When not to use an AI agent: nine cases

Nine cases where an AI agent is the wrong tool, what to use instead (delete, a script, a form, a person, one model call), and a test that settles each one.

By Nolli research, The Nolli team · Oct 20, 2026 · 9 min read

Short answer

Skip the agent when the step should not exist, a script can follow the rule, the agent cannot reach its inputs, you cannot define correct, a wrong answer is costly and unreviewed, the work is a relationship, volume is too low, cost or delay is too high, or you cannot test it first. Use a deletion, script, form, person or single model call instead.

On this page

You get: Nine cases, each with what to use instead and a test
Read: Anthropic’s and OpenAI’s agent guides, plus our four-bucket method
Checked: October 19, 2026

Key takeaways

Five of the nine cases end with a person doing the step, which is the honest result: an agent fits only the narrow band of work that needs judgment you can check.
Anthropic and OpenAI both say to start with the simplest thing that works, and OpenAI adds that a deterministic solution may suffice when the use case does not clearly need an agent.
Every case has a test you can run on one real example in an afternoon, and where the test is ours and not a published one, the text says so.

Skip the agent in nine cases. Each has something better to use instead, and each has a test that tells you in an afternoon whether you are in that case. “An agent is not right here” is easy to feel and hard to prove to a team that has already bought the licenses, so the tests are the point.

An agent, in OpenAI’s practical guide, is a system that uses a model to control how a workflow runs, with tools to act. A single model call or a classifier is not one. Anthropic’s Building effective agents says to find the simplest solution and add complexity only when needed, and that this “might mean not building agentic systems at all.” It also says agentic systems often trade latency and cost for better task performance. That page is from December 2024 and flags its tooling advice as dated, so we use only its advice on when to use agents. We read both pages on October 19, 2026, and tested no product.

The order below is our four-bucket method turned into reasons to stop: delete, code, person, and only then an agent. Where a case has no published test, we say so and give ours.

Nine questions, in the order to ask them

Ask them top to bottom. The first one that says stop decides, and you go to the thing in its note instead of an agent.

Before you build an agent, the first “stop” wins

01Should the step exist?If nothing breaks without it, delete it.
02Is the rule fixed?If it fits as if-then on one page, write a script.
03Can it reach its inputs?If not, add a form or move the data first.
04Can you say what correct is?If not, a person keeps it until you can.
05What does a wrong answer cost?Costly and unreviewed: a person decides, the agent drafts.
06Is the work a relationship?If yes, a person. The agent prepares the brief.
07Does it pay back?Run cost and upkeep against hours saved.
08Is one model call enough?If cost or delay is high, use a single call or a router.
09Can you test it first?If not, wait. If yes, run shadow mode.
10Then an agentStart with one narrow step.

Our reading of the four-bucket method, Anthropic’s “Building effective agents” and OpenAI’s practical guide. The order is ours: cheapest and most certain answers first.

Cases 1 to 3: the step, the rule and the inputs

1. The step should not exist

A workflow has a re-export, a chase for a missing update, a handoff that exists because two systems do not talk. An agent does these faster, and they are still waste. Michael Hammer argued this in the Harvard Business Review in 1990, and the title says what to do: “Don’t Automate, Obliterate”. A recent number, with its limits: Varick Agents’ CEO told The Startup Ideas Podcast that a redesign of one client’s accounts-payable process took it from 17 steps to 7, and raised the share of invoices handled with no manual touch from 18% to 87%. That is one firm’s report on one client, not a benchmark.

Use instead: delete it. The test: ask what breaks if the step vanishes, and who notices. If the answer is “nothing, and nobody”, you were about to build an agent for waste. Ask about the step as it should run, not as it runs today. One disagreement with OpenAI: its guide tells you to prioritize workflows that have resisted automation. That is a fair filter for case 2, and a poor place to start. Some workflows resist automation because they are a pile of handoffs.

2. The rule is fixed

If a person applies the same rule every time, a script applies it better: same result on every run, easy to test, close to free. OpenAI reserves agents for workflows where rule-based approaches fall short, and names three signs: nuanced judgment, rulesets so tangled that updates are costly or error-prone, and heavy reliance on unstructured data. If the case does not clearly show them, it says, a deterministic solution may suffice. The field behaves that way. In Zapier’s analysis of 20 million live workflows (data as of June 2026), 4.2% include a step done by an AI assistant, and four jobs that move information make up 81%: capturing a lead, sending a notification, creating a record, updating a record. Varick Agents says the systems it ships run five to ten deterministic steps with one or two model calls.

Use instead: a script, a rule, or a plain automation tool. The test: write the rule as if-then on one page. If it holds no “use judgment”, “as needed” or “appropriate”, you have a script. If the page balloons into exceptions nobody can maintain, that is OpenAI’s second sign, and the step may earn an agent.

3. The agent cannot reach its inputs

Watch a person do the task once. They read the order in one system, glance at an email thread, remember that this supplier is slow, and ring someone. The agent gets only what it was handed. Anthropic describes agents as models using tools in a loop, getting results back from the environment at each step to judge progress. Agent Process, as we reported in our SOP post, says to send a missing input to a person rather than let the agent invent a success. A missing API is not an automatic stop: OpenAI notes that agents can drive screens through computer-use models when a legacy system has none. A missing fact is a stop.

