The integration you didn’t need takes three weeks to undo. The pre-flight check that would have prevented it takes ten minutes.
I’ve wired AI into workflows that produced no net benefit. Twice in the last year. Looking back, the warning signs were visible before I started. Both times I skipped the questions that would have surfaced them. The cost wasn’t catastrophic – it was three weeks of fiddling, then a quiet rollback, then the slightly embarrassed realisation that the workflow I had before was the one I should have kept. Or at least just tinker with, instead of throwing out the window.
Five questions, asked honestly, would have saved both rounds.
The questions below are not about AI capabilities. They are about whether AI belongs in this specific workflow, in this specific solo business, this week. The answers don’t change much when the next model ships. The questions still hold.
Question 1 – Would I do this job by hand if AI didn’t exist?
The first filter is the most uncomfortable, which is why it gets skipped.
If the answer is no – you wouldn’t do this job manually, in your own time, with your current capacity – then AI isn’t fixing a real workflow. It’s hiding the fact that the job isn’t worth doing.
A common example: “I should write five LinkedIn posts a week.” You wouldn’t do this manually. You don’t have the time or the inclination. AI feels like the unlock that makes it possible. What it actually does is automate a job that didn’t earn its place in your week. The output is a thinner version of the manual version that wasn’t going to happen anyway. Three months later you have 60 LinkedIn posts and no engagement, because the work was generic the moment AI got involved.
The reason this filter works: AI is most useful when it accelerates a job you already do and value. It’s least useful when it substitutes for a job you don’t do and shouldn’t.
This isn’t a productivity question. It’s a strategic one. The piece on what AI is actually good at for solo founders draws the same line at a category level. This question applies that line to a specific workflow on a specific Tuesday.
Question 2 – Can I verify the output cheaply?
The cost of AI in a workflow is not the cost of generating the output. It’s the cost of checking the output.
A workflow where verification is fast – a list extracted from a transcript, a reformatted summary of a document I wrote, a translation of my own text – is a workflow AI improves. I read the output in 30 seconds. If something is wrong, I fix it in another 30. The integration is net-positive.
A workflow where verification is slow – fact-checking AI-generated research, validating AI-suggested citations, confirming AI’s interpretation of a client’s email – is a workflow AI degrades. The verification takes longer than the work would have taken by hand. The integration looks productive on the input side and is net-negative on the output side. The 2023 Brynjolfsson, Li and Raymond paper Generative AI at Work found measurable productivity gains for AI-assisted customer-support agents, but the gains depended on the task having an immediately verifiable outcome. The same pattern shows up in solo work. Fast verification produces real gains. Slow verification produces apparent gains that disappear on review.
The honest version of this question: if the AI is wrong, how long does it take me to notice and fix it? If the answer is “longer than doing the task by hand,” the integration is the wrong move.
Question 3 – Will I use this every week, or once?
A wired-in workflow has setup cost. Prompts to refine, connections to configure, edge cases to handle, the slow build of trust in the output. That cost amortises over use.
A workflow you run weekly amortises the setup over 52 cycles a year. The hours of pre-flight planning and the first three rough weeks pay back many times.
A workflow you run once or twice does not amortise. The setup cost exceeds the savings, even if the savings per use are real. The solo-founder version of this trap is the slow accretion of one-off integrations that each took half a day to wire up and now sit unused, costing only the maintenance of remembering they exist.
The rule that works: wire in the recurring jobs. Use AI ad-hoc for the one-offs. The ad-hoc version of the one-off is faster, simpler, and doesn’t accumulate.
This connects directly to designing capacity instead of chasing it. Workflow integrations are a capacity cost as much as a capacity tool. The floor of integrations that earn their setup is smaller than most solo founders assume.
Question 4 – Does this job need my judgment or my hands?
AI is good at jobs that need hands – sorting, formatting, extracting, generating drafts. AI is bad at jobs that need judgment – deciding what matters, picking the angle, naming the position.
A workflow that requires hands-work AI handles is a workflow AI improves. The work was always going to come out the same. Letting AI do the mechanical part frees attention for what only attention can do.
A workflow that requires judgment-work AI handles is a workflow AI degrades. The output sounds done. The thinking it skipped surfaces as drift in your archive across the next six months.
The test: if you outsourced this job to a competent stranger who knows nothing about your business, your archive, or your audience, would the output be useful? If yes, AI is fine. If no, AI is not the right tool – not because it can’t produce output, but because the output won’t be the right shape for what you’re actually doing.
Most workflows contain both hands-work and judgment-work mixed together. The pre-flight question isn’t “does this workflow have any judgment in it.” It’s “where does the judgment live in this workflow, and am I keeping that part?”
Question 5 – What happens when the AI is wrong?
The last question is the one that protects the system from the integrations that look fine in week one and break in week six.
A workflow where wrong AI output is visible and recoverable – the headline is wrong, you write a new one – is safe. The cost of an error is contained.
A workflow where wrong AI output is invisible and propagates – a fact you didn’t catch is now embedded in a piece you’ve already sent to your list – is dangerous. The cost of an error compounds. By the time you notice, the wrong thing is downstream of multiple decisions.
The Anthropic research note on building effective agents makes a parallel argument at the systems-engineering level: the value of an AI workflow is bounded by your ability to detect when it’s wrong. The same applies to solo workflows. The integration is only as good as the verification step it includes. A verification step you can’t run reliably means the integration is taking on hidden risk you can’t see.
The honest version of this question: what’s the worst thing that happens if I don’t catch the error? If the answer is “an embarrassing public mistake” or “a wrong fact in my archive,” the verification has to be airtight before the integration goes live. If you can’t make it airtight, don’t wire it in. Use AI ad-hoc for that job instead, where the same risk exists but you’re actively watching for it.
When the answers don’t all come back clean
The five questions don’t have to produce a perfect five-for-five score for the integration to be a good idea. They have to surface what you’re trading off.
A workflow that fails on Question 2 (verification is slow) but passes the other four can still be wired in – if you accept that the verification cost eats most of the gain, and the speed-up matters anyway. A workflow that fails on Question 3 (you’ll only use it once) can still get the AI treatment, just without the wiring – run it ad-hoc, keep the prompt in a file, don’t build infrastructure.
A workflow that fails on Question 1 (you wouldn’t do it by hand) shouldn’t get the AI treatment at all. That’s the workflow whose existence isn’t earned. AI doesn’t fix the strategic problem. It hides it.
The cluster’s next piece on prompt patterns that survive the AI plateau covers what to do when an integration that passed the pre-flight stops producing the quality it used to. The pre-flight is the entry gate. The prompt patterns are the maintenance.
The rest of the AI & Workflows archive lives at the AI & Workflows category.
Ten minutes, five questions
The pre-flight is small on purpose. Ten minutes. Five questions. No flowchart, no spreadsheet, no integration audit.
The cost of running it is low. The cost of skipping it is three weeks of fiddling, followed by a quiet rollback, followed by the recognition that the workflow you had before was the one you should have kept.
Ten minutes. Five questions. Then you can wire it in.
Build Your Content Machine. A free 5-part email course on building an AI content system that sounds like you – the pre-flight discipline above is where it starts. Start the free course →
