AI Engineer World's Fair 2026

Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg

Dor Sasson· Stigg20:08

Loading slides…

Read the talk

Every AI Company Is Accidentally Building a Bank

Selected presentation frame from Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg at 34 secondsOpen full source frame
Title slide: “Every AI company is accidentally building a bank.”

Dor Sasson explains why AI spending needs a decision before inference, a reservation while work runs, and settlement afterward—and why shared credits, autonomous agents and enterprise budgets make this an infrastructure problem.

From a talk by Dor Sasson

At a glance

Ideas worth remembering

  • Check entitlement synchronously before inference; reconcile and settle actual usage asynchronously afterward.

  • Concurrent consumers need coordinated reservations, because several requests can otherwise authorize spending against the same balance.

  • Preserve credit sources and consumption ownership: a single total cannot express drawdown priority or team, user and agent spending controls.

  • Recheck and reserve as agents take on more work, because the full cost may be unknown when the task begins.

  • Consumption behavior can change effective spending even when the per-unit price stays fixed, making pricing and runtime infrastructure closely connected.

A subscription can hide an uncontrolled compute bill

In April, a series of AI product changes looked like pricing emergencies: Anthropic restricted third-party agents such as OpenClaw from using subscription plans, OpenAI changed its Pro offering, and GitHub restricted free Copilot access. Dor Sasson, co-founder and CEO of Stigg, uses these episodes to introduce his claim that every AI company is accidentally building a bank. Selling access to software now means deciding who can spend a shared resource, how much they can spend, and when to stop them.

Selected presentation frame from Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg at 138 secondsOpen full source frame
A slide lists April changes from Anthropic, OpenAI and GitHub.

The OpenClaw example makes the mismatch concrete. Under Claude Max, a subscription payment covered agent-driven consumption that could impose much greater costs on Anthropic. In Sasson’s account, the system could not effectively distinguish consumption appropriate to a subscription from consumption that belonged on the API, and the response was to stop access. These incident descriptions and the diagnosis of the companies’ internal controls are Sasson’s account; they do not establish the precise costs or implementation failures inside those companies.

The same exposure can appear inside a customer organization. Sasson cites Uber exhausting its annual AI consumption budget within weeks and a Replit example in which three users could consume an organization’s entire credit pool. The important relationship is between a large shared allowance and a small number of consumers capable of spending it quickly. A contract approved for a year does not, by itself, make the resource last a year.

0:120:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Move permission checks ahead of inference

The architectural failure is a timing problem. A user or agent receives access, burns tokens, and generates usage records. Only afterward does the system reconcile the spending against what that consumer was entitled to use. By then, the compute has already been consumed. An invoice can describe the loss, but it cannot prevent it.

That ordering creates problems on both sides of the sale. Customers face sticker shock or unexpectedly exhausted budgets. Vendors face overages, compute spikes and potentially lower margins when consumption exceeds what they planned to supply. Usage logging remains necessary, but a record of completed work cannot serve as permission for work that has already happened.

The proposed ordering checks access synchronously on the request’s hot path, before inference starts. Reconciliation and settlement happen asynchronously afterward. The ATM analogy captures why: permission to withdraw cash must be checked before the machine dispenses it. AI inference has the same irreversible-spending problem, even though final usage accounting can finish later.

Selected presentation frame from Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg at 448 secondsOpen full source frame
A two-column slide contrasts checks after inference with a synchronous check before inference and asynchronous settlement afterward.

Where does the decision have to move to prevent spending? The comparison below separates the permission decision from the later accounting step. Its useful distinction is temporal: settlement can follow inference, while authorization has to precede it.

Permission before consumption; accounting afterward

Run inferenceChecks afterward — Compute is consumed before entitlement is checked.
Log usageChecks afterward — Record the consumption that already happened.
Check and settleChecks afterward — Discover overspending after resources have been used.
Check synchronouslyChecks before — Decide whether the request may consume resources.
Run permitted inferenceChecks before — Consumption follows the access decision.
Reconcile asynchronouslyChecks before — Settle usage after the work completes.
Checking completed usage discovers an overage. Checking the request first can prevent unauthorized consumption.
Read the diagram as text
  • Run inference. Checks afterward — Compute is consumed before entitlement is checked.
  • Log usage. Checks afterward — Record the consumption that already happened.
  • Check and settle. Checks afterward — Discover overspending after resources have been used.
  • Check synchronously. Checks before — Decide whether the request may consume resources.
  • Run permitted inference. Checks before — Consumption follows the access decision.
  • Reconcile asynchronously. Checks before — Settle usage after the work completes.
  • Run inference → Log usage: Consumption produces usage.
  • Log usage → Check and settle: Checks follow logging.
  • Check synchronously → Run permitted inference: Allow consumption.
  • Run permitted inference → Reconcile asynchronously: Account for completed work.
5:005:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:00 · section reference included

One request needs several policy decisions

Sasson points to an architecture published by OpenAI’s financial engineering team as an example of this request-time approach. He describes a decision waterfall that considers the client, user or agent and evaluates several conditions before granting access. A balance alone cannot answer whether a particular request is allowed.

The conditions address different parts of the decision:

  • Feature access: Is this product or capability included for the consumer?
  • Rate limits: Does the request fit the limits of the consumer’s plan?
  • Trial and promotional status: Do special program rules apply to this consumption?

Together, these conditions determine whether the user or agent can draw down resources now.

Selected presentation frame from Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg at 517 secondsOpen full source frame
A decision-waterfall diagram shows runtime evaluation with supporting policy panels.

