Operations / Economics & ROI
Economics & ROI
The unit economics of agents — pricing, cost attribution, ROI measurement, build-vs-buy, failure modes.
- Build vs Buy vs OrchestrateFive of the six biggest agent frameworks ship MIT or Apache-2.0 and OpenAI retires its own visual Agent Builder on 30 November 2026 in favour of its SDK, so "build" no longer means writing the loop — the runtime is $0.085 per vCPU-hour, and per-seat prices ($19 Copilot Business with 1,900 credits at a cent each, $30 M365 Copilot) are the floor rather than the bill now that products meter per outcome at $0.99 a resolution or $0.10 an action. What is still unpurchasable is the eval set, the labels and the verifier. And the deflation rate is not constant: a capability costs 66% less per quarter while it is new and only 32% less two years on, so the layer never passes through a window where building it pays.
- Agent Unit EconomicsCost per successful task is the right unit, and the figure you compute for it is only as good as the verifier that decides what counted — a judge with an 8% false-positive rate moves the unit by 4%, systematically and in the wrong direction, which is more than most token savings ship in a quarter. Hand-labelling 300 attempts leaves a ±5-point interval and a $0.69–$0.80 band around a $0.74 unit, so publish the point estimate, the interval and the verifier together, load escalation, eval, escaped errors and the standing bill into the numerator, and underwrite at p95.
- Per-customer economicsWhole-system unit economics hide which customers cost you money — a per-tenant cost view, what drives the heavy-tail user, and the levers you actually have.
- Cost Attribution & BudgetsAn estimated cost carries error that ranks your features in the wrong order: the same 8,000-token prefix costs 10,000 units cold at the 1.25x cache-write rate and 800 warm at 0.1x, so a busy feature is overstated tenfold and a quiet one understated by a fifth — and the one you kill is the cheap one. Record the provider's returned usage object, not your tokenizer's estimate, because cache state, reasoning tokens and downstream retries are reconstructible from nothing you sent. Then store counts rather than dollars, reconcile against the invoice monthly and publish the gap, tag the model and effort setting as dimensions, and make the budget a counter in the request path with a velocity breaker.
- Measuring Agent ROIMcKinsey's 25 August 2026 survey found 37% of firms reporting positive EBIT impact — unchanged in a year, while agent scaling at large firms went 27% to 40%. What is scarce is measurement: the MIT 95% is a measurement statistic read as a failure statistic, and the baseline stops being observable the day you ship. One worked quarter scores +6.0% or -26.9% on one number — the deflection rate you did not record before launch.
- Pricing & Packaging Agent ProductsEvery agent price is a decision about who carries the cost variance: seat pricing leaves it with the vendor, where a 5% heavy-user tail can drive more than half of cost; usage pricing hands it to the buyer, as Cursor’s 2025 switch showed; and outcome pricing turns the definition of “resolved” into the real contract.
- Where the Economics BreaksEvery inversion here moves the denominator while cost per token holds flat, and the largest multiplier is a default nobody chose: openai-python and anthropic-sdk-python both ship DEFAULT_MAX_RETRIES = 2, and Temporal retries Activities with Maximum Attempts unlimited. The break-even escalation rate is the all-human average task time over the escalated task time — the token bill does not appear in it at all, which is why a cheaper model is not a response. Measure cost per accepted outcome at the p99.
- Provisioned Throughput & CommitmentsA 30% discount means breaking even at 70% sustained utilisation, which bursty agent traffic never reaches — so reserved capacity is a latency guarantee you can put in front of a customer, not a saving, and a term commitment quietly freezes your model choice in a market that moves quarterly.
- The Cost of EvaluationEval spend scales with change rate, not traffic, and its price is set by the smallest regression you insist on catching — which costs quadratically, so a two-point threshold buys roughly 7,700 runs where a five-point one buys 1,400.
