When the agent delivers the function

In September 2026 Apple released a Siri built on Google's Gemini models, and Amazon began blocking Meta's shopping agent, Muse, twelve days after its launch.¹² Apple rents its intelligence from its principal rival and, so far, keeps its position. Amazon holds the data and the checkout an agent must pass through, and closed them. I read both events through functional integration, Barry Wacksman's idea, which I took up in 2015 to ask why good innovations get lost between start-ups and the corporations that need them.

In 2015 I began a doctorate at Cass Business School with that question.³ A start-up builds something a large company plainly needs, fails to reach scale alone, and never makes it inside. A corporation has a good idea of its own and lets it die because no business unit owns it.

The two classic directions of corporate growth explain why. A firm grows horizontally by taking more of a market and vertically by owning more of its supply chain, and an innovation that is neither a bigger share nor a cheaper input shows up in neither. Diversification can take it in only as a new business. Functional integration gave me another way to see it, as a function, something a user the firm already serves wants done. Seen that way, the innovation is worth taking in, because it raises what an existing user gets. The firm can add it, connect it to what it already offers, or open its platform so that someone else can.

I set that reading down in 2017.⁴ Until recently a firm added a function by building it, by backing or buying the firm that built it, or by hosting it on a platform. An agent can now call it through a published contract, on the user's behalf, without the firm owning it. I think that either closes the gap I started from or moves it somewhere else, and by 2028/29 the evidence should show which.

The idea

The idea is Barry Wacksman's. As chief growth officer of R/GA he presented it at Cannes Lions in 2011, and with Chris Stutzman he made it a book, Connected by Design, in 2014.⁵ Horizontal integration had driven the top line and vertical integration the bottom line, he wrote, and "these two dimensions" had "directed most if not all of corporate competitive strategy" until Amazon, Apple and Google began to grow differently.⁶

He called what they did functional integration. Ecosystems, he wrote, "grow in terms of function (adding value to the ecosystem) and integration (adding connections within the ecosystem)".⁷ He drew the two as the axes of a chart, the level of functional value against the level of integration, placed the functionally integrated ecosystem where both are high, and used the chart to decide what to add, connect and open next.⁸

As far as I can find, the term fell out of use after about 2015.⁹ The subject did not. A large literature now studies ecosystems and platforms under other names, mostly from the firm's side; a few papers start from the consumer, and I borrow from them openly.¹⁰

The user's seat

I use "ecosystem" narrowly, for one firm's connected offering as a single user meets it, where most of the academic literature uses the word for a set of partner firms aligned around a proposition.¹¹ Corporate strategy builds the corporation. What the corporation offers, when it is an ecosystem, needs a strategy of its own, and functional integration is its logic.

Draw one user in the middle of a page and put around them the functions they expect: listen to music, pay, find the way home, keep their photographs. A function is something the user wants done. It is close to what Clayton Christensen called the job to be done, read for function only; jobs also have "powerful social and emotional dimensions", as he insisted, and nothing here captures them.¹² A person can enter the ecosystem through any product, the earphones or the phone or the watch. The framework is about what happens after the first piece.

It has two variables, one for each side. The user's is User Functional Value, what the pieces a user holds do for that user. Its founding proposition is that it rises with the number of pieces held: the earphones, the streaming service and the phone are each worth more to their owner because of the others.¹³ Functional value comes at two levels. Available value is what the pieces could do; realised value is what the user actually gets. Where the gap has been measured it is wide. Users of business software rarely or never use 80 per cent of its features, and at least a fifth of new-car owners had never used 16 of 33 technology features.¹⁴

The firm's variable is the Business Integration Ratio, the share of a new function that runs on what the firm already owns: the account, the operating system, the chips, the data, the distribution. Ron Adner has a name for this on the firm's side: carryover, the reuse of what one ecosystem built to launch the next.¹⁵ Neither variable claims new ground in theory. Richard Priem's consumer view of value and David Teece's complementary assets are their parents.¹⁰ Both are read for one user and one function at a time, so a firm can put them on a single page.

That page is the map that comes before Wacksman's road map. It shows which of the user's functions the firm serves, how well, from which of its assets, and how much of what it offers the user actually uses. Only then should the firm plan the next function, and two questions decide it. Does the user expect this function from us? Can we serve it from what we already own? Where the first answer is yes and the second is no, the firm can back the start-up that has it, buy it, or host it on a platform.

