Before using an agent-AI market forecast in a valuation, decide what the number measures. Software revenue, spending that might move between vendors and the value of purchases influenced by software are different economic quantities. Putting them in the same spreadsheet column can make a business opportunity look much larger without adding a single paying customer.
A September 14 contributed analysis at The Next Web draws attention to the gap between agent-market forecasts. The useful next step is to return to the original research summaries. They corroborate a large difference in projected market size, but do not provide a licence to average the estimates or treat them as independent measurements of one precisely defined industry.
A similar growth rate does not make the markets identical
Grand View Research projects a $24.5 billion enterprise agentic-AI market in 2030 and reports a 46.2% compound annual growth rate for 2025–2030. Its summary estimates the 2024 market at $2.6 billion. These are the research firm's market estimates and projections, not audited revenue totals or a guaranteed trajectory.
MarketsandMarkets projects a $52.62 billion AI-agents market in 2030, with a reported 46.3% growth rate for 2025–2030. Its summary gives a 2024 base of $5.26 billion and a 2025 projection of $7.84 billion. The dates belong with the figures: a historical estimate, an intermediate projection and an endpoint should not all be presented as current market sales.
The reported growth percentages are close, while the scopes and starting values differ. That does not establish independent agreement about future demand. A compound rate summarises a path between a chosen starting value and an endpoint; it does not reveal the quality of the underlying assumptions. Similar rates can emerge from different definitions, customer populations or modelling choices.
The public summaries alone do not allow a complete reconciliation of every included product and revenue category. That is a material limit, not something to fill with an invented explanation. An investor should identify whether a particular supplier's revenue actually falls inside each definition before using either total to calculate market share. Adding the two totals would be especially difficult to justify because the categories may overlap.
Spending at risk is not a new revenue pool
Gartner offers a third kind of number. Its July 1 analysis says up to $234 billion of enterprise application spending is exposed to agentic arbitrage between then and 2030. It describes agents completing work across systems and weakening the link between user interaction and traditional software-seat revenue. This is an estimate of exposure to disruption, not a forecast that agent startups receive that entire sum.
If a customer replaces part of an existing software contract with an agent service, the new supplier may gain revenue while the old one loses it. Some savings may stay with the customer. An incumbent may also supply the agent itself. The gross budget affected therefore need not equal incremental industry revenue, let alone profit for a particular vendor.
Contracts determine who captures the value. Charging per seat, per task or per completed outcome distributes volume and execution risk differently. A supplier promising a completed result may have to absorb retries or human review. A customer paying for usage may carry more of that variability. A large addressable market does not answer those contractual questions.
Project attrition changes who earns the forecast
In June 2025, Gartner predicted that more than 40% of agentic-AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. That remains an attributed forecast over a stated horizon; it is not a measured cancellation rate for September 2026.
High attrition and strong aggregate spending can coexist. Organisations can spend on experiments that never reach durable operation, while a smaller group of successful deployments expands. Revenue earned during a pilot may therefore be genuine but less repeatable than revenue from a renewed production contract. Neither a large market forecast nor a cancellation forecast identifies that mix for an individual business.
The counterargument to excessive scepticism is that experimentation can create reusable learning, and successful suppliers may gain share as weaker offerings disappear. That is possible. It still requires evidence of retained customers, repeat usage and costs that improve with experience. A failed project should not automatically be counted as proof that the entire technology lacks value.
Build the investment case from a paid workflow
The practical starting point is a workflow for which someone pays: what is delivered, how frequently it is needed and what it costs to complete reliably. From there, revenue can be compared with computing, integration, support and exception-handling costs. This produces a narrower estimate than a market headline, but one that can be tested against actual contracts.
Evidence that would strengthen the case includes renewals after initial experimentation, disclosed customer concentration and contribution margins that remain sound as usage grows. Evidence that would weaken it includes recurring manual intervention, expensive integration repeated for each client or revenue that vanishes when pilot budgets expire. These are evaluation criteria, not assertions about any named supplier's present performance.
The forecasts are useful maps of possible opportunity. Their most important labels are the unit being measured, the market boundary, the base year and the forecast horizon. Once those are preserved, investors can ask the harder question: which part of that possible market can a business serve repeatedly and profitably?
