When the example is the indictment
Y Combinator’s Spring 2026 Request for Startups gave three examples of what an AI-powered agency might look like. Read them straight:
Think of a design firm that uses AI to produce custom design work for clients upfront, to win the business before the contract is even signed. Or an ad agency that uses AI to create stunning video ads without the time and expense of setting up a physical shoot. Or a law firm that uses AI to write legal docs in minutes, rather than weeks.
Read them again and ask the question they invite.
Why are those not just SaaS businesses?
A design tool that produces custom design upfront is a design tool. An ad-creation tool that produces video without a physical shoot is an ad-creation tool. A legal-document tool that writes contracts in minutes is a legal-document tool. The framing dresses each one in agency clothing — a firm, a roster, a relationship — but the unit of value is the SaaS underneath. The agency layer is decorative.
This essay is about that decoration. About why dressing a SaaS up as an agency creates a business that prints money for a window and then doesn’t compound. About what the RFS got right that nobody is saying out loud, and what it got wrong that everybody is repeating.
The thesis comes in two pieces.
The first is an indictment. The agency that builds itself around a productized AI workflow is not a competitor to SaaS. It is a feeder to it. Every contract such a shop signs is free demand validation and free workflow standardization for the SaaS company that will eventually ship the same thing and own the category. The productized-services shop is the R&D arm of its eventual replacement, paid for by clients who don’t yet know they’re funding it.
The second is what survives. The agencies that compound from here are the ones flipping the unit of sale entirely — selling judgment, opinion, and accountability under ambiguity. The work that doesn’t survive a price card. The position that sits above where AI keeps climbing, because the climb is structurally uneven.
Both are true at the same time. The indictment is the engine of the argument. The alternative is the way out.
The feeder thesis
Picture the design firm from the RFS. Founders have built a tool that produces high-quality custom design output for a prospect upfront. The pitch is the product. The product wins the contract. Then the firm uses the same tool to deliver more design across the engagement, billed by the deliverable rather than the hour, with humans hovering over the AI to keep it on-brand and on-spec.
This works. It might work very well for a window.
Then watch what happens next.
The firm signs forty clients. Each client is a different brand context — automotive, healthcare, consumer fintech, B2B SaaS. Each engagement teaches the firm’s tool how to handle a different vertical. Each set of feedback rounds teaches it what “good” looks like in a different domain. Each contract is, mechanically, a labeled training set for the firm’s workflow — what works, what fails, what clients reject, what they sign off on.
The firm now has something valuable. A tool, plus the operational knowledge to run it, plus contractual relationships that prove it produces commercial outcomes.
The firm also has a problem. The same value is visible to Figma, to Canva, to Adobe, to Framer, to Anthropic, to OpenAI, and to a hundred well-funded startups whose business it is to be visible to that exact opportunity. The firm has, by virtue of doing the work, shown all of them where the demand is, what the workflow looks like, and how much customers will pay for it. The firm has performed the most expensive parts of SaaS product discovery — at its own cost, with its own capital, on its own roster of clients — and published the answer.
The diagram makes the loop visible. Demand flows from clients to the shop. The shop’s existence flows, as legible market data, to every SaaS company looking at the category. The SaaS company ships parity. The customers route to the SaaS company. The shop is left holding workflow expertise that the SaaS company has now generalized and embedded.
This is roughly the history of the no-code consulting market between 2018 and 2023. It is roughly the history of the WordPress build-and-run market between 2010 and 2020. It is the shape of every productized-services arbitrage on a maturing platform. The pattern repeats because the economics are structural.
The naive read of the RFS is that productized services lets a firm capture SaaS-shaped margin by removing humans from delivery. The structural read is that productized services lets a firm transfer the value it creates — demand, workflow, willingness-to-pay — out of its own books and into the books of whichever SaaS company is paying attention. The arbitrage is real, and it is routinely mistaken for a business model.
Why the moat moved upstack
The historical agency moat sat in a particular slice of work: standardized output produced repeatedly for paying clients. Logos, websites, landing pages, video edits, simple legal documents, monthly content calendars, basic media buys. This was the work that filled the calendar and paid the salaries. The moat was the relationship plus the production capacity, and both held for decades.
SaaS ate that slice. Logo design went to Looka and a thousand AI-image variants. Landing pages went to Framer and Webflow templates. Simple copy went to Jasper and a parade of GPT wrappers. Basic media buys went to programmatic platforms with self-serve UIs. Each new SaaS that won a category did the same thing: took the standardized output, automated the production, and priced against the SaaS curve rather than the agency curve.
The moat moved upstack.
