The 60-second test: should this workflow become an agent?

When teams start using Avery NXR, the most common question after the first couple of agents is: "What should I automate next?"

The space of possible agents is huge. Not all workflows are good candidates. We've watched many users automate things they shouldn't have (and skip things they should have).

Here's a 60-second test we use internally to evaluate whether a workflow should become an agent. It's not perfect. It's good enough to make most decisions correctly.

The test

Answer these 5 questions about the workflow:

1. Does this workflow happen at least once per week?

If yes (or daily, or multiple times per day): proceed. If no: stop. Agents are best for recurring workflows. One-off automation rarely pays back the setup cost.

2. Does the workflow have a predictable input shape?

If yes (forms, emails with consistent structure, scheduled triggers, etc.): proceed. If no: stop. Agents need consistent inputs to produce consistent outputs.

3. Does the workflow have a clear "done" state?

If yes (output is created, message is sent, file is saved, etc.): proceed. If no: stop. Workflows without clear completion criteria are hard to automate reliably.

4. Could I describe the workflow in writing to a junior employee in <500 words?

If yes: proceed. If no: stop. If you can't describe it concisely, you probably don't have a clear enough mental model to configure an agent for it.

5. Would I be okay if 5% of executions had to be reviewed by a human before final action?

If yes: proceed — this becomes the agent's behavior, and the agent is a good fit. If no (need 100% autonomous, no review): the workflow may not be ready for agent automation in 2026. Wait, or design around the constraint.

If you get to "proceed" on all 5 → strong candidate for an agent.

If you stop at any question → either not a good fit, or you need to refine the workflow before automating.

Why these specific questions

Question 1 (frequency) filters out one-off work. The setup cost of an agent (15 min for a template, 30-90 min for custom) doesn't pay back for tasks done once or twice.

Question 2 (input shape) filters out unstructured chaos. Agents need consistency. Workflows that have wildly different inputs each time produce inconsistent outputs.

Question 3 (done state) filters out open-ended exploration. Agents are for completion-driven workflows, not for novel reasoning where "done" is hard to define.

Question 4 (describability) filters out workflows you don't actually understand. If you can't describe it to a human, you can't configure an agent to do it.

Question 5 (5% review tolerance) filters out workflows where 100% autonomous accuracy is required and you can't accept any review overhead. Most workflows can handle 5% review. Some can't.

Workflows that pass the test (and you should automate)

Common workflows that pass cleanly:

→ Meeting follow-ups (weekly+, predictable input, clear done, describable, review OK) → Resume screening (varies, predictable, clear done, describable, review essential) → Support ticket triage (daily, predictable, clear done, describable, review fine) → Daily pipeline digest (daily, scheduled trigger, clear done, describable, review fine) → Invoice processing (regular, predictable, clear done, describable, review fine) → Competitor monitoring (weekly, scheduled, clear done, describable, no review needed for monitoring) → Customer onboarding sequences (every new customer, predictable, clear done, describable, review optional)

These are the kinds of workflows that templates (Anna, Sophia, Marcus, Priya, Carlos, Yuki, Liam) cover. They pass the test by design.

Workflows that fail the test (and you shouldn't automate, yet)

Common failures:

One-off project work. "I need to research this market segment once for our board meeting." Question 1 fails. Use Claude or ChatGPT for one-offs. An agent is overkill.

Strategic decision-making. "I want an agent to decide which features to prioritize." Question 3 (clear done state) and question 4 (describability) both fail. Strategic decisions don't decompose into agent-able steps.

Creative brainstorming. "I want an agent to help me brainstorm campaign ideas." Question 3 fails. Brainstorming doesn't have a "done" state — it has good ideas and bad ideas, and the assessment is human.

Complex unique communications. "I want an agent to draft my response to this complex partnership offer." Question 1 (frequency) and question 2 (input shape) fail. Unique communications need human judgment.

High-stakes irreversible actions. "I want an agent to send our customer renewal emails automatically." Question 5 fails. Customer-facing communications usually need 100% review, not 95%.

Edge cases worth talking about

"It happens at least weekly, but each time is different." This is the most common edge case. The frequency is there but the input variability is high.

Solution: see if you can find the consistent SUBSTRUCTURE within the variability. Maybe the inputs vary but the OUTPUT structure is consistent. In that case, build the agent around the output structure.

"It's predictable but only happens 2-3 times per year." Frequency fails.

Solution: don't build an agent. Use a checklist or runbook. Agents are overkill for low-frequency work.

"It's done once a week but I'm the only one who does it, and I'll be doing it less in 6 months as we grow." Future-state frequency.

Solution: think about whether the workflow will need to be done by multiple people in the future. If yes, build the agent now and let it scale. If no, manual work is fine.

"It passes the test but I don't trust an agent to do it." Question 5 in disguise.

Solution: build it but require 100% human review for the first month. Gradually relax review as trust builds. Don't try to be 100% autonomous on day 1.

Using the test for prioritization

Beyond pass/fail, the test produces a rough priority:

→ Highest priority: workflows that pass + happen daily + take >30 min per execution + you do them → High priority: workflows that pass + happen weekly + take 1-2 hours per execution → Medium priority: workflows that pass + happen monthly + take a half-day → Low priority: workflows that pass but are quick (10 min) and infrequent

We'd build agents in roughly this order. Start with highest impact, work down.

What this saves

Teams that use this test (or something similar) avoid two common mistakes:

Mistake A: Building agents for the wrong workflows. This wastes setup time, produces unreliable agents, and burns trust in the platform.

Mistake B: Missing workflows that should be agents. This leaves value on the table.

The 60-second test takes a minute per workflow. It's fast enough to evaluate 10-20 candidates in a single session. The portfolio of "agent-able workflows" emerges from the screening exercise.

What we tell new users

If you've installed Avery NXR and built 2-3 agents from templates, the next step is to identify what to build next.

Don't browse the platform looking for inspiration. Look at YOUR work. List every recurring task you do. Run each through the 60-second test. The ones that pass become your agent backlog.

This grounded approach beats "what cool agents could I build" because it ensures the agents you build will get used. The pain you have is the pain the agent solves. The agent gets adopted because it solves real pain, not because it's clever.

The principle

The biggest mistake in AI agent adoption isn't picking the wrong platform. It's picking the wrong workflows.

The 60-second test is a small tool to avoid the big mistake.

Use it before building. Use it before pitching new agents to your team. Use it when teammates suggest "what if we built an agent for X."

The test takes a minute. It saves hours of wasted setup. It also saves the cost of having an agent that nobody uses sitting in your platform reminding you of a misjudgment.

→ avery.software — Free Desktop tier. The platform that lets you build agents fast — and the 60-second test that helps you build the RIGHT agents.