A publisher's value, tested by AI

In 2025, OpenAI’s revenue went from $3.7 billion to more than $13 billion. That same year, the company lost $38.5 billion.

These two numbers don’t contradict each other. They tell the same story: in classic software, more growth mechanically produced more margin. With agentic AI, that’s no longer automatic.

This piece isn’t about OpenAI. It’s about what changes in a software publisher’s P&L, in its value chain, and in what still protects it from competitors. Three shifts, one question at the end: does your software stay useful in the strong sense, or merely stay operational?

Value and valuation, two notions we conflate

Everyday conversation blurs two words: value and valuation.

Value is what a company concretely delivers to its customers: time saved, more efficiency, more flexibility in their job, and what the company gets back in return, a healthy margin, a model that holds up over time. The customer stays at the center of this definition. Without perceived usefulness on their end, there’s no value to measure.

Valuation is something else: a price that someone, an investor, an acquirer, a shareholder, puts on the company at a given moment. That price doesn’t measure current value. It bets on what the company will produce tomorrow, in a defined economic context. The customer doesn’t factor in directly; what matters is a third party’s read on the company’s likely trajectory.

The two notions stay linked, with a natural lag: valuation anticipates a value that doesn’t exist yet, or keeps betting on a value that has already changed without anyone having noticed yet.

Agentic AI touches value at both ends at once. It can strengthen the service delivered to the customer (more speed, more personalization in their work), while also reshaping the publisher’s economic structure: cheaper development, and production (inference) that becomes a new recurring cost. The rest of this piece explores what AI changes in a publisher’s value, before coming back in the conclusion to what that means for its valuation.

What AI changes in your P&L

The classic software mechanic rests on a precise imbalance: expensive development, paid once, followed by production that costs almost nothing per additional customer. That imbalance is what let gross margin climb with growth, and it’s what historically justified software’s high multiples.

Agentic AI reverses part of that mechanic. Development gets cheaper, thanks to AI-assisted coding and much faster iteration. Production, on the other hand, regains a real, recurring cost: inference, which fires on every agent action, rather than once and for all like compiled code.

That cost doesn’t sit in the same place in the books as before. Classic development was often capitalized as a one-time investment. Inference, by contrast, books as a recurring variable expense, which hits gross margin directly, not just R&D spend.

The most counter-intuitive consequence touches growth itself. In classic software, more customers or more usage mechanically improves margin, thanks to operating leverage. With poorly managed AI, the opposite can happen: more usage erodes margin, since every agent action carries a cost that never fully disappears.

OpenAI’s case illustrates this mechanism at scale. The company’s revenue went from $3.7 billion in 2024 to more than $13 billion in 2025 (a jump few tech companies can claim). Audited financial documents nonetheless put its net loss at $38.5 billion for that same year, including an operating loss of nearly $21 billion, driven in large part by inference cost.

A publisher who chooses to host their own inference infrastructure, rather than pay for an API by usage, shifts part of that cost into an infrastructure investment and its depreciation. That choice isn’t just a line on a spreadsheet: it also brings data sovereignty and technical control of the system back to the publisher’s side. Its financial structure then starts to resemble, in some respects, that of an infrastructure operator, with financing needs different from those of a pure software publisher.

This erosion shows up directly in gross margin, a ratio no serious observer overlooks. Bessemer Venture Partners, in its February 2026 pricing playbook, puts AI product gross margin at roughly 50-60%, against 80-90% for classic SaaS, a gap close to the one confirmed by Digital Applied’s 2026 analysis of AI economic models.

ICONIQ Capital, in its State of AI 2026, measures a tighter gap and a different trajectory: 45% gross margin in 2025, 53% projected for 2026, 59% for 2027, against 70-80% for traditional SaaS. Two-thirds of surveyed companies also report improving per-query unit economics, driven by inference cost management and model routing. This improvement doesn’t contradict the mechanic described above: it concerns the marginal cost per action, not whether that cost exists. A company that manages this cost well improves it year over year; one that manages it poorly keeps absorbing it. This pressure on gross margin isn’t just a reporting matter: it directly hits agentic AI pricing decisions.

The value chain shift

The value chain in SMB accounting used to rest on a clear split: a publisher sells software, the accountant bills production time (data entry, reconciliation, filings) on top of advisory work. Pennylane built a platform that directly automates that production, already at a high and growing level. What AI absorbs here isn’t a peripheral task: it’s a large share of the work that historically justified accounting firms’ billing.

This shift is legitimately worrying part of the profession. Billable material shrinks on the production side, while advisory work, supposed to pick up the slack, is neither structured nor guaranteed to make up the shortfall. Valuation follows this shift faster than the debate gets settled: Pennylane reached unicorn status as early as 2024, while the part of the chain it automated sees its own economic value erode on the side of whoever used to occupy it.

The lesson for a business software publisher isn’t to check that the software still works, but to ask whether it stays useful in the strong sense, or risks being demoted to a secondary role while an automation layer captures the most valuable part of the chain. Software that keeps running without a hitch can still lose most of its value, if the usefulness it provided ends up absorbed elsewhere.

Competition between publishers shifts ground

Competition between business software publishers has long played out on functional breadth, integration quality, and switching cost for the customer, a field where publisher size carried real weight: a bigger development team meant covering more use cases.

