The Long-Term GaaS Market-Size Projections, Scrutinized

Table of Contents

• Why GaaS Market Forecasts Deserve Suspicion
• What the Big Numbers Actually Count
• The Substitution Problem at the Heart of Every Model
• Three Forecasting Methods and Where They Break
• The Bottlenecks Forecasts Underweight
• A Saner Way to Size the Market
• What Would Make the Bull Case Real
• Insights Most People Overlook
• References

Why GaaS Market Forecasts Deserve Suspicion

There is a familiar pattern to any new technology category. A few credible firms publish a number. Competing firms publish bigger numbers so as not to look bearish. Vendors quote whichever figure flatters their pitch deck, and within eighteen months the largest number has become "consensus" even though no one can reconstruct how it was built. Agentic AI-as-a-Service is deep into this phase right now.

I am not arguing the category is small. It is plainly going to be large, and within this cluster I have argued elsewhere that the marginal-cost-near-zero workforce thesis is directionally correct. What I am arguing is that the specific dollar figures being thrown around, the ones that put GaaS at two, three, or five trillion dollars by the early 2030s, are far softer than their precision suggests. A forecast that reads "$4.7T by 2033" implies a confidence the underlying model does not earn. The honest version of most of these numbers is "somewhere between a few hundred billion and a few trillion, depending on assumptions we can't yet test."

The reason to scrutinize this matters beyond pedantry. Founders raise against these TAMs. Enterprises budget against them. Public-market analysts will eventually price equities against them. When the input is a casually sourced hockey stick, the downstream decisions inherit that fragility.

What the Big Numbers Actually Count

The first problem is definitional, and it is bigger than it sounds. Ask three analysts what "the GaaS market" includes and you get three different boundaries.

Some count only the revenue paid directly to agent vendors, the per-task and per-outcome fees that define the agent economics of the category. That is the narrowest and most honest definition, and it produces the smallest number. Others count the total contract value of any deal that touches an agent, including the human services, integration work, and software seats bundled around it. The broadest definitions count the value of the labor displaced or augmented, they take the wage bill of every knowledge worker an agent might touch and call that the addressable market.

That last move is where the trillions come from, and it is a category error. The wage bill of a job an agent can do is not the price someone will pay for the agent to do it. If anything, the whole premise of GaaS pricing is that the agent costs a fraction of the labor it replaces, that is the buyer's incentive. So the more disruptive agents are, the more the price compresses relative to the labor displaced. A forecast that sizes the market at the labor it replaces is implicitly assuming agents capture the full value they create, which is the opposite of what competitive markets do. This connects directly to the question of who captures the productivity gains from agents, and the answer is usually "the buyer, not the vendor."

When you see a GaaS TAM north of a couple trillion dollars, your first question should be: is this counting revenue to agent providers, or is it counting the GDP of the work agents touch? Those are different by an order of magnitude.

The Substitution Problem at the Heart of Every Model

Every market-size model for a labor-replacing technology has to answer one question, and most answer it badly: when an agent does work that a human used to do, what happens to the money?

There are three possibilities, and they produce wildly different market sizes.

Money transfers. The buyer pays the agent vendor roughly what they paid the human. This is the assumption baked into the biggest forecasts, and it is almost never how it plays out. Vendors don't get to price at labor parity for long, because a competitor will undercut them.

Money compresses. The work still gets done, but it gets dramatically cheaper, and most of the savings flow to the buyer or to end customers as lower prices. This is the deflation of professional-services pricing scenario, and it is the most historically common outcome for automation. Under this model, the GaaS revenue market is a fraction of the displaced wage bill, but the economic impact is enormous and shows up as cheaper services everywhere, not as agent-vendor revenue.

Money expands the pie. Cheaper agentic work makes previously uneconomic tasks worth doing, so total spend rises even as per-task prices fall. This is the genuinely bullish case, and it's real, think of how cheap cloud compute created workloads that never existed before. But it's also the hardest to forecast, because it depends on demand elasticity nobody has measured yet for autonomous work.

McKinsey's research on generative AI's economic potential is careful about this distinction, it sizes value created across the economy, not vendor revenue, and the two are routinely conflated when the headline number gets recycled. A serious GaaS forecast has to pick which of these three regimes it believes and defend the choice. Most don't even acknowledge the fork exists.

Three Forecasting Methods and Where They Break

The numbers floating around are produced by a handful of methods. Knowing which one generated a figure tells you a lot about how much to trust it.

Top-Down Labor Substitution

Start with total knowledge-worker spend, estimate the fraction of tasks agents can do, apply an adoption curve, multiply. This is the method that yields the scariest and least reliable numbers. Its fatal weakness is the substitution assumption above, it almost always implicitly prices agents at labor parity. It also tends to treat "tasks an agent can do" as "tasks an agent will be trusted to do unsupervised," which collapses the distinction this whole cluster cares about: the gap between augmentation and replacement.

Bottom-Up Vendor Aggregation

Sum the actual and projected revenue of agent vendors, extrapolate. This is the most grounded method and produces the most defensible near-term numbers, but it badly undercounts the long tail and the embedded case, agents sold inside other software, where the GaaS revenue is invisible because it's bundled into a SaaS seat. It also struggles with the consolidation endgame: if a handful of winners eventually capture most of the market, summing today's fragmented vendors misreads the trajectory.

Analogy-Based Sizing

"GaaS will follow the cloud/SaaS adoption curve." Analysts anchor to the trajectory of AWS or Salesforce and scale by some multiple. The appeal is obvious and the danger is subtle: cloud sold capacity with near-perfect reliability, while agents sell outcomes with probabilistic reliability. The adoption curve for something you can't fully trust looks different from the curve for something that just works. a16z's writing on the shift toward AI-driven services revenue gestures at this, but the analogy gets stretched well past where it holds.