- The Cost of Human ReviewThe reviewer costs ten to fifty times what the tokens do, and it is the only line that does not shrink when the agent gets better — because a reviewer has to read the correct outputs too; only calibrated selective review removes it.
- Scaling Back an Agent DeploymentNearly half of enterprise leaders cut an agent deployment last quarter over cost, and "scaled back, narrowed, delayed or paused" is four names for one blunt instrument — inside a single deployment one workflow can clear its manual baseline by 55× while another loses money every run, so cut the task class instead, carve out the tail before condemning the class, and write the resume condition down as a number.
- Free Tiers & Trial EconomicsA SaaS free tier is capped by human boredom; an agent free tier is capped by nothing, because the user is a loop and the marginal cost is real — so price the tier off the maximum a single account can consume rather than the average, meter work instead of days or seats, and enforce the ceiling at request time rather than in a monthly review.
- Forecasting Agent SpendAn agent’s per-task cost is heavy-tailed, so the mean forecasts nothing and is biased low — worse at scale, because more volume draws more of the tail. Forecast a sum over task classes carrying each class’s p95, cap every class so the tail has a finite number, drive it off retry rate, context growth, route mix and cache hit rate, and reconcile each month into volume, mix and per-task drift.
- Price Deflation & Cost per TaskFrontier input tokens cost about a sixth of GPT-4’s 2023 price and a cheap capable model a three-hundredth, yet almost nobody’s agent got six times cheaper, because agentic workflows burn 5–30× the tokens of a chat completion. The deflation lands on a unit you do not buy — which makes token-shaving a depreciating asset, a term commitment a directional bet against a three-year trend, and cost per successful task the only line worth putting on the dashboard.
- The Cost of Being WrongYou can read the token bill to four decimal places and nobody has ever computed the error bill, which on most deployments is one to two orders of magnitude larger — so price it as a product of error rate, escape rate, unit remediation cost and amplification, source the unit cost from the incidents you already had, and notice that detection latency is a multiplier you can buy down. The resulting figure is what sets autonomy per action and what finally prices a reviewer honestly.
- Fixed Costs & the Pilot TaxAgent economics are modelled as pure variable cost, so a pilot divides a total that is mostly standing bill — index, eval suite, trace retention, capacity floor, review roster, on-call — by a tiny task count and reports a number that says nothing about the agent. Name the six fixed lines, note that they move in steps triggered by audits and deprecations rather than by traffic, and report the breakeven volume instead of the cost per task.
- Pricing LatencyForty seconds of p50 on a task that blocks a $60-an-hour employee costs $0.67 against a four-cent token bill, so the cheaper model that runs twice as long is a sixteenfold cost increase wearing a discount’s clothes. The cost is convex — free under a second, linear to ten, a step change after that when the person context-switches away — so budget the p95, classify each deployment into one of four regimes before spending anything, and notice that streaming, which removes no latency at all, is often the largest dollar win on the list.
- Fractional Time SavingsForty people saving twenty minutes a day is thirteen FTEs on a spreadsheet and zero dollars in any budget anybody controls, because a fifth of a person is not a line item you can cut. Freed capacity becomes money only where it relieves the binding constraint, which gives exactly four auditable shapes — absorbed growth, cycle time on a priced clock, avoided external spend, a named deferred req. Meanwhile the cost side is fractional too and only one side gets instrumented, so delete the hours-times-rate line and see what survives.
- Energy & Carbon per RunPublished figures for a single text prompt run from 0.10 Wh to 14.6 Wh, and one paper puts the same median prompt at 0.10 Wh for active accelerators and 0.24 Wh for the serving stack around them, so every figure you publish is a boundary choice before it is a measurement — and all the public ones were measured on a single chatbot turn, which an agent multiplies by ten to forty. Report Wh per completed task with the boundary stated, build it from tokens-per-run in your own traces divided by the success rate, give location-based and market-based carbon side by side, and note that the top four levers all attack the multiplier while choosing a smaller model is last.
