AI Delegation Shifts Costs, Control and Accountability
OpenAI’s ChatGPT sign-in could let app users cover model inference with an existing allowance, while Lockheed Martin says one F-35 could control as many as eight autonomous Vectis aircraft. In consumer software, military operations and digital finance, the cases raise different questions about who authorizes delegated systems, retains control and answers when something goes wrong.
When compute costs move from developer to user
AI features can be expensive to offer even when customers are willing to use them. OpenAI’s “Sign in with ChatGPT” could change who pays for model inference: instead of a developer covering each request through its own API account, a customer could authorize an app to use an existing ChatGPT allowance. TBPN’s John Coogan and Jordi Hays describe the arrangement as “bring your own compute.” For a small team, shifting that cost could make a product viable that would otherwise be difficult to price.
The pressure is not hypothetical in their account. Hays recalled apps that gained users quickly and then faced API bills as high as $30,000. A developer charging a modest subscription may struggle if heavy use of an AI feature makes the underlying model costs unpredictable. Coogan’s example of a language-learning app captures the opening: its developer, Peter Yang, had not planned to launch it because of API costs, but said the ChatGPT sign-in option could make it possible. The same logic could apply to existing products adding model-intensive features, rather than only to new AI-native apps.
That number belongs to a very different kind of delegation, but it points to a broader question: what changes when an AI-enabled system can draw on resources or act beyond the person or organization that first paid to build it? In consumer software, BYOC could reduce a developer’s exposure to inference costs and remove friction for users who already pay for model access. It does not make the rest of the product free. Coogan notes that a user’s model credits could cover generating a manuscript, for example, but not printing and shipping the book. Developers still have to pay for the non-model parts of a service, and they still have to build something people want to use.
The model provider’s role could expand along with the convenience. A direct integration might help independent apps reach users who already have an account and a paid plan. But if the integration becomes a distribution channel, the provider could gain influence over discovery, billing and the customer relationship. Coogan raises the possibility of an AI app-store economy, while treating it as a question rather than an established outcome. The trade-off would be familiar: developers might gain reach and demand while surrendering some leverage over customers or accepting a platform fee.
There is a difference between letting a user authorize model inference and handing a provider the whole commercial relationship. Coogan describes a possible “AI front door” that could eventually consolidate more billing, but the current arrangement he discusses is narrower. The distinction matters to developers deciding whether to integrate: using a customer’s existing allowance may lower one cost without resolving how the app itself earns revenue, handles other expenses or builds loyalty.
For users, the appeal may be simpler access and fewer awkward steps. Coogan describes agent-driven workflows that can navigate sites and enter payment details, but says they remain unreliable around sensitive information and consequential actions. A direct sign-in could avoid some of that friction. Yet easier access is not neutral: it may make users more dependent on the provider that controls the account and the route into third-party tools. Whether that arrangement becomes useful infrastructure or a platform with its own tolls depends on how the integration develops and what control developers retain.
From assistant to operational system
Lockheed Martin’s Vectis proposal moves delegation from a software product to a military operation. Aeronautics president OJ Sanchez says the company has demonstrated that one F-35 could control as many as eight autonomous Vectis aircraft. In Lockheed’s account, the crewed fighter would extend its reach by directing additional aircraft for sensing and missions including air-to-air and air-to-ground operations. The proposal pairs autonomous aircraft with pilots; it is not a claim that crewed fighters are about to be replaced.
The operational argument begins with information and reach. Sanchez says fifth-generation aircraft such as the F-22 and F-35 can give pilots a view of the battlespace, while their ability to act on that information may be limited by weapon capacity or range. Vectis is intended to carry that decision advantage farther and across more aircraft. Lockheed’s case is therefore not simply that an autonomous aircraft can perform tasks on its own. It is that connecting it to a crewed platform could extend what that platform can sense and do.
But the reported demonstration and the proposed force structure are not the same thing as deployment. Sanchez describes a capability Lockheed says it has shown and a wider “family of systems and platforms” it wants to build. The aircraft is not described as already operating in this configuration in service. Nor has the Pentagon set formal requirements for Vectis. Lockheed is investing before those requirements are defined, betting that future customers will want capabilities that combine existing fighters with autonomous aircraft.
The timing makes the company’s position notable. Sanchez cites about 1,350 F-35s in service with the United States and its allies, with the fleet growing by roughly 160 a year. Lockheed sees that installed base as something Vectis could augment. Its argument is partly about the aircraft, but also about planning ahead: build a system that could work with multiple platforms before a formal procurement process fixes the shape of demand.
That differs from BYOC in both scale and consequence. A consumer app may shift who bears the cost of inference and who controls the billing route; a military system raises questions about how a force distributes sensing and action across crewed and autonomous platforms. The common thread is delegation, not equivalence. In either setting, technical capability does not by itself settle who authorizes the system’s actions, who is responsible for failures or how much independent control the delegated component has. The supplied account of Lockheed’s proposal describes its intended role and claimed capability, but does not establish answers to those questions.
More access still requires trust and recourse
Stefano Scarpetta’s account of digital finance offers a separate test of what happens when more activity moves through automated systems. The OECD chief economist sees AI as potentially useful both to financial institutions and to consumers: banks already use it to help detect fraud, while tailored, timely guidance could help people understand a financial product at the moment they are considering it. But Scarpetta warns that the same technology could make illicit activity harder to identify. Greater capability can improve detection and complicate it at once.
His example of consumer guidance is deliberately practical. Rather than rely on long courses that people may not attend, an AI-enabled service might offer a short explanation when someone is about to buy a particular product. That could make relevant information easier to reach at the point of decision. It does not establish that users will understand the risks, or that a brief lesson can substitute for protections when something goes wrong. The value depends on whether guidance is accurate, accessible and connected to a system that gives consumers meaningful recourse.
Model inference is only one part of this accountability problem. In finance, responsibility can be hard to locate when systems involve multiple providers, platforms or jurisdictions. Scarpetta’s wider argument about digital financial infrastructure emphasizes clear liability, consumer protection and coordination across borders. Those conditions matter to AI’s practical use too: detecting a suspicious transaction is not enough if consumers cannot understand what happened or identify who must respond.
The burden is not evenly shared. Scarpetta says fraud and scams are increasing, with one in seven adults affected, and cites data indicating that 57% of adults have very little financial literacy. People who have limited financial or digital literacy may be less able to assess the risks of new tools or respond when a transaction appears suspicious. AI might help deliver guidance in a more accessible form, but Scarpetta’s warning about harder-to-detect illicit activity means the same deployment could also increase the demands on fraud controls.
The three cases therefore turn on different kinds of trust. Developers considering a ChatGPT integration must weigh lower model costs against possible dependence on a platform for distribution and billing. Lockheed presents autonomous aircraft as an extension of a crewed force, while formal requirements remain unsettled. In financial services, AI may help consumers and institutions navigate digital activity, but those benefits sit alongside fraud risks and questions of accountability. In none of these settings does access or capability alone answer who remains responsible. The relevant test is whether the systems surrounding delegation preserve meaningful control, safeguards and recourse.
