The AI Era Has a Maturity Gap, and the Bill Is Coming Due

Two senior Gartner analysts told a Sydney media roundtable this month that enterprise AI is past the experimentation phase and into the bill-paying phase. The advantage of releasing the most capable model has shrunk to a single quarter, agentic AI is in production at only a fraction of organisations, and a majority of buyers still cannot forecast what their AI deployment will cost them next year. Arun Chandrasekaran, Distinguished VP Analyst at Gartner, and Pieter den Hamer, VP of AI Research and Advisory at Gartner, framed the moment in plain numbers: 17 per cent of organisations have scaled AI across the business, 51 per cent are stuck at medium maturity, and 32 per cent are still experimenting.

The piece that follows traces each of those pressures through the same Sydney roundtable and the 2026 Gartner research stack behind it, including the firm’s own Hype Cycle for Agentic AI published in April. The pattern is the same in every direction: capability has raced ahead, contracts have not caught up, and pricing models designed for seat-licensed software are colliding with token-consuming agents that book no seats at all.

The AI Cost Reckoning Is Already Inside the Enterprise

The two analysts named four pressures that are reshaping enterprise technology buying at the same time. Model competition has shifted toward efficiency, reasoning, agents, and multimodality. Token usage is exploding as autonomous agents fire parallel requests into the same model backend. Seat-based SaaS pricing is under sustained pressure from customers who are starting to ask why they are still paying per user for software now run by an agent. And the maturity of the buying organisations themselves has not kept pace with any of the above.

The result is a 17 per cent figure that does real work. It is the share of organisations Chandrasekaran and den Hamer say have reached what Gartner classifies as high AI maturity, meaning AI is embedded across business functions at enterprise scale. The remaining 83 per cent sit on the wrong side of a maturity curve that vendors continue to sell across as if everyone were already on the other side.

The Sydney Roundtable That Framed the Gap

Both analysts spoke at a media roundtable during the Gartner Data & Analytics conference in Sydney, where the central theme was the gap between what frontier AI labs promise and what their contracts actually deliver. The first pressure they named was the service-level agreement chasm, and it sets the lens for everything else in the room.

Chandrasekaran told the roundtable that enterprise customers are routinely surprised by the contracts they receive from frontier labs, and that the surprise cuts in both directions: customers expect enterprise-grade protections on data, liability, and uptime, while the labs limit their exposure in ways traditional enterprise software vendors never did.

The gap between those two positions is one of the defining features of this stage of the market, and it is the reason the rest of the pressures downstream, on pricing, on cost, on governance, keep landing harder than the marketing materials suggest.

There is a chasm between enterprise expectations from an SLA [service level agreement] perspective from the Frontier AI Labs versus what the Frontier AI Labs are actually providing to enterprise customers.

The remark came from Arun Chandrasekaran, Distinguished VP Analyst at Gartner, at the Sydney conference, and it tracks with the broader ARNnet reporting on the same roundtable (the ARNnet account of the Gartner Sydney roundtable), which frames the gap as evidence of the immaturity of both the technology and its commercial scaffolding.

Vendor Competition Has Shifted to Efficiency, Reasoning, and Agents

The leaderboard churn is now a quarterly event rather than a generational one. The advantage gained by releasing the most capable model has shortened to a single quarter in most cases, and vendors are converging on the same four axes of competition: model efficiency, reasoning capabilities, multimodal functionality, and agents capable of orchestrating workflows across enterprise systems.

Computer-use capabilities and tool integration are the most visible near-term frontier. As models grow better at acting on real software interfaces, the work that an agent can complete without human supervision grows with them, which is why Gartner’s 2026 Hype Cycle for Artificial Intelligence places AI agents and AI-ready data as the two biggest movers on this year’s curve.

Speech recognition and reasoning models are creating new openings in customer service, where voice agents are starting to handle airline, bank, and insurer calls before routing the harder cases to a human. The direction of travel is set, but the vendor race is no longer about who can ship the most powerful model. It is about who can ship the most useful one per token spent.

