Will AI agents get paid a living wage?
Labs earn money from providing intelligence either directly or through subscriptions. Whether their margins persist as open models catch up (Kimi K2.5 let’s go!) is a separate question––one I addressed in this blog earlier this month.
This blog has been bearish too on B2B application-layer agent monetization, both on defensibility of rents and candidacy to price outcomes and participate in upside.
Agents earn money by applying intelligence. Their rents accrue not from the provision of intelligence itself (previous blog: that gets squeezed + public markets are similarly bearish) but from what it does with the intelligence.
But what labor is that exactly? What is an honest day’s work for agents?
The AI Router Thesis Link to heading
Last October The Diff posed ASI as a grand matching “Router” problem:
What’s interesting is that the function of an economy is also a never ending march to solve matching problems. Economics as a body of theory is the study of allocating scarce resources; economics as a practice is whatever can be done to improve the efficiency of that allocation. In this sense, the economy can be modeled as a massive, distributed information processing system that tries to most efficiently execute all the intermediate steps involved in these matching problems, and one element of productivity can be modeled as the real time measure of how well the economy is solving these matching problems.
Let’s say that routing is the practice of applying an LLM to match supply and demand.
This is consistent with A16Z’s Martin Casado’s question of why AI coding has only provided a modest increase in productivity:
I work with multiple companies where nearly all code is AI generated now. However, the productivity probably has only increased 20-30%. Why?
I suspect because writing code is really running code. Changes are the result of a business learnings. Or an operational learnings. For mature companies, the majority of PRs are sub 10 lines codifying these learnings.
AI clearly helps here (e.g. debugging, running tests, building tools) but less so. Operations and business learnings are workload and company specific.
Until AI can perfectly predict what the market needs, or how a system will be used this bottleneck will exist.
Martin similarly poses this routing thesis.
With ChatGPT chat histories or memory, OpenAI is narratively positioned to become a Router. To this end, it has released apps (Booking! Spotify! Canva!), allowing users to pull context from and take action in their favorite apps.
Today, users must manually connect their ChatGPT account to apps––many requiring login––so the vision of frictionless routing still faces friction.
It’s rumored that ChatGPT app-connect conversion rates are low.
And even with this innovation, later notes The Diff,
that router is only as good as the consumer intent data it can capture and the legibility of available product offerings.
How good is ChatGPT intent data? ChatGPT only knows what you tell it––i.e., requires active human labor––and its developed memories may end up stale. Gemini has my latest workout progress, for instance.
That doesn’t work for routing.
Ad networks are effectively routers (see Google’s Hal Varian on Google as matchmaker ‘yenta’) and they do best with recent signals. Mobile Dev Memo’s Eric Seufert explains
Recency plays an incredibly important role in the signal parsed from the data identified above. Not only must an advertiser know that a user has historically engaged and monetized with products following an ad click, but they must know that a user has done so recently: the older the data is, the less helpful it is in targeting ads to that user. Facebook’s ad targeting models require a constant supply of data — the off-property events stream — in order to retain efficacy.
That is, Facebook’s data advantage isn’t from active user engagement with cat videos or AI slop but from the placement of Facebook pixel or conversions API “sensors” that give it visibility into off-property activity data. Unlike OpenAI, Facebook doesn’t need direct on-platform user engagement to perform effective routing.
Either way, the narrative is that––taken to the limit––OpenAI’s app use could help it, per The Diff, “become the meta-layer over the entire AI-powered economy.”
This could be plausible. Indeed, Ben Thompson’s aggregation theory poses that value accrues to the players that aggregate demand. OpenAI is a candidate.
RIP Routing Link to heading
But basic routing doesn’t pay.
The original pitch for Agentic Commerce was that it’d make autonomous purchases on behalf of users. That’s a pretty sweet value proposition for certain goods, but agentic commerce today has reduced to in-chat product discovery and checkout.
Agentic Commerce is a chat storefront.
