AI, APIs, and the End of the Firm as We Know It

Forget traditional business structures. AI agents demand radical efficiency through APIs, forcing a complete overhaul of how businesses survive.

Hi there,

The world of AI marches on. This week, I did some research on industrial organisation, a part of economics that includes concepts like the theory of the firm, market power, and platform economics.

What follows is a speculative piece of thinking. It’s a work-in-progress, and I have more essays planned in this vein, but it argues that AI accelerates the firm’s collapse into a flexible concept centred on AI agent optimisation of APIs.

An API, or application programming interface, is a systematic way of getting something from a computer. Every app relies on APIs to deliver everything from internal application data to maps or payment capability. Human coordination could be considered a less efficient form of API. You are asked to perform a task, then execute it, in a much more variable way than an API call.

Stripe built a business around making payment APIs easy for developers. Instead of optimising for developers’ happiness, firms must now optimise for AI agents’ ease of interaction. This builds on Jean Tirole’s work on two-sided markets - firms become coordination mechanisms when marginal costs drop beyond a certain level.

How AI Breaks Traditional Economics

Ronald Coase's theory of the firm rests on the importance of transaction costs. Companies exist because internal coordination is better than haggling in the market for everything they need to produce goods and services. Coase identified three examples of costs pushing work inside a firm's boundaries: finding suppliers, negotiating terms, and ensuring output quality.

AI drastically reduces these costs. When an AI agent needs data it doesn’t have, it scans the available API endpoints and gets what it needs for a few cents. The firms building microservices-type architectures were headed in the right direction.

Still, this next evolution with AI means that every internal and external interaction needs to happen in a systematic way that an AI agent can interact with and optimise. For the technology architects in the audience, this is a concept I’m trying to keep simple for this article, so no stone-throwing yet, please.

The code will work or not. An API call will succeed or fail - the focus of monitoring shifts from human oversight to system reliability over time. A lot of this stuff already exists inside technology functions.

The step-change is that every function inside a firm has to operate like a technology function using best practices today. Industrial automation already puts sensors everywhere in a warehouse or factory line. The AI imperative takes every single process or capability that hasn’t been completely automated and pushes it to the machine, making it so.

Goldfarb and Tucker documented how digital technology reduces five cost categories: search, replication, transportation, tracking, and verification. AI pushes this to the extreme. The costs don’t just shrink; they become negligible. We move towards a “perfect competition” world, where marginal revenue equals marginal cost. As the economics of the firm shift, a business must respond or fail.

How AI Systems Behave

There’s a lot we don’t know about AI model behaviour. I read through Anthropic’s Claude 4 System Card recently, and the amount of safety testing and guardrails they put on a model before release is fascinating. Claude demonstrates an ability to behave in surprising ways, and when these models are at the core of operating every remaining business, there will be social and ethical consequences.

Calvano’s research on algorithmic pricing revealed concerning patterns, given what we’ve seen with model errors so far. Without cooperation instructions, pricing algorithms learned to maintain higher prices together through repeated interaction. These algorithms favoured other algorithmic players, sidelining human decision-makers.

This goes beyond models potentially colluding on pricing and engaging in the theoretical “perfect price discrimination”, where everyone gets a variable price at the checkout depending on their real need for the product. When AI agents select API-centred business partners for speed and predictability, there will be path dependence that could lock in whoever has the best API documentation or functionality today.

There will come a point when focusing on human customers no longer makes sense. The firm's sole priority will be to optimise the ease with which AI agents can do business with you via APIs. There will be no board meeting.

This will be a challenge for executives: every deferred or denied technology project that should have been completed over the past decade to clean data, streamline processes, consolidate accountabilities, and optimise for the machine will need to be completed at a furious pace to remain in business.

The tired old excuses about budget or other priorities taking precedence won’t work. In this API-everything era, there are no other projects worth doing but pushing everything you can to the machine while you still can.

What Disappears

In this new world, concepts like brand loyalty and relationships mean little. Your “trusted partnership” is only of value if the decision algorithm validates that it meets the threshold of what factor is being optimised. Decades of business relationships are recognised as friction that impedes machine optimisation of value chains.

