The 30-Second Rule for AI Coworkers

Key Takeaways

• The 30-second rule: if a task takes you under 30 seconds, do it yourself. The cognitive cost of delegating to an AI coworker, reviewing the draft, and approving the action is more than 30 seconds. Anything under that threshold loses money.
• Recurring beats complex. A 5-minute task you do every Monday is a better candidate than a 60-minute task you do once a quarter. The runbook compounds; the one-off does not.
• The trust ladder has four rungs. Read-only queries graduate first, then internal-only drafts, then external drafts, then auto-execute on low-stakes actions. Skipping rungs is how teams end up rolling back a coworker after one bad email.
• Four task shapes pay back fastest. Recurring digests, multi-tool reconciliation, triage classification, and meeting-prep aggregation. Everything else is a maybe.
• Anything that touches money or customers stays manual. Auto top-up, refunds, plan changes, customer-facing emails. The cost of one wrong send is higher than the time saved by auto-execute.
• The wrong question is "can the AI do this?" The right question is "is this worth a runbook?" The model can do almost anything in 2026. The bottleneck is whether the workflow is shaped well enough to delegate.

A founder messaged me last week with a question I get every few days. "We just deployed Viktor in our Slack. What should I have it do first?" He was looking for a checklist. I gave him a question instead: "What did you do this morning that took longer than 5 minutes and you knew the whole time it was a waste of your senior brain?" He had three answers in 20 seconds. Pipeline reconciliation against HubSpot. Reading overnight Pylon tickets. Drafting his Monday revenue post.

That is the actual checklist. Not "what can the AI do?" The model can do almost anything in 2026. The real question is which of your tasks is worth the cognitive cost of delegating, reviewing, and approving. Most teams get this wrong on day one. They throw the AI at exciting problems (building a new product feature, drafting a strategy doc, replacing the social media manager) and end up frustrated when the output is mediocre. They miss the boring middle: the recurring work nobody on the team actually wants to do, the work that has a clear shape and gets done the same way every week.

This post is the decision framework. The 30-second rule, the trust ladder, the four task shapes that actually pay back, and the categories that should stay human. Use it the next time you are about to ask "what should the AI coworker do?"

The 30-second rule

The simplest filter: if a task takes you under 30 seconds, do it yourself.

Why 30 seconds is the floor

Delegating to an AI coworker has a fixed overhead. You write the request. You wait for the draft. You review the draft. You approve or correct. Even on the fastest workflow with the cleanest runbook, that loop is at least 60 to 90 seconds of your attention.

If the task itself was 20 seconds of typing, you just turned a 20-second task into a 90-second task and felt clever doing it. This is the trap most people fall into in the first week: delegating "send a Slack message to Marcus" or "add this to my calendar." The AI does it. You feel productive. You spent more time than just doing it.

The math, rounded

A typical knowledge worker spends a meaningful share of every week on repetitive tasks they recognize as low-value while doing them. The 30-second rule says target the tasks within that pool that are also at least a few minutes long. The Monday revenue digest. The Friday reconciliation. The morning ticket triage. Not "send a one-line Slack ping."

This is also why "AI saves you 10 hours a week" headlines are usually wrong. They count the task time saved without counting the delegation overhead. The honest number is closer to 4 to 6 hours a week for most operators, and only if the workflow shapes are right.

The trust ladder: four rungs

Once you have picked a task that clears the 30-second rule, the next decision is how much autonomy to give. The general shape is: the agent acts, the human audits. The trust ladder is the practical version.

Rung 1: Read-only

The AI can pull data and post a summary. It cannot write to any tool. Cannot send a message that creates a record. Cannot trigger an action.

This is where every workflow starts. The Monday revenue digest reads from Stripe and HubSpot, posts to Slack. If something is wrong, the worst case is a wrong number in an internal post and a "huh, that looks off" reply. No customer is affected.

Rung 2: Internal-only drafts

The AI drafts a Slack message, an internal Notion page, a Linear ticket. A human approves before it goes live. The blast radius is still inside the team.