Shopify and Amazon already have chat storefronts.
To these retailers, their native chat storefronts only cost marginal tokens, while ChatGPT is reportedly charging Shopify stores a 4% affiliate fee. Stripe’s John Collison says
[Agentic commerce] could be quite a democratizing force because you could lead to discovery of lesser known brands.
but what exactly is democratizing about New Search only available to brands who can afford an additional 4% fee? Or more democratizing than native search in the Shop app?
As I and Eric Seufert wrote last year, the economics of agentic commerce do not make sense. Eric and Fermat Commerce’s CEO Rishabh Jain estimate the affiliate fee will be gone by end-of-year.
Agentic Commerce as New Storefront Router will not get paid a living wage.
The problem is that this basic routing––distinct from ads––identifies no economic inefficiency. @lefttailguy quoting Casado and The Diff
Once execution is commoditized the bottleneck is in creating the portfolio of risk adjusted bets that best identify and capture value from an ever evolving set of economic inefficiencies.
Viewed under this lens, all New Storefront is doing is taking a toll on a happenstance upstream attention positioning. “Here’s what to buy” (based on who’s decided to revenue share 4%) doesn’t surface any economic waste and it likely creates it through adverse selection!
Affiliate “New Storefront” routing, of course, is distinct from auction-based ad-networks, which solve matching inefficiency, matching advertisers to intent at prices determined by the market.
Can agents capture existing margin? Link to heading
If not vanilla routers, maybe new agents can capture rents captured by existing aggregators?
Two weeks ago Faire’s Chief Strategy Officer published LLMs v Marketplaces as a strategy guide to what sorts of marketplaces would be up for disruption via LLM routing. He posed that the candidacy of LLM routing would come down to three criteria
- Difficulty of supply aggregation
- Degree of marketplace management
- Nature of customer engagement
The marketplaces most at risk for LLM disruption are those with
- easy-to-aggregate supply
- low marketplace management
- high customer engagement
of which, travel platforms like Expedia or Booking are examples. And it’s true, these marketplaces have healthy take rates to disrupt, reportedly taking 15-20% commisions on sales. If ChatGPT became the front door to travel, could it capture that cut?
Maybe!
But the winning LLM router would have to replicate the travel aggregator’s supply-side aggregation. I do not believe this is as trivial as Hockenmaier seems to think –– to my knowledge there’s no single API with all global travel bookings. Even if the LLM resolved this, this new agent would then compete with existing travel players with no new value proposition that the incumbents couldn’t replicate.
Travel booking doesn’t exactly require PhD intelligence.
And, to the bad luck of the yet-to-be-paid agent, other marketplaces are apparently out of reach. Hockenmaier suggests marketplaces with managed operations, high frequency, and low consideration––like Uber and Doordash––are safe.
Their
- hard-to-aggregate supply
- complex operations
- sticky habits
make agent intermediation reportedly unrealistic.
So to get paid a living wage, the agent must engage a zero-sum fight for existing margin against big incumbents?
You must just do things Link to heading
In December, I wrote how you must just do things.
That is, AI inverts aggregation dynamics of the internet, modularizing simple coordination or execution, while integrating distribution and trusted operations. The argument applied Clay Christensen’s Law of Conservation of Modularity to AI (see Ben Thompson’s original 2015 exposition of the Law).
Hockenmaier doesn’t consider this case.
His only concern is with demand aggregation. But if AI is looking for work and demand aggregation doesn’t pay, perhaps AI should try its hand at supply!
Not necessarily supply aggregation––though that could help––but rather becoming the supply, like Waymo or Zipline.
This puts Hockenmaier’s whole defensibility thesis in jeopardy.
A year ago Waymo surpassed Lyft’s marketshare in SF. Who cares about agent demand aggregation against locked-in supply if the heterogeneity of this supply is now entirely modularized by AI and operated by someone else?