What survived in agency form is the work above where SaaS could reach. Brand systems that have to hold across years of campaigns. Monetization strategy for publishers with idiosyncratic audiences. Discovery work where the client doesn’t yet know what they’re buying. Negotiated outcomes that require accountability for ambiguous deliverables. Senior judgment applied to messy organizational situations. These categories share a property: they resist being priced against a deliverable. The value sits in the framing, the choice, the bet on what to do. The artifact at the end is incidental.
SaaS bounced off this work because price-cardable means commoditizable. The moment a clean number can be put next to “brand strategy for a category-creating fintech,” somebody has either already productized it or is about to, and the productized version trades at productized prices.
The productized-services shop operates in the layer SaaS already eats. It automates inside that slice and sells the automation as if it were upstack work. The slice below the moat, dressed in moat clothing, priced as if the moat were still there. The pricing power lasts as long as customers don’t notice the dress.
The ladder AI climbs unevenly
AI capability climbs unevenly. The unevenness is structural, built into who builds the models and which pain those builders feel acutely.
Code has been eaten fastest. Front-end implementation, scaffolded back-end work, infrastructure as code, test generation, code review on familiar patterns — all of it has been pulled forward dramatically in the last eighteen months. Junior engineers shipping at mid-level pace on familiar stacks is common. The productivity step-change in code is the largest in the AI era so far.
Design has moved more slowly. AI produces competent design output and stops short of design judgment — when to break a pattern, why a brand should resist a trend, what a piece of work means in the context of a client’s category position. The handoff between AI design output and a usable production artifact still requires a human in the middle, and the middle is the work.
Strategy and monetization have moved slowest of all. There are reasons, but the most useful one is mechanical. AI models are built by engineers, for engineers, against problems engineers feel acutely. The model labs are full of people who write code every day and feel the pain of code being slow. The pain of running pricing experiments on subscription products, negotiating brand-coherence concessions across multi-year client relationships, or sitting with publishers as programmatic CPMs decline — that pain lives somewhere else, in a population the labs are not training against.
The ladder will move. The order will reshuffle. Strategy will be eaten too, eventually. “Eventually” matters here. A productized-services shop betting on AI’s current capability has eighteen to thirty-six months before the layer it’s operating in is fully absorbed. A POV-led firm betting on judgment and accountability under ambiguity is betting on a layer the model labs are not currently pointed at. The latter is a longer bet on a more defensible position.
The productized-services shop is taking the short bet on the layer where the climb is fastest. The math is brutal.
Distribution economics and the worst of both worlds
The case for SaaS-margin agencies leans implicitly on a misreading of where SaaS margins come from.
SaaS margins come from distribution economics. One product, sold to many buyers, with marginal cost approximately zero on the next sale. The same code that serves the first customer serves the thousandth. Customer acquisition cost amortizes across an account that auto-renews for years. Gross margin compounds because the cost structure scales sub-linearly with revenue.
The SaaS-margin agency runs on agency distribution.
Each engagement is sold individually. Each engagement involves a sales cycle measured in weeks or months. Each engagement requires account management, scope negotiation, and the kind of senior-led pitch process that doesn’t scale. The shop’s customer acquisition cost looks like an agency’s — proposals, decks, conference attendance, network maintenance, senior partner time burned on pitches.
What the productized-services shop has is SaaS-shaped delivery costs. Software, compute, model calls, the engineering to keep the workflow integrated. These costs are real and scale, just not as steeply as labor.
Stack the two together and the picture clarifies. Delivery costs that look like SaaS. Distribution costs that look like services. The combination is worse than either pure model, because neither curve is the one that produces SaaS margins. The shop is paying SaaS infrastructure costs to deliver while paying services-business GTM costs to sell. The margin window between those two cost structures is the arbitrage. It is real. It is also narrow and closing.
The SaaS company that ships parity has the right distribution economics for its delivery cost structure. The productized-services shop has neither.
Roadmap debt
Year one looks good. Margins improve. Headcount stays roughly flat against revenue. The pitch lands because the proof is in the work.
Year two is where the structural cost arrives.
The SaaS company that owns the category — or the well-funded startup positioning to own it — observes the productized-services shop and ships parity. The SaaS company ships it embedded in the platform clients already use, with integrations the shop never had, with a roadmap of adjacent features the shop cannot afford to build, and at a price that reflects SaaS economics rather than services economics.
The productized-services shop has defenses. They are worth steel-manning.
Switching costs are real. A client who has integrated the shop’s workflow into procurement, creative approval, and project management does not switch on a dime. Embedded process is sticky.
Brand matters. A client who trusts a specific firm with a specific partner does not always re-evaluate when a cheaper option ships.
