Expound/The economics
What does it cost when nothing computes reliance?
When systems cannot compute whether a result may be relied on, people compensate. They rerun tests, reconcile dashboards, compare versions, request approvals, reconstruct evidence, hold exception meetings, reopen tickets, repeat model calls and rebuild audit trails after the fact. That work is the trust layer of modern operations. It is enormous, it is paid at expert rates, and it does not appear on any budget line.
The size of it
A modeled $0.8–1.4 trillion a year.
Verification, attestation, review, reconciliation and audit-evidence work is performed by a workforce Expound models at tens of millions of people worldwide. It spans auditing and assurance, compliance and risk, quality assurance and testing, clinical and pharmaceutical review, underwriting and claims adjudication, inspection and certification, contract and title review, and the internal review time of engineers, analysts and managers who spend part of every week confirming that somebody else’s work is what it claims to be.
These are Expound’s modeled estimates, not measured markets. The technical paper states every input (the published employment and wage series, the analytic verification share per occupation, and the displacement share) so the arithmetic can be rerun against other assumptions. Five separate economic models were run and reconciled to reach the headline band.
Why nobody tracks it
No single budget line reads “trust.”
The cost is real and rarely visible, because it is fragmented across engineering, security, legal, compliance, finance, operations, support and vendor spend. Each department pays a slice; none of them sees the whole. That is what makes the trust bottleneck invisible on the books, even where it adds a lot to cycle time and risk.
Invisible cost produces the wrong optimization target. A cheaper model call, a faster workflow, a higher ticket-closure rate or a lower token count gets celebrated before anyone knows whether the result became usable.
The apparent saving comes back later as verification, retry, rework, incident, audit, remedy or liability cost, and by then it is attributed to a different team and a different budget.
Computing the answer converts that reconstruction from an unmeasured recurring cost into a measured one. Once it is measured, it can be removed.
Saved at generation≠Saved at reliance
The sourced anchors
The adjacent markets are large, real and published.
None of these is a size for the category Expound describes. Together they show where the money already moves, and why the relevant budget is the verification labor and attestation activity those markets sit beside.
| Market or spend line | Published figure | Publisher | Boundary |
|---|---|---|---|
| External audit fees, United States public companies | $21.7 billion across 6,656 registrants, fiscal 2024 | AuditUpdate, analysis of SEC filings via Ideagen Audit Analytics | An adjacent spend line, not a size for this category. |
| Testing, inspection and certification services | $263–418 billion, 2025 | Grand View Research; Mordor Intelligence | An adjacent market; the two publishers disagree, hence the range. |
| Financial-crime compliance operations | $206 billion globally per year, 2023 study | LexisNexis Risk Solutions, True Cost of Financial Crime Compliance Study | An adjacent operating cost, not a size for this category. |
| Enterprise governance, risk and compliance | $72.4 billion in 2025, projected to $203.7 billion by 2033 | Grand View Research | An adjacent software market; the 2033 figure is the publisher’s projection. |
| Software testing | $54.4 billion in 2026, projected to $99.9 billion by 2031 | Mordor Intelligence | An adjacent market; the 2031 figure is the publisher’s projection. |
| Compliance software | $35.4 billion, 2025 | Mordor Intelligence | An adjacent software market, not a size for this category. |
| Digital transaction management | $20.3 billion in 2025, projected to $59.2 billion by 2030 | Mordor Intelligence | An adjacent market; the 2030 figure is the publisher’s projection. |
| Federal improper payments, United States | $162 billion in fiscal 2024 and $186 billion in fiscal 2025 | U.S. Government Accountability Office | A measured loss, not a market; reported by the auditor, not modeled by Expound. |
| Poor software quality, United States | approximately $2.41 trillion, 2022 | Consortium for Information & Software Quality | A cost estimate by its publisher, not a market and not a size for this category. |
| Enterprise generative-AI spend | $37 billion in 2025, up 3.2× from $11.5 billion in 2024 | Menlo Ventures | Vendor-sponsored survey; adjacent spend, not a size for this category. |
Each figure is the publisher’s, cited by exact report or release in the technical paper. Ranges reflect genuine disagreement between sources.
Attestation events
Billions of events a year, each vouched for by a person.
Every document or decision that crosses from one authority to another currently relies on a person vouching for it. The volumes below are rough orders of magnitude, and they are large.
| Attestation class | Events per year |
|---|---|
| Global invoices | approximately 560 billion |
| Notarial and electronic-signature acts | over 1 billion |
| Know-your-customer and anti-money-laundering case closures | hundreds of millions |
| Public-sector eligibility and benefits determinations | over 100 million |
| Automobile loans and leases, United States | roughly 25 million |
| Control tests under financial-reporting and service-organization attestation regimes | 10–20 million |
| Machine-to-machine acceptance events, early 2030s | on the order of a trillion |
A worked illustration makes the arithmetic concrete. Fully loaded human attestation of a consumer credit contract plausibly runs to tens of dollars. A computed attestation priced at a few dollars is a large reduction per event. At roughly one hundred million such originations a year in one country, one attestation class in one jurisdiction is a fee pool in the hundreds of millions of dollars.
