AI Task vs. Workflow: Run the Work Shape Test Before You Rebuild the Same Request
AI Task vs. Workflow: Run the Work Shape Test Before You Rebuild the Same Request
Say you run a small software company, and every Friday you ask AI for a summary of what your three closest competitors changed that week: new prices, new features, anything they announced.
By the fourth Friday you notice how much of the job is explaining things you've already explained. You name the three competitors again and paste in the pages to check. You remind it that you decided to ignore one competitor's enterprise tier, because you don't sell to that market. You say again that your sales lead wants price changes at the top, since she updates the comparison sheet on Monday. And you describe, one more time, what counts as a change worth mentioning.
The request is the same every week. Everything you worked out around it disappears when the chat ends.
We call that repeated rebuilding the reconstruction tax. You've now explained the competitor list, the enterprise-tier decision and your sales lead's prices-first format four Fridays running, and they still live in your memory or an old chat, with nowhere for the work to pick up where it left off.
The obvious fix is to put the whole thing on a schedule. Sometimes that's right. It can also mean running last month's method on this week's facts, or sending a report nobody has read in a month.
The Work Shape Test is four questions that help you avoid both mistakes. What actually repeats? What has to stay connected? What's exposed if the work is wrong? What could stop being true before the next run? The answers point you to the smallest structure that holds the work together: a task, a project, a recurring task or a workflow.
What's the difference between an AI task and an AI workflow?
An AI task is one piece of work with one finish line: you brief it, the result comes back, and you judge it in one review. A memo comparing your prices with your competitors' before Tuesday's pricing meeting is a task. So is bigger or repeated work, as long as one brief and one review hold it together.
A project is one finite outcome with connected work along the way. Changing your public price could be one: the decision, then the page edit, then a word to your sales lead so the comparison sheet matches.
A recurring task is a task on a schedule, judged against the same standard each time, like a Friday summary your team only reads.
A workflow keeps the method as well as the output. Once the Friday report feeds the comparison sheet, the method is which pages to check and in what order, the enterprise tier you skip, what your sales lead gets, and who checks each price before the sheet changes.
Automation is a separate decision, about how a single step runs. Pulling the competitors' pages could run on its own. Deciding whether to answer a competitor's price cut shouldn't.
Where the rebuilding hides
Go back to the Friday report. When you start it from scratch, you're rebuilding five things:
- the outcome, and what a finished report looks like
- the sources, including which ones are current and which you've stopped trusting
- the decisions you've already made, like skipping the enterprise tier
- the coordination: your sales lead's Monday update, and who decides whether a price change needs a response
- the review standard, meaning what you check before the report goes out
Some of that setup is real work, though. A competitor's new product has to be researched from scratch, and prices have to be checked again every week. Calling that a tax would give you an excuse to skip the judgment the work needs. The tax is the part you rebuild even though nothing about it changed.
Coming back to work costs something. Across three experiments, the longer and more demanding an interruption was, the slower people were to pick up what they'd set aside. Those were lab studies of people resuming a task, not of anyone re-briefing an AI. The Friday report gets set down for a week at a time, and each restart begins with the same explaining.
The Work Shape Test
1. What actually repeats?
How often a request comes back matters less than what will still be true about it next time. For the Friday report, ask these separately:
- Does the same kind of outcome come back each week?
- Do the steps, and the rules for what counts as a change, still apply?
- Are the sources still the right ones?
- Does the review standard still hold?
If the outcome repeats but you're still changing the method every week, keep it as a task for now. You're still working out how the job is done. Once the steps and rules have settled, they've earned a place somewhere more durable than your memory.
Google's Site Reliability Engineering book makes a related point about toil, the operational work its engineers try to reduce. It describes toil as work tied to running a production service that tends to be "manual, repetitive, automatable, tactical, devoid of enduring value" and that grows as the service grows, then adds that "not every task deemed toil has all these attributes." Repetition is one trait on that list, and the Friday report has it. What makes the report worth keeping, though, is that its competitor list and its rules for flagging a change have stopped moving.
2. What has to stay connected?
Count the inputs, decisions, permissions and people that shape what happens next, not the number of steps. For the Friday report, ask:
- Which pages have to be pulled before the summary can be written?
- Does a price change come to you before it reaches the comparison sheet?
- Can the three competitors be checked at the same time?
- Who decides when a price is unclear or a page has gone missing?
If the report is a summary your team reads and nothing more, it has few dependencies. If your sales lead updates the comparison sheet from it on Monday, and any price change comes to you to decide whether to respond, the report is part of something bigger. It still looks like a single task, but two people's decisions now depend on it.
