Use AI to question a workflow before you ask it to speed the workflow up. Here are six prompts, run in order on one worked example, and what to check at each step.
Before you automate a recurring task, use six prompts in order: name the job, list and sort the assumptions, argue against the task, redesign from a blank page, bring the constraints back, and pressure-test the result. A filled-in context sheet, a sorted assumption list and a four-question pre-automation check make each step usable, and the model drafts while a named person decides.
Most teams asked to use AI on a recurring task start by speeding up the steps they already have, and the steps themselves rarely get questioned. The weekly update gets drafted in half the time. The approval email gets a summary. The work moves faster, and the same people receive the same document they were already skimming.
An operations analyst who owns a weekly sales deck is a typical case. Thursday afternoon goes on matching account names across three exports, pasting figures into a slide template and writing a page of commentary. Told to use AI, the analyst tries a prompt that drafts the commentary. It works, the deck is finished earlier, and it still goes to the same readers, most of whom stop after page one.
The fix is an order of operations. Before you ask AI to improve a workflow, use it to question the workflow: what it’s for, what it assumes, whether it should exist, and what you’d build today. This guide gives you six prompts to do that, run in order on one worked example, with a note on what a good and a weak answer look like.
None of this is a new idea. Michael Hammer’s 1990 Harvard Business Review article carried the headline “Don’t Automate, Obliterate,"[1] and Jason Averbook has made a version of the same argument about AI: organizations add AI to broken or outdated processes and mistake faster execution for progress.[2]
One thread runs through all six prompts. The AI does the challenging, and a named person owns every answer.
First principles thinking sounds grander than the work needs. In practice it’s three questions, asked in order. What is this actually for? What’s true here that would stay true whatever we chose? If we were starting today, with today’s tools, what would we build?
Jensen Huang, NVIDIA’s chief executive, describes the habit in a Stanford Graduate School of Business interview: he goes back to first principles “all the time” and asks, given how things have changed, how he’d redo something.[3] He puts the test as a question about the underlying facts. Did physics change? Did gravity change? If not, the direction may still be sound. If the tools changed, the way you get there may not be.
That second half is the useful one for AI. Many workflows were shaped when drafting, matching and summarizing were slow for a person. A step that exists because “someone has to write it up” rests on an assumption AI changes. A step that exists because finance needs a signed-off number doesn’t.
AI helps with the questioning in four ways:
It’s weak in two places. It doesn’t know your facts: the contract term, the regulation, the thing your finance partner won’t move on. And it tends to agree with whoever is talking, which is why prompt 3 below is built the way it is.
The sequence this guide follows. Each prompt uses the previous answer, so run them in one conversation.
Every prompt below starts from the same one-page description of a piece of work. The prompts are only as good as what you paste in, and without context a model fills the gaps with plausible generalities, which gives you an answer that could be about anyone’s deck.
| Field | What goes in it |
|---|---|
| The work | Weekly sales deck |
| Who does it, and when | One operations analyst, Thursday afternoon, most of the afternoon |
| Inputs | A CRM pipeline export, a billing system export, and a forecast spreadsheet kept by the regional leads |
| Steps today | Pull the three exports. Match account names by hand, because the systems spell them differently. Paste figures into a slide template. Write a page of commentary from memory of what was said in calls. Save as PDF and email it. |
| What comes out | A long deck. The first two slides carry the numbers. |
| Who gets it | Regional leads, the sales director and a finance partner |
| What we think they do with it | Leads skim page one. Finance checks one table. The director reads it before the Monday forecast call. |
| Known pain | Account names don’t match. The commentary is thin. The figures are a few days old by Monday. |
| Constraints we know are real | Finance needs a signed-off number at month end. The CRM is the record for pipeline. |
| What we don’t know | Whether anyone reads past page one. Whether finance needs it weekly. |
Paste this above each prompt, with your own work in place of ours.
The last two rows are the ones people tend to leave blank. Fill them in first: prompt 3 leans on what you don’t know and prompt 5 on what you do.