The case where the user expects a function the firm cannot serve is the one I started from. On this map, a start-up's innovation that would raise what an existing user gets has a known place. The App Store was Apple's answer. Another company's function reaches the user inside the ecosystem, and Apple's own filing says that "decisions by customers to purchase its hardware products depend in part on the availability of third-party software applications and services".¹⁶ A firm without a platform had two answers, to back the start-up or to buy it. The agent may give it a third.

One function, two ecosystems

Take one function, listening to music, and put it in each kind of ecosystem.

In Apple's closed ecosystem, Apple's own service is built in, and its rival has to get past what Apple controls. Apple Music comes with the phone, Siri and the HomePod play it natively, and Apple sells it inside the Apple One bundle. Spotify has to be installed. Siri can learn it as a preference, but no setting makes it the default; the HomePod plays it only when it is relayed from the phone; and until a court order in 2025 Spotify could not link to its own subscription from its iPhone app in the United States.¹⁷ In Google's open ecosystem Spotify meets the owner's service on more equal terms. A Nest speaker lets the user choose the default music service, and Google's assistant reaches Spotify as it reaches any connected app.¹⁸

The results split by ecosystem. Inside Apple's, Apple's own service leads: among owners of Apple devices in the United States, 40 per cent use Apple Music and 29 per cent Spotify.¹⁹ Across all ecosystems, the service open to every one of them leads: Spotify has 300 million paying subscribers worldwide, against an estimated 116 million for Apple Music.²⁰

Both results are what the framework predicts: inside the ecosystem that owns it, the integrated function leads; across ecosystems, the open one does. What the figures cannot show is which of two things gives Apple Music its lead on the iPhone: that it works better with the user's other Apple devices, or that Apple controls what Spotify must get past. On the iPhone both push the same way. The prediction holds either way; the claim that integration creates value for the user is not yet shown, and showing it means measuring the value each user gets, which belongs to the open work.

Spotify is also the start-up that was not lost. In the terms of figure 1 it had high consumer value, traction and the resources to reach scale alone, and built a business of one function across every ecosystem. That is the successful end, and it is rare. Most start-ups with a valuable function never get there. For them the better outcome is to become a function inside someone else's ecosystem, and that is the route this essay asks whether agents can open.

The same closed ecosystem shows where integration pays across functions. Apple's Services line grew from $32.7 billion in fiscal 2017 to $109.2 billion in fiscal 2025, about 3.3 times, while total sales grew 1.8 times. It earns a gross margin of 75 per cent, against 37 per cent for the products, on more than 2.5 billion active devices.²¹ Not all of it is integration. Part is rent for a position: Google paid Apple about $20 billion for 2022 to be the default search engine on Apple's devices, by the court's estimate in the United States' case against Google.²²

My reading is that integration paid only where the firm controlled what the user passed through for that particular function, the device, the operating system or the checkout, and that control in one place did not carry over to another. Three firms that tried to add functions on the phone without controlling the phone fit it. Nike built a running service on Apple's devices, then a wristband of its own, the FuelBand, and in 2014 it was reported to be dismissing most of the team behind it.²³ Nike had the brand, the runners and even a device of its own; Apple had the phone the runners carried. Amazon, which owns a checkout, and Microsoft, which owns the PC's operating system, each tried to build a phone of their own, and both withdrew.²⁴ Neither lost its services elsewhere, because those sit on a checkout and an operating system each controls; what neither could do was carry that control onto the phone. I have not tested this reading. A firm that adds functions across a services business, on devices and checkouts it does not control, and keeps a high margin would show me wrong.

Two agents, four gates

In 2017 the unit of a function was an app the user opened. In 2024 Apple told developers that "Siri will take actions inside apps on your behalf", and in 2026, "Think of domains as categories of contracts between your app and Siri."²⁵ Anthropic's Model Context Protocol, an open standard for connecting assistants to outside services, counted more than 10,000 public servers by December 2025, and OpenAI and Google had adopted it.²⁶ The firm that offers a function still builds it, but software now calls it, through a contract. For the lost innovation this means a firm can integrate a start-up's function by calling it, without buying it.

