OOOMETA/OS
← Back to Insights

November 2026 · 5 min read

AI costs a cent per decision; the wait is 7 days

AI costs a cent per decision; the wait is 7 days

Key Definitions

Decision service time The elapsed time from a decision entering the queue to its close. It is set by the people who can sign and by the items ahead, not by model speed.

90th percentile cycle time The close time at the 90th percentile of a period. It shows the tail that an average hides.

WIP limit The cap on open decisions one person or unit carries at once. Work beyond the cap waits in line instead of running in parallel.

On the morning of November 16 the Discord thread held 742 messages from the previous thirty days. They came from the same set of scheduled jobs: price checks, signal scans, drafts, builds. On the same morning the decision queue held 31 items still open, and the oldest had been sitting there for eleven days.

Those three numbers together are the clearest comparison I have seen this year. The production end is too cheap to matter. The decision end is still congested, and it congests in a way that does not respond to how much is produced.

The production end: two dollars thirty one cents

Between October 17 and November 16 the scheduled jobs completed 990 runs; nine missed their scheduled time and twenty deliveries failed. Behind those runs sit 1,370 model sessions, 874 million input tokens, most of them cache re-reads, and 15.3 million output tokens. At our billing rates that is USD 2.31. Per session it is 1.7 cents, and per message that reached a person it is a third of a cent.

That is the marginal cost today. Ten more workers on the floor add cents to the bill. Halving model calls saves less than a cup of coffee. Optimisation aimed at saving money here does not move the account, and a meeting about it is worth holding only if someone has already shown that cost is the constraint.

Production is cheap for a plain reason. Calls run concurrently, caches reuse the prefix, retries need no person standing by. Those three together hold the cost curve almost flat as volume grows. Every commercial conversation about seats, discounts and monthly packages happens on this layer.

The decision end: two hours median, seven days in the tail

In the same window the queue opened 359 items and closed 352; 31 are open now. For the closed items the median service time was 2.2 hours. The 90th percentile was 178 hours, more than seven days. The top tenth of decisions carries most of the elapsed time, while the median looks entirely healthy.

The 31 open items are not unworked. They are unserved. Unworked is a resourcing problem and more people fix it. Unserved is a queueing problem, and more people fix it only if the people added are review seats. Most enterprise AI accounting treats the second shape as if it were the first.

One more observation deserves attention. The 203 closed items marked high priority took an average of 84.9 hours. The 141 marked medium took 32.3 hours. The ones labelled high were 2.6 times slower. The eight critical items averaged 17.1 hours, so the labels do carry information; the problem is what they encode. High is where work crosses a unit boundary, needs an outside document, or needs a person to look back later. The label measures importance, and importance is not the same property as waiting time.

Two throughput laws, not one

Production obeys a linear law. Add money, add machines, add concurrency, and output rises in proportion. Decisions obey a queueing law. The time from entering the queue to closing equals the items ahead divided by the service rate. The service rate is set by people who can read, judge and sign, and it does not move when the model gets faster.

Put the two laws together and the ratio is fixed by structure. Raising production adds arrivals faster than departures, so waiting time grows. Raising model speed leaves departures unchanged. A doubled investment with flat returns is therefore not evidence that the models are weak. The numerator is held by the queue while the denominator grows with spend, and switching model vendors does not change either term.

Our own numbers test this. Over thirty days production went from zero to 990 runs for USD 2.31, while decisions went from zero to 352 closes at a service rate of roughly twelve per day. Double model speed and the same 31 items stay open this morning, because they wait for people, not for compute.

This also explains why pilots look good and scale looks bad. During a pilot the arrival rate is small, the median is beautiful, everyone signs off. At scale arrivals multiply by batch while the departure capacity stays at those few seats. The median may barely move and the tail breaks first. The report shows an average, the person experiences the tail, the two disagree, and the argument becomes a debate about whether AI works when the real subject is the queue.