Leave out customer names, employee details and anything your organization’s AI policy keeps out of these tools. If you’re unsure what that covers, ask whoever owns AI governance first. The prompts were tried on this sheet before they went in, but your output will differ with your context and your tool. Read it as a draft from a sharp colleague who hasn’t met your company.
The first prompt does one job: it makes you say what the work is for without describing a single step.
| Here is a description of a recurring piece of work: [PASTE THE CONTEXT SHEET] Write what this work is for in one sentence, without mentioning any of the current steps, tools or file formats. Then tell me who uses the result, what decision or action it supports, and what would happen if it stopped for a month. Where you’re guessing, say so, and list what I’d need to find out. |
A good answer names a person and a decision, such as helping the sales director agree the forecast before Monday. A weak one says something like “provides visibility into sales performance”, which has no decision in it. Reply: “Name the decision.”
Write down what the work is for before you describe how it’s done.
Once the purpose is fixed, the second prompt goes after everything you’ve been taking for granted. It asks for sorting as well as listing, because what comes next depends on which assumptions are open.
| Using the purpose we just agreed, list every assumption built into how this work is done today: about timing, format, who does what, who receives it, and which tools are involved. Aim for at least ten. Label each one FACT (true whatever we choose, such as a law or a contract), POLICY (a choice someone made and could change), HABIT (nobody decided it, it grew) or TOOL LIMIT (it exists because of how a system works today). Add one line on how I could check it. Don’t redesign yet. |
If the list is short or generic, add your known pain points to the context sheet and run it again.
| Assumption | Type | How to check it |
|---|---|---|
| The deck goes out weekly | Habit | Ask the director which decisions need a weekly view and which a monthly one |
| Account names have to be matched by hand | Tool limit | Ask the CRM and billing owners whether a shared account ID exists |
| The analyst writes the commentary | Policy | Ask who decided that, and whether leads add their own |
| Everyone gets the same document | Habit | Ask each group what they use |
| Finance needs a signed-off figure at month end | Fact | Confirm with the finance partner. This one stays. |
| The CRM is the record for pipeline | Policy | Check with the sales director. Changing it is a bigger decision than this deck. |
The type label is the model’s first guess. The person who owns the rule decides whether something is really a fact.
Several labels will be wrong, and the cost is uneven. A habit that’s really a contractual commitment gets cut by mistake, and a fact that’s really a habit rarely gets challenged. Check the fact and policy rows with whoever owns them before you go on.
Facts stay, policies need their owner, and habits and tool limits are open to change.
Asking whether a process should exist is the hardest question to put to your own team, which makes it a good job for a tool with no stake in the answer. The catch is that the tool has a lean.
A 2023 paper led by Mrinank Sharma found that five AI assistants showed sycophancy, a tendency to match what the user seems to believe, and that in existing human preference data, responses agreeing with a user’s views were more likely to be preferred.[4] In practice, if you ask “is this deck still worth keeping?” in a tone that signals you like it, the answer tilts your way.
So prompt 3 doesn’t ask for a verdict. It asks for the two strongest cases, one after the other, and tells the model not to be polite.
| Act as a skeptical colleague with no stake in this process. First, make the strongest case that this work should be stopped, merged into something else, or cut to a fraction of what it is now. Use only what’s in the context sheet and the assumptions list, and point to the assumptions you’re leaning on. Then make the strongest case for keeping it as it is. Finally, give me the one question I should ask the people who receive it, and say what answers would settle the matter. Don’t flatter the current process, and don’t pretend to know facts you haven’t been given. |
A good answer points at specifics from your sheet, such as pages two onward being a cost nobody has defended, and ends with a question you can ask, such as “If this stopped for four weeks, who would notice first?” A weak answer lists pros and cons in balance. Ask for the strongest case only, and add a fact it can use.
Treat the output as a hypothesis. The model has only your sheet, so the real test is the question it hands you. Take it to two or three of the people who receive the deck, as a question and not a pitch. What they say outweighs the model.