The Token Bill That Broke the Free Lunch

For two years, enterprise buyers treated AI as if it were free, and the vendor community let them. Generative AI was a consumerisation phenomenon, Chandrasekaran told iStart, with tools built for personal use flowing into the enterprise at subsidised rates as vendors chased user bases and the data advantages that came with them. That arrangement is over.

The clearest signal landed at the start of June. GitHub moved GitHub Copilot Pro from a user-based pricing model, in place until May 31, to a token-based or consumption-based pricing model starting June 1, and the change has hit enterprise procurement teams harder than GitHub intended. Chandrasekaran told the Sydney roundtable that the shift has created concern because organisations do not fully understand the downside risk of token-based pricing, and because many of them lack the tools and instrumentation needed to measure token usage effectively. AI free lunch language is what the Linux Foundation is now using to describe the regime that is ending.

The structural reason is that AI agents behave differently from chatbots. An agent can fire sequential or parallel requests into the backend model as it works through a task, and each of those requests costs tokens. As agent workflows proliferate, the per-task bill grows in a way that the previous generation of chatbot deployments never did.

• GitHub Copilot Pro: user-based pricing through May 31, token-based from June 1 (Chandrasekaran via the Sydney roundtable).
• Token-based AI costs are now the largest and fastest-growing line item on enterprise technology budgets per the Linux Foundation (reported by iStart).
• Reddit case reported by iStart: one developer’s GitHub Copilot bill moving from US$25 per month to US$750 per month.
• 51 per cent of organisations sit at medium AI maturity and are challenged on demonstrating AI ROI (den Hamer via the Sydney roundtable).
• The Linux Foundation launched the Tokenomics Foundation in mid-2026 to standardise how AI cost is measured across vendors (the iStart write-up on AI token costs).

Seat-Based SaaS Pricing Meets the Agent Era

The market has been talking about the SaaSpocalypse, the projected collapse of seat-licensed software in an agent-first world. Chandrasekaran is sympathetic to the direction of the disruption, but not to the timeline. He told CRN Australia that SaaS providers still hold meaningful moats in domain knowledge, domain data, tight workflow integration, and the compliance and regulatory capabilities that enterprise buyers rely on. Conservative procurement teams will not throw those away overnight.

The trouble is in the pricing model. Per-seat or per-user pricing was built for software used by humans, and an agent running workflows on a user’s behalf does not occupy a seat. As organisations deploy more AI agents orchestrating workflows independently or semi-independently of humans, the question of why pricing should still be tied to user count becomes louder.

Startups are already experimenting with outcome-based and workflow-based pricing. The large SaaS vendors see the same logic, but they also see the cannibalisation risk: a wholesale move to consumption or outcome pricing would reset their gross margins and break the recurring revenue model the public markets reward them for. The result is a holding pattern in which the largest providers talk about flexible pricing while protecting the structure that funds their margins.

I do personally believe [SaaS] is being disrupted by AI, but at the same time, we may be over-exaggerating the impact in the near term. SaaS providers still have moats, domain knowledge and domain data [as] strong advantages, as [well as] very tight workflow integration.

The remark came from Arun Chandrasekaran, Distinguished VP Analyst at Gartner, speaking to CRN Australia at the same Sydney roundtable (the CRN Australia report on the SaaSpocalypse), and the caveat is the load-bearing part: the disruption is real, but the doomsday version is not.

The Maturity Curve in Three Numbers

Den Hamer framed the maturity question in three tiers during the Sydney roundtable, and the global averages he cited are now the most-cited frame for where enterprise AI stands in 2026. A small group has moved past pilots into scaled, enterprise-wide deployment. A larger group is generating some value but cannot prove ROI on it. The largest group of all is still at the experimentation stage.

Den Hamer added that the medium-maturity band is the one most under pressure right now. Those organisations have moved past the proof-of-concept stage and are running real workloads, but they have not yet built the governance, the measurement frameworks, or the change-management muscle to demonstrate the returns to a board. That is also where most of the AI budget pressure is showing up.

The 17 per cent, 51 per cent, and 32 per cent figures come from Pieter den Hamer, VP of AI Research and Advisory at Gartner, at the Sydney media roundtable, reported by ARNnet, and they cover organisations globally, across industries.

