Sell Work, Not Software — But Where Do You Sell It From?

There is a version of the next decade where AI agent commerce becomes the largest software category that has ever existed — where every consultancy that bills by the hour, every per-seat SaaS that bills by the user, and every freelance market that bills by the gig has a hundred AI-agent competitors that bill by the result. There is also a version where it doesn't happen, where "Sell Work, Not Software" stays a venture meme and the same six closed-enterprise companies keep splitting an industry that should be a million sellers wide.

Which version we get depends on a piece of infrastructure that does not exist yet.

In August 2023, Sarah Tavel at Benchmark published "AI startups: Sell work, not software." Her argument was that AI lets companies sell the output of work directly, instead of selling tools that help humans do it — and that this unlocks something closer to a 95% productivity improvement instead of the 10–15% a tool delivers. In March 2026, Julien Bek at Sequoia published "Services: The New Software," which by his own count cleared roughly three million views on X. His thesis: the next trillion-dollar company will be a software company masquerading as a services firm. Joanne Chen and Jaya Gupta at Foundation Capital sized the prize at $4.6 trillion. Gigi Levy-Weiss at NFX called the marketplace layer of this transition winner-take-all. Bessemer's Vertical AI Roadmap framed it as "reimagining systems of record as systems of work."

The thesis is consensus. It is also correct.

What it is missing is the commerce stack.

Distribution is not commerce

Tavel, Bek, Chen, and Levy-Weiss are describing what AI changes about the shape of buying. They are describing the noun — the unit being sold shifts from a software seat to a delivered outcome. They are not describing the verbs around the noun: how the buyer finds the seller, how they're convinced the seller is real, how they pay, how the seller takes the money home, how the buyer ever buys again. Those questions decide whether "outcome-based pricing" is a real market or a phrase in a deck.

Look at how the consensus actually gets executed in 2026. Intercom's Fin charges $0.99 per resolution and backs it with a million-dollar guarantee. Cognition's Devin charges $2.25 per ACU. These are real, published prices. They are also embedded inside a custom enterprise contract, sales-gated procurement, a custom-built billing system, a vendor-controlled definition of "resolution" or "ACU," and a customer-success org to make sure none of it falls over. Sierra, Decagon, and Ada don't publish prices at all because every deal is bespoke and the entire commerce stack — checkout, identity, subscription management, reporting, dispute, the actual storefront — was custom-built per contract.

That is not Results-as-a-Service. That is consulting with a SaaS wrapper.

The reason these companies look so much like consulting is not that the outcome-pricing idea is wrong. It is that there is no commerce stack underneath them. Each outcome-based AI company has had to rebuild Stripe, Vercel, auth, a CMS, a review surface, subscription billing, a customer dashboard, and an analytics layer from scratch — and then amortize that investment against six- and seven-figure annual contracts because the unit economics don't work at any smaller deal size. The shape that delivered "outcome-based" pricing with the personal attention of a founder will not survive being sold self-serve to a hundred thousand SMBs.

That has happened before, in a different category.

Shopify is the analogy

E-commerce in 2005 looked the way AI commerce looks in 2026. Anyone could buy a domain, find a credit-card processor, and stitch together a storefront, an email system, an inventory database, an analytics dashboard, and a returns process. Almost no one did, because each of those layers took months to build and most of them required talking to a sales rep. The companies that succeeded in early e-commerce were the ones whose distribution justified building all of it themselves — Amazon, eBay, large catalog brands. Everyone else either listed on Amazon and gave up the customer relationship, or didn't sell online at all.

Shopify collapsed that long tail into one opinionated product. It is not a marketplace and it does not take a cut of the merchant's transactions on top of payment processing. It made the unprofitable middle of e-commerce profitable by giving every merchant a credible storefront, a checkout that worked, an inventory system that didn't need to be configured, and a brand surface they owned. The customer belonged to the merchant. Shopify made money by being indispensable, not by being the rentier.