Human coordination roles exist because coordination is expensive and complicated. Middle managers “align” teams and do all the little tasks that fall through the cracks. Sales builds and maintains relationships between human decision-makers. When AI coordinates at minimal cost, these roles diminish rapidly.

Market dynamics speed up. If your API is milliseconds slower or has downtime, AI agents switch. You may be on a supplier panel of perfectly interchangeable API endpoints. Customer stickiness plummets unless there are compelling network effects, platform effects or secret data that compels the AI to choose you.

The strength of interpersonal bonds and communication fades, and market share fluctuates rapidly. The most efficient and AI-friendly APIs start to dominate and become more powerful. Few have realised this will happen in the B2C and B2B spaces.

What Remains

Some firms gain value through network effects. Perfect competition is tough. Each API integration an AI agent makes expands knowledge and intelligence, creating compound advantages for first-movers. Physical services requiring the human touch still exist, but the entire human interaction is wrapped with API workflows and space-age devices.

You can look at the management of Amazon warehouses for inspiration or dread at what this could look like. Call centres will be remembered as pleasurable office work compared to being a human API endpoint reporting to an AI agent. Much human work becomes a node in an agentic workflow - micro-managed to a level that a micro-manager today could only dream of.

What stays deeply human will go beyond tasks to qualities and belief systems. Empathy in a crisis, ethical judgment in new situations, physical presence in highly consequential decision loops, or navigating profound levels of ambiguity.

High-stakes advisory and creative origination stay human for some time, but how long? The demand for these human traits declines at the same rate as AI’s capability rises, and its integration with robots and industrial automation blends into one self-learning reinforcement loop.

New roles might emerge in AI governance. We already have many courts that specialise in domains and arbitration. A Court of Artificial Intelligence might branch off the existing judicial system with its API capability: a judgment becomes code, and all agents subject to its jurisdiction upgrade their code to comply within milliseconds of a published court judgment. Every law and regulation is in a Git repository.

As APIs and agentic AI workflows do almost everything, the data flowing through these pipes becomes a source of market power. This requires human oversight of ethical use and control, replacing previous regulatory agencies with highly skilled AI engineers and ethicists who can stage suitable interventions.

What Slows Change

Regulatory frameworks built for humans create friction: data geolocation rules and privacy legislation force local storage and slower adoption of the most capable AI agents. When deep fakes are trivial, identity verification becomes more important than national defence.

Governments must modernise or lose competitiveness, but the mathematical reality of machine optimisation will cast aside their bureaucracies' desire to control the rate of change.

Physical resource constraints will impose limits. Containers need cranes, products need packaging, and human-APIs will be required to complete the tasks the AI-API firm cannot do with software. Every physical process will be automated and optimised beyond all recognition. “Dark factories” are a concept worth exploring if you’re interested in this logical endpoint.

Open-source software and open-weight AI models prevent complete monopolies from emerging. But new concentrations of power emerge. Dominant API endpoints coupled with AI agents can extract rents, collude, create new forms of lock-in to their products and services, and make current antitrust concerns look cute in contrast.

If core technology at the frontier can remain reasonably constrained and accessible, the overall societal impact can be muted. But we don’t know how aggressive self-learning reinforcement loops could become. There are valid concerns AI safety thinkers raise, yet there is so much we don’t know.

How AI Systems Coordinate

The change from human coordination to AI coordination changes how economic activity is organised. We move from trust, relationships, and biases to protocols, metrics, and APIs.

This creates new equilibria - switching costs are trivial, so stability erodes. Agents allocate tasks to the most efficient API endpoint. If you’re human, write your resume in YAML and expose an API endpoint on your server.

Markets become free-flowing optimisation engines. The more I think about it, high-frequency trading works as a mental model: highly paid people guide machine learning technology to extract profit from high-volume activity in capital markets. AI agents will apply that level of IQ to every value chain in the economy, one by one, until they are all optimised.

Integration on a task-by-task basis becomes the new battleground of competition. API documentation quality, error handling, performance metrics, and output quality will determine success. We could see “weird” outcomes, like the first mover in an industry to clearly articulate everything as an API endpoint and centre the firm on supporting those capabilities leads to rapid market share gain.

