When to use AI agents, and when a simple workflow is better

You’ve probably had this conversation already. Someone saw a demo, or a vendor pitched one, and now there’s a proposal to “use an agent” for a piece of work: handling supplier emails, chasing unpaid invoices, answering customers. The demo was impressive. The real question is whether an agent is the cheapest reliable way to do that work, or whether a plain workflow with one AI step would do it better.

It’s worth asking early. In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls. Many of those problems start with choosing an agent for work that didn’t need one.

This guide gives you six questions to decide, the trade-offs to expect and the patterns that work in practice. If you need a refresher on what an agent is, start with What is an AI agent?

Autonomy is a dial, not a switch

“Agent or no agent” is the wrong way to frame it. There’s a whole range between a fixed rule and a fully independent agent, and most good systems sit somewhere in the middle.

How much freedom to give

1. Workflow Same path every time 1
2. AI workflow AI reads or writes inside fixed steps 2
3. Agent, person approves Plans and drafts, a person signs off 3
4. Agent within limits Acts alone on low-risk actions 4
5. Autonomous agent Plans and acts alone 5

← Fixed rules Full autonomy →

1. 1 Workflow Same path every time
2. 2 AI workflow AI reads or writes inside fixed steps
3. 3 Agent, person approves Plans and drafts, a person signs off
4. 4 Agent within limits Acts alone on low-risk actions
5. 5 Autonomous agent Plans and acts alone

Take a physiotherapy clinic:

• Fixed rule: an SMS reminder goes out 24 hours before every appointment. No AI needed.
• AI workflow: AI reads incoming patient emails and sorts them into rescheduling, questions about treatment and billing, then a workflow routes each one to the right person. The steps are fixed; AI only reads.
• Agent, person approves: for rescheduling requests, an agent checks the calendar, finds slots with the same therapist and drafts a reply. Reception approves it with one click.
• Agent within limits: the agent moves appointments on its own if it’s the same therapist in the same week. Anything else goes to reception.
• Fully autonomous: rarely sensible in a clinic, or in most companies, for anything that touches customers or money.

Our earlier post on workflows, AI workflows and agents describes the three main levels with many examples. This one is about choosing between them.

Six questions to ask before you choose an agent

1. Can you write the steps down?

Ask the person who does the work today to describe it. If they can write it as “if this, then that” and the list fits on a page, build a workflow. If the honest answer is “it depends: I look at X, then maybe Y, sometimes I call the supplier”, the task may need an agent.

2. Does the work vary, or only the input?

Messy input on its own doesn’t call for an agent. An invoice in an unusual layout is messy, but the steps after reading it are the same as always. That’s an AI workflow. Agents are for variation in what needs doing, not only in how the input looks.

3. What does a mistake cost, and can it be undone?

A wrong draft that a person catches costs a minute. A wrong payment, a cancelled order or an email to the wrong customer costs a lot more. The higher the cost and the harder it is to reverse, the more you need fixed rules or a person approving each action.

4. How many cases, and how fast?

Agents are slower and more expensive per case than workflows. That’s fine for 50 complex cases a week that each take a person 20 minutes. It’s a poor deal for 5,000 simple ones that need an answer within two seconds.

A rough check: multiply the cases per month by the minutes a person spends on each. Take an illustrative accounting firm with 120 clients, where preparing month-end questions takes 25 minutes per client. That’s about 50 hours a month, enough to justify an agent that does most of the gathering. Five cases a month at 10 minutes each is not.

5. Will you need to explain every decision?

In finance, healthcare, hiring and anything regulated, someone may later ask why a decision was made. A workflow takes the same path every time, so the answer is easy. An agent can be audited too, but only if you log every step, tool call and result. That has to be designed in from the start.

6. Do the tools and data exist?

An agent is only as useful as the systems it can reach. If the order system has no API (a way for other software to talk to it), or the documents live in someone’s personal inbox, fix that first. Often that work alone removes most of the manual effort.

