There's a requirement in one of our June planning documents that isn't about what the product does. It's about how a delay feels:
"Latency framed as craft, never apology; primed up front."
The document is otherwise plumbing: what has to finish before what, what happens when a step fails, the unglamorous machinery of getting a report built. Requirement six is the odd one out. It exists because our AI strategy reports take real time to build (the plan budgets the wait at 20 to 40 minutes), and at no point in that document, or any document since, did we try to make the wait shorter.
The title of this essay rounds up, and I'll own that now: forty minutes is the top of the window our own plan writes down. But the number isn't the point. The point is a decision that I think matters to anyone who buys strategy work in the AI era, on either side of the transaction: we treated the wait as a design surface instead of a performance bug.
A strategy document that arrives instantly is telling on itself
If you run a consumer brand, you have probably already been pitched an AI-generated audit, scorecard, or strategy document. You will be pitched many more. Here is one of the fastest honest tests you can run on any of them: notice how long it took to make, and ask what the time was spent on.
I put a conviction on record back when we were designing our delivery flow, before a line of it existed:
"If you're getting a multi-hundred or thousand dollar report you don't want it instantly… you want it to feel like there is a ton of work going into it."
Every engineering instinct fights this. We measure everything else we build in milliseconds: dashboards, editors, internal tools. Speed is oxygen there, and I will happily burn a day shaving seconds off a screen our team opens fifty times a shift.
The customer's premium artifact is the one place the logic inverts. A strategy document is not a search result. When the deliverable claims to carry judgment (research weighed, claims checked, contradictions resolved), the delivery speed is part of the claim. Instant contradicts it. Nobody opens a forty-page strategy document that arrived in under a second and thinks thorough. They think template, and they're usually right to.
So the design principle became: match the delivery speed to the value story. Which sounds like marketing until you see what it demands from the engineering. It demands two things that pull in opposite directions. The wait has to be told well, and the wait has to be true. The second one is the hard one.
Telling the wait well
Here's what requirement six turned into, in the product.
When a report is requested, the buyer gets a message at minute zero, before they've left the page. The transactional email's job, in the words of the note we wrote at the top of the part of the system that sends it:
"Primes the buyer for the multi-minute wait so the latency reads as craft, never silence that looks broken."
The copy itself: "Your strategists have started on the report… Real strategic work takes a little time — we'll email you the moment it lands." While the build runs, the status page says what's actually happening in one plain sentence: "We're studying your brand — this takes a few minutes." No percentage bar pretending to know things it doesn't. No spinner apologizing for existing.
And when the work is done, the completion email arrives under a design instruction I'd frame and hang on the wall:
"premium, no apology, no mention of wait time."
"Your strategists have assembled the work." That's it. The wait is never mentioned again, because craft doesn't bill you for the hours twice.
The plan is blunt about why the minute-zero message is load-bearing. Our system tracks a report mid-build under a status it calls "attention" (parked, working, not yet done), and the plan's own words are that the craft framing only works if the buyer is primed before they leave; without it, a 20–40 minute silent "attention" reads as broken.
That's the whole thesis in one sentence. The same forty minutes is either evidence of work or evidence of breakage, and the difference is whether anybody told you at minute zero what the machine was doing and when it would be done.
Silence converts craft into outage at a fixed exchange rate.
| Told at minute zero | Silence | |
|---|---|---|
| The same 20–40 minutes reads as | evidence of work | evidence of breakage |
| What the buyer sees while the build runs | We're studying your brand — this takes a few minutes. | a 20–40 minute silence that reads as broken |
The wait has to be true
Now the hard half, and the half that separates a real report from a slow one. Framing a wait as craft is a confession with a deadline: eventually someone asks what the minutes actually bought. If the answer is "a queue," you've built the other kind of lie — slower than the work, which is just as dishonest as faster than the work, and more cynical.
So here is the inventory of what our clock is spent on, because this doubles as a checklist for evaluating anyone's AI-built strategy work, ours included.