Two properties make the waterfall useful: it runs as a single synchronous evaluation at the request, and its rule priority is deterministic. The system knows which rules take precedence before it encounters the request. Sasson treats this as a financial decision system that operates ahead of billing: an invoice records a commercial outcome, while this evaluation controls whether the outcome may occur.

8:008:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:00 · section reference included

A balance check alone cannot stop double spending

Consider Sasson’s account with ten dollars and two parties entitled to withdraw from it. If both inspect the account at the same time, both can see the same ten dollars. Each request looks affordable in isolation. The failure appears when both act on that observation: one balance has supported two spending decisions.

A hold changes what the next consumer can spend. Once one withdrawal reserves the funds, those funds are committed to that work rather than freely available to another withdrawal. Settlement follows when the actual consumption is known. Applied to AI, the reservation must affect subsequent access decisions before other requests can spend the same pool. With ten thousand agents reaching for a shared balance, checking early helps only if concurrent checks and reservations cannot all authorize the same money.

Sasson groups hold-and-settle, double-entry accounting, idempotency and auditability among the banking properties AI products need. They address related concerns: reserving spend, recording its movement, avoiding duplicate effects from repeated requests, and retaining a history that can be inspected. Double-entry records alone do not specify how competing reservations are coordinated; the talk identifies the required properties without developing a transaction or locking implementation.

10:2911:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:29 · section reference included

Credits have sources, and spending has owners

Even a correctly coordinated balance can discard information the product needs. Sasson’s example is a pool of 1,700 credits assembled from different sources. Some credits might have been granted earlier; others might be promotional. The total answers how much exists, but it cannot answer which credits should be consumed first. That requires preserving the sources and defining a drawdown order.

Selected presentation frame from Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg at 759 secondsOpen full source frame
A balance slide shows 1,700 credits alongside three category cards.

This is why storing credits as a single integer becomes limiting. Once sources have different treatment, consumption is also an allocation decision. An oldest-first policy and a promotional-first policy can produce different remaining pools after the same amount of work. The talk presents those as choices to resolve, rather than prescribing one universal ordering.

The ownership model also grows beyond a flat organization-to-user relationship. An enterprise may approve a seven-digit annual pre-commitment, then distribute its consumption across teams, users and agents. Each level can require allocations, budgets and spend caps. The person approving the contract needs to see both how much remains and who is consuming it over the contract’s life.

The two dimensions belong together: credit provenance determines which funds a request consumes, while organizational attribution determines whose activity consumes them. A top-level balance cannot supply either form of fine-grained control on its own.

12:0612:36
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:06 · section reference included

Four patterns for selling AI at scale

The architecture comes together in four recurring patterns. They cover the timing of spending, the location of accounting data, the uncertainty of agent work, and the customer’s ability to understand consumption.

Selected presentation frame from Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg at 897 secondsOpen full source frame
A slide introduces “Four enforcement patterns” in four panels.
  • Reserve before inference; settle actuals afterward. Permission and reservation precede costly work. Accounting then reconciles the reservation with actual usage. This connects request handling to the concurrency problem: simultaneous workloads must not each claim the same available credits.
  • Keep metering and ledgers inside the VPC when needed. Sending usage events to an external service raises latency, cost and data-sovereignty concerns. Sasson expects more companies to place these systems within their own virtual private cloud. The placement choice responds to those concerns; it does not remove the need for metering or settlement.
  • Recheck balances as agents continue working. An agent’s eventual actions and total cost may be unknown at the start. As it takes on further work or spawns more activity, the system needs additional checks and reservations, followed by asynchronous reconciliation.
  • Make consumption visible across the business. CFOs and CIOs need views across models, features and products, alongside attribution to teams, users and agents. Sasson expects this visibility to become a prerequisite for enterprise purchases.

Agentic checks add an important limit to the initial ATM analogy. A cash withdrawal has a known requested amount; an agent task may keep discovering more work. Admission at the beginning cannot authorize an unknown amount of future spending. The check-reserve-settle pattern therefore has to recur as the task progresses.

14:3615:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:36 · section reference included

Consumption rate can change the price customers experience

Near the end, Sasson describes an April change in which the per-unit cost stayed the same while token consumption increased. His spoken account names OpenAI, while the displayed slide identifies Anthropic and a Claude tokenizer change. That provider attribution is inconsistent across the presentation; the general mechanism is the supported point: the unit price alone does not describe what customers spend when the same input consumes more tokens.

The mechanism is straightforward: at the same cost per token, consuming more tokens increases total cost. Executing such a change therefore involves the infrastructure that governs consumption, rather than only editing a price list. Rate of use, credit drawdown and the customer’s budget are connected parts of the product’s economics.

Sasson closes by comparing the opportunity to the early-2010s “Stripe moment,” when reusable payment infrastructure made checkout easier for developers shipping internet products. His proposed equivalent for AI would make reservations, settlement, ledgers and spending controls easy to adopt. Coming from Stigg’s co-founder, this is also a product thesis: teams should be able to keep shipping AI without independently reconstructing these financial mechanisms.

The practical warning is to recognize the bank-like responsibilities while the system is still being built. Shared pools, concurrent consumers and uncertain agent costs become harder to correct after revenue and usage scale. You may not have set out to build a bank. Once your product grants access to spendable resources, its spending rules are part of its architecture.

2:1317:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:25 · section reference included

Transcript download
Full transcript

Loading transcript…