Economists have begun to sort agents by whom they work for. Shahidi, Horton and colleagues separate the agent a person brings along, which "carries the user's instructions and data across sites", from the agent a platform provides, which has "deep integration, privileged signals" and a risk of lock-in.²⁷ They study what the platform's agent does to a single purchase. I look at what it adds to everything the user already owns. In this essay the two are the individual's agent and the ecosystem's agent.

The ecosystem's agent is the one in the phone, shipped by its maker as a new function, and its first effect is on realised value. In 2024 Apple said Siri "can now give users device support everywhere they go, and answer thousands of questions about how to do something"; the 2026 release adds "personal context understanding" and "even more systemwide app actions".²⁸²⁹ An agent that knows every setting and makes the change can raise realised value without the firm shipping a new product, for everyone who uses it and not only for the expert. It also touches every other function, and the makers sell it that way: "ask a question on your iPhone and pick up where you left off on your iPad".³⁰ I expect it to raise both variables at once. Users get more of what their pieces can do, and because every new function can be delivered through it, more of each new function rests on what the firm already owns.

The individual's agent works for the person across every ecosystem they touch; today it is usually a general assistant made by a model company and used across services. It closes the same gap between what the user's pieces could do and what they get, everywhere at once, and it treats each ecosystem as one supplier among several. The user states the intention and the agent chooses. Early experiments find that agents choosing among products concentrate demand "on a few 'modal' products".³¹ Ben Thompson thinks agents are "the ultimate Aggregators" and that "most people and companies will only have one agent, not multiple".³² I expect most people to have two, their own and the one in their phone, and the phone's to hold where its owner has many of its maker's pieces. The first proposition below will show which of us is right.

What decides between the two agents is what any agent must pass through. I count four gates. A gate is where a firm collects on the value and the reuse it has built, and two of the four sit on each side of the framework.

The default and the user's history sit on the user's side, even when a firm holds them. The default is which agent speaks first on the device in the user's hand; the history is what that agent has learned about the person across their pieces. Both shape what the user gets, not what the firm owns. Functional value rises with the pieces a user holds, and with an agent it also rises with what the agent knows across them. The makers know what the history is worth. When Google launched tools in March 2026 to import a person's memory from rival assistants, it said that "starting over with an assistant that doesn't know you can feel daunting".³³ The default is already contested. Apple's iOS 27 contains a framework, not yet open to outside developers, that would let a rival's model answer through Siri.³⁴ On Android, the European Commission ordered in July 2026 that users be able to summon a rival assistant by voice, as they summon Google's, and let it act in apps for them, from July 2027.³⁵

Data and rails belong to the firm. Data is what the agent must read and no one else holds, such as inventory, prices and records; rails are what it must use, such as the payment, the licence and the checkout. Both are integration in the framework's sense, what the firm owns that a new function has to reuse. Amazon showed in September what holding them means. When it blocked Muse, it said that agents buying for customers "should operate openly and respect service provider decisions about whether or not to participate".² In March 2026 OpenAI stepped back from the checkout it had built inside ChatGPT and let merchants use their own, so even the largest agent host left the rails with the merchants.³⁶

I leave the model off the list. Apple's new Siri runs on models "custom-built in collaboration with Google and its Gemini models".¹ Renting works while every rival can rent the same model and there is enough of it to rent. If either stops being true, owning the model becomes an advantage again.

The lost innovation, again

On the framework's own terms, agents could close part of the gap I started from.

The map shows a firm an outside innovation as a function, with a user and a place, and that gives the firm a reason to want it. A firm takes it in only if its users expect it and its own assets can carry it. The contract makes taking it in cheaper. Once a start-up publishes its function to agents, the ecosystem's agent can deliver it through the account, the device and the distribution the firm already has, and I expect the integration ratio of an outside function to rise towards that of one built inside. And the ecosystem's agent can find the function for a user who would never have looked for it, which is realised value the start-up could not have reached alone. In 2015 innovations failed for lack of traction or resources. I expect agents to supply some of both.

When the ecosystem's agent calls the start-up, the firm has integrated the function into its own offer, which answers the loss between start-up and corporation. When the individual's agent calls it, the innovation reaches the user without any corporation. It is saved, but not for the incumbent.