If you’re the analyst, have that conversation with your manager first, because retiring a deck is their decision. If you’re the manager, you can ask the recipients without it sounding like a threat.
With a purpose, sorted assumptions and a challenge on record, you can build from a blank page. The prompt gets only the purpose and the facts. Everything else stays out until the next step.
| Forget how this work is done today. Using only the purpose and the items we labelled FACT, design three different ways of achieving the purpose from scratch. Make them genuinely different from each other: one that uses no new software, one that removes a recurring human step, and one that doesn’t produce a document at all. For each, describe in plain steps what happens each week, who does what, and what would be lost compared with today. Don’t keep a step just because it exists now. |
A good answer gives three options that differ in kind. A weak one gives three versions of the same deck in different formats. Reply: “Option B is today’s process with a new name. Start again without a document.”
| Now bring the constraints back. Here are the POLICY items and TOOL LIMITS from our list, plus anything I’ve added: [PASTE THEM] For each of the three designs, tell me which constraints it breaks, which it works around, and what I’d need to find out before trying it. Rank the designs by how little they depend on changes I don’t control. Don’t pick one for me. Tell me what would make you choose each. |
Check that it names a constraint from your list, not a generic risk. If not, ask it to cite the row.
| Option | What happens each week | What’s lost | What it needs |
|---|---|---|---|
| A: One live view | Pipeline and billing sit in one shared view that updates itself. The analyst checks the account match once and fixes exceptions. | The narrative. Nobody explains why a number moved. | Billing data reachable from the CRM view. Check with whoever owns both systems. |
| B: A change note | A short Thursday note lists only what changed since last week, flags a handful of deals, and carries one table for finance. | The full picture each week, and context for someone new. | A stable account ID to see what changed. A source for the month-end figure. |
| C: A standing slot | The director and leads walk one table in a short slot on the Monday call. The analyst brings the questions. | A written record, and anyone who can’t attend. | The director agreeing to give up the document. |
The kind of options a good response produces for this sheet. Yours will differ.
A finance controller reads the one-page change note and asks where the signed-off month-end number comes from, because the old deck had a table for it. Nobody had put that on the constraint list. It was a missing fact, and this is the cheapest place to find it, before anything is built or automated.
Pick the option that needs the fewest changes you don’t control, and run it alongside the old deck for a few weeks. The model can rank the options, but it can’t know which colleague will say no.
Only now does automation become the question. A redesigned process can still be a bad one, and the quickest way to find out is to make the model run it for a year.
| Here is the design we’re leaning towards and the purpose it serves: [PASTE THE DESIGN AND THE PURPOSE] Play an operations lead who has to run this for a year. Tell me how it fails in month three, what it quietly depends on that nobody has written down, and which part is most likely to produce a wrong output that looks right. Then answer this: if we automate it as described, are we speeding up a good process or a bad one? Name who would catch the first wrong output and how long it would take them. If you can’t tell from what I’ve given you, say what’s missing. |
A good answer names failure points tied to your design, such as a renamed CRM field silently dropping a deal from the change note, or a note that hides the one quiet deal that matters because it didn’t change. A weak answer says “data quality” and “change management”. Paste the actual steps and ask for the single likeliest failure.
This is a close cousin of the pre-mortem prompt in our piece on running a pre-mortem before a big decision. Use that one for a decision and this one for a process. For the order of work in a teardown, see fixing broken workflows before you collect more tools. Before anyone automates anything, score the design on four questions.
| Question | Score 2 if | Score 0 if |
|---|---|---|
| 1. Purpose | You can state what the work is for in one sentence that names a decision | You can only describe the steps |
| 2. Habits tested | Every habit assumption has been put to the people who receive the output | The habits are still assumed |
| 3. Constraints | The chosen design survives the full fact and policy list | A constraint is unchecked or unknown |
| 4. Catching errors | A named person can tell within minutes that an output is wrong | Nobody would notice until a reader complains |
Score the design, not the idea. Use the bands below.
| Question | Score | Why |
|---|---|---|
| 1. Purpose | 2 | It names the director’s Monday forecast decision |
| 2. Habits tested | 1 | Regional leads were asked. Finance wasn’t. |
| 3. Constraints | 1 | Where the month-end figure comes from is still unconfirmed |
| 4. Catching errors | 1 | The analyst would notice, but nobody checks on a schedule |
Total 5, so fix before automating: confirm the month-end source with finance, then name a reviewer and a weekly check.
