Resource
Every term on the Yield Model page: what it is, what moves it, and how to read it. The same text the "?" buttons show, in one place.
AI alone is the cheapest lane, no judgement call caps or blocks it, and the saving clears the bar to clear.
Moves it It flips to Augment or Keep human if acceptance drops below break even, review time passes the ceiling, or a gate caps the lane.
Read it A clean case is rarer than the market implies. Treat marginal gates as the real project risk, not the model.
A person using AI is cheaper than a person alone and than AI alone, or AI alone is not permitted.
Moves it It depends on minutes with AI staying well below minutes unaided. The person's own checking sits inside those minutes, so budget it there.
Read it Measure minutes with AI after the novelty wears off. The checking inside them is what erodes first, and then the errors arrive.
A person alone stays: AI is not permitted here, the saving is too small to justify a project, or a person is simply cheapest.
Moves it It flips when a gate answer changes, or when the saving of a permitted lane clears the bar to clear.
Read it This is not an AI problem. The effort you were about to spend here will earn more somewhere else.
The economics work, but one specific thing blocks the build: the quality bar is not met yet, or the process keeps changing.
Moves it Fix the named blocker and the verdict becomes Automate or Augment without any other number moving.
Read it Do not start the build. Start the thing the chip names, then come back.
The cost of one unit of work done three ways: a person alone, a person using AI, and AI alone.
Moves it Each bar is people time, model cost, build and upkeep spread over the volume, and the cost of fixing bad outputs.
Read it Cheapest is gold. The verdict is not always the cheapest bar, because the judgement calls can rule a lane out.
One person doing the unit unaided: their loaded hourly rate times the minutes it takes.
Moves it Only the hourly rate and the minutes unaided move it. It is the baseline every other lane is measured against.
Read it If this bar is already the cheapest, no amount of model progress changes the answer until people time changes.
AI drafts, a person finishes and checks: minutes with AI at the person's rate, plus the model cost and the build.
Moves it It wins when minutes with AI are far below minutes unaided and the review does not eat the difference.
Read it The most common honest answer. Its weak point is the checking hidden inside minutes with AI, which nobody budgets for.
The model does the unit, a reviewer checks it, and rejected outputs are reworked at the rework factor.
Moves it Acceptance dominates it, then the rework cost (minutes unaided, hourly rate, rework factor), then review minutes and the reviewer's rate. Model cost is usually the smallest term.
Read it Below break even acceptance this lane loses money on every unit. Above the supervision ceiling the saving is gone.
The three lines the AI alone lane must clear: break even acceptance, the supervision ceiling and payback within the horizon.
Moves it They are computed from your numbers, so every input you change redraws them.
Read it How far you sit from each line tells you how fragile the verdict is, not just what it is.
The acceptance rate at which AI alone costs exactly what a person alone costs. Below it the AI lane loses on every unit.
Moves it It rises when review time, the reviewer rate or the rework factor rise, and falls when minutes unaided or the hourly rate rise.
Read it Compare it with your measured acceptance, not with a vendor demo. Two points of headroom is not headroom.
The most minutes of review per unit the AI lane can carry before it costs as much as a person alone.
Moves it It grows with the person's minutes unaided and shrinks with the reviewer's hourly rate and the rework cost.
Read it If your review time sits above the ceiling, the AI lane is a more expensive way of doing the same work.
How many months of saving it takes to earn back the one-off build cost. The page judges it against a 12 month horizon.
Moves it Build cost and monthly volume decide it. A lane that saves little per unit never pays back at low volume.
Read it Over 12 months means the project does not pay for itself within the horizon this model uses.
The money at stake per year: what the chosen lane saves, or what the refused lane would have saved. When a person is simply cheapest, the cost per unit.
Moves it It is the per unit difference times the monthly volume times twelve, after build and upkeep are spread over the year.
Read it Left on the table is not a recommendation. It shows what the judgement calls are costing you, deliberately.
