If the Agents Stop, Who Still Knows How to Do the Work?

The coming AI IPOs may tell us more about the financing of intelligence. They should also make us ask how a customer keeps working when a provider cannot.

Imagine arriving at work on Monday to find that an agent which handles an important process is unavailable. It has not gone rogue. Its model provider has changed its terms, its compute supplier has lost capacity, or the service has simply stopped. The team can see yesterday’s completed cases, but cannot easily reconstruct why the agent made each decision or move the unfinished work elsewhere.

That is a business-continuity problem. It is also a useful way to read the finances behind AI.

OpenAI and Anthropic have each announced a confidential draft S-1 submission, giving them the option to go public. Neither announcement sets an offering date or price. The public documents, when available, should bring more detail about risks and obligations. For now, we need to distinguish company disclosures from reported expectations about the IPOs. OpenAI; Anthropic

The numbers already disclosed are striking. OpenAI says its latest funding round closed with $122 billion in committed capital. Anthropic says it raised $65 billion in its Series H round. Those figures describe financing, not profit, and they should not be casually relabelled as debt. Nor do they tell us whether any particular customer is earning a return from using an agent. OpenAI funding announcement; Anthropic funding announcement

Follow the money through the chain

The model company is only one part of the system. Someone has to build the data centre, obtain power, buy the GPUs, finance their useful life and keep capacity ready for demand that may vary by the hour.

Some of that risk is visible in public suppliers. CoreWeave reported $35.6 billion of total indebtedness at 30 June 2026. Oracle said it raised $43 billion in debt financing during its 2026 fiscal year. These are obligations of different companies; adding them together would tell us little about OpenAI’s or Anthropic’s own debt. They do show why “how much did the model company raise?” is an incomplete question. CoreWeave filing; Oracle results

The links between customers, suppliers and investors matter too. CoreWeave disclosed an agreement under which OpenAI committed to pay up to approximately $11.9 billion through October 2030, subject to delivery and service-availability requirements. Oracle says customer prepayments or customer-supplied GPUs form part of its large AI contracts. A long contract can support investment in capacity. It is not the same as cash already earned, and the strength of one company’s plan may depend on another’s ability to keep paying and delivering. CoreWeave filing; Oracle results

This does not make the investment irrational. Infrastructure often has to be built ahead of demand. It does mean we should ask where the risk sits: with the model company, the cloud provider, a financing vehicle, a customer making a capacity commitment, or investors expecting future growth to cover today’s costs. We should also ask how much demand comes from customers outside that financing chain.

A cheap run can still be an expensive job

For an organisation buying AI, the useful unit is not a token or a successful demonstration. It is an accepted outcome.

What did it cost to take a request through retrieval, model calls, tools, retries, human checks, correction and final approval? How many attempts failed? What capacity had to be reserved whether or not the agent was busy? Did the result save someone time, or did it create a new queue of review and repair work?

A cheaper model may need three attempts and a lengthy check. A more expensive model may complete the job once. Either could be the better purchase; the invoice for the first model call cannot settle it. The comparison also changes when the rules, source data or definition of “done” change.

That is the practical connection to our Digital Labour Exchange work in Agent Control. A job budget, routing among providers, verification and an auditable transaction record give us a way to compare cost per accepted outcome in bounded tests. They do not prove the economics of the wider market. The point is to make the buyer’s measure as rigorous as the supplier’s account of capacity and growth.

Software escrow solved part of an older problem

There is a precedent for asking what survives supplier failure. In conventional software procurement, an escrow arrangement can place source code and supporting materials with an independent party, to be released under defined conditions such as insolvency or failure to maintain the product. UK government guidance treats escrow as a possible protection for critical technology and stresses the need for an exit plan covering data, knowledge, dependencies and transfer of responsibility. UK supplier guidance; Digital, Data and Technology Playbook

An agent needs a broader continuity test. Source code alone will not recreate a hosted model, the capacity to run it, the rights to use it, or the current state of a live workflow. And a folder of prompts is little help if nobody knows which version was authorised, what data the agent used, which tools it could call, or which cases it left unfinished.

For a consequential process, I would want the organisation to retain a dated record of its work definitions and acceptance rules; the agent’s tool permissions and their owners; source and decision history; open cases and handoff points; and tests that a replacement process can run. The replacement might be another provider, a smaller local model or a human team. It need not produce identical wording. It must meet the organisation’s current standard and preserve the evidence needed to explain past decisions. NIST’s AI Risk Management Framework likewise calls for safe decommissioning and contingency processes for failures of high-risk third-party AI systems. NIST AI RMF

This is where the financial and operational questions meet. A company may be well funded today and still change a product, price, model or capacity allocation tomorrow. A customer should be able to absorb that change without losing its process or its institutional memory.

An IPO prospectus can help us examine the financing chain. It cannot answer the customer’s most immediate question on its own: if this agent stopped tonight, what work would stop with it, who could take over, and could they tell what had already been done correctly?

Also published on Medium ↗.

This article reflects the evidence and development status at its original publication date.