Use instead: a form that captures the input where it starts (a required field, a dropdown), or a plain automation that moves the data. The test: take the last real case and list every screen, file and person the human used. Mark each one the agent can read. An unmarked item that changes the answer is a gap. This test is ours, not a published one.

Cases 4 to 6: correct, costly and personal

4. You cannot say what correct looks like

Anthropic’s two worked examples of good agent work are customer support and coding, and it says both share clear success criteria and feedback loops: support success is measurable as a resolution, and code can be checked by tests. Agently says that if you cannot fill in the verification, you do not yet have an SOP. Our four-bucket rule says the same: an agent earns a step when it needs judgment and you can check the result.

Use instead: a person, until you can write the check. The test: write “done means ___” in terms someone else can verify, such as a status, a number or a matching record. Then collect past cases with known right answers. Our cost per task guide, adapting Varick’s process, suggests 20 to 50, including the awkward ones. OpenAI says to set up evals for a baseline before tuning anything. With no cases and no check, every output is an opinion.

5. A wrong result is costly and nobody reviews it

OpenAI says actions that are sensitive, irreversible or high-stakes should trigger human oversight until confidence in the agent grows, and gives canceling orders, authorizing large refunds and making payments as examples. It also says to rate every tool low, medium or high on read-only versus write access, reversibility, permissions and financial impact. Anthropic warns that autonomy brings higher costs and the chance of compounding errors. On consistency, Ramp Labs and Mercor’s APEX-Accounting gave models 160 realistic month-end close tasks: no model finished more than 2.6% of tasks correctly on all eight attempts. On four simpler, well-defined close tasks in a separate Mercor study, frontier models scored perfectly, so read the first number as a warning about long, messy work and not about every task.

Use instead: a person decides, and the agent drafts. Make the approval specific, as our four-bucket guide says, so the reviewer sees what was checked and which decision is theirs. The test: if this answer is wrong and nobody looks, what is the worst thing that lands, and how long to undo it? If the answer is a payment, a customer message or a record you cannot restore, and no named person sees it first, no agent.

6. The work is the relationship

A renewal talk with a customer of ten years, telling an employee their role is changing, a negotiation with a supplier. Our four-bucket guide keeps a person on negotiating, approving discounts and signing, and notes that an agent can still prepare the work. Anthropic’s support example works because success is a measurable resolution. Relationship work has no such measure. We found no guide that tests this case directly, so this one is our position.

Use instead: a person, with an agent writing the brief: the history, the open issues, what they asked last time. The test: would the other side judge the work differently if they learned who, or what, did it? If yes, a person does it.

Cases 7 to 9: money, time and testing

7. The volume is too low to pay back

An agent has a one-off cost: write the SOP, collect cases, test, name an owner. It has an upkeep cost, because AWS notes that an edit to an SOP the agent reads acts like an untested code change, so the tests run again. And it has a run cost on every task. We found no published break-even figure, and we would distrust one: it depends on your hours and your rates. The rule from our prioritization matrix post is that weekly run cost (cost per task times tasks per week) should sit well below the value of the hours the agent takes over. Add the one-off and upkeep to it.

Use instead: a person, a checklist, or a saved prompt someone runs by hand. The test: hours per month on the task times the loaded hourly cost, against build hours once plus upkeep plus run cost. If the first number is small, leave it. OpenAI advises re-scoring each quarter because high-effort ideas can become low-effort. Put a date in the calendar and move on. This test is ours.

8. Each task costs too much or takes too long

Anthropic says that for many applications, optimizing single model calls with retrieval and in-context examples is usually enough, and that an agent’s autonomy means higher costs. OpenAI’s definition leaves out single-turn calls and sentiment classifiers: they are not agents, and a small, faster model can handle simple retrieval or intent classification. Cost per task counts every attempt, the review time and the cost of fixing mistakes, not the price per token.

Use instead: one model call with the right context, a script around it, or a router that sends easy cases to a small model, which is Anthropic’s own routing example. The test: price one task all in with the cost per task calculator. Then run the same task as a single call on 20 past cases. If the single call passes your check as often as you need, you are done. If a person waits on the result, time it too.

9. You cannot test it before go-live

Anthropic recommends extensive testing in sandboxed environments, with guardrails, before an agent runs. Shadow mode is the practical form: the agent gets the same inputs as your team, sends and changes nothing, and is scored against what your people did. Agent Process, as we reported in the SOP post, adds that time passing is not evidence. Agree in advance how many false passes and false fails you can accept.

Use instead: a person does the work while you build the cases and run the agent beside them. The test: can you run it beside live work, with nothing sent or changed, and score it against what people did? If the only way to learn is to let it act, wait.

The nine cases, by what to use instead

A person is the main answer in five of the nine. That is not a verdict against agents. It shows how narrow the band is where an agent is the right tool: judgment needed, result checkable, mistakes cheap or reviewed, volume high enough to pay for the build.