What one unit costs in the chosen lane, all in: people time, model runs, build and upkeep per unit, and rework.
Moves it Every input touches it. The lane breakdown under the bars shows which term dominates.
Read it Compare lanes on this figure. Compare projects on the yearly figure and the bar to clear.
The yearly saving a project must exceed to be worth running: the larger of a fixed floor and the build cost.
Moves it A bigger build raises the bar. A saving under it produces Keep it human on materiality, even when a lane is cheaper.
Read it A real saving that is too small to matter is still too small to matter. That is the point of the bar.
The list price of the picked model in USD per million tokens, input and output, from the vendor's pricing page.
Moves it Changing the model changes both prices. The checked date says when we last verified them.
Read it Published means you can open the source. Vendor prices change; the date tells you how fresh ours is.
How often the AI output is good enough first time, as a share of all units.
Moves it Raise it and the AI lane gets cheaper through rework alone: fewer rejected outputs to redo. Review minutes are paid on every unit regardless.
Read it Measure it on real units before you trust it. It is the number people overestimate most.
Where to find it Sample twenty recent outputs and count how many needed no edit. Your review queue or approval log usually holds this already.
How long one person takes to do one unit without AI, start to finish.
Moves it It sets the scale of everything: the person lane, the ceiling and the saving all grow with it.
Read it Use an average over a normal week, not the best case. Include the interruptions.
Where to find it Time a few real units, or take the team's own estimate. A ticket or helpdesk system with a handling time field already holds the average.
How many minutes someone spends checking one AI output before it is used.
Moves it Every minute here is paid at the reviewer rate on every unit, accepted or not. It moves the AI lane fast.
Read it If it is near the ceiling, the saving is fragile. Review time rarely appears in a business case, and it should.
Where to find it Ask whoever checks the output, or time three reviews. Count the checking that happens, not the checking in the process document.
The fully loaded cost per hour of whoever checks the AI output.
Moves it Often a more senior person than the one replaced, so the AI lane buys a dearer hour than it saves.
Read it Review minutes outweigh acceptance only when review time per unit, at the reviewer's rate, exceeds the expected rework cost. Check the ranking.
Where to find it Name the reviewer first, then take their rate from payroll the same way as the hourly rate. Guessing the reviewer is where this goes wrong.
How long a person takes on one unit when AI drafts it first and they finish and check it.
Moves it The gap to minutes unaided is the whole hybrid saving. Close the gap and Augment disappears.
Read it Measure it after the novelty wears off, not in the first week.
Where to find it Nowhere yet, unless a pilot ran and logged it. Otherwise run ten units with the AI draft in front of the person and time them.
How much more it costs to fix a bad AI output than to do the unit fresh, as a multiple.
Moves it Above one, every rejected output costs more than never having used AI. It multiplies with low acceptance.
Read it Usually above one: the error has to be spotted before it can be fixed. Keep it honest.
Where to find it Ask someone who has fixed a bad one. Spotting the error, deciding, and redoing it is usually 1.2 to 1.6 times fresh work.
The fully loaded cost per hour of the person doing this unit today.
Moves it It scales the person lane and, through the assisted minutes, the hybrid lane. Higher rates make AI lanes look better.
Read it Salary plus employer costs, tooling and overhead. Not the gross salary divided by hours.
Where to find it Payroll has the gross salary. Your accountant or finance system knows the employer add-on; about 1.3 to 1.5 times gross, over roughly 1,600 working hours.
How many of these units happen per month.
Moves it It multiplies every per unit saving into the yearly figure and decides whether the build ever pays back.
Read it Low volume is the usual reason a real per unit saving is still not worth a project.
Where to find it Your ticket, CRM or invoice system counts these already. Use last month, or an average of three when the work is seasonal.
The model and tool cost of one run, in your currency. Derived from tokens and the list price when you pick a model.
Moves it Usually the smallest term in the whole model. Doubling it rarely changes the verdict.
Read it If the verdict hinges on this number, look at acceptance and the rework cost first. They dominate.