Where the highest-cost failures live
The spaces between the green checks.
The old model assumes a successful approval, a running process and a completed status are connected by safe intervals. Modern automation breaks that assumption. The highest-cost failures occur in the seams.
| Hidden seam | What happens today | The consequence |
|---|---|---|
| After admission, before completion | Authority expires, evidence goes stale, the executor exceeds scope, or the world changes while the task keeps running. | Damage builds up until the change is committed or a person notices; investigating afterward replaces stepping in during. |
| At interruption | Long-running work is either allowed to continue unsafely or stopped and restarted from zero. | Operators resist stopping when they should, because stopping destroys valid progress. |
| Across many small actions | Every action passes its individual limit while aggregate exposure runs away. | A desk, fleet, campaign or set of agents exceeds the risk the board intended to authorize. |
| At human escalation | A request sits indefinitely, is worked around informally, or defaults through because nobody responded. | Finality dies in an inbox; delay, liability and accountability become invisible. |
| Between verified and asserted | A result is held until convenient, a claim keeps streaming after its evidence fails, or a tentative result is promoted. | Overclaims and exposure emerge after the underlying work was technically correct. |
| After invalidation | Teams know a verdict is wrong but not every decision, report, payment, product or counterparty that relied on it. | The dominant cost becomes finding and unwinding the blast radius, not correcting the original error. |
| Across autonomous iterations | A loop cites itself, spends cumulatively, degrades gradually, or keeps running because the original authorization had no bound. | Cost and exposure compound quietly. |
What computing reliance changes
From an unmeasured recurring cost to a measured, shrinking one.
Under computed reliance, the evidence is produced at the moment of change rather than reconstructed at audit time, so proving the ten-thousandth decision costs about the same as proving the first. Audit becomes a query rather than a project. Review effort moves from confirming items to judging the cases evidence cannot settle.
Human review and attestation
Routine proof and reconstruction shift from people to retained evidence and computed finality. The $240–700 billion band above is the modeled displaceable total across every channel at maturity.
Audit reconstruction
External audit, SOX, SOC 2, ISO and evidence work form a large recurring proof-spend pool. Decision packages produced during the work turn weeks of reconstruction into a query and a replay.
Rework and duplicate validation
AI-heavy delivery increases churn, repeated review and multi-agent duplication. Blocking work that has not proved it is complete forces the gaps to be closed before downstream work builds on the defect.
Token, model, CI and runner spend
A material share of agent compute goes to failed, duplicate, retry and over-provisioned routes. Governed routing optimizes cost per accepted outcome rather than tokens per attempt. On a fixed 36-task set it reached the same outcomes with 24 model calls against 36; that set does not show lower total cost.
Compliance and risk
Improper payments, claims leakage, outages, recalls and wrongful actions dwarf the cost of a governed verdict. Proof scaled to how much is at stake is priced against the risk it removes, not per software seat.
Cross-domain attestation
Computed attestations can replace part of what is spent on human attestation, with better evidence that can be replayed.
An illustration of the routing arithmetic alone, with every input assumed for the example. An organization processes ten million verifiable model tasks a year. An all-frontier route costs $0.02 per task; a governed lower-cost route costs $0.002 including its verification allocation. Routing 70% of eligible tasks away from the frontier path reduces annual variable inference cost from $200,000 to about $74,000, before avoided retries, lower data egress, reduced vendor premiums and lower rework. At one billion eligible calls, the same unit economics produce about $12.6 million a year in variable-cost avoidance. The larger benefit is freedom of choice: work can move among frontier providers, open models, on-device models and plain rule-based computation without giving up the acceptance standard.
Measure it in your organization
Seven channels of value, each with the measure that tracks it.
| Channel | What you would measure |
|---|---|
| Avoided loss | Incident, rework and remedy cost against baseline |
| Expert time won back | Expert minutes per outcome |
| Lower cost per outcome | Fully loaded cost per accepted outcome |
| Latency and opportunity value | Time to accepted work inside the decision window |
| Accepted-Work throughput | Accepted units per period, quality held |
| Work you can newly hand off | Work newly eligible to delegate under computed acceptance |
| Trust and revenue premium | Audit effort, dispute rates, counterparty friction |
Keep every input metric you already track: tokens, model calls, GPU hours, seats. They stay in the denominator. What changes is the numerator: not activity, but work the organization is entitled to rely on. That ratio is Accepted Work per Dollar, and it is the measure that still means something when producing output costs almost nothing.
Next steps
Start where the cost is largest.
Which line above describes work your organization performs, and how many events a year is it?
Accepted Work per Dollar
The measure that counts what was accepted, not what was produced.
Explore →Field validationOne consequential decision
Freeze the baseline, write the list of requirements, measure the full cost per outcome on your own systems.
The program →FoundationsThe technical paper
Every figure above, with its inputs, so the arithmetic can be rerun.
Read the research →