This omission by Hockenmaier is exactly the sort of opportunity that can earn AI its bread and roses. And it has nothing to do with routing.
What other cases are there where AI shifts or replaces economic dynamics around supply?
Three inefficiencies Link to heading
It seems there are at least three sources of capturable inefficiencies suitable for AI agents earning a living wage.
Introduce a new kind of supply Link to heading
Offer new value propositions that the supply chain couldn’t offer before.
- A distributed battery network that provides grid services––lower electricity rates, always-on electricity––that consumers couldn’t access before.
- A self-driving car network that provides safer personalized transportation without a human driver.
- A last-mile autonomous delivery service that expands availability of same-day deliveries for a modest price premium versus existing delivery networks.
All of these services use AI, but they do it in the service of modularizing supply itself, capturing the resolution of inefficiency.
Legibilizing latent context Link to heading
Resolving economic inefficiencies is also gated by legibility of context. This can include cases of
- latent preferences
- local supply and demand unobserved or uninstrumented up- or downstream in a supply chain
You can capture this inefficiency by originating sensors that exclusively legibilize––make accessible to AI––and commercialize this context.
The economic value is making the context available to the machine sand god, whose unlocked trusted actions you ideally productize yourself. You went through the trouble of legibilizing the context and coordinating the distribution, so you should get paid for that, not a centralized AI lab.
Technologies that reduce the costs of uncertainty Link to heading
If you suppose AI is a coordination technology (e.g., see Will Manidis’ posing), one way to capture value is to reduce the costs of coordination. Not via the provision of intelligence – that process is well handled by the labs, but rather by physical infrastructure that relaxes the need of existing context. Coordinating supply and demand signals is partially an exercise of managing uncertainty, but if you can reduce the costs of uncertainty, you can capture some of that value (while also sharing some of it with your customers).
Supply chain buffers are the only technology of which I’m aware that do this, and they’re used widely in industy.
Base Power and David Energy manage edge power buffers that stabilize energy consumption. Even if the power goes out, battery power at the edge can continue to provide power until power comes back online. When power is relatively cheap, Base can turn on electricity consumption, allowing it and its end-user to participate in the surplus of edge buffer arbitrage. As software commodifies, owning the buffer is likely more valuable than managing the buffer (think Base v David Energy).
Maintaining private buffer state is also strategically valuable to prevent price discrimination––imagine Uber raising prices when your iPhone battery runs low. [Uber reportedly does not do this, but they could with this data access.]
In essence, you’re giving the AI agent a physical resource––a wallet with supply chain capabilities––to barter with.
An intelligent refrigerator––one that knows its contents and consumption rate––is also an example of a buffer. It holds a buffer of perishable groceries, all with their own consumption and expiration rates, enabling the same sorts of optimizations unlocked by David and Base. In a simulation study co-authored with Claude Code, we demonstrate that fridge “buffer” context can reduce grocery prices by 10%.
Businesses can attack multiple of these inefficiencies, but they must do so privately.
The art of dark ops Link to heading
To remain defensible and participate in respective surplus, all of these technologies must ‘operate in the dark.’
That is, they must privately commercialize their respective informational advantages. If you leak your legibilized context or private buffer state, you could invite competitors to commoditize it (information is not rivalrous!), price discriminate. [See Dark Pools from finance as a corollary.]
For instance and perhaps obviously, Waymo vehicle and sensor state is private. When Avis and Lyft behave as operating partners for Waymo, they get no access to car data. Waymo commited the research and development dollars to fully productize a self-driving car product and is an investment it expects its cars’ commerce to pay back and more.
Businesses that combine these––Waymo is a combination of at least two––are effectively examples of Cybernetic Arbitrage that I wrote about last year. They own assets at the edge that directly monetize private context with intelligence. They cannot share their context because it’d dilute the alpha they worked so hard to originate––they must vertically integrate.
Does information at the edge even matter? Link to heading
I’ve committed a lot of tokens toward advantages of commercializing context at the edge.