The takeaway: when a forecast cites a single clean number, find out which method produced it. Top-down gives you the trillions; bottom-up gives you the billions; the truth lives uncomfortably between them.

The Bottlenecks Forecasts Underweight

Hockey-stick charts assume adoption is gated by capability, that as models get better, the market simply fills in. In practice, the binding constraints on GaaS revenue are mostly non-technical, and forecasts that ignore them overshoot badly.

Liability and trust. An agent that's right 95% of the time is a productivity miracle and a procurement nightmare. Enterprises don't buy unsupervised autonomy until someone can answer "who's accountable when it's wrong." This is why agent reliability and agent security aren't side topics in the GaaS story, they are the actual gating function on revenue. The public-trust perception gap shows up directly in sales cycles.

Integration drag. Agents create value by acting inside a company's real systems, and those systems are a mess of legacy software, undocumented processes, and permission boundaries. The work of wiring an agent into a live enterprise is slow, expensive, and deeply unglamorous, and it caps how fast adoption can compound regardless of model quality.

Pricing-model immaturity. Per-outcome pricing sounds elegant until you try to define an "outcome," attribute it cleanly, and price it before you know your own cost to deliver. Until the agent economics of outcome-based billing stabilize, a lot of deals stall in pilot purgatory, generating zero recognized revenue.

Backlash and regulation. The cultural backlash against agent-everything and inevitable regulatory friction will slow adoption in exactly the high-wage, high-stakes verticals (law, healthcare, finance) where the top-down models assume the biggest gains.

None of these kill the category. All of them flatten the early years of the curve, which is precisely where most forecasts assume the steepest growth.

A Saner Way to Size the Market

If I were forced to put numbers on this, I'd refuse to produce a single one. The useful output is a scenario band, not a point estimate.

Build three explicit cases. A conservative case assumes money compresses, integration drag is heavy, and trust gates limit unsupervised deployment to low-stakes work, GaaS vendor revenue lands in the low hundreds of billions by the early 2030s. A base case assumes a mix of compression and pie-expansion, with outcome-pricing maturing in a few verticals, call it several hundred billion to roughly a trillion. A bull case requires the pie genuinely expanding, trust mechanisms maturing fast, and a few category winners achieving cloud-scale economics, that's where the multi-trillion figures live, and it should be labeled as the tail scenario it is, not the midpoint.

The discipline here is to separate economic impact from vendor revenue and report both, clearly labeled. The impact number is legitimately enormous and worth citing. The vendor-revenue number, the thing a GaaS founder or investor actually cares about, is much smaller and much more uncertain. Collapsing the two is the single most common sin in this literature. Gartner's repeated cautions about the hype cycle for emerging AI are a useful corrective: peak-of-expectations forecasts have a poor track record, and agentic AI is sitting near that peak right now.

What Would Make the Bull Case Real

I want to be clear that I'm not bearish on the category, I'm bearish on the precision of its forecasts. The multi-trillion outcome is genuinely possible, but it depends on a specific chain of things going right, and watching those is more useful than watching the headline number.

The bull case becomes real if outcome-based pricing stabilizes into something auditable, if liability frameworks emerge that let enterprises deploy unsupervised agents in high-stakes work, and if demand proves elastic enough that cheaper agentic work expands total spending rather than just cannibalizing existing budgets. It also depends on what agent-native companies look like at scale, if a generation of firms is built from the ground up around agent labor, that's a structurally larger market than retrofitting agents into existing org charts.

Track those mechanisms, not the dollar figure. A forecast is only as good as the assumptions you can interrogate, and in GaaS the assumptions are doing all the work the decimal points pretend to.

Insights Most People Overlook

The biggest GaaS revenue may never be labeled "GaaS." As agents get embedded inside ordinary SaaS products, the per-task economics disappear into seat pricing. The category could "underperform" its own forecasts on paper while actually winning, because the revenue migrated into line items nobody counts as agent revenue. Pure-play GaaS TAM may shrink even as agent usage explodes.

Forecasts assume agents are bought; many will be built. The cheaper agent-building gets, the more enterprises assemble their own agents instead of buying them as a service. That's a direct tax on the GaaS vendor market that top-down models entirely ignore, every internally built agent is displaced labor that generates zero GaaS revenue.

Deflation is the tell that the category is working, not failing. If GaaS succeeds, professional-services prices fall, which means the dollar value of the work shrinks even as the volume of work explodes. A market-size model denominated in dollars will look like it's stalling at exactly the moment the technology is most transformative. We may need unit-of-work metrics, not revenue, to see what's actually happening.

The winner-take-most dynamic cuts the TAM, not just splits it. If consolidation produces a few dominant agent platforms with real economies of scale, they'll compete on price, which compresses category revenue below what a fragmented, higher-margin market would show. A healthy competitive endgame is bearish for aggregate vendor revenue even as it's bullish for adoption.

The labor-substitution framing flatters the wrong party. Sizing the market by displaced wages implicitly credits agent vendors with value that, historically, flows to buyers and consumers. The most likely outcome is that society captures most of the gain through cheaper services, a great result that shows up nowhere in a GaaS vendor-revenue chart.

References

• The economic potential of generative AI: The next productivity frontier, McKinsey & Company
• The Emerging Landscape of AI Agents, Andreessen Horowitz (a16z)
• Hype Cycle for Artificial Intelligence, Gartner