The same shape outside

Domino Data Lab published its fifth enterprise AI report in July, based on a survey of 639 senior enterprise AI leaders. Fifty seven per cent say AI returns fail to outpace investment, a figure unchanged from 2025 to 2026. In the same survey 93 per cent report improved ability to move models into production, up from 88 per cent in 2025. Production capability rose five points and the returns line did not move. Domino, fifth enterprise AI report

The second half of the finding is closer to mechanism. In 40 per cent of organisations at least one route for business users to reach AI conclusions is mediated: a scheduled report from a data science team, or a request filed with an analyst who runs the work and returns the result. Another 34 per cent use a different method per business unit, with no consistent delivery layer anywhere. In a large share of companies the model finishes and the conclusion still passes through one more pair of hands before it reaches the person who decides.

One hand-off is one queue. The analyst is not idle; the analyst can take a limited number of requests per day. The report is not late; the schedule is weekly and this week request ships next week. Every gain upstream turns into waiting time here, and that waiting time never appears on the model dashboard. It appears in a calendar.

The same report carries a governance finding. Among organisations whose governance fully keeps pace with AI activity, 75 per cent report significantly improved delivery velocity, against 23 per cent where governance falls behind. Those organisations are 3.9 times more likely to have reached governed agentic production. This is worth reading as correlation rather than cause: well governed shops have usually also fixed who signs and when the work counts as done, and once that is fixed the delivery does not need to be renegotiated in each pair of hands. That is still the language of a queue.

MIT NANDA reached a colder conclusion in 2025: 95 per cent of organisations running AI pilots saw no measurable profit and loss impact. Domino cites the study in a footnote, authored by Challapally, Pease, Raskar and Chari. The GenAI Divide: State of AI in Business 2025

Why priority labels measure the wrong thing

Back to our own numbers. High priority took 84.9 hours, medium took 32.3. The mechanism is not hard to infer. Work that closes inside one unit is one item and one owner, done the same day. Work that crosses units adds a queue at each signature plus a context rebuild each time. The person who marks an item high is saying the item matters, and the label does not encode how many hands it must pass. The number of hands is what decides how late it comes back.

That yields a criterion anyone can recompute: whether a decision is slow depends on how many owners it changes, not on what colour it carries. It is also easier to manage, because the number of owners is readable from a record while colour is a person subjective call, and the two are different kinds of data.

The third thing measured wrong is the close itself. Many teams treat a closed ticket as a closed decision, while the ticket closes on whether somebody replied rather than on whether anyone changed their behaviour. When closing needs only one sentence, service time looks excellent and the tail hides by design. Our rule is one acceptance sentence per item, checked once at close, with a vague close sent back.

Four buyers, four numbers to move

For the C suite, change the unit of account from models deployed, runs executed and seats bought to decisions closed per week, plus the 90th percentile of their service time. The first grows linearly with budget and the second is where value sits. Cost items are measured in cents and cutting them produces no return. Attention items are measured in working days, and moving those changes the return.

For the transformation lead, the number to move is hand-offs per decision. Each hand-off is a queue. Cutting three to one lowers the 90th percentile before any model quality work pays off. The workable rule is one owner per class of decision rather than one owner per stage, because the second rule manufactures queues.

For the AI application lead, the requirement on a deliverable is not that it reads correctly but that it can be decided on arrival. Without an evidence chain and the specific question that needs a yes, the reviewer rebuilds the reasoning from scratch and the time the model saved moves intact into the review seat. That transfer is invisible in the accounts, which hold tokens and not hours of attention.

For the compliance lead, governance is not the opposite of delivery. The survey data above says organisations that fixed signing, logging and recomputation first report significantly better delivery velocity far more often. Compliance work is itself a decision and it occupies the same queue. Turning a review into a checklist decided once rather than a discussion held every time moves it out of the tail.

What we think

First, the flat return is a queue phenomenon, not a model phenomenon. The claim is falsifiable. If an organisation main bottleneck really is model quality, cutting hand-offs will not move its 90th percentile. If the bottleneck is a queue, adding ten agents will raise message volume and leave the 90th percentile where it is, and we would see both facts at once. That is an experiment any team can run this quarter, not a position to argue.

Second, accounting by token hides the resource that is actually scarce. On our books thirty days of tokens cost USD 2.31, while one person read 742 messages and opened 359 decisions in the same window. What was tight was those 742 reads, not those 2.31 dollars. A budget meeting held about tokens optimises a variable that is not the constraint, and success at that optimisation makes the real problem look solved.

