessive permissions and excessive autonomy. A workflow step can only do what it was built to do, so an agent is where all three need limits set on purpose.

## The hybrid: a workflow with one agent step

A business does not have to pick one shape for a whole process. A workflow can carry the work from system to system and hand the single step that needs judgment to an agent.

### The workflow as the backbone

The workflow receives the input, moves it between systems and enforces approvals. The agent handles only the step that needs judgment, with a narrow set of tools, then hands its result back. The rest of the process stays testable step by step, and only one step needs outcome testing across a set of real cases.

### Example: inbound sales email

A workflow picks up each new message, logs it in the CRM and routes it. One step in the middle, working out what the person wants and drafting a reply, goes to an agent that can look up the account and the product catalog. The draft returns to the workflow, which sends it to a sales rep for approval. Fixed where it can be, flexible only where it must be.

### Moving a step from agent to workflow

The handoff also works in reverse. If an agent's traces show it taking the same path on nearly every run, that path can become fixed steps, which cuts the model calls spent on deciding.

## How to tell what a vendor is selling

The word agent appears on products of both shapes. One question about who controls the sequence sorts the label from the product, and three more show whether it is ready for real work.

### The question that cuts through the label

Ask whether fixed logic or the model decides what happens next. If a person or a rule set defines every step and the AI only reads, classifies or drafts inside them, the product is a workflow with AI steps. That can be exactly what the task needs.

### Three questions before you buy

What does the tool do when it is unsure? Can you see a log of each step it took? Which actions wait for your approval? A product that cannot answer all three clearly is not ready for work that touches customers or money.

## Four questions to decide

Four answers, about the steps, the judgment involved, the cost of a mistake and the volume, point a task toward a workflow, an agent or a workflow with one agent step.

### The questions

1. Can you list every step in advance? If yes, start with a workflow.
2. Does any step need reading, research or judgment? Use AI inside that step, or an agent for it.
3. What happens if the AI gets it wrong? The higher the cost, the more you want fixed steps and approvals around it.
4. What is its monthly volume? High volume pushes toward workflows; rare, complex tasks can carry agent costs.

### Reading the answers

If every step can be listed and the task runs at volume, build a workflow. If one step needs reading or judgment, put AI or an agent in that step and keep the rest fixed. If the whole task changes shape from case to case and its value covers the extra tokens, an agent fits, with approvals around anything costly to undo.

## FAQs About AI Agents vs Workflows

Short answers on whether a workflow counts as AI, what an agent costs per run, whether one shape can turn into the other, and whether you need a framework to build either.

### Is an AI workflow still real AI?

Yes. A workflow can use AI models for reading, classifying and drafting. What it does not do is let the model choose the sequence of steps.

### Are agents more expensive to run?

Per run, yes in Anthropic's data: agents used about 4× the tokens of a chat interaction. For low-volume tasks that need judgment, the extra cost can be worth it.

### Can a workflow become an agent later?

Yes. A project can start as a workflow, then hand one difficult step to an agent once the rest is reliable.

### Do I need a framework to build either one?

No. Anthropic's guide says frameworks can help you start quickly, then advises fewer abstraction layers as you move to production.

## Start with the right shape

A task with clear steps starts as a workflow. A task that needs judgment in one step adds an agent there, with limits, tests and approvals around it, and keeps the rest fixed.

If your process has clear steps, our [workflow automation](https://www.doitwithfriday.com/services/workflows/) service is the place to start. If one step needs real judgment, our [AI agents and skills](https://www.doitwithfriday.com/services/agents-skills/) service builds that piece with tests and limits. Not sure which fits? [Tell us how the work runs today](https://www.doitwithfriday.com/contact/) and we will map the right shape with you.

Last checked: November 16, 2026, against Anthropic's "Building effective agents" (December 19, 2024) and "How we built our multi-agent research system" (June 13, 2025), the Claude API pricing and tool-use documentation, and the OWASP Top 10 for LLM Applications entry on excessive agency. Prices and models change; check the linked pages before relying on a figure.

The workflow and agent distinction, token findings and prices come from Anthropic's published guidance and documentation, and the risk causes from OWASP, checked November 16, 2026.