What to use instead of an agent, by case

	Delete it	Script or rule	A form	A person	Plain automation	One model call
1. The step should not exist	fully	none	none	none	none	none
2. The rule is fixed	none	fully	none	none	partly	none
3. The agent cannot reach its inputs	none	none	fully	partly	partly	none
4. You cannot say what correct is	none	none	none	fully	none	none
5. A wrong result is costly, unreviewed	none	partly	none	fully	none	none
6. The work is a relationship	none	none	none	fully	none	none
7. Volume is too low	none	none	none	fully	none	partly
8. Cost or delay is too high	none	partly	none	none	partly	fully
9. You cannot test it first	none	partly	none	fully	none	none

none partly fully · Our reading of the sources named in each case. A full ball is the main thing to use instead, a half ball is a second option, an empty ball means we would not suggest it. “Plain automation” means a trigger-based tool that moves data between apps. This is our judgment, not a measurement.

Where costly mistakes meet fixed rules

Two questions place most steps: how bad is a wrong answer, and how fixed are the rules. Fixed rules go to a script. Costly mistakes go to a person. An agent gets what is left, the bottom left.

Cost of a wrong answer against how fixed the rules are

Our reading, not a measurement. The example steps come from our four-bucket guide (quotes, discounts, negotiating, payments, inbound email) and from Zapier’s list of common automation jobs. Where each sits is our judgment.

Five questions before you build

If you have time for five questions, ask these. Each one folds in the cases above.

Before you build an agent

Would anything break if this step vanished, and would anyone notice? If not, delete it.
Can the rule be written as if-then on one page, with no “use judgment”? If so, write a script.
Can the agent reach every input a person uses today, and can you write “done means ___” so someone else can check it?
If a wrong answer lands unreviewed, what is the worst result? Does the other side care who did the work? If either is bad, a person decides.
Does the run cost plus upkeep sit well under the hours saved, is one model call not enough, and can you test it beside live work first?

Where Nolli fits

Where Nolli fits, and where it does not

Nolli’s aim is to sort every step of a workflow into the four buckets (delete, plain code, agent, person) from how the work really runs, using your people’s work on company devices, your email and calendar, and the logs of the AI agents they use, and to show what each change is worth. That covers cases 1 and 2, and gives hours and volume for case 7. Its limits are real. It sorts only the steps it has seen. It cannot tell you whether work is a relationship (case 6) or what correct looks like (case 4), so your people still decide those. It does not replace a test, which is case 9. Your company decides what Nolli may learn from. You can do the sort by hand with the free four-bucket sorter. See the Decide pillar. No public price. Request a demo.

Request a demo

Questions people ask

When should you not use an AI agent?

When the step should be deleted, a script can follow the rule, the agent cannot reach its inputs, you cannot say what correct looks like, a wrong answer is costly and unreviewed, the work is a relationship, volume is too low to pay back, cost or delay per task is too high, or you cannot test it before go-live.

Is an AI agent better than automation?

No. Automation follows rules written in advance and gives the same output every time. OpenAI says agents suit workflows where rule-based approaches fall short, and that a deterministic solution may suffice otherwise.

What can I use instead of an AI agent?

Delete the step, write a script, add a form, keep a person on it, use a plain automation tool, or make one model call with the right context. Anthropic says optimizing single model calls is usually enough for many applications.

How do I know if a task is too risky for an agent?

Ask what the worst result is if the answer is wrong and nobody looks. OpenAI rates each tool low, medium or high on read-only versus write access, reversibility, permissions and financial impact, and wants human oversight for sensitive, irreversible or high-stakes actions.

Can an agent still help when a person must decide?

Yes. The agent can draft the answer or write the brief, and a person makes the decision. Make the approval specific, so the reviewer sees what was checked and which decision is theirs.

Sources

Erik S. and Barry Zhang. “Building effective agents.” Anthropic, December 19, 2024 (read October 19, 2026).
“A practical guide to building agents.” OpenAI, accessed October 19, 2026.
“Identifying and scaling AI use cases.” OpenAI, accessed October 14, 2026.
Michael Hammer. “Reengineering Work: Don’t Automate, Obliterate.” Harvard Business Review, July–August 1990.
Greg Isenberg, with Vasuman Moza of Varick Agents. “Become a $1M/yr FDE (Full Course).” The Startup Ideas Podcast, October 1, 2026.
Daniel Kornum. “If AI is so great, why isn’t it working?” Varick Agents.
Tim Hey for Zapier. “The most common business automation use cases in 2026, according to 20 million live workflows.” Stacker, via AOL, September 21, 2026.
“APEX–Accounting: Testing AI on the work accountants actually do.” Ramp Labs and Mercor, 2026.
Aden Barton. “Human Baselines for Benchmarks: AI Now Outperforms Junior Accountants.” Mercor, October 1, 2026.
“How to Write Agent SOPs That Give AI Agents Space to Think.” AWS Builder Center, August 11, 2026.
Omar Ghandour. “How to Document SOPs for AI Agents.” Agently, April 15, 2026 (updated August 7, 2026).
“Best practices for process authors.” Agent Process, accessed October 16, 2026.

Plain-text copy: /blog/when-not-to-use-an-ai-agent.md