Independent resource · not affiliated with or endorsed by Clemson University

AI Workshops by ProvenLabs

Resource · Worksheet

The process-mapping worksheet.

“What would you like to automate?” asks you to already know what’s automatable. Start from a process you already run instead, and let the steps show you where AI fits. This page walks the same framework the workshop teaches, so you can work through it on your own with a blank doc open.

Step 1: Pick a process

Not your hardest problem. A process: something recurring, with a beginning and an end, that you could explain to a new coworker. Planning an event, onboarding a student worker, producing the monthly report, triaging the shared inbox. If it annoys you every time it comes around, even better, because you know it well and you’ll use what you build.

Write it down

The process, one line. Who it serves. Roughly how often it runs and how many hours it eats per cycle.

Step 2: Break it into steps

Write down the phases first. Then break each phase into the steps that really happen, which are rarely the ones written down. “Email the vendor list” hides four real steps: find last year’s list, check who left, draft the email, chase the non-responders. The decomposition is where the insight lives; a step you can’t describe concretely is a step you can’t make a call on.

For a fully worked example, this is exactly what the process maps do to a campus job fair.

Step 3: Make the call on every step

Automate

The step runs without a person in it. You build it once, put it on a schedule or a trigger, and check a sample of the output instead of reviewing everything.

The right call when

The output is quick to check, a single miss is cheap and catchable downstream, and the step doesn't need context that lives only in people's heads.

Augment

AI produces the draft, a person makes the call. Nothing ships unseen. The blank page disappears; the judgment stays.

The right call when

You add real verification or context the AI lacks. Research is blunt about this: a person in the loop only improves results when the person genuinely could have done the step. Augment has to earn that.

Keep

A person does the whole step, on purpose, because the value of the step is the person doing it.

The right call when

Trust is the deliverable (relationships, negotiations, hard conversations), accountability attaches to the step, you couldn't verify an AI version anyway, or doing the rep is how you stay sharp enough to catch AI errors elsewhere.

The questions that make the call for you

  1. 1. Do I need to verify the output regardless of who generates it?

    If you'd check the work no matter who did it, delegating costs you nothing new. If checking means redoing it, don't hand it off.

    For example

    Your dishwasher dies on a Sunday.

    Find 10 local plumbers and submit each contact form so quotes start coming in

    You'll hear directly from every plumber on timing and price either way. Who made the first contact doesn't matter.

    Automate

    Upload the policy and ask which page the water-damage language is on

    Now the answer is a pointer you can check in ten seconds. Caution: it can find the clause and miss the exclusion two pages later.

    Augment

    Ask whether your homeowner's policy covers the water damage

    Whatever it says, you're reading the policy yourself before you file the claim.

    Keep

  2. 2. Is this step creating something or deciding something?

    AI is at its best creating and at its worst choosing. Let it draft the options; the decision stays yours.

    For example

    Two things on this week's list.

    The family reunion email you've put off for three weeks

    AI drafts it in a minute. You fix Aunt Carol's name and send it.

    Augment

    Picking which contractor rebuilds the deck

    Have AI line up the three quotes item by item and flag what the cheap one left out. Then you pick, because you live with the deck and the person building it.

    Keep

  3. 3. What happens if it's wrong once?

    Some AI outputs have room to fail; the stakes are low enough. Others aren't worth the risk.

    For example

    Hiring a student intern, with AI in every seat but one.

    AI drafts the interview questions and the scorecard

    Sharper questions than you'd write at 11pm the night before, and every candidate gets measured against the same bar. If one question lands flat, you feel it live and skip it. A cheap miss.

    Augment

    AI analyzes the interview transcripts against the scorecard

    It catches the quote you'd forgotten and scores all ten interviews with the same consistency. A missed nuance here still gets caught, because you were in the room.

    Augment

    AI makes the hire

    A transcript can't carry body language, tone, or how someone handled the question they didn't expect. Get the hire wrong once and you spend a semester managing the miss, which is why AI sets the table and a person makes the call.

    Keep

  4. 4. Does it need context the AI can't see?

    It only knows what you've shown it. Supply the missing context, or keep the step.

    For example

    Same request, two ways.

    “Plan a week of dinners”

    A lovely plan for someone else's family.

    Keep

    Add the onion rule, the twenty-minute Tuesday, and the budget

    The plan gets genuinely good. The AI couldn't see your kitchen until you showed it.

    Automate

  5. 5. Could I have done this step myself?

    You can only catch mistakes in work you could have done yourself. If the step is beyond you, so is reviewing it.

    For example

    AI writes the Python script that cleans your survey data, and it's good.

    You read code

    You skim it, catch that it drops rows with blank emails, and fix one line. AI just saved you an afternoon.

    Augment

    You don't

    The script runs without errors and looks done, and you can't see what it quietly excluded. Have someone who codes glance at it, or ask for the row counts before and after and check those instead.

    Keep

  6. 6. Have I tried it on this exact step?

    AI is great at some tasks and useless at others that look identical. The only way to know which is which is to try it.

    For example

    Both feel like “plan my trip.”

    The Rome itinerary

    Trains, neighborhoods, the restaurant two streets from the Pantheon. Rome is documented to death, and it nails it.

    Automate

    The national park trip

    It invents a trailhead and a shuttle that doesn't exist, because backcountry conditions aren't documented. Test each new kind of task on ground you can check.

    Augment

Step 4: For every Automate and Augment step, fill in the panel

A classification without an implementation plan is a sticky note. Every field below is required before the call counts:

Mechanism
What runs it: a custom GPT, a scheduled task, a connector, a plain saved prompt?
What you build once
The instructions, the knowledge files, the template that makes it repeatable.
The literal prompt
Word for word. If you can't write it, the step isn't understood yet.
What comes back
The exact output and where it lands: an email, a draft in a doc, a row in the log.
Where a person still looks
On Augment steps nothing goes out until a person reads it. On Automate steps a person just checks a sample once in a while.
Depends on
What must exist first: an approved connector, a shared log, an account, a decision from someone else.

Step 5: Build one this week

Pick the single step with the best ratio of hours saved to build effort and build it: a custom GPT if the step needs your documents and standards, a scheduled task if it runs on a clock. One working step beats a beautifully classified map, and the first build teaches you more about the other calls than any amount of planning.

What the research says

  • People are unreliable narrators of their own AI gains. In one randomized study, experienced developers estimated AI made them 24% faster; measured, they were 19% slower. Measure the step instead of polling your feelings about it.
  • A large meta-analysis found human-plus-AI combinations often performed worse than the better of the two alone, with losses concentrated in decision tasks and gains in creation tasks. That’s why Augment has to earn its place.
  • Watch for Augment steps quietly becoming Automate: review effort drops as trust in the tool grows, and a nominal gate stops being one. Re-check your calls after a few months of evidence, which is exactly what the recurring process map shows.

Sources