Third, limiting work in progress beats limiting spend. One decision per person with a cap on how many stay open is the lever that reliably lowers the 90th percentile. We carry 31 open items today, and that number is neither better larger nor better smaller. Too low and production piles up outside the queue. Too high and the tail lengthens without anyone noticing.

We build our own products on this. Every deliverable carries a decision object, one owner and a close time. Remove any of the three and the output is a message rather than a decision, and it should not be counted as value.

Where this does not apply

When the decision is machine verifiable, as in reconciliation, compilation, staged rollout or risk blocking, it never enters a human queue. There output is value and scale converts directly into return. That is the common property of the best returning AI cases today and worth hunting first, before forcing human intensive decisions onto a model.

When volume is small, do not build this apparatus. With a few dozen decisions in thirty days, the cost of recording the times exceeds the queueing saved. Adopt the accounting once daily volume reaches the teens and someone starts complaining about the wait.

One case needs separate handling. When review quality itself is the constraint, the way to lower the 90th percentile is not more reviewers but a written standard. Without a standard, extra reviewers only approve more weak judgments faster. The queue shortens, rework lengthens, and total time does not move.

Three steps today

First, make the decision an object with an owner. Not a line in a task list but a record with an entry time, a close time, one owner and an acceptance criterion. We use a ticket store, and any record can be asked how long it has been open and whose hands it passed through.

Second, measure the 90th percentile rather than the average. Averages are pulled down by the many items closed the same day they arrive, which hides the real wait. The 90th percentile is the number the next budget meeting should discuss.

Third, give each person a work in progress limit and let the excess wait rather than run in parallel. A limit governs the total, while priority only reorders the same pile, and the limit decides the 90th percentile.

One question for the buyer. Last quarter, how many AI outputs were actually used by a person to make a decision, and how long did those decisions take from entry to signature? If that cannot be answered, the unit of account is still on the production side, and doubling the investment will leave the returns line exactly where it is.

References

Internal figures come from our own run records for October 17 to November 16, 2026: 990 completed scheduled runs, nine missed, twenty failed deliveries, 742 messages delivered; 1,370 model sessions, 874 million input tokens, 15.3 million output tokens, USD 2.31; 359 decisions opened, 352 closed, 31 open, median 2.2 hours, 90th percentile 178 hours, 203 high priority items averaging 84.9 hours, 141 medium averaging 32.3 hours, eight critical averaging 17.1 hours.

External figures are vendor reported survey results, quoted as reported. Domino Data Lab, fifth enterprise AI report, July 21 2026

MIT NANDA, The GenAI Divide: State of AI in Business 2025

FAQ

Thirty days cost only USD 2.31, so why are returns still flat+

Because cost is not the constraint. In the same window one person read 742 messages and opened 359 decisions, and attention was the scarce item. Token cost sits in cents, and saving it changes no value.

What is the relation between 990 runs and 352 decisions+

990 runs completed, 742 messages were delivered and 352 decisions closed. Production and decisions are different layers: the first scales with concurrency, the second queues behind review seats.

Why is high priority slower than medium+

The high label lands on work that crosses a unit, needs an outside document or needs a person to look back, and every hand-off is a fresh queue. The label never encodes how many hands the item passes.

How do you measure the 90th percentile+

Each decision needs an entry timestamp and a close timestamp. An average is pulled down by the many items closed the same day, which hides the tail.

Should machine verifiable decisions be measured this way+

No. Reconciliation, compilation and staged rollout never enter a human queue, so output is value and scale converts straight into return.

Does governance slow delivery down+

In the Domino survey, 75 per cent of organisations whose governance fully keeps pace report significantly better delivery velocity, against 23 per cent where it falls behind. Governance fixes who signs and when work counts as done, which is exactly the queue rule.

Run the same check on your own execution log

Send a redacted 30-day agent execution log and get an eight-row table in 48 hours: one row per technical item, each cell carrying its numerator, denominator and criterion, gaps sorted by whether a schema change is needed. First three engagements are ¥8,000.

See the field check →