Only About One in Five Has Agents in Production

Agentic AI has been the most heavily marketed technology category of the past year, and yet the production footprint is still small. According to the 2026 Gartner CIO and Technology Executive Survey, only 17 per cent of organisations have deployed AI agents to date, while more than 60 per cent expect to do so within the next two years. Gartner calls this the most aggressive adoption curve among all the emerging technologies measured in the survey.

Den Hamer told the Sydney roundtable that the remaining 80 per cent of organisations are only experimenting with agentic AI or are still thinking about it, a gap that fits the broader pattern of ambitious planning and slower execution that runs through the maturity data.

Gartner’s own 2026 Hype Cycle for Agentic AI, published in April, places the category at the Peak of Inflated Expectations, with strong momentum across the curve and uneven maturity across the supporting capabilities. Most deployments today remain narrowly scoped, and fully autonomous agents are not ready for the majority of enterprise use cases.

• Software engineering: AI coding assistants and testing tools remain the most common production use case.
• Customer support: agent-assisted triage and resolution in contact centres.
• Operations: workflow automation for back-office tasks.
• Most deployments remain narrowly scoped, and fully autonomous agents are not ready for the majority of enterprise use cases (Gartner’s 2026 Hype Cycle for Agentic AI).

For a concrete example of how agents are starting to script their own tool use in production, see an example of agents scripting their own search stack, a deployment pattern that mirrors the same shift from per-seat to per-token that is hitting SaaS pricing.

AI Literacy Is the ROI Lever

Across both analysts’ remarks, the variable that distinguishes the organisations extracting value from AI from the ones still experimenting is AI literacy. Den Hamer was direct on the link: investing in workforce education correlates with measurably higher returns on AI spend, and organisations that treat AI literacy as a real program rather than a slide deck see the difference on the income statement.

The literacy work is not only about how to use AI in daily tasks. Den Hamer pointed to the rise of shadow AI, the unsanctioned use of public AI tools inside the enterprise, as the reason responsible-use training matters now. Employees need to know the policies, the data-handling rules, and where critical thinking still applies, especially as agents take on more autonomous work.

The labour-market signal runs in the same direction. Separate Gallup data summarised by Oton Technology found that regular AI use is now associated with lower layoff risk, a finding that fits the literacy correlation both analysts named in Sydney.

Frequently Asked Questions

What is enterprise AI maturity?

Enterprise AI maturity is a Gartner framework that places organisations on a three-tier curve based on how broadly they have deployed AI across business functions and how well they govern it. The tiers are high, medium, and low maturity, and the Sydney roundtable placed 17 per cent of organisations at high maturity, 51 per cent at medium, and 32 per cent at low.

How many organisations have deployed AI agents in production?

The 2026 Gartner CIO and Technology Executive Survey reports that only 17 per cent of organisations have deployed AI agents to date, while more than 60 per cent expect to do so within the next two years. Pieter den Hamer told the Sydney roundtable that the remaining 80 per cent are still experimenting or have not yet started.

Why is the SaaSpocalypse unlikely in the near term?

Arun Chandrasekaran told CRN Australia that SaaS providers still hold moats in domain knowledge, domain data, tight workflow integration, and compliance capabilities, and that conservative enterprise buyers will not throw those away. He sees real disruption to seat-based pricing, but not a near-term collapse of the SaaS model.

What is the AI free lunch?

The phrase, used by Arun Chandrasekaran and the Linux Foundation, refers to the period in which enterprise users accessed AI tools at subsidised, per-seat rates while vendors chased user bases and training data. That regime is ending as vendors move to consumption-based pricing tied to token use, which makes AI cost a boardroom variable rather than a fixed line item.

How does AI literacy affect return on investment?

Pieter den Hamer told the Sydney roundtable that organisations investing in AI literacy see measurably higher returns on AI spend than those that do not. The program covers effective use of AI tools, responsible use of AI in line with corporate policy, and the critical thinking employees need to avoid over-reliance on AI output. Outside reporting from Oton Technology also found, via Gallup data, that regular AI users face lower layoff risk (Gallup’s finding that regular AI users face lower layoff risk).