Where to find it Your provider's invoice, or leave it to us: pick your model in All numbers and we derive it from the published price.
How many model runs one unit really takes, counting retries and prompt iteration.
Moves it It multiplies the model cost per run. Assuming exactly one run per unit flatters the AI lanes.
Read it Check your logs. One and a half runs per unit is common; one is rare.
Where to find it Your provider's usage log shows calls per day. Divide by the units done that day: retries and re-prompts are the difference.
The one-off cost to build the AI lane: engineering time, integration and the pilot.
Moves it It sets the bar to clear and the payback. Spread over the volume it becomes a cost per unit.
Read it Include the pilot and the people who will do it. A cheap build with an expensive pilot is not cheap.
Where to find it Ask whoever would build it for a rough quote, or price your last comparable integration. Nobody else has this number; it does not exist in a system.
Monthly upkeep of the build: prompts, monitoring, fixes.
Moves it Spread over the horizon like the build cost. It is left out of the ranking: its measured effect is small and in practice it tracks the build.
Read it Plan for it from day one. A model that is never maintained is a model that quietly degrades.
Where to find it Nobody has this yet. Ask whoever will run it what a month of upkeep costs; a day or two of an engineer is typical.
Seats and tools the build needs every month, added to maintenance before the engine runs.
Moves it A flat cost spread over the volume. It hurts low volume cases most.
Read it Count only the seats the AI lane needs, not the ones people already had.
Where to find it Your card statement or the vendor's billing page, which lists every seat by month. Finance can tell you which of them existed before this build.
What one unit earns, for revenue roles. It turns cost per unit into a return on the money spent.
Moves it It does not change which lane is cheapest. It adds a reading: what the unit brings in as a multiple of what it costs.
Read it Use it for sales and revenue roles. For cost centres leave it empty.
Where to find it Revenue divided by units, from your CRM or invoices. For a lead, the average deal value times the conversion rate.
The prompt, the context and the input itself, in tokens, for one run.
Moves it Multiplied by the input price per million tokens. Long context and big documents push it up fast.
Read it Measure a few real runs. Typing your own count makes the model cost yours.
Where to find it Whoever runs the model reads it off the usage block in any API response. Your provider's usage page has the monthly totals. Roughly 750 words is 1,000 tokens.
What the model writes back, in tokens, for one run.
Moves it On most models an output token costs several times an input token, so a verbose model costs more than a long prompt.
Read it Short structured outputs are cheap. Long reasoning traces are not.
Where to find it The same usage block, output tokens. With nothing built yet, take a typical answer's length in words and multiply by 1.3.
Your own input price in USD per million tokens, negotiated or measured.
Moves it Replaces the list price for the input side of the model cost.
Read it Only use it when you actually pay it. Otherwise pick a list model and keep the published tag.
Where to find it Your contract or your provider's invoice. If you pay list price, pick the model instead and the published figure is used.
Your own output price in USD per million tokens.
Moves it Replaces the list price for the output side of the model cost.
Read it Same rule: only a price you actually pay.
Where to find it Your contract or the provider's invoice, the same place as the input price. Output is billed separately and usually costs more, so never copy one into both.
The rate that converts the USD list price into the page currency. Only the model price converts; your amounts never do.
Moves it It scales the model cost per run and nothing else. Switching currency drops a typed rate.
Read it The ECB reference rate is used unless you type your own. After 90 days a published rate becomes an assumption.
Where to find it The ECB publishes euro reference rates every working day, and the page names the one we used. Type your own only when your provider bills you differently.
Five questions the arithmetic cannot answer. They can cap a lane or block a build, however cheap the lane looks.
Moves it Filled disc: clear. Half disc: a warning. Ring with a bar: caps the lane or blocks the build.
Read it Answer them honestly first. A cheap lane that is not permitted is not an option, and the page says so.
Whether the AI output meets the standard the person is held to today.
