Agentic Workflows: When Not to Use Them

Most business processes don't need an agentic workflow.

That might be counter-intuitive for an AI consultant to write, but it is the conclusion that keeps surviving contact with real operations.

Automation wins tend to come from deterministic pipelines with a few well-placed AI calls and explicit human gates, not from autonomous agents deciding their own next steps.

Agents earn their keep in a narrower band than the current AI enthusiasm suggests, and knowing where that band ends is worth more than knowing how to build one.

This is the decision framework we use before any agent gets built: what agentic workflows are, the five signals that say "do not build one here," and where they pay off.

What is an agentic workflow?

An agentic workflow is an AI system where the model decides the sequence of steps at runtime: it plans, calls tools (search, code, databases, other systems), evaluates the results, and loops until it judges the task done.

The defining feature is model-directed control flow: in a traditional automation a person decided the steps in advance; in an agentic workflow the model chooses them while the work is happening.

That puts agents at one end of a spectrum with three distinct tiers:

1. Deterministic automation: fixed steps, no model in the control flow. A script, a scheduled job, a Zapier or Make scenario.
2. AI-assisted steps: a fixed pipeline where one or more steps call a model (classify this email, draft this summary), but the pipeline's shape never changes.
3. Agentic workflow: the model owns the control flow, deciding which tools to use, in what order, and when to stop.

Each tier up buys flexibility and costs you predictability, auditability, and token cost per run. The mistake we see most is starting at tier 3, the exciting tier, for work that tier 1 or 2 handles better.

This isn't a contrarian position: Anthropic's own engineering guidance on building agents opens with the advice to find the simplest solution possible and add complexity only when the task needs it.

When should you not use an agentic workflow?

The short answer: when the steps are knowable in advance, when errors are expensive and quiet, or when nobody owns verification. If any of the five signals below applies, build something simpler first.

1. You can write the steps down. If a competent person can list the procedure ("pull the report, filter to X, recompute Y, flag gaps over Z"), a deterministic pipeline will do it faster, cheaper, and identically every time. Agents suit work where the path varies per case; an undocumented procedure is still a procedure.

2. Errors are costly and hard to notice. Agent mistakes rarely announce themselves; a wrong judgment at step 2 compounds silently through step 9 and arrives looking finished and believable.

If a bad output reaches a customer, a regulator, or your accounting before anyone would catch it, the process needs deterministic rails or a human gate, not more autonomy.

3. Volume is high and variance is low. An agent run is many model calls, and a loop that searches, reads, and re-evaluates burns more of them as context accumulates. Thousands of near-identical tasks a day is the shape where a fixed pipeline with one model call per item wins on cost and speed.

4. You cannot define "done." Agents need a verifiable end state to terminate against: the test passes, the data reconciles, the claim matches the source. Give an agent an open-ended goal with no checkable finish line and it will produce something confident rather than something correct. In my own work with Claude, the runs that paid off most started with a light, checkable goal and let the model pick the path.

5. Nobody owns verification. The measured wins in AI process automation come with a human engaged as an error-catcher; every number below shares that shape. Put the reviewer's name and hours on the plan before anything gets built. Remove the gates and what you have is unaudited output at scale.

The pattern behind all five: agents trade predictability for flexibility. That trade is only worth making when the flexibility gets used on every run.

Where do agentic workflows actually pay off?

They pay off where the inputs vary case by case, each step requires judgment, the output can be verified, and a wrong answer costs review time rather than real damage. Research and triage work is the ideal fit. One client case: years of vendor documents in inconsistent formats, an agent that extracts and flags its doubts, and a person who verifies the flagged rows before anything downstream consumes them.

A worked example: a growth marketing agency I work with needed more SEO content without adding headcount. They built an agentic content engine with Claude Code. It researches and drafts, takes its tasks from ClickUp, reports to Slack, and stops at a human validation checkpoint at every step: a person approves each output before the workflow advances. The checkpoints are the workflow.

Our own Amazon recovery service follows the shape too: agentic steps read and match remittance evidence, and the claim that goes out is assembled by fixed rules and checked by a person. We wrote about the failure pattern on the other side of this line in why AI pilots fail at mid-market companies.

What do the measured numbers say?

My largest first-hand number comes from a tier 2 pipeline. At a travel company where I led product and commercial, we replaced a defective email parser for flight itinerary imports with a Llama LLM call, plus a second prompt for the roughly 5% of edge cases the first pass missed. The parsing error rate came in under 2%. The organization logged 5,900+ hours a year saved across all of its automation; the parser replacement was a subset of that total. Both figures are internal measurements, never independently verified.

The agency measured a 75% reduction in employee time on content creation from the engine above. One vendor review of human-in-the-loop editorial pipelines puts realistic cycle-time gains at 25 to 40 percent and credits the human gates for holding quality.

On the fear that AI-assisted work gets penalized: a vendor study, an Ahrefs analysis of 600,000 pages, found roughly 86.5 percent of top-ranking content used some AI assistance, with near-zero correlation between AI use and search penalties. What Google penalizes is scaled output with no human editorial judgment. Beyond content, removing the humans is the risk.

The five-minute decision test

Five questions, answered in writing, decide it. A blank on any one means the process is not ready for an agent, whatever the diagram says.

1. Can someone list the steps today? If yes, build the pipeline, not the agent.
2. What is the worst plausible error, and who would catch it, and when?
3. How many model calls does one run cost at your real volume?
4. What is the verifiable definition of done?
5. Who reviews the output, and is that review on their calendar or just in the diagram?

If the answers are "no, nobody, too many, unclear, and nobody," you have your architecture decision, and it is not an agent. If the answers are solid, an agentic workflow may be the highest-value system you build this year.

The test is older than agents. At the same travel company, the fraud control everyone recommended was 3D Secure. I sized what it would prevent: six chargebacks across three quarters, worth about $3.2K, at two airlines in high-3DS-adoption geographies. The cost-benefit said no, so we did not build it. Apply the same test to an agent: size what it would prevent before anyone builds it.

We run this assessment as part of a scored AI readiness diagnostic: data, workflows, team, and governance, with a prioritized use-case map and an honest list of what not to automate yet. If you are weighing where agents fit in your operation, a 45-minute working session on your actual process list is the fastest way to find out. Get in touch to book one.