Put the second and third questions together and most tasks fall into one of four boxes:

Where an agent fits

Cost of a mistake →

Workflow with checks

Agent drafts, person approves

Workflow or AI workflow

Agent can act alone

• 1 Payment runs
• 2 Invoice approval routing
• 3 Appointment reminders
• 4 Sorting incoming email
• 5 Delivery claims
• 6 Month-end client questions
• 7 Research before a sales call

How much the cases vary →

1. 1 Payment runs
2. 2 Invoice approval routing
3. 3 Appointment reminders
4. 4 Sorting incoming email
5. 5 Delivery claims
6. 6 Month-end client questions
7. 7 Research before a sales call

The trade-offs nobody shows in the demo

Same task, three ways to build it

	Workflow	AI workflow	AI agent
Model calls per case	None	One or two	Many, one per step
Time per case	Seconds	Seconds	Up to minutes
Same input, same path	Yes	Yes	No
Copes with unplanned cases	No	Partly	Yes
Easy to test like normal software	Yes	Partly	No
Typical failure	A rule doesn’t match	A field is read wrongly	A wrong plan or tool choice

Cost. An AI step in a workflow makes one model call. An agent makes a call for every step of its loop, and the text it sends grows with each step. Anthropic’s engineers reported in 2025 that agents typically use about four times as many tokens (the units models are billed in) as a chat interaction. At volume, that adds up.

Speed. Each round of the loop takes seconds. A task a workflow finishes in two seconds can take an agent a minute or more.

Reliability. Small error rates multiply. If an agent gets each step right 95% of the time, the chance it gets 10 steps in a row right is about 60% (0.95 multiplied by itself ten times). That’s illustrative arithmetic, not a measurement, but it explains why long autonomous chains disappoint. Fewer steps, better tools and checks between steps help more than a cleverer model.

Testing. A workflow given the same input takes the same path, so you can test it like normal software. An agent can take different paths on the same input. You need a set of real past cases, a run through all of them after every change, and someone reviewing the results.

Patterns that work

Start with a workflow, add one AI step

In most processes, most of the value comes from connecting systems and ending the retyping. Build that as a workflow first. Then add AI where a step needs reading or writing: pulling fields from a document, sorting an email, drafting a reply.

Let an agent handle only the exceptions

Rules and AI workflows handle the common cases. The unusual ones, the ones that land on someone’s desk today with “can you look into this?”, go to an agent. It gathers the facts and prepares a proposal. A person decides.

An agent only for the exceptions

1. 01 Request arrives Email, form or ticket
2. 02 AI sorts it Type, order number, urgency
3. 03 Rules handle known cases Order status, invoice copies
4. 04 Agent looks into the rest Gathers facts, drafts a resolution
5. 05 Person decides
6. 06 Done Reply sent, systems updated

The support team at a wholesale distributor is a good example. “Where is my order?” and “please resend the invoice” are handled by a workflow that reads the email, looks up the order and replies. A claim about a damaged pallet goes to an agent. It collects the order, the delivery photos and the carrier’s record, then drafts a resolution for a team lead.

Agent prepares, person decides

For anything with real consequences, the agent’s output is a draft: a reply, an order, a credit note, a shortlist. A person approves it. Over time you’ll see which drafts are always approved unchanged, and you can let the agent act alone on those. Our post on keeping people in the loop covers how to design that review step so it stays quick.

Give the agent narrow tools

An agent with a tool called run_any_database_query is a risk. An agent with find_order, check_stock and draft_replacement_order is predictable. Narrow tools limit what can go wrong. How AI agents use tools explains how to design them.

Good and poor candidates

Good candidates for an agent:

• Accounting firm: preparing month-end questions for clients by comparing bank transactions with the documents received. Every client’s gaps are different, and the output is a list a person checks before sending.
• Distributor: looking into delivery claims across the order system, delivery notes and carrier tracking. The path depends on what each record shows, and the agent only drafts the resolution.
• Recruitment agency: preparing candidate briefs from CVs, interview notes and past placements for a recruiter who makes the decisions. The agent reads and summarises; it doesn’t rank or reject anyone.
• Operations team: answering “why is this number different from last month?” by querying several reports and explaining the gap. Read-only tools mean a wrong turn costs time, not money.

Poor candidates, and what to use instead:

• Routing invoices for approval by amount: a rule.
• Appointment reminders: a rule.
• Sorting incoming email into five categories: an AI workflow.
• Approving payments: a rule plus a person.
• A task that happens five times a month: probably nothing, or a good checklist.

Agent-readiness checklist

If you can tick most of these, an agent is worth a pilot. If not, start one step lower on the scale.

Is this task ready for an agent?

• The steps genuinely vary from case to case
• Each case takes a person 10 minutes or more today
• Mistakes can be caught or undone Or a person approves every risky action
• The systems have APIs or exports The agent can reach the data it needs
• You have 30 or more real past cases To test with before and after every change
• One person owns the result And reviews the early runs
• You know the budget per case Model usage, time and review effort
• You’ve ruled out a workflow with one AI step

We build internal tools and automations at every point on that scale, from plain rules to agents with approvals. If you’re not sure where your process belongs, book a 30-minute call and you’ll leave knowing what to build first.