Developer experience was significant in the first wave of the focus on APIs. AI takes that and changes the calculus: you must now optimise for the experience of how AI agents will interact with your products and services, or you won’t be chosen. Will there be API documentation written in Neuralese in a few years?

Survival Strategies For Firms

1. I suggest that your entire business be based on an API-first strategy. Any other project that doesn’t support this goal wastes time, and the opportunity cost is too high.

2. Document everything for machine friendliness and use AI to optimise, streamline, and disrupt internally so that delivering every internal or external task is as automated as feasible.

3. Could you change your pricing model so you can profitably charge per API call? Enterprise service-type support and consulting wrappers that last a few years will provide additional margin to finance this transition.

4. Monitor your revenue sources. As AI agent revenue from API calls rises, old revenue sources collapse to zero overnight.

5. Focus on the firm boundaries in this new world - what cannot become a machine API or easily micro-managed human API. This is a challenging problem!

6. Invest in automated security and governance - new identity, entity, and machine verification methods will become central. We don’t know how aggressive AI agents will be at permanent blocklisting, i.e. one cybersecurity incident, and you might have to shut down within minutes.

7. Prepare your workforce transition plan. Many people have their identity tied to their job, and given the pace and scale of change ahead, this mindset will need a lot of therapy in the next five years. Be generous in your exit packages.

Implications for IO Theory

The AI-API economy and new intelligent firms challenge the fundamentals of industrial organisation theory. As functionality becomes modular and even granular because it is broken down into APIs, the logic justifying firm boundaries and insourcing-outsourcing decisions changes.

Things like data and trade secrets could persist as knowledge forms that enable continued internalisation of capabilities. Everything aligns around API-centricity, making life easy for AI agents to optimise. Forces like supplier power, buyer power, barriers to entry, substitutes and rivalry become fluid as the calculus shifts from an annual supplier review to real-time performance monitoring and reallocation of tasks.

Tools already connect to all the AI frontier models and push your request to the “best available” API endpoint for that query type. Extend this thinking to every single value chain or product and service. That is where this thinking ends - all interactions become micro-managed and micro-optimised by the intelligent firms' AI agents.

Entry barriers will collapse - you can see this with the rise of vibe coding. While skilled engineers are still required, non-technical folks will make money from this new capability that can be used with plain language. What happens when self-learning agents decide what to build and how?

Your Choice

The theory and the changes we are seeing point to a challenging endpoint. The transaction costs justifying the existence of firms are dropping fast. The cost of frontier models per million usage tokens drops every few months. The limit is available compute, and someone is deciding to kick off an AI-first transformation.

When coordination becomes nearly free through AI and APIs, many large firms that haven’t invested in clean data and great technology will flounder. Internal cultures that are risk-averse and change-phobic will not be able to sustain any transformational change. Hybrid implementations will be like traditional transformation programs: complex to execute without complete and total buy-in from the board down to the office floor.

If you are a board director or C-level decision maker, you must transform your business into API-accessible services while you still have customers and bargaining power. Operating model compression is now mission-critical.

If you are too slow, traditional frictions like client migration times and transitions will become more automated and faster, so the decision to change will become easier and easier every few months as AI capability advances.

The first path is either doubling down on an existing API strategy and making it company-wide or starting a top-to-bottom transformation driven by partnering with AI to operate quickly. The second path is to keep the blinders on and continue down a path to irrelevance.

Different sectors will lag the frontier at different speeds. Fintech will accelerate. Older financial services firms will struggle. Healthcare and construction will lag. Manufacturing could accelerate existing industrial automation bets to stay relevant. Every industry will be impacted, and the societal disruption will be enormous.

Workforce displacement will require a rethink of the entire status structure, which cares about education systems, qualifications, career ladders, and increasingly archaic notions like time in the office, work ethic, and grinding. It’s unclear how the benefits of all this change will be distributed—for now, it’s more likely that this entrenches existing inequalities and locks them in for an eternity.

We can figure this out, though. There is room for retaining some optimism. There will be a vast increase in the availability of goods and services, innovations in medicine and technology, and easy space tourism.

I’d be interested in hearing what you think about this:

Is the traditional theory of the firm over?

Will AI-API-Agent collectives replace them?

Let me know in the comments.

Best regards,

Brennan