Scope a job by decisions, not deliverables
Deliverable counts stopped predicting effort once variants got cheap. Scope by approval decisions instead, and price each one by what it locks.
Two jobs land in the same week. The first is one 30-second spot. The second is twelve 15-second cutdowns for paid social. Under a per-deliverable model the second is the bigger job by an order of magnitude. In practice the first one ate three weeks and the second was done Thursday afternoon.
The reason is not that the cutdowns were easy. It is that they were twelve outputs of one set of decisions — one concept, one look, one voice, one product treatment — and the spot was one output of about nine. Output volume stopped tracking effort somewhere around the point where generating a variant became cheaper than the meeting about whether to generate it.
So count the thing that is still expensive.
Why the deliverable count broke
Hourly billing broke first, and for a well-understood reason: if AI cuts a 20-hour deliverable to 5 hours and you still bill by the hour, you have cut your own invoice by 75%. Worse, an hourly line item puts your cost structure on the page for the client to negotiate against. Most shops moved to per-deliverable or fixed-scope retainers as a defensive measure rather than an opportunistic one.
Per-deliverable is better, but it inherits the same defect one level up. It assumes output volume proxies for effort, and that proxy is now unreliable in both directions. A single hero shot with a specific character action can burn more attempts than an entire family of aspect-ratio variants. Community-reported figures put usable output at roughly one clip in four, or three to five attempts per usable clip, with the number rising sharply for specific character action and complex motion. Those are aggregated anecdote rather than measured data, but the shape is consistent enough to plan against, and the operating principle it implies is sound: usable rate is the benchmark nobody publishes, and it is the one you should be instrumenting on your own jobs.
Meanwhile the pressure has not eased. Superside's Breakpoint research puts 80% of creative teams at or beyond capacity and 70% of creative leaders burnt out after AI adoption rather than before it. Whatever replaced generation as the constraint — filtering, brand governance, rights, taste — is work that scales with decisions, not with outputs. A deliverable count cannot see any of it.
What a decision is
A decision is a point where someone with authority chooses between directions and closes the others. It has three properties worth writing down:
- It is irreversible-ish. Reversing it invalidates work downstream.
- It is owned by a name, not a team.
- It locks something specific, and the lock is what makes the following work cheap.
A deliverable, by contrast, is what falls out of the decisions once they are made. Twelve cutdowns fall out of a locked concept in an afternoon. One spot with an unlocked concept does not fall out of anything.
The decision ladder for a video job
Here is the ladder for a typical branded video, ordered by how much each one costs to reverse. Adjust the list for your work; the ordering is the part that transfers.
| # | Decision | What it locks | Cost of reversing it |
|---|---|---|---|
| 1 | Proposition | What the piece is arguing | Everything. Restart. |
| 2 | Format and duration | Shot count, structure, platform fit | Restart from script |
| 3 | Script or beat sheet | The words and the order | All voice, all timing, most shots |
| 4 | Look | Palette, lens language, grade | Every generated frame |
| 5 | Cast and character | Faces, wardrobe, continuity anchors | Every shot with a person in it |
| 6 | Product treatment | Angles, geometry, packaging fidelity | Every product shot |
| 7 | Voice | Casting, read, pace | All audio, all lip timing |
| 8 | Music | Track, tempo, edit rhythm | The whole cut's timing |
| 9 | Final cut | The version that ships | One export |
Nine decisions. Note where the money is: reversing decision 3 is expensive and reversing decision 9 is not. Under a deliverable-count model those two look like the same category of "revision."
Decisions 4, 5 and 6 hold the rest of the job together, and they are held by concrete mechanisms rather than good intentions. Reference images, named reusable characters and products, and reference stacking each work in different conditions and fail in different ones — four consistency mechanisms and when each fails is the honest map, and named assets across scenes is the version you can set up in an afternoon. A lock with no mechanism behind it is not locked, it is agreed.
