Skip to main content
Architecture

The Economics Model

As of August 2026. Phase 4 synthesis. Two levels: what the invoice measures, and what the decision requires.


Why two levels

Every platform meters per run: tokens, actions, credits, agent seats. Every business case is decided per outcome. The gap between those two units is where agent programmes are won and lost, and it is not a rounding error. A conversation that ends without escalation is one run on the invoice whether the customer was helped or gave up.

So this model has two levels and insists on both. Level 1 reconciles to the invoice. Level 2 decides whether to keep going.

Level 1: cost per run

The unit the platform charges for. It exists to reconcile with finance, to attribute spend to a sponsor, and to catch anomalies early.

Anchors, all as of mid-2026 and vendor-published unless noted. Outcome-priced customer service agents list between $0.99 and $2.00 per automated resolution against a $6 to $12 human-handled comparator. One major CRM platform meters at $0.10 per agent action. Where outcomes are not priced, action or credit metering supplies the unit.

The regime decides who absorbs the multiplier. This is the structural question, and it shapes architecture more than it shapes budget:

Most enterprises run all three at once. That is the two-estate problem expressed in money: your gateway meters only what flows through it.

Level 2: cost per resolved outcome

The unit the decision requires. Three terms, and the second and third are the ones business cases omit.

Cost per resolved outcome
    =   run cost / resolution rate
      + supervision minutes per item x loaded supervisor rate
      + failure rate x cost of a wrong outcome

Term 1: divide by resolution, never by containment. Containment counts conversations that never reached a human, whether or not the customer was helped. Resolution counts problems actually solved. Every documented reversal of an AI-first service programme set a containment or headcount target. Track repeat-contact rate alongside it, because re-contact multiplies true cost per issue while the containment dashboard improves.

Term 2: supervision labour is a real cost and it is usually zero in the model. The oversight capacity gate on A4 exists because supervision has a capacity limit; this term exists because it also has a price. Instrument supervision load, intervention rate and wait time per item (R12 carries the mechanics), and put the resulting minutes into the business case at a loaded rate. A programme whose unit economics only work when supervision is free has not been costed.

Term 3: wrong outcomes cost more than no outcome. Where statements bind externally, this term dominates: a tribunal has held a company responsible for what its chatbot told a customer, and a 2026 appellate decision held that general disclaimers do not provide sufficient protection. Internally the term is smaller but not zero, and R08 supplies the uncomfortable version: in observed-task testing, spreadsheet data analysis was slower and less accurate than working without the tool, and slide generation was faster and markedly worse. A negative-return task does not become positive at scale.

The who-pays-the-loop worksheet

Six questions, answered per workload rather than per enterprise. The answers determine the regime, and the regime determines a great deal of the architecture.

  1. How many runs per unit of business value? High ratio favours metered with hard caps. Low ratio favours seats.
  2. Is the work bursty or steady? Bursty work under a seat regime pays for idle capacity. Steady work under metering is predictable and often cheaper.
  3. Does the agent need to persist? Durable agents with accumulated memory fit agent-seat pricing. Ephemeral fleets do not, and treating them as pets is how the meter runs away.
  4. Which estate does it run in? Licensed platform work is not metered by you and not visible to your gateway. Budget for it separately and extract its telemetry deliberately.
  5. Who is the sponsor? Every agent needs a named business sponsor holding the envelope. This is recommended design rather than established practice: the construct is productised across identity, workforce and control-plane products, and no enterprise has yet published running per-agent budgets.
  6. What does a wrong answer cost here? This sets term 3, and it decides whether the workload belongs in a customer-facing lane at all.

Budget envelopes, not salaries

The control is a sponsor-owned envelope with unit-economics targets, hard per-run caps and variance alerting, backed by anomaly detection and showback maturing to chargeback. It is deliberately not framed as a salary: agent consumption is volatile in a way payroll is not, so the useful analogy is a budget with alarms rather than a fixed cost per head. Central funding stays on the shared platform layer; per-agent envelopes sit with sponsors.

Where tokens are the wrong unit entirely

R05 supplies the corrective. In line-of-business and operational-technology settings, value is measured in specialist hours recovered, avoided truck rolls and deferred asset replacement. A model that reports token cost per run in a utility control room is measuring the cheapest input in the system. Use Level 1 for control and Level 2 in the currency the function actually manages.

What the guide will not claim

There is no peer-reviewed unit-economics comparison of self-hosted model serving against metered APIs; the ranges in circulation come with heavy caveats and are published as such in R01. There is no published head-to-head cost comparison of embedded versus external agents on identical tasks. And no enterprise has published running per-agent budget envelopes. Where an economics claim in this guide rests on vendor pricing pages rather than on outcomes, the text says so.

Sources

research/R12-observability-and-finops/, research/R09-experience-and-channels/, research/R08-productivity-and-collaboration/, research/R13-operating-model/, research/R14-agent-data-engineering/, research/R01-infrastructure/, research/R05-lob-and-ot/. Loop multiplier figures are vendor-published (2025). Outcome and action pricing are vendor list prices as of mid-2026 and sit on the quarterly re-verification list.


Source: synthesis/economics-model.md in the evidence repository behind this site.

On this page