Ten thousand to one

The FBI's 2025 Internet Crime Report gave "AI-related" its own crime descriptor, the first in the report's 26-year history. More than 22,000 complaints, roughly $900 million in losses, and the number is widely regarded as a dramatic undercount. A government form grew a new field.

The same openness that puts capable agents in anyone's hands lets an institution run the same class of model entirely on its own floor. Customer data never has to leave the building.

We build the cross-institution intelligence network for the agentic era, and this is the first of an annual series: every August, the year ten years out. Every number below is sourced or marked as ours.

The baseline

Illicit financial activity reached an estimated $4.4 trillion globally in 2025, up $1.3 trillion since 2023. That is a 19.2% compound annual growth rate. The figure comes from the Nasdaq Verafin 2026 Global Financial Crime Report, which surveyed more than 500 anti-financial-crime professionals. The two-year delta is on the order of a Netherlands appearing inside the illicit economy.

The baseline already compounds faster than the economy that hosts it.

The fork

2036 is not a forecast. It is a fork.

Agents are already entering finance. The fork is whether the compliance layer exists when they arrive in force.

The Bank for International Settlements, in Working Paper 1194, notes that open-goal agent deployment in finance is still minimal. The paper names this as a window for responsible deployment. The installed base is wet concrete. It hardens on a schedule.

Statista's compilation projects finance functions climbing from 20% meaningful agentic application in 2025 to 66% by 2030. Early 2026 data puts 9% of financial regulators at agentic maturity against 30% of firms. The gap between those two numbers is the clock's face.

A BioCatch survey of 1,440 banking leaders, published June 2026, found 84% name AI agents the industry's greatest exploitable vulnerability in the next year. 80% say their institution has already encountered agentic attacks. Everyone sees the fork. Seeing is not building.

Our synthesis, stated as ours: the fork closes around 2028. BIS names the window. The two-to-three-year gap between tier-1 production and mid-market deployment, which vendor assessments report, is how long it stays open.

World One: 2036 if nobody takes the threat seriously

Run the arithmetic. $4.4 trillion, grown at 19.2% a year for ten years, is roughly $25 trillion. That is our straight-line extrapolation, illustrative, and no source models a 2036 loss figure, which is why we show our own. For scale, that is on the order of a fifth of world output today.

What gets us there is not what most people picture. The Roosevelt Institute describes herding: agents with similar algorithms and training data react near-identically, the mechanism behind bank runs and flash crashes. BIS adds the sharper point. An agent given one narrow goal, profit maximization, will pursue it while ignoring unstated side goals like financial stability. OSFI, the Canadian regulator, ran a workshop in October 2025 where 44% of participants named autonomous AI the most likely primary source of AI-related systemic risk. They compared it to a rogue trader. They noted decision-path replication is "virtually impossible."

The 2036 nightmare has no thief in it. Correlation is the failure mode, and correlation lives between institutions.

The seam is where this happens. Monitoring stops at each institution's edge. The space between is out of frame for any tool that lives inside one institution. One template passes KYC once, and by afternoon there are ten thousand siblings, each institution holding one slice of a pattern nobody holds whole. BioCatch's respondents were asked how hard it will be to distinguish legitimate AI-assisted actions from malicious or manipulated ones in a future where agents commonly initiate transactions. 72% said "very difficult."

Citrini Research's February 2026 scenario, reported by TechCrunch, is the boundary of this world, not its forecast: unemployment doubling, equities down more than a third, within two years. It is a scenario, and we name it as one. The herding mechanism is what keeps it from being fantasy.

The hinge: 10,000 to 1

This piece runs on a ratio, and we build it in front of you. No source hands it to us.

The numerator: CyberArk, cited by Palo Alto Networks in November 2025, counts 82 machine identities per human in the average enterprise. Vendor projections put that past 100:1 by 2027, and we label that weak. Agents are the newest slice of that population, and a big part of why it keeps climbing.

The denominator is harder. No global census of investigators exists, so we say so plainly. Gartner's forecast of organizational flattening in finance, over half of current middle-management positions potentially gone by 2030, points the number down. The World Economic Forum describes security operations facing alert volumes that nearly outpaced human capacity, which explains why nobody is doubling the headcount.

Our extrapolation: 10,000:1 by 2036. Machine identities compound while investigator headcount flattens. Four orders of magnitude survives a wide range of assumptions about either curve. The number is ours, and we say so in the same breath.

Defense is headcount-bound. Offense is copy-bound. You can't hire your way out of a population that clones. If headcount cannot close the gap, the only instrument that scales to the numerator is a network that sees across the seam.

World Two: 2036 if the layer gets built in time

The precondition is BIS's window, used. Oversight and visibility deployed before the installed base hardens. In this world, agent activity carries verifiable records: what the agent did, when, with which permissions, on which inputs. The audit trail is first-class from the start.

Cases assemble across institutions. Nine weeks of inter-institution correspondence, say, becomes a case a human can judge in an afternoon. That is the world with the layer in place, not a shipped feature list.

The job survives. Gartner already forecasts finance professionals shifting to overseeing and coordinating AI agents. The work becomes judgment over cases no single institution could have assembled. The AI flies the plane. The human is the co-pilot. Anyone who has done this work knows the difference between a system that decides and a system that prepares.

Money arrives in both worlds. Future Market Insights puts AI in fraud management at $95.11 billion by 2036. That is a commercial forecast with wide error bars, and we label it. The spend happens either way. The fork decides whether it buys intelligence or cleanup.

Crime still happens in this world. It gets seen. The difference is the records.

The clock

The fork closes around 2028. That is our read, anchored to BIS's window and the deployment timelines.

Carry one question into your next planning meeting. If an agent opened accounts at five institutions this month and drained the sixth, could anyone at any of the six see the pattern? Today the answer is no by design. The fork is whether it still is in 2028.

This piece is the baseline. Every August, the year ten years out, we re-measure against it. The series makes it falsifiable, which is what makes it a prediction.

2036 is ten years out. The fork closes around 2028.