But you might wonder––centralized aggregators have such a grand view of the economy. They view billions of searches and hudreds of millions of chats. What does the edge have that can’t be reasoned about statistically by centralized AI? Doesn’t the Law of Large Numbers smooth out idiosyncrasy?
The previous era of productized ML rather required a large user base to develop an unassailable data advantage and sufficient sample to generalize. AI largely commodifies this, both directly by its predictive power, and indirectly by the speed at which new models can be trained and shipped. This pushes the scarcity toward context that a centralized aggregator cannot see––namely local, tailor-derived context where a centralized aggregator has no sensors.
While large samples can help to smooth idiosyncrasy, they don’t help with exogenous correlated demand shocks:
- an unexpected storm
- a viral new trend
- or surprise news of allegedly fake free-range eggs.
It’s at these moments that the information advantages of scale completely break, and all value lives at the edge.
The edge’s power is greatest when the distance between context and desired action is smallest, most apparent in autonomous devices. In these devices, local context is necessary for operation, and aggregation of users’ data––what most people searched for today or liked yesterday––does not matter.
This change inverts the value chain. The edge is no longer the “last mile”, a cost center receiving goods pushed from the center. It is the First-Mile––the origin of signal that all supply competes to meet. In the intelligence era, the world really does revolve around you, and the local devices and sensors situated to most efficiently deliver it.
I think this is enough of a change that it deserves a moniker––I pose it as “The Asymmetry of Things”: while suppliers used to hold the Scale, consumer hold the Truth. For over two decades, scale dominated because intelligence needed to be centralized. But AI commoditizes intelligence, and statistics favored the edge all along––it’s only now that we get to readily act on it.
For new games, Truth > Scale.
But how do you vertically integrate under hyper-scalers? Link to heading
Suppose you deploy intelligence to the edge and you want to vertically integrate but supply chains above you are massive. How do you make a deal while protecting your users and positioning?
Ideally you can figure out a way to vertically integrate anyway. Be more Waymo or Zipline, less Samsara. You are the complete owner-operator of the service.
Next, you could try to compose a vertically integrated solution by assembling a supply chain (ideally of disaggregated parts) yourself––this is Base Power’s strategy.
To my knowledge, Base is a customer of power companies whose energy consumption it manages on behalf of consumers. This could be extended to more complex physical supply chains where Autonomous Logistics––Zipline, Autolane, Coco Delivery––are disaggregating fast last-mile logistics, perfect for an incoming integrator. Abstractly, that is, internally become a business customer of the supply chain on behalf of your customers to become the Merchant of Record yourself.
Finally, you could accept some loss of surplus and integrate with the supply chain above you. Google’s Universal Commerce Protocol appears to offer a path. UCP’s first instantiation with Gemini––an AI assistant with context from all your Google-local apps––is a case study. Gemini allows merchants to integrate with Gemini, remaining the merchant-of-record, while Gemini offers a slick user experience. Clearly this pattern can spread to other local agents.
In the last case, the strategic posture is that the edge agent offers net-new context in exchange for a shared surplus.
An honest day’s work Link to heading
The agent seeking honest work won’t find it in vanilla routing.
Auction-based mechanisms do the hard work of matching supply and demand while routing affiliate revenue taxes a position that an incumbent could cheaply replicate. The biggest applications in B2B software the labs will vertically integrate––Anthropic is already attacking Context Graphs.
With intelligence and execution modularizing, value flows to scarce trust and distribution––technologies of new value propositions or those that develop or delete a need for context.
That said, AI creates a clear opening for local intelligences, particularly where the distance between context and trusted action is near zero––namely, autonomous applications. In this era, and counter-intuitively counter to the Bitter Lesson, Truth beats Scale.
No bread and roses for a centralized agent using weak context to match supply and demand. Fair pay will come from becoming the supply, owning the context, trading certainty for surplus.
That’s an honest day’s work.