3. What's exposed if it's wrong?
A wrong price in an internal summary costs a correction on Monday. The same wrong price copied into the comparison sheet your sales team sends prospects is a public mistake, whether the report stays a task or becomes a workflow. Either way, someone needs to check each price against the competitor's own page before the sheet changes, and they need to know they're allowed to stop the update. Writing "human review" into the process doesn't help much if nobody knows what the reviewer is supposed to look at.
4. What could stop being true?
For the Friday report, the competitor list and the rules for flagging a change might last for months. The prices won't last a week. Fast-changing context can still belong in a workflow, as long as the workflow's first job is to check it again. Ask:
- Which source is the right one for this run?
- Has a decision or permission expired?
- Is the priority still the same?
- Who owns the report now?
If your sales lead moves to another role and someone new takes over the comparison sheet, the report still runs on Friday, but the person it was shaped for is gone.
Match the structure to the work
Pick the lightest structure that keeps what the work needs. Anything heavier has to be looked after: someone keeps the competitor list current and notices when the comparison sheet changes hands.
| Work shape | Where to start | What the structure keeps |
|---|---|---|
| One result, few dependencies that matter | Standalone task | One brief, one finish line, one review |
| One finite result with connected work or important handoffs | Project | The decisions along the way, dependencies, owners and coordination |
| One contained output on a schedule, with a steady review standard | Recurring task | The cadence, the unit of work and the acceptance standard |
| A connected method that repeats and has settled | Workflow, or a project that recurs | Inputs, sequence, handoffs, exceptions and how it's reviewed |
| High stakes or fast-changing context | Whichever structure fits above, plus evidence, a refresh step, approval, a stop condition and a way to recover | Human authority and current context |
The Friday report could go either way. On Supanova, it can start as a standalone task set to run every week. Once it feeds the comparison sheet and your pricing decisions, make it a project and set that to repeat weekly. The finished report comes back for your review, and that's where the price check goes.
A workflow isn't the same as automation
Decide what the Friday report needs to keep before you decide which steps software should repeat. Google's SRE Workbook, written for teams that run production systems, says "you rarely want to literally transcribe a human workflow into a machine workflow" and, for problems with many edge cases, suggests "a partially automated approach as an interim step toward full automation." For the report, a first step could be software pulling the three competitors' pages while a person still checks every price before the sheet changes.
Before you automate any step of the Friday report, make six things visible:
- the outcome: what changed at the three competitors, prices first
- the steps that stay the same: which pages, in what order
- the inputs that change: this week's prices and announcements
- the exceptions, and when to stop: a page that's down, a price you can't confirm
- who's allowed to let it go ahead: whoever checks prices before the sheet changes
- the review standard: what counts as a change worth mentioning
If the method still isn't clear, and another run is safe, run it again as a supervised task. That run is how you find out what's worth keeping.
Three requests from the same week
Picture the same company in the same week.
Before Tuesday's pricing meeting, you need one memo comparing your prices with your three competitors'. It happens once, it doesn't depend on much, and the public evidence will go stale quickly. Make it a standalone task with source links and the date each price was checked. Keep the decision you make in the meeting. The method can go.
The Friday report is the second request. The competitor list and the rules for flagging a change stay the same, the prices have to be checked again, and your sales lead works from it on Monday. Make it a workflow whose first step is refreshing the sources and whose last step is recording what changed.
After the meeting, you decide to change a price on your public pricing page. It's a small change you'll rarely make, but customers will see it. Make it a task or a small project with a named owner, the evidence behind the decision, an explicit approval before anything changes, and a plan for putting the old price back, including what happens for anyone who paid the new one in the meantime. Being public calls for those safeguards, but it doesn't make the change something you'll do every week.
Won't this create bureaucracy?
It can. A workflow you didn't need adds upkeep and delay, which is why the test often points back to a task. To decide, compare the whole cycle. For the Friday report, put four Fridays of re-explaining the competitors, the enterprise tier and the prices-first format against the work of keeping the competitor list current. If the list barely changes, that upkeep is small.
Does the process have to be perfect first?
No. It needs a stable core. On the Friday report, the competitor list and the enterprise-tier decision settle early, while working out which pages to trust for each competitor may take longer. Keep the settled parts, keep checking the rest by hand, and supervise the early runs. You can write down a brief, a record of decisions or a list of trusted sources long before you automate a single step.
Start with next Friday
Before you open a new chat next Friday, run the four questions on the report. What about the outcome will still be true? Which decisions and handoffs should carry over? What happens if a price in it is wrong? Which sources, permissions and owners need checking again?
Whatever survives those questions is worth keeping somewhere the work can find it. The rest you check fresh each week, or let go.
When you're ready to hand that work to an autonomous workforce, start here.