On this page
Six weeks on how much authority we hand AI inside our own company. Here's the version you can run on yours in an afternoon, and the one paragraph it leaves you holding.
Skip: The homework we promised
Six weeks ago Chuck and I laid out four rungs: reads, drafts, acts and decides. We said we'd walk our own company up them one rung at a time and close with a version of the ladder you could run on your own business in an afternoon.
We've walked it. The company memory that gave us a clean, cited answer that was wrong. The drafts that got every fact right and still didn't sound like me. The writes our systems make without asking, and the one approval we kept with a person. The phone agent that decides who a caller is before anyone at InTech hears about it, and the second version that came out of reading what the first one said.
This is the afternoon. It's the same shape as the six questions we gave you in September, an instrument with a deliberate order, built to tell you no as easily as yes. Most businesses that run it honestly will end with fewer things on the top two rungs than they started with. That's a passing grade.
Chuck: What each rung turned out to be about
When we started, I described each rung by what the system is allowed to do without a person. That's still the definition. What the last six weeks added is the thing that actually breaks at each one, and that's what you'll be checking for this afternoon.
Reads fails on coverage that looks complete. Nothing changes, so nobody watches. The danger isn't a wrong fact. It's a record that stops partway through a story and reads exactly like one that didn't. The question to ask at this rung is what the system didn't see.
Drafts fails on voice before facts. Every detail checks out and the draft still doesn't sound like the person whose name is on it, or it answers in the wrong place. A person reading before anything goes out catches that, which is why this rung is safe and why it's slow.
Acts is gated by whether you can put it back. The first rung where nobody reads first, so every write has to be one you can find and undo, through the same permissions and the same audit trail a person clicking in the interface would use. It's also the rung where we drew a line we didn't expect to draw. Our time-off system can do a lot on its own, and approving a request still stays with a person.
Decides is gated by the boundary and the record. The choice reaches a live caller before anyone at InTech knows it was made, so you can't put it back. What holds it is a short written list of what it may decide, and a record someone reads soon after. The phone agent's second version added one more line to that: it only makes promises someone wrote down ahead of time.
One thing held across all four. Every real lesson came out of the record of what the system actually did. None came from the demo, and none from a count of how many times it ran.
Skip: The afternoon
Five steps, in order. Each one changes what the next one is worth.
Step one: list the work, not the tools
Write down every job AI does in your business today, plus the ones you're about to turn on. Name each one as a task: "summarize sales calls," "answer the after-hours line," "update the deal after a meeting." Not the product name.
This matters because rungs belong to tasks. The same tool sits on all four rungs at our company in a single afternoon. If your list reads like a software inventory, you'll grade the vendor instead of the work.
The trap answer is a short list. Look at the inbox, the calendar, the phone and the spreadsheet somebody keeps current by hand. AI work hides there more than in the tools you bought for it.
Step two: put each task on the rung it actually runs on
Not the rung you meant to put it on. The one it runs on when nobody's looking. Does anyone read it before it goes out? Does it change a record? Does it reach a customer before a person knows?
Then ask the question from our opener for each one: did someone decide it belongs there, or did it drift there because that's where the tool started? Drift is the most common answer, and it's the one worth writing down.
A tool whose demo sits on decides tends to get switched on at decides.
Step three: look for skipped rungs
Circle anything on acts or decides that never spent time on reads or drafts in your business. No history of watching it, no record of how it fails, no paragraph saying what it may do.
Those get demoted today, without a meeting. Move them down until someone has watched them long enough to know how they break. That isn't a verdict on AI. It's how you avoid the version where something goes wrong, nobody can say why, and the whole idea gets switched off.
Step four: run the promotion test
For everything left, run Chuck's three conditions. Do you know how it fails at its current rung, and is that rare enough that the person reviewing it is mostly confirming instead of correcting? Is the step up reversible or tightly bounded? Has someone written the paragraph of what it may do alone?
Most tasks fail the first condition, not because they fail often but because nobody has been looking. That's the finding.
A task you haven't watched isn't ready to move, however well it demos.
Step five: pick one
One promotion, not a roadmap. Choose the task that passes the test, is cheapest to undo if you're wrong, and saves the most time if you're right. If nothing passes, your one move is to start reading the record of whatever matters most, and come back in a month.
Then write its paragraph. That's what Chuck covers next, and it's the part that takes most of the afternoon.
Chuck: What goes in the paragraph
The paragraph is the third promotion condition, and it's the one people try to skip by writing a feature list instead. A feature list says what the system does. The paragraph says what it's allowed to decide, and it has to answer six things.
What it may do alone. Name the decisions, not the capabilities. "Sort incoming calls into new business, billing, support, urgent or vendor" is a decision. "Uses natural language understanding" is not.
What goes to a person. The line where it stops and hands off. For our phone agent, that's anything that commits the company to a price, or to a promise nobody wrote down ahead of time.
What it may promise. This is the line our first version of the phone agent didn't have, and it's the one I'd put at the top of any list now. Left alone, an agent will fill silence with commitments that sound reasonable. A Monday-morning callback nobody agreed to. A follow-up from a person who doesn't work here. Write down the promises it's allowed to make, and treat everything else as one it isn't.
What it says about itself. Disclosure has run through this whole series. Our phone agent now tells every caller at the start that it's InTech's AI assistant and that the call may be recorded. Skip still hasn't decided the version for drafts, whether something written by AI can go out unread with a line that says so. Either way, the answer belongs in the paragraph, decided ahead of time, not left to whatever the tool does by default.
Who reads the record, and how soon. At decides, the record is the only safeguard you have after the fact. Put a name and a timeframe on it.
A record nobody reads is the same as no record.
What drops it a rung. Decide now what kind of surprise moves the task back down, so that when it happens it's a routine step and not a debate.
If you can write all six in a paragraph, you're ready to promote. If one of them won't come, you've found what you need to watch before you try again.
Skip: The longer version of the same afternoon
If you run this across a whole department instead of a single list, it stops being an afternoon. That's the work we do in a Business OS Blueprint. We take one function at a time, map who does what and what each role waits on, then answer the ladder's questions for it. What the organization knows that an agent would need. Which roles get an agent, what stays human and what gets no agent at all. Where a person reviews. What the numbers are today. And the order things move up, starting with the safest promotion that's most worth making.
The client keeps all of it whether they build with us or not. But you don't need us to start. Everything above is the part you can do yourself.
If the plan only works with us attached to it, it isn't much of a plan.
Skip: Where we land
We started this series by admitting we now do something we said in April we don't do: let a system decide with nobody approving in the moment. We do it in one place, on purpose, with a paragraph behind it and a record that gets read. Everywhere else, the honest answer is a lower rung than the tool would have picked.
That's what I'd want you to take from the afternoon.
Most of the ladder's value isn't in climbing it. It's in knowing which rung each task is actually on, and who decided.
Here's what I'd like to hear in the comments. Tell us the one task you picked to promote, and which of Chuck's six lines was hardest to write for it. If you found something that drifted to a rung nobody chose, tell us that too. Those are the answers that will shape what we write next.