Nothing gets written until the brand's own record is fresh: the first checkpoint parks the entire build rather than write a report on top of a stale foundation. That's the "studying your brand" stage, and it's real study.
Some report types then wait on that live measurement of how AI engines see the brand today.
Research is split across separate AI researchers, one per angle, each walled off from the others. The instruction is explicit that "each researcher gets ONLY its own angle (no sibling context)," so findings can't contaminate each other. And every source they retrieve is kept verbatim.
Then the claims get checked one at a time against the sources they cite, under a rule I'd put on a poster: an unreadable or errored judgment counts as unsupported, because "a claim never passes the gate by accident."
Then the report argues with itself: contested points across sections get resolved before you ever see them, which is its own essay.
And only then does composition start, where the arithmetic gets visible: each document runs four to six writing passes, ten to thirty seconds each, one after another (one to three minutes per document), and a complete engagement is seven documents plus a closing strategy letter.
One flagship run made nearly ninety separate AI calls, each one logged to our cost ledger, before the document existed.
Stack it up and the 20-to-40-minute band is a sum, not a policy.
Which is why the receipt I care most about is a negative one: there is no artificial delay anywhere in this process. I checked before writing this. The only pause in the build is the thirty-second interval between checks on a real measurement that genuinely isn't finished yet. Nothing pads. Nothing holds a finished report back to make the wait feel more impressive.
The rule cuts both directions
That negative receipt matters because this conviction has a failure mode on each side, and the comfortable essay would only name one of them.
Instant premium is an oxymoron. That side's easy, and it's the side that sells. The other side: models get faster every quarter, and the honest build time of this report will shrink whether I like it or not. The day it drops to twelve minutes, the principle does not entitle us to a hidden twenty-eight-minute nap buried in the system. The conviction was never "make them wait." It was "never let the delivery speed tell a different story than the work" — and that sentence cuts both directions. If we ever pad the wait to protect the feeling, delete this essay; we'd have become the thing the walls in this pipeline exist to keep out.
I'll also concede what the conviction can't prove yet. This product hasn't had years of customers ratify the psychology; the design is conviction plus craft, not conviction plus data. Maybe buyers of premium AI reports will turn out to want them instantly and trust them anyway. I doubt it — twenty years of watching people value what signaled effort says I'm right — but I hold it as a bet, not a law.
Here's what I'll defend as law, and what you can carry into any vendor conversation: the wait and the story about the wait have to be the same thing.
Ask what the minutes bought.
A vendor who can walk you through the sum — study, research, verification, composition — is selling work. A vendor who can't is selling either a template with a spinner on it or a queue with a story on it, and you deserve neither.
The pipeline contains no artificial delays and no apologies. Both facts are the product.
See what the minutes buy
Study, walled-off researchers, claims checked one at a time against the sources they cite, seven documents plus a closing strategy letter.
Walk through the buildWhy is there no progress bar while the report is being built?
Because a percentage would be pretending to know things it doesn't. While the build runs, the status page says what is actually happening in one plain sentence — we're studying your brand, this takes a few minutes — and no spinner apologises for existing.
How does a buyer know the wait will not run forever?
Every dependency the build waits on carries a time limit. The live measurement of how AI engines see the brand is checked thirty seconds at a time, twenty checks maximum, and then the buyer is told plainly that it's still waiting rather than being left to watch a spinner. Copy about time is only credible when a counter enforces it.
If AI models get faster, will the report be slowed back down to protect the feeling?
No. The honest build time will shrink whether we like it or not, and the principle does not entitle anyone to a hidden delay buried in the system. The conviction was never make them wait; it was never let the delivery speed tell a different story than the work.
How can a buyer tell a real build from a template with a spinner on it?
Ask what the minutes bought. A vendor who can walk you through the sum — study, research, verification, composition — is selling work. A vendor who cannot is selling either a template with a spinner on it or a queue with a story on it.