Rung 3: External drafts

The AI drafts a customer email, a CRM update, an external-facing post. A human still approves. The blast radius now reaches a customer or partner.

Rung 4: Auto-execute

Drafts have been correct three to five times in a row. The action is low-stakes (internal post, calendar block, internal CRM tag). The AI runs without review.

What never graduates past Rung 3

In our experience, almost none of the AI mishaps we hear about are "the model went rogue." Almost all are a human flipping a workflow from draft to auto before it was ready. The trust ladder is the discipline that prevents that.

The four task shapes that actually pay back

Across hundreds of customer deployments, four task shapes consistently pay back in the first two weeks.

1. Recurring digests

The Monday revenue digest. The end-of-day customer success digest. The Friday pipeline summary. Each one runs on a schedule, pulls from two to four tools, posts to Slack with a clean format. The runbook for these is short, the success criteria is obvious, and the AI gets to Rung 4 (auto) within two weeks.

2. Multi-tool reconciliation

"Did every closed-won deal in HubSpot actually get a Stripe charge this week?" "Are all our Linear high-priority tickets reflected in the Notion engineering roadmap?" The AI is unreasonably good at cross-referencing. A human doing this manually has to switch tabs, copy IDs, hold context. The AI just queries both tools and posts the diff.

3. Triage classification

Overnight Pylon tickets need to be sorted into P0 / P1 / P2. New inbound leads need to be tagged ICP-fit / not-fit. New PRs need to be assigned a reviewer based on the file paths touched. The AI applies the rule consistently every time. A human applies the rule consistently for the first 10 items and then gets sloppy.

4. Meeting-prep aggregation

Before every customer call, you want a one-page summary: who they are, last touch, current ARR, open tickets, recent product activity. Building this manually takes 20 minutes. The AI builds it in 90 seconds, draws from CRM + product analytics + support tool + LinkedIn. You read for 60 seconds before the call.

Every weekday, scan my Google Calendar 24 hours ahead. For
every external meeting (any attendee outside our domain),
build a prep card and post to my DM the night before.

• Card structure:
• Company name + meeting purpose (from invite + recent emails)
• Stripe ARR (if customer) or Apollo enrichment (if prospect)
• Last 3 emails between any of our team and any of theirs
• Open Linear or Pylon tickets tagged with their account
• Any LinkedIn activity from the attendees in last 30 days
• Three suggested talking points based on the above

Post 8 PM the night before, in my DM. If a meeting was added same-day, post it 1 hour before the meeting. ```

This single workflow saves most operators 4 to 6 hours a week. Almost everyone we deploy with for under 50 employees runs some version of it inside the first month.

Task shapes that look promising but are not

The wrong fits are subtler than the right ones. These all sound like good candidates and consistently disappoint.

Strategy and writing from scratch

"Write our 2026 strategy doc." "Draft the keynote announcement." The AI can produce a 90% version fast, but the last 10% is where strategy actually lives, and you spend more time editing than you would spend writing. Use the AI for research aggregation and outline drafts, not the final artifact.

One-off complex projects

"Pull every customer who churned in Q1, classify by reason, and build a retention playbook." This is a great use of an AI assistant in chat. It is a bad use of a runbook because it does not recur. Recurring is where the runbook compounds. One-off is where chat is the right interface.

Anything dependent on tribal knowledge nobody wrote down

"Do the thing Lena does every Monday." If Lena cannot describe in five sentences what she does, the AI cannot do it either. This is not a model problem. This is a missing-runbook problem (covered in How to write a runbook for your AI coworker).

Tasks where the cost of an error is much higher than the cost of doing it manually

Sending an apology email to your biggest customer. Writing a board update. Approving a refund over $5,000. The AI can draft. A human always sends. There is no shame in keeping these on Rung 3 forever.

How to apply the framework on Monday morning

A practical sequence. Open a fresh Notion page. List every task you did this week that took longer than 5 minutes. Beside each one, mark:

• Recurring (R) or one-off (O)
• Tools touched (Stripe, HubSpot, Slack, Linear, Notion, Pylon, etc.)
• Cost-of-error (low / medium / high)

Now sort. Recurring tasks at the top. High tool count next (multi-tool reconciliation is the AI's sweet spot). Low cost-of-error promoted, high cost-of-error demoted to "draft only forever."