Integrations are non-trivial. The shop has built bespoke pipes into client systems that the SaaS company has to reconstruct.
These defenses buy time without changing direction. The SaaS company closes the gap on each one — switching cost erodes the moment the SaaS company’s parity feature gets cheap enough; brand erodes the moment a procurement team’s CFO asks why the line item is twice the alternative; integration debt is exactly the kind of work AI is best at compressing. The defenses degrade on the same curve that SaaS adoption follows everywhere else.
Customers, asked to choose between an agency that wraps a workflow and a SaaS platform that owns it natively, choose the platform. They get more from it. They pay less for it. They absorb the switching cost once the gap is large enough to notice.
The roadmap debt is paid in irrelevance.
The eighteen-to-thirty-six month window
The arbitrage window has a shape, and the shape is observable.
Clients buying AI-augmented services today are eighteen months into a particular conversation. The conversation began with “should we use AI to make this work faster” and has matured into “the work happens with AI now, and we are paying for the result rather than the process.” That maturation took roughly eighteen months from first-meaningful-deployment to baseline expectation.
The next eighteen months follow the same arc, on the same kind of conversation, with a different question. Clients now paying for AI-augmented services will start asking the second question. The second question is: if the AI is doing the work, why are we paying you to operate it rather than operating it ourselves. That question is already being asked in private. It will be weight on every contract negotiation soon.
The math of the window is straightforward. The arbitrage opens when the SaaS doesn’t yet exist for a given workflow and the client is not yet sophisticated enough to operate the AI directly. It closes when either condition flips. SaaS coverage is expanding monthly. Client sophistication is expanding monthly. The window is open in both directions and closing in both directions.
Eighteen to thirty-six months is a generous read. A specific shop in a specific category may have less. A shop in a less mature category may have more. The shape of the window — open now, closing on two sides — is the structural fact. The exact length is the variable.
Take the arbitrage when it appears. Refuse to mistake it for a business model. Capture the window, take the cash, do not build an organization that assumes the window stays open. The shops that compound do something different.
What survives
The agencies that compound from here are the ones flipping the unit of sale.
The unit of sale that survives is judgment, applied to ambiguous problems, with accountability for outcomes. What clients buy is a point of view on what they should do, why, and at what cost — and the willingness of the firm to stake its name on the answer being right. The deliverable is a decision, supported by artifacts, taken on the client’s behalf.
The shops still selling discovery-as-billable-hours and calling it strategy will not survive the next thirty-six months. The shops building a public point of view — pricing it, publishing it, betting compensation structures on it — will. The first kind operates below the moat that AI is steadily relocating upward. The second kind operates inside the moat and will keep operating there because the moat is not, structurally, where the next generation of foundation models is being pointed.
Three things mark the difference between a shop with a point of view and a shop performing the trend deck.
The first is pricing. A shop with a point of view prices against the decision. The price reflects the value of being right rather than the cost of the hours or the artifacts that produced the work. A shop performing strategy prices the way it has always priced and labels the result strategic.
The second is publication. A shop with a point of view publishes it. Essays. Talks. Public positions on what its category is getting wrong. A shop performing strategy circulates trend decks privately and keeps the position vague enough to retreat from when a client disagrees.
The third is compensation. A shop with a point of view structures compensation around the strategic call — partners and senior operators paid against the decisions they made, the bets that paid off, the clients they walked away from. A shop performing strategy pays on utilization and revenue per head, on the metrics that grade execution.
A shop hitting all three has a moat. A shop hitting one or two is in transition. A shop hitting zero is operating below the line where AI is climbing, and the climb is faster than the lease on the office.
The right read of the RFS
Y Combinator was looking at something real. The delivery model is broken. Hours are the wrong unit. AI changes the labor math underneath every services category in a way that hasn’t fully landed in P&Ls yet. The RFS is correct in diagnosis and the funding will produce companies.
The RFS is wrong in prescription. The companies it will produce are the feeders. They will validate categories. They will standardize workflows. They will publish, by virtue of doing the work, the answers the next generation of SaaS companies need to build inside their core platforms. Some will be acquired by those SaaS companies in the kind of bolt-on deal that pays the founders well and ends the firm. Most will simply run out of distance.
The agencies that compound — that grow into durable services businesses operating at margins approaching but never matching SaaS — are doing the opposite of what the RFS asked for. They concentrate senior judgment in the firm. They price judgment explicitly. They refuse the work that doesn’t carry strategic stakes. They operate above the climbing line. They publish a position the founder is willing to defend. They build compensation around the decisions they make rather than the hours they bill.
The meter is the bug. The position below the moat is the other bug. SaaS-margin agencies fix neither; the agencies that compound from here fix both.