Writing the scope this way
The scope document stops being a list of files and becomes a list of gates. Three parts.
One: name the decisions and their owners. Literally a table, in the proposal, with a column for who signs each one. Most scope disputes are actually authority disputes discovered late — the person who approved the look was not the person who could approve the look.
Two: state what each decision locks. This is the sentence that does the real work later. "Approving the look locks palette, lighting and grade for all shots in this job" is what converts a week-six request into an obvious change order rather than an argument. It also tells the client something true and useful: the lock is what buys them cheap variants afterwards.
Three: state what is unlimited once locked. This is the part clients like, and it is honest. Once concept, look, cast and voice are locked, additional cutdowns, aspect ratios and text variants are cheap to produce, and pricing them as if each were a fresh production is not defensible. Say so, and price the volume tier accordingly. What you are selling is the direction; the outputs are the artefact.
A worked example of the sentence in a proposal:
This quote covers nine approval decisions, listed below, each signed by the named owner. Once decisions 1–8 are signed, the deliverable set is produced against them: 1 master, 12 cutdowns, 3 aspect ratios, and 2 text variants per cutdown. Additional outputs against the same locked decisions are quoted at the volume rate. A change to any signed decision is a change order, priced against the decisions it reopens.
Note that the deliverable list is still in there. It has just moved out of the position of pricing driver and into the position of a schedule.
Pricing against the ladder
Do not try to derive a price per decision from first principles. Price the job as you always did, then use the ladder to decide two things.
Where the buffer goes. Buffer belongs against the decisions with the worst usable rates, not spread evenly. If the job hinges on a specific character action or an exact product geometry, that is where the attempts pile up, and that is where the contingency belongs. Reroll rates and the shots you throw away is the production-side version of this arithmetic, and estimating credit cost before you dispatch a batch is how to sanity-check it before you commit.
What a change order reopens. Price the reopen by depth in the ladder rather than by output count. Reopening decision 3 is a different order of magnitude to reopening decision 8, and a flat per-round fee prices them identically, which is why flat per-round fees feel wrong to both parties on big changes. Fixed-price or cost-plus quoting covers the base structure this sits inside, and a revision policy that stops scope creep covers the rounds mechanics for changes that do not reopen a decision at all.
One thing the ladder is not for: justifying a discount. The survey data suggests the fear here is larger than the phenomenon. Reporting aggregated by TechBullion puts 73% of agencies as never having been asked to cut prices despite adopting AI, and of the 27% who were asked, only 13% actually lowered rates. Clutch's 2024 figure of 61% of clients raising AI at renewal is about the subject coming up, not the price coming down. Faster delivery is margin you have earned. A decisions-based scope makes that easier to hold, because it moves the conversation onto direction, which the client can see is the expensive part, and off unit count, which they can see is not.
FAQ
Does the client actually need to see the decision ladder?
Yes, and it usually helps rather than hurts. Clients are not trying to get free work; they are trying to get to a result and have no visibility into which of their requests are cheap. A table showing "changing the text is free, changing the cast is not" gives them the ability to make the cheap request on purpose, which is what you wanted anyway.
How many decisions is the right number?
Enough that each one has a distinct owner and a distinct lock, and few enough that the client will sign them all inside the schedule. Nine works for a video job. A single-image job might have four. If two rows on your ladder always get signed by the same person on the same day, merge them.
What if the client will not name an owner per decision?
Then name one approver for the whole ladder and put it in writing. The single-approver fallback is worse for you than distributed ownership on big jobs, but it is far better than an unstated assumption. The failure mode you are avoiding is not "the wrong person approved it" — it is "nobody thinks they approved it."
Does this work for retainers?
It works better for retainers than for one-off jobs, because a retainer's real economics are about how many fresh directions the client wants per month, not how many assets. A retainer priced at "one new locked direction plus unlimited variants against it" is easy to deliver and easy to defend. Per-minute, per-deliverable or retainer pricing covers where each structure fits.