M.L. Sebastian is now br8n.

AI exit plan

Could your team run this without the current AI?

Pick one job your team asks AI to help with. Imagine opening it tomorrow in a different tool. Could someone get the work done from the material you have saved?

Use the workflow inventoryAll resources

That is a useful place to start an exit plan.

In my original exit-plan essay, I described a sequence for moving the pieces of my own setup. The useful question underneath it was what I could actually take with me and what I would have to rebuild. I want to make that question easier for someone else to answer about their own work.

Start with a job small enough to check. Preparing a weekly stock list is manageable. Moving the whole company at once gives you too many places to hide an assumption.

Write down what a correct result looks like

Choose a result someone can judge. If the job is a stock list, a correct result might be the number of packs to order, with the calculation shown. If the job drafts a report, you might require each factual statement to point to its source and unanswered questions to remain unanswered.

Write that down before trying another tool. Otherwise, a polished answer can persuade you that the move worked even when it quietly changed the job.

Use the companion inventory. Start with the “What should happen?” field, then give each dependency its own row. Record where it lives, what you have copied, and what happened when you tried to use the copy. Include the person who can resolve a missing permission or explain a rule. Keep passwords and client information out of a worksheet you plan to share.

Look for the rules outside the file

A useful export might include the working document and still leave out what makes the work correct.

Someone knows which source is current. An instruction explains how to handle missing information. A connection brings in the stock count. Your team may have changed a rule in chat and never moved it into the instructions.

Those details belong in the inventory. If you cannot find one, record that gap. Do not fill it with a guess to make the row look finished.

Here is a fictional example you can work through without connecting an account or buying anything.

A workshop needs eight markers and two rolls of tape. There are three markers and one roll of tape on hand. The stock file gives you those numbers. A separate rule says markers come in packs of four and tape in packs of two. There are already enough cards.

The order is two packs of markers and one pack of tape. Ordering five markers and one roll of tape would describe the shortage but miss the purchasing rule.

In the internal walkthrough, copying only the stock file stopped the checker: the pack-size rules were missing. Adding the original rules reproduced the expected order. Changing the marker pack size then changed the order, even though every file still opened.

That leaves an evidenced dependency and a specific next action: preserve the pack-size rules with the stock file and check the result again. This was a constructed example, not a failed customer export.

Try the copy somewhere separate

Use a harmless sample. Keep the original in place. Recreate the job in a separate location using only the material you intend to take with you.

For the workshop example, a person can check the arithmetic directly. The supplied stock and pack-size files let you check the calculation by hand. The example uses no AI service. It checks whether a defined rule and its input survive the copy; it says nothing about whether two models exercise the same judgment.

For an AI-assisted job, you would still need to run the same permitted sample in the proposed replacement tool. Check the result against the requirements you wrote down. Pay attention to omitted facts, invented answers, source references, and situations where the tool should stop for help. Record what happened instead of relying on how familiar the answer sounds.

Give the unfinished part a next action

An export you can open is evidence of a readable file. A process you can repeat from that copy is evidence of a restore. Whether it behaves well enough on ordinary work needs its own checks.

Keep those results separate in the inventory, and fill in “What did we not test?” before calling the row complete. You may have the files and still be missing tool access. You may reproduce a calculation and still need a person to review the judgment in a written recommendation. One successful example supports a narrow conclusion about that example.

Write the next action so somebody could do it: recover the missing rule, confirm permission to export a source, or compare the replacement's answer against the same evidence. Assign it to the person who can resolve it. Do not invent a completion date before that person has assessed the work.

You can decide to stay with your current tool after doing this. You will have a clearer reason for staying and a record of what a move would require.

That is what I want from an exit plan: enough evidence for someone else to take the next step without needing me to explain the whole system again.

Use the workflow inventory

Choose one small job. Write down what should happen, what the job needs, and what your first check leaves unanswered.

Download the spreadsheet · Download the plain worksheet · Get the inventory and worked example

The inventory is free to use. No email or account is required.

Could we repeat this job somewhere else?

Copy this page and fill it out for one small process. A spreadsheet version is available in portability-inventory.csv. Use harmless sample information for the first check.

The job
[What work are we trying to repeat?]
What should happen
[Write the result or requirements before the test.]
What the job needs
What the job needsWhere it lives nowWhere our copy is
[Source material][Reference][Reference or not copied]
[Rules and exceptions][Reference][Reference or not copied]
[Any required tool access or permission][Reference, never a password][Available, missing, or not checked]
Export check
[What did we copy? Could we open it? Cite the file or observation.]
Restore check
[Could we repeat the job using the copy? What was missing?]
Behavior check
[Did the result meet the requirements above? Record the actual result.]
What did we not test
[For example: another AI model, a real integration, edge cases, or a person's ease of use.]
One dependency we have evidence for
[What does the work rely on? Point to the observation.]
Next action and person who can resolve it
[A specific action; do not invent an owner or permission.]

Use not checked where evidence is missing. A readable file is an export result; a repeated job is a restore result; meeting the written requirements is a behavior result. Record each separately.

Workshop stock example

This is fictional workshop stock. There are no customers, account connections or AI-provider calls. The incomplete copy described below is deliberately constructed for the demonstration; it is not evidence that a real export tool failed.

Check it by hand

The stock file shows 30 cards when 24 are needed, three markers when eight are needed, and one roll of tape when two are needed. The separate rules say markers come in packs of four and tape in packs of two.

Stock counts
ItemOn handNeeded
Cards3024
Markers38
Tape12

There are enough cards. Five more markers require two packs. One more roll of tape requires one pack. The expected result is in expected-order.csv.

Expected order
ItemShortageOrder packs
Cards00
Markers52
Tape11

Now imagine that the separate rules were missing. You could calculate the shortage, but you could not establish how many packs to order. That is the dependency the inventory should make visible.

Download the example inventory

Want help checking a workflow your team depends on? Talk with Michael.