The first three runbooks you should write are usually obvious by the time you finish sorting. They are recurring, multi-tool, and low cost-of-error. The Monday digest. The pipeline reconciliation. The morning triage.

Write those runbooks. Run them at Rung 1 (read-only) for a week. Promote to Rung 2 (internal draft) the second week. Promote stable workflows to Rung 4 (auto) by week three. Do not skip rungs.

The wrong question and the right question

Most teams ask "can the AI do this?" The answer is almost always yes. The model is capable. The integrations exist. The tools are connected. That is not the bottleneck.

The right question is "is this worth a runbook?" Is the workflow shaped well enough to write down in five sentences? Does it recur often enough to repay the cost of writing it down? Is the cost-of-error low enough to graduate to Rung 4? If you cannot answer yes to all three, it is not a runbook candidate. Keep doing it manually, or do it in chat.

This is the difference between operators who get value from AI coworkers and operators who churn. Not the model, not the integrations, not the prompt-engineering. The discipline of asking the second question.

Frequently Asked Questions

Is 30 seconds really the right threshold?

It is the conservative floor. For some workflows the threshold is closer to 60 seconds (when the review step is more careful) or closer to 15 seconds (when the AI can run on auto without human review). Use 30 seconds as the default. If a task is under 30 seconds and not recurring, do not delegate.

What about creative work like writing a blog post?

Mixed bag. The AI is great at research aggregation, outline drafts, and FAQ generation. It is bad at the final artifact for any post where voice or strategy matters. Use it for the boring middle (gather sources, draft headers, generate counterpoints). Write the actual prose yourself.

How do I know when a workflow is ready to graduate to Rung 4 (auto)?

Three to five consecutive correct drafts in production, with a human reviewing every one. If any of the three are wrong, restart the count. The runbook is not ready until the AI has been right three times in a row on real data.

Should I delegate emails to the AI?

Drafts only, never auto-send. Customer-facing emails stay on Rung 3 forever. Internal Slack messages can graduate to Rung 4 once the runbook is stable. The asymmetry is the cost of an error: a wrong internal Slack message gets a "lol fix that" reply. A wrong customer email costs the relationship.

What if my workflow does not fit any of the four shapes?

It is probably not a great runbook candidate. The four shapes (recurring digest, multi-tool reconciliation, triage classification, meeting-prep aggregation) cover roughly 80% of high-value AI coworker workflows. If your task fits none of them, ask whether it is recurring at all, and whether the cost of the AI being wrong is acceptable. If the answer is no to either, keep it manual.

How does this differ from "AI agents" advice I read elsewhere?

Most AI-agent content is written by people who have never deployed one in production. It focuses on capability ("look what the agent can do!") not on shape ("which workflows actually pay back?"). The 30-second rule is shape-first. The capability is assumed; the question is what to point it at. For a comparison of agent products, see Viktor vs ChatGPT Agent.

Where do I start if I am brand new?

Pick one recurring task you do every week that takes 10 to 30 minutes. Write the runbook in five sentences. Run it at Rung 1 for one week. If the AI's output matches yours, promote to Rung 2 the next week. The whole loop takes three weeks. By week four you have a workflow on auto and an honest read on whether the AI coworker fits your team.

Closing thought

The teams that get value from AI coworkers in 2026 are not the ones with the best models or the most integrations. They are the ones who learn to ask the right question. The 30-second rule, the trust ladder, the four task shapes. None of it is exciting. All of it compounds.

Pick one workflow this week. Run it through the framework. If it clears all three filters (recurring, multi-tool, low cost-of-error), write the runbook. If it does not, keep doing it manually and pick a different one. The discipline is in the picking, not in the prompting.

For the runbook template once you have picked, see How to write a runbook for your AI coworker. For an operator's view on the first week, see The first 7 days with an AI coworker.