How to know which parts of your job to hand over to AI (the plumber rule)
You don't pay the plumber because he knows how to bang a pipe. You pay him because he knows where to bang it. The same line decides what AI can do for you and what you still have to.
Summarize this article with:
- The plumber rule from Suits, as a working framework. 'How' is mechanical. 'Where' is judgment. AI handles 'how' well. AI is bad at 'where' because 'where' requires context AI doesn't have.
- The 3-question test for any task: is the right answer knowable from data? Does failure compound? Is the cost of a wrong answer reversible? Three yeses means hand it to AI. One no means keep it human.
- The mechanical tasks AI should run for you: research, summarization, first-draft writing, structured data analysis, mechanical SEO checks, ad copy variants.
- The judgment tasks AI should not run for you: ICP decisions, channel-mix calls, attribution interpretation, brand voice strategy, hiring, pricing.
- The hybrid pattern: AI for the 80%, human for the 20%, plus the editing pass between them that prevents AI's confident wrong answers from shipping.
The wrong question is 'can AI do my job?'. The right question is 'which parts of my job is AI good at?'. The plumber rule answers it cleanly. AI bangs pipes. You decide where to bang them. Get the split wrong and you either underuse AI (slow) or overuse it (you ship the wrong answer fast).
You don't pay the plumber because he knows how to bang a pipe. You pay him because he knows where to bang it.
That line is the cleanest working framework I have for thinking about AI inside an SMB.
The how-to-bang-the-pipe layer is mechanical. Run the audit. Crawl the site. Compress the image. Draft the email. Summarize the doc. These are tasks where the right answer is knowable from data and the steps are repeatable. AI handles them well, and that pattern is exactly what the 2023 Harvard / BCG study of 758 consultantslabeled the "inside-the-frontier" zone, where AI lifted productivity 25% with 40% higher quality.
The where-to-bang-the-pipe layer is judgment. Pick the channel. Position the offer. Decide whether to ship the redesign or kill the project. These are tasks where the right answer depends on context AI doesn't have. Your specific ICP. Your specific competitive position. Your specific risk tolerance. AI gives you a confident generic answer when these require a specific informed one.
The split is the whole game. Find the split for every task in your business. Run the mechanical side through AI. Keep the judgment side human.
The 3-question test
For any task you might hand to AI, run these three questions. Three yeses means AI. One or more nos means keep it human.
1. Is the right answer knowable from data?
A task is data-knowable if running the same inputs through the same logic produces the same output. "Compress this image to under 200KB" is data-knowable. "Decide whether to compress images more aggressively" requires judging your specific quality bar and audience expectations, which is not.
Most operational tasks are data-knowable. Most strategic tasks are not.
2. Does failure compound or stay local?
A failure compounds if shipping the wrong answer makes the next 10 decisions worse. Picking the wrong channel for 6 months compounds. Drafting a mediocre email blast stays local.
AI is fine for local-failure tasks. AI is risky for compound-failure tasks.
3. Is the cost of a wrong answer reversible?
A wrong subject line on one email costs you that email's open rate. A wrong pricing decision costs you 6 months of underpriced contracts. The first is reversible. The second is not.
AI for the reversible. Human for the irreversible.
The mechanical-tasks list (AI handles these)
From running 50+ engagements, here are the tasks where AI saves us 60 to 90% of the time without compromising quality:
- Research summarization. Read 20 articles, return a 1-page summary. AI is faster than reading.
- First-draft writing. Outlines, briefs, internal docs. AI gets you from blank page to 70%.
- Mechanical SEO checks. Missing meta, duplicate titles, thin content. AI walks the punch list.
- Ad copy variants. Take the winning hook, produce 20 variants in the same voice.
- Structured data analysis. Pivot tables, gap analysis, anomaly detection in CSV exports.
- Translation and localization. First-pass language conversion. Human proofs.
- Code refactors and basic feature work. Inside Cursor, with senior review.
- Meeting transcripts and summaries. Save the 30 minutes of post-meeting writeup.
- Email drafts. Templates, replies, sequences. Founder edits to add voice.
- Slide outlines and first drafts. Designer or founder polishes.
The judgment-tasks list (keep these human)
From the same 50+ engagements, here are the tasks where handing to AI hurts more than it helps:
- ICP decisions. Who is the customer, who is excluded. AI gives generic answers.
- Channel-mix strategy. Which two channels to over-invest in. AI cannot read your specific competitive position.
- Attribution interpretation. The numbers don't speak for themselves; someone has to interpret what they mean. AI will confidently misinterpret.
- Pricing. Especially price increases. The right move depends on customer relationship history AI doesn't have.
- Hiring. AI screening reads as cold and high-error. Founders should hire founders' hires.
- Brand voice strategy. AI can apply a voice. AI shouldn't define it.
- Crisis response. Customer complaints, PR issues, internal team conflicts. AI tone is wrong.
- Investor conversations. The judgment, framing, and trust-building can't be drafted.
- Customer interviews (the actual interview). AI codes transcripts; it does not converse.
- Killing a project. The decision to stop is harder than the decision to start. Human only.
The hybrid pattern
Most tasks split. A marketing audit has both mechanical and judgment components. A redesign has both. A product launch has both. The pattern that works:
- AI for the 80%. Mechanical tasks. Speed matters more than nuance.
- Edit pass. Human reads everything AI produced. Flags the confidently-wrong outputs. Cuts the generic phrasing.
- Human for the 20%. The conclusions, the recommendations, the calls. These are what the work actually gets paid for.
- Ship. The hybrid output is faster than human-alone and better than AI-alone.
A worked example: my own writing workflow
Take this article. Here is how it split.
AI handled: outline generation from the title, draft of the 3-question test framework, first-pass draft of the mechanical-tasks list, FAQ candidate questions, related-reads cross-link suggestions.
I handled: the opening hook, the Harvey Specter quote choice, the position-taking sentences in each section, the specific numbers (50+ engagements, 60-90% time savings, 80/20 split), the founder-voice anchors, the final editing pass.
Time split: AI did roughly 60% of the keystrokes. I did 40%. The 40% was the part that made it sound like me. The 60% was the part that would have taken 4x as long without AI.
Frequently asked questions
How do I decide if a task is mechanical or judgment?
What if my whole job is judgment? Should I worry about AI?
Can junior people use AI for everything to speed up their learning?
What about creative work — can AI do that?
How do I tell when AI is confidently wrong?
What's the cost of getting this split wrong?
Related reads
- How to learn AI the right way. The 90-day curriculum that builds the calibration you need to spot the split.
- The Foxpro lesson. The skills that transfer across tool eras vs the ones that don't.
- How to run a digital marketing audit using AI. The split applied to one specific workflow.

Maddy
Maddy runs every WeActive8 engagement personally. Nine years working on growth across SMB and funded-startup stacks. Builds the 8CRM, Team8s, 8Host, and 8Automations products.