Sell the Outcome, Not the Tool
A thesis keeps surfacing in nearly every conversation I’ve had with founders and investors since the start of the year. It comes from an essay by Julien Bek (Sequoia, March 2026), “Services: The New Software”, which has become one of the most-discussed investment frameworks of 2026. The headline: “the next $1 trillion company will be a software company disguised as a services company.”
I initially read it as VC positioning. Then I looked at the numbers. This isn’t a slogan — it’s a documented value shift already underway. Here is how I read it: the argument, and especially the two flaws that rarely get cited.
The Shift, in One Idea
The core idea is a simple opposition. The biggest winners of the AI era will not sell tools to professionals — they will sell the work itself, directly to the end client.
The economic logic is the real pivot. If you sell the tool (the copilot), every model improvement at Anthropic, OpenAI, or Google threatens to turn your product into a mere feature. If you sell the outcome (the autopilot), every model improvement makes your service faster, cheaper, and harder to compete against.
The canonical example: you spend $10K/year on accounting software, then $120K hiring an accountant to use it. The next great company won’t sell you the software. It will close your books for you.
This is no longer theoretical. In customer support, Sierra charges around $1.50 per successfully resolved ticket, and Decagon built its entire offer on the same pay-per-resolution principle. You’re no longer paying for access to a tool — you’re paying for a problem solved.
Why the Market Is Structurally Larger
The strongest argument in the thesis isn’t technological. It’s arithmetic.
For every dollar spent on software, six go to services. Tool vendors compete for the software budget; outcome vendors target the services budget — six times larger — with margins that expand as models improve.
This is the mechanic behind Harvey, the legal AI unicorn valued at $11B in March 2026, already deployed in nearly half of Am Law 100 firms. Rather than selling software seats, the company deploys “embedded legal engineering teams” at clients and signs revenue-sharing agreements: it captures the “delegated legal work” budget, not just the “licence” budget.
And this is no longer a lone Sequoia position. a16z, YC, and Bessemer are reaching the same conclusion: the next wave of value creation will come from AI companies that sell the work, not the tools.
Sizing the Disruption
Recent Gartner data quantifies the shift.
Scope of displacement. Up to $234B in enterprise application spending is exposed to “agentic arbitrage” by 2030 — roughly 20% of enterprise SaaS spend. The mechanism: AI agents complete tasks across multiple systems, reducing the need to interact with traditional software interfaces.
Agentic market growth. Spending on agentic AI software is projected to reach $985B by 2030, at a pace of +62.7% annually between 2025 and 2030.
Nature of the change. Gartner captures the key point: “you’re no longer buying software first for humans — you’re increasingly buying it for agents.” Consequence: as agentic AI takes hold, the user interface stops being a differentiator, which erodes the historical advantage of self-service SaaS — namely, a polished, easy-to-onboard UX.
Gartner’s framing stays measured, not apocalyptic: “it’s less an apocalypse than a metamorphosis. SaaS won’t be destroyed; it will re-emerge in a different form.”
The Structural Flaw: Margins
This is where the “AI-boosted” thesis hits accounting reality — and it’s the most serious counterpoint.
Classic SaaS rested on brutally efficient economics: write the software once, replicate the binary across thousands of customers, and let the marginal cost of an additional user tend toward zero. AI breaks this mechanic. Every query relaunches the model; every execution consumes GPU, memory, and energy. There is no “build once, sell forever.”
The numbers bear this out. Bessemer places gross margins for “LLM-native” companies at around 65%, well below the 80–90% ceiling that defined the cloud decade. For truly service-oriented models it’s worse: a16z observes revenue that tends toward non-recurring, lower gross margins (30–50%), and linear scaling at best.
The sharpest critique goes further: some “AI recurring revenue” is disguised consulting — integration and retainer work billed as a software subscription. The reason is structural: large models require continuous human alignment, client by client, to remain useful amid imperfect enterprise data. The conclusion is stark: selling an AI agent is economically closer to hiring a worker than buying a per-seat licence.
There’s even a trap specific to the service-as-software model: the rented AI stack powering each outcome looks like an operating expense — until it becomes a margin ceiling. And the vendors setting that ceiling raise prices precisely when you’re most locked in.
A counter-response from a16z does exist: low margins at a given moment don’t preclude a sustainable model, and part of the problem gets solved by routing expensive queries to smaller, cheaper models. The debate isn’t settled.
The Failure Rate Nobody Mentions
The investment narrative masks high attrition.
According to IBM’s 2025 CEO study, only 25% of AI initiatives delivered the expected ROI. Gartner expects more than 40% of agentic AI projects to be cancelled by 2027. And while 72% of organisations have at least one AI workload in production, the gap between deployment and captured value remains wide.
This is the void that players like Ode are trying to fill. But it also confirms that “making AI work in production” remains difficult, people-intensive, and far from guaranteed.
What This Changes, Depending on Where You Sit
For a software editor
The strategic question is no longer “what features to add” but “what work can I deliver on behalf of the client.” This pivot expands the addressable market but degrades the margin profile and reintroduces headcount. It’s not “AI replaces everything” — it’s “AI + experts capture a larger budget.” Crescendo illustrates this with its hybrid model: AI agents paired with a global human CX team, billed per outcome rather than per seat.
For an investor
Traditional SaaS multiples (10–12× ARR) no longer reflect the risk profile. AI-driven ARR, gross margin gains from automation, and the share of revenue anchored to outcome-based contracts become critical metrics.
For a pure tool vendor
The clock is running. Any company selling per-seat software in a category where AI can deliver the outcome directly is on a countdown — multiples and seat counts compress every quarter spent as a simple tool vendor.
Verdict
The trend is real, documented by serious market data, not just VC marketing. The economic mechanism is sound: sell the outcome rather than the tool to capture a budget six times larger and align with model progress. The cross-fund consensus is striking.
Two caveats temper the enthusiasm, however.
First, unit economics: the “service” model sacrifices precisely what made SaaS so profitable — near-zero marginal cost — and the risk that “AI ARR” is disguised consulting, structurally margin-capped, is very real.
Second, execution: with 40% of agentic projects expected to be cancelled by 2027, the done-for-you promise remains far harder to deliver than to pitch.
In short: this is an authentic and large-scale value shift, but one that trades an exceptional economic model (pure SaaS) for a larger, less profitable, and more fragile one. The funds’ bet is that market size and customer lock-in will offset margin compression. That bet isn’t won yet.
And at ARGO?
I’m not going to pretend we have the perfect answer — nobody does, and this analysis makes clear why. But the shift is happening in front of us, and we’re factoring it concretely into how we approach client projects.
Rather than only asking “what feature to deliver”, we’re asking more and more “what outcome can we take responsibility for on behalf of the client.” It’s a different conversation, and it changes both our deliverables and our approach to pricing and support. We’re moving into this territory with as much conviction as lucidity about what remains to be invented.
The figures and positions in this article reflect the first half of 2026 and are evolving fast. Reasoned disagreements are welcome.