Moves it Not yet blocks every AI lane until the bar is met, however cheap it is.
Read it Sample real outputs against the real standard, not a demo against a friendly one.
What happens when the output is wrong: reversible cheaply, recoverable at a cost, or not at all.
Moves it Irreversible caps the lane at a person using AI: a person stays in the loop.
Read it Think of the worst plausible error, not the average one.
Whether regulation or a contract requires a named person to answer for each output.
Moves it Required caps the lane at a person using AI. Informal accountability only warns.
Read it If someone signs it, someone must have read it.
Whether the knowledge the work needs is documented and reachable, partly written down, or in people's heads.
Moves it In people's heads caps at a person alone. Partly caps at a person using AI and warns.
Read it AI can only use what it can read. Tacit knowledge is the most common hidden blocker.
How often the process itself changes: rarely, a few times a year, or constantly.
Moves it Constantly blocks the build when the build cost is above the 5,000 floor. Below it nothing costly is riding on the process, so it only warns.
Read it Stabilise the process first. Automating a moving target is how budgets disappear.
Every figure on the page says where it came from: your input, our assumption, or a published price with a source.
Moves it Typing a number or importing a file makes it yours. Picking a list model makes the price published.
Read it The card only unlocks when the two figures that decide this role are yours. Ours are starting points, not benchmarks.
A figure you typed, chose with Use this value, or imported.
Moves it It counts toward the honesty gate when it is one of the two inputs frozen for the role.
Read it Only tag as yours what you would defend in a meeting.
Our starting figure for this role. Plausible, deliberately round, and not a benchmark.
Moves it It stays until you replace it. Nothing on the page pretends an assumption is measured.
Read it If the verdict rests on an assumption, that is the number to go and measure.
A figure taken from a vendor pricing page or the ECB, with a source link and the date we checked it.
Moves it The ECB rate becomes an assumption 90 days after its date. A model price stays published with its checked date until you switch to your own price.
Read it Open the source. If it moved, tell us.
The inputs ranked by how far a 20 % change in each moves the answer, using your current numbers.
Moves it It reshuffles as you enter numbers. Yours are marked; ours are the ones worth measuring next.
Read it The top row is the number to check first. If it is still ours, the verdict is still ours too.
The one input that would most change the answer right now, asked with a slider, a field and Use this value.
Moves it Enter it and the page asks the next one. The two frozen inputs for the role always come first.
Read it Type what you know. Use this value when our figure is close enough to yours.
The Yield Card stays locked until the two inputs that decide this role are yours, whatever else you enter.
Moves it The two are frozen when you pick the role, so the pips count down 2, 1, 0 and never chase a moving ranking.
Read it It is not there to collect emails. It is there so nobody shares a card built entirely on our assumptions.
Every input the model reads, in four groups, plus the model picker, the derived cost per run and the exchange rate.
Moves it Edit any row and the whole page recalculates. The summary counts which figures are yours.
Read it Start with the two the page asks for. Then fix anything that looks wrong for your team.
A one page summary of this case: verdict, lanes, thresholds and every figure with its provenance, shareable by link and printable.
Moves it It reflects the page exactly. Change a number and the card changes.
Read it Share the link, not a screenshot. The link carries every figure and where it came from.
What to do with this verdict, and where in the AIOS material it leads.
Moves it It changes with the verdict: a clean Automate points at the gates, Keep it human points at the real constraint.
Read it If you want a second pair of eyes on the case, say what you are looking at.
The currency your amounts are shown in. Your figures stay exactly as typed.
Moves it Only the model price converts, at the rate shown in the drawer. Switching currency drops a typed rate.
Read it Pick the currency your rates are in. Nothing is converted behind your back.
Import a CSV or JSON of inputs, export the current numbers as a template, or copy a prompt that asks Claude to fill it from your data.
Moves it Imported figures are tagged yours and count toward the card. Nothing is uploaded.
Read it Export, fill in what you know, import. Check anything an assistant estimated before you trust it.