Against these, the friction moves to the gates. A start-up whose function an agent can call still reaches the user through a default it does not own, a history it cannot see, and data and rails that belong to someone else. Regulation can open a gate, as the Commission has begun to do on Android; a holder can close one, as Amazon did to Muse, and then the innovation is lost again at a different point. Sangeet Paul Choudary adds that the market itself changes only when sellers have agents as well as buyers; until then, a start-up meets the buyer's agent on the buyer's terms.³⁷ My answer to the question I asked in 2015 is therefore conditional: an agent saves a lost innovation only where the gates it must pass through are open to it.

The gates also sort the firms in between, which earn a fee for sending customers to someone else, such as a hotel-price comparison site, a loan marketplace, and two firms that sell shoppers' requests to insurers and lenders. They live by standing where the customer chooses, and they pay for that position: they spend between about 70 and 90 per cent of their revenue on advertising and other paid traffic to bring in the shoppers they pass on.³⁸ An agent that chooses for the customer does not need to pass through them. I read the latest annual reports of eight listed firms of this kind. Four were set aside because they hold a gate or never meet the consumer: Groupon takes the customer's payment and NerdWallet and MONY Group hold licences to sell financial products, all three of them rails, while MediaAlpha trades between businesses. The other four, trivago, LendingTree, EverQuote and QuinStreet, hold almost nothing an agent must pass through.³⁸ Only LendingTree holds a little of one gate, the user's history: people keep an account with it to check their credit score, and those users already bring it about 2 per cent of revenue. That history matters less as revenue than as a reason to come back without paid traffic. If it counts, LendingTree's cost of traffic should rise less than the other three's once agents do the choosing.

What I expect to observe

The framework makes five propositions that can be checked from outside, each with a public source and a stated way to be wrong.

Proposition 1 is that the ecosystem's agent holds on its maker's devices: a person with more of one firm's pieces gets more from that firm's agent, and the firm can show it. It fails if a person's own agent becomes the default on a large maker's devices and the device set stops mattering.

Propositions 2 and 3 are that the gates get priced. The default that decides which agent speaks first will be bought and sold, and holders of data will charge agents for access or block them. The first fails if no one pays for placement, the second if agents can read and act on any service freely.

Proposition 4 is that the user's history stays where it was made, with assistants letting a person bring their history in more readily than take it out. It fails if a common format or a regulator makes an assistant's memory portable both ways and people move freely.

Proposition 5 is that firms holding no gate lose the customer to the agent. trivago, EverQuote and QuinStreet will report a lower take or a higher cost of traffic and attribute it to agents acting for buyers, not only to answers on a search page, and LendingTree will hold up better than they do. It fails if they keep their economics while selling through agents.

I will check them in 2028/29 by three rules. Each score quotes its source. "Wrong" is a permitted score. A proposition whose source disappears is scored "cannot be scored", never "right".

If functional integration does not survive agents, 2035 looks like this. A person's own agent does every job equally well whatever they own, is the default on every maker's hardware, carries its owner's history, and picks suppliers on price, rating and availability. Owning more of one firm's products buys the user nothing the agent cannot get elsewhere, and integration survives only inside the firm, as a saving on cost. If that is the world of 2035, the framework described the app era, and I will say so.

The open work

Functional integration began as a practitioner's idea, and this is a practitioner's reading. The academic work it needs is plain. The two variables have to be measured on real data, functional value per user and the integration ratio per function, and the propositions tested on ecosystems that failed as well as on Apple's. An agent that can read filings and usage data at scale makes that work cheaper than it was in 2015, and a firm can do the first part on its own customers tomorrow.

For a company the order is the framework's. Map first: one customer in the middle, the functions they expect, and what they actually use of what you already offer. The gap between what they expect and what they use is the cheapest value you have, and an agent will close it, yours or theirs. Then ask which of the four gates the next function must pass through, and whether you hold it. If you do, integration can still pay, including for a function someone else invented. If you hold none, my position is that you are building an interface, and that an agent will be a better one.

Decide what you hold before you decide what to add.

Revisited: not yet. First checkpoint 2028/29.

Sergio Mottola is a partner at Tyche.Black. He studied how value is built at Bayes Business School, under Charles Baden-Fuller, and designed and chaired the Republic of San Marino Innovation Institute.