Automate a process only after you’d choose to build it that way today.
Here is where the model’s job ends and yours begins, step by step, for the weekly sales deck.
| Step | What AI does | What the human still owns | How it gets checked |
|---|---|---|---|
| Prompts 1 and 2 | Drafts the purpose and sorts the assumptions into four types | Decides what the work is for, and confirms facts and policies with their owners | Each fact and policy row is confirmed by its owner before prompt 3 |
| Prompt 3 | Argues both cases from the sheet and drafts the question for recipients | Asks the recipients, then decides whether to stop, shrink or keep the work | Two or three recipients answer. The manager who owns the deck agrees the result. |
| Prompts 4 and 5 | Proposes three different designs and shows what each one strains | Chooses one to trial and finds out the unknowns | The owner of each constraint, such as finance or IT, reads the chosen design |
| Prompt 6 | Plays the skeptical operator and names likely failure points | Decides whether to automate, and who reviews the outputs | A named person spot-checks one output a week at random |
The model drafts and challenges. A named person decides.
Who does what depends on the role. The analyst who owns the deck runs the prompts. The sales director, as the person who uses it, answers the question from prompt 3. The finance partner confirms the facts and policies that touch finance, and whoever owns the CRM and billing systems confirms the tool limits. For a recruiter reworking interview scheduling, the prompts stay the same and the people holding the facts change.
When you run the redesign, judge it on three things, in this order:
Counting how many people ran the prompts measures access, not whether the work changed.
A team can run this pass on its own, and many do. Structured training with an experienced facilitator shortens the learning curve and builds the habit of questioning a process across a team instead of in one enthusiastic person (our corporate workshops are built around that). In the terms we use at Future Factors, training builds the capability to question a process. The context sheet and the ownership table are workflow design, which gives that capability somewhere to be used. Reinforcement is a manager asking in week four what the pass turned up. Adoption is what you see when that holds, as covered in our guides on driving AI adoption on non-marketing teams and adding AI to your workflows. For more prompts that argue back, see the best prompts from the internet, tested and explained.
This week:
Any general chat assistant can run these prompts, including ChatGPT, Claude and Microsoft Copilot Chat. The context sheet matters more than the tool, and so does whether your organization allows the tool for work content.
Generic answers usually mean the model is missing a fact. Add the known pain and constraints to the context sheet, then ask: “Which sentence in your last answer would be false for my situation?”
No. Facts stay, and policies stay until their owner agrees to change them. In the sales deck example, the month-end figure survived every redesign. Often what changes is the format, the frequency or who receives it, not the purpose.
Describe the work in plain words and leave out names, figures and anything your AI policy restricts. If you don’t know the policy, ask whoever owns AI governance.
Start with whoever owns the recurring work, because they hold most of the context. Bring in the recipients for prompt 3 and the people who own the facts for prompts 2 and 5.
Sources were read on October 16, 2026: the Stanford Graduate School of Business Insights interview write-up of Jensen Huang (April 2024), Jason Averbook’s Now to Next essay (July 2025), the abstract of Sharma and colleagues’ paper on sycophancy in language models, and the Harvard Business Review page for Michael Hammer’s 1990 article. For Hammer we confirmed the headline and date only, and the article itself sits behind a paywall, so we don’t describe its content. Averbook’s piece is opinion, not research. The six prompts and the sample deck are our own, and the prompts were run on that deck with an AI assistant before going in. Outputs vary by tool and by what you paste in. The sorting of assumptions, the pre-automation check and its score bands are Future Factors working views, not validated models. The scenes in this guide are illustrations, not client stories. We left out productivity and ROI figures because we couldn’t confirm any against a primary source.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.