- Fewer Tokens vs Cheaper TokensA model that writes three times fewer tokens and one that charges three times less per token are the same discount on a rate card and nowhere near it on a twenty-step loop: the price cut pays the full 3x, brevity pays about 1.7x, because a generated token is billed once at the output rate and again as input at every later step while the prefix and the tool results shrink for neither. One figure decides which lever to buy — the share of the bill your model's own writing controls — and prompt caching raises it from 0.63 to 0.79 on the worked example, which is why caching is what makes a terse model worth a premium.
- Tokenizer Changes & Your Cost BaselineA 20% price cut paired with a tokenizer that emits ~30% more tokens for the same text is a 4% price rise: on October 2026 list prices, Opus 4.6 to Opus 5.5 reads as a discount and lands near break-even, while Opus 4.7 to Opus 5.5 is the same cut and delivers it. The token is a vendor-defined unit, so the change also re-scaled context capacity (~23% less text), TPM limits, max_tokens caps and length-priced thresholds without a deploy. Measure your own per-content-class ratio, report cost per completed task, and quote the price change and the unit change separately.
- Agents Priced in CreditsA credit with no peg to a currency is not a price, it is an exchange rate the vendor can reset without touching the rate card. Notion's October 2026 rebuild gives Business and Enterprise seats 1,500 routed credits plus 500 user-choice credits a month, no rollover, no pooling, and no published conversion — while its own router decides how many credits your task burns, so the supplier sets price and quantity both. GitHub went the other way on 1 June 2026 and pegged one AI Credit to one cent. Measure credits per completed task on your own traffic, size each seat against its own peak rather than the team average, and ask in writing what can change the number and whether you will be told first.
- The Per-Step Cost of LookingA screenshot costs ceil(w/28) x ceil(h/28) visual tokens — 1,036 at the recommended 1024x768, 4,784 at the documented ceiling — and an agent loop re-sends it, so a 40-step task pays for 820 image reads, not 40. Caching an append-only transcript turns that quadratic back into a line and is worth 6.4x; dropping the frame to 1024x768 is worth 2.6x; halving the model's input price is worth 2x; and trimming to the last three frames on top of a working cache is worth 1.09x while destroying the cache that paid for everything. Buy detail with a zoom on the region you are about to click (1,201 tokens) rather than with a bigger frame (4,784).
- The Cost of a Voice MinuteVoice is the one agent workload billed by wall-clock, and the same second is metered in parallel by the carrier, the platform, STT and TTS while the token line — the only one anyone optimises — is about 4% of a worked 4-minute call. So a 700 ms endpointing timer over twenty turns costs more than the whole model bill, and the τ-Entity confirmation scaffold's 21–28 extra seconds breaks even if it prevents one failed capture in seventy. Manage cost per resolved call: fifteen points of containment is worth roughly three times halving every rate you pay.
- The Cost of an Idle AgentOnce a run outlives one request the bill measures resources held, not work done, and the longest hold is a human approval nobody costed: a 2 vCPU / 4 GiB sandbox left alive through a four-hour review bills about $0.66 of nothing per run, roughly thirteen times the model spend. Providers differ on whether idle is billed at all, so the idle policy and not the per-vCPU rate is what you are buying - make waiting a state of the run rather than of a machine, and meter held seconds separately or you will never see the line.
- The Price Cliff Above the Context ThresholdLong-context pricing is a step function, not a slope, and the step selects a rate for the whole request rather than the overage: as of November 2026 Gemini 3.1 Pro preview bills $2 per million input tokens under 200K and $4 above, output $12 to $18, so a 205K request costs 108% more than a 200K one. The cached prefix you added to save money counts toward the threshold and its own rate doubles too ($0.20 to $0.40), which makes a 180K cache a permanent position 90% of the way up the cliff. Trimming 3% of tool output halves the call, so cap request size below the threshold, fire compaction at the price edge rather than at a fraction of the window, and record each model's threshold as metadata — vendors differ and change.