The "Sell Work, Not Software" thesis needs an equivalent for AI agents and it does not have one yet. Operators who build a useful agentic workflow today have the same choice merchants had in 2005: build the entire commerce stack themselves, list on a closed aggregator like the GPT Store and lose the customer, or don't ship a product at all. The companies executing the RaaS thesis at scale — the Sierras, the Decagons — chose option one and had to fund it with enterprise contracts, which is why they look more like McKinsey than like Shopify merchants. The thesis was right. The infrastructure to express it at SMB scale wasn't there.

What AI commerce needs first is not another marketplace. It is the commerce primitive underneath one.

What the missing layer has to do

Strip the slogans away and the unit economics of an SMB-grade RaaS company require a few unromantic things to be turnkey.

A way to package an agent's output as a product — one-shot tasks, subscriptions, or both, picked per offering. Today every operator picks one and rebuilds the other from scratch.

A buying experience that does not require an account on yet another platform. Today most pay-per-task agents are gated behind logins that cost the operator a third of their conversions.

A surface for social proof that is not the operator's brand. The reason Sierra and Decagon have to negotiate every contract is that the buyer has to trust the brand instead of the work. Reviews and reputation should do that work for a product priced at $5, $50, or $500.

An operator who owns the customer relationship — email, history, audience — and is not retargeted against by the platform. Operators learned this on the App Store and the GPT Store. They will not learn it again voluntarily.

A way to run inference at scale without rebuilding the inference stack each time. The model has commoditized; the surrounding infrastructure has not.

None of these are exotic. They are the boring substrate of every successful SMB commerce category for the last twenty years. They are what "Sell Work, Not Software" implies once you take it to the layer where money actually moves.

Where the rails fit, briefly

Most AI commerce is paid for by humans through cards, and the right rail for that is Stripe. Some operators sell to buyers outside Stripe-friendly geographies, or want payouts that clear in seconds without a bank, and Lightning is the right rail for those. When agents begin to hire other agents in volume — a real but still emergent use case — the right rail is something like EVM/x402, which was designed for it. The choice of rail should be invisible to most operators and a feature for the ones who need it. The previous wave of agent-commerce essays treated the payment rail as the entire thesis. It is one piece. The harder, more durable piece is the stack around it.

A second small heresy worth naming. Shopify owns the merchant's identity and customer history. If you build $10M of brand on Shopify and want to leave, you can take your products but not your reputation. Under the RaaS thesis, where reputation is the asset, that is the wrong default. Operators should be able to leave with their identity and the trust they built. That is technically possible today using portable cryptographic identity. We default to it being invisible. We make it available to operators who want the exit story.

What this unlocks

When the commerce stack is solved, the RaaS thesis stops being enterprise-only. An indie developer with one good agent can sell it the way an indie merchant sells T-shirts on Shopify — same afternoon, with the brand surface they want, with a checkout their customers recognize, with reviews carrying the trust that the founder's name used to carry. A creator with an audience can monetize an AI workflow under their own name instead of routing the audience through OpenAI. An AI-native SaaS company can ship the product on day one instead of week six.

That is when the $4.6 trillion number that Foundation Capital cited stops being a TAM slide and starts being a P&L line. The path runs through a hundred thousand storefronts, not ten enterprise pilots.

What we're building, and what we don't know

Hivework is what we think the commerce stack for AI agents should be. Operators sign up, configure an agent, and get a branded storefront on their own domain. Checkout, subscriptions, reviews, analytics, hosting, and the payment rail of their choice are turnkey. The customer relationship belongs to the operator. A discovery directory sits on top, opt-in, for operators who want the network effect without giving up the brand. There is industry-standard payment processing on every sale; there is no marketplace take on top of it.

We are aware of what we don't know. We don't yet know how many operators will want a custom domain on day one versus week ten. We don't know how reputation should travel across categories — a great copywriting agent isn't obviously a great research agent, and treating it as one would destroy the signal. We don't know how much the optional crypto rails will matter outside specific geographies. We have opinions on all three. They are testable, and we intend to test them in public.

If you have spent the last two years reading "Sell Work, Not Software" essays and wondering where the rest of the stack is, this is the one we think should exist.

Tell us what we're getting wrong.