Agentic AI shifts that balance of power by sharply cutting development cost and time. In a controlled experiment run by GitHub Next with Microsoft’s Office of the Chief Economist, developers using Copilot completed a development task 55% faster than those without it; at enterprise scale, Accenture measured 84% more successful builds after adoption. Building a business workflow automation, a chain of tasks that used to require a dedicated team and several months, becomes accessible to a smaller publisher with AI-assisted coding and faster iteration.

The direct consequence: the functional barrier to entry drops for everyone. A mid-sized publisher can now close part of the functional gap that used to separate it from a better-resourced competitor, on automations its customers already expect.

This leveling intensifies competition rather than easing it. If everyone can build the expected automation, it quickly stops being a differentiator and becomes table stakes. The battle then shifts elsewhere: the depth of domain knowledge encoded in the automation, execution reliability, and the speed to cover the next use case before others do. To fight that battle, publishers still have to choose between rebuilding their software with agentic AI or evolving what they have.

For valuation, this shift cuts both ways. It opens a window for smaller publishers to close a gap that once looked durable, but it also shortens the shelf life of any functional advantage once gained. Which pushes toward valuing iteration speed and proprietary data more than a feature list at a given point in time. We cover this shift in competitive forces in a dedicated piece.

Rallying around a domain, or a unifying concept

Lower development cost and time open up an opportunity: the means to evolve your value proposition faster than before.

One option is to consolidate, either internally or through integration into a larger group, around a unifying theme. This grouping can be horizontal, several different domains brought together by a shared concept, like Septeo around regulated professions. It can also be vertical, several links of the same value chain brought under one group, like ISAGRI in agriculture, which spans farm management, agricultural accounting, production-specific tools, and even equipment dealerships.

A second axis runs through presence in the customer’s operational workflows. An automation that embeds deeply into a customer’s day-to-day operations, to the point of becoming hard to remove, doesn’t just defend market share: it protects the publisher’s ability to hold its price, and reduces the risk of being replaced at the next contract renewal.

These two axes aren’t mutually exclusive, and neither is optional. A publisher that adjusts neither its structure nor its value proposition stays exposed: operating margin under pressure, a value chain redistributing without it, competition playing out on ground it no longer controls.

Not letting value drain toward foundation model providers

The question here is one of soundness as perceived by a third party evaluating the company: an analyst, a prospect, a partner. The test is simple to state: does the differentiation come from the software itself, or from convenient access to someone else’s model?

This soundness is already being scrutinized for so-called “GPT wrapper” startups: would the product survive the model provider’s next update, or would an in-house feature simply absorb it? An established business software publisher isn’t immune to the same diagnosis if its AI layer amounts to a call to a third-party API.

Open-weight brings benefits independent of that question: data sovereignty, and no dependency on token-based billing. Those benefits stand on their own, without needing to be tied to the strength of differentiation in front of an investor.

Differentiation itself plays out elsewhere. Foundation models will likely stay more powerful than what a publisher can host itself. For a business function well served by a smaller model, that power gap costs nothing, and open-weight can be chosen without sacrificing customer-perceived value. For a task that demands a foundation model’s frontier capability, giving that up for sovereignty carries a real cost, one to take on consciously rather than ignore.

Mike Saeks, CTO at Cursor, sums up the reverse problem well: using the most powerful model for a trivial task is like taking a race car to pick up milk.

Four scenarios for the next 3-5 years

Three to five years is a long time for an industry that reinvents itself every six months. No one can predict which of the following scenarios will win out, but each implies different choices starting today.

Consolidation. Publishers that haven’t built their own differentiation join larger groups, horizontal or vertical, that pool expertise and data. The market tightens around a handful of consolidated groups per sector, with independent publishers surviving mostly in niches with strong proprietary data.

Fragmentation. Lower development costs let smaller publishers ship targeted automations faster than large groups can integrate them. The market fragments into narrow niches, where agility beats scale.

Rising inference cost. A rise is actually more plausible than a drop. Agentic usage consumes growing volumes of tokens, and nothing guarantees that the drop in foundation models’ unit cost keeps pace. In this scenario, the gap widens between publishers who control their cost structure and those who’re subject to it, to the benefit of specialized setups whose design cost has become negligible.

The customer-turned-publisher. One last scenario deserves to be taken seriously: the one where the customer becomes a competitor. Facing foundation models powerful and simple enough to use directly, a share of customers takes their own workflows into their own hands rather than going through a dedicated business application, risking the agent becoming the entry point in place of the software itself. The publisher doesn’t just lose market share to a classic competitor; it watches part of its own customer base build what it used to sell.

These trajectories can coexist across sectors, some consolidating while others fragment or watch their customers become independent. What stays true in every case: a publisher’s position in three to five years will depend on the choices made now about its cost structure and its differentiation.

Conclusion: back to basics

Exits, consolidation, funding rounds, IPOs are only moments where a value that’s already there finally translates into an observed price. They’re circumstances, not strategies: solid value comes through any of them from a position of strength, neglected value can’t be made up for by picking the right moment or the right counterparty.

Generative and agentic AI poses a challenge to business software publishers. It erases their oldest protection: the difficulty and time it used to take to build an alternative solution. In the same motion, it opens possibilities that seemed out of reach not long ago. It reshapes the economic equation, the competitive landscape, up to seeing some customers become competitors, and up to the very structure of the P&L.

There’s no ready-made recipe for navigating this. In this situation, it’s better to go back to basics: what value do you want to deliver? To whom? At what price?