Three years in, the enterprise AI journey has shifted character. The early questions were about possibility, then changed to proof of deployments and now the questions are about finding ROI.
Sustainable AI happens only when enterprises and vendors can see value for growing investments and viable, fundable business models. That is not easy given the challenges in budgeting costs (the iceberg below the obvious waterline) and realising value (the Fog that surrounds the cost iceberg). Yet, enterprises can rely on fundamentals for navigating the AI led transformation.
In Mindvista Edition 56 (October 2025), we observed a structural shift with Enterprise AI shifting to Enterprise wide AI moving from isolated experiments to embedded operating architecture across the enterprise.
In Mindvista Edition 66 (February 2026), we revisited AI is eating software (Mindvista Edition 35( February 2025) to find that AI is not devouring software. To the contrary, it is sharing the table by shaping an emergent new three-layer architecture of Systems of Record, Execution, and Fabric.
In Mindvista Edition 67 (February 26) , we mapped the business model friction as vendors and enterprises wrestle over how to price and capture the value of these new layers.
The feedback from those editions raised a fundamental question: How do enterprises actually model costs, make the business case, distinguish between incremental and transformative approach?
This edition is a response to questions and also a critical exploration of the variables to model ROI for sustainable AI for continued investments by buyers and vendors. The variables covered are AI costs, AI value, the role of accounting rules, the risk and liability gap and reconciling all these for an AI ROI.
Enterprise AI Cost Structure: Pricing Models, Hidden Costs & Implementation Debt
Above the waterline, vendor cost dynamics are in a state of unprecedented volatility and natural friction for a market searching for equilibrium. The fundamental unit of a per seat cost from the value of a human using software, is being replaced by AI and Agentic AI capabilities which do not have settled pricing conventions as yet.
The empirical data proves the scale of this shift. PricingSaaS tracking of 500 SaaS companies recorded roughly 1,800 pricing changes in 2025, an average of 3.6 per company, a level of volatility unseen in traditional subscription era SaaS. (reported via Growth Unhinged).Consumption credit-based models grew 126% year-over-year. The PricingSaaS500 Index further showed AI leaders made 28% more feature changes , 21% more plan changes and 20% more price change with a clear move to add AI add-ons to base tiers.
Vendors are navigating severe unit economics pressure, shifting away from predictable subscriptions toward metered pass throughs to protect their margins from compute-heavy AI workloads.
Pricing is also emerging as a distinct competitive posture, than me too models. Consider the approach to the APIs that connect the AI Fabric to the Systems of Record. Vendors like Salesforce utilize “Flex Credits” that act as API tollbooths, charging per-action. In a powerful counter-signal, Veeva another System of Record vendor has introduced fast, explicitly free Direct Data APIs for AI and Agentic workloads. This demonstrates use of open data access as a competitive wedge to win enterprise trust, reducing customer friction and signaling alignment rather than extraction.
Below the turbulent pricing waterline lies the iceberg of hidden, structural costs. First, token consumption for heavy developer and BI roles is acting as a practical budgeting threshold. In enterprise budgeting discussions, inference usage frequently flags at ~$100K per year for power users, prompting CFOs to implement strict per-session usage caps, role-based rollbacks, or shifts to open-source models.
Compounding this variable cost is massive “Implementation Debt. Integrating the AI Fabric into legacy Systems of Record requires heavy data cleansing, context engineering, and observability tooling. This debt is commonly reported as multiples of the initial software license cost, turning what looks like a straightforward software purchase into a multi-year engineering lift.
This debt compounds in a specific way in the three-layer architecture. The Systems of Record carry decades of technical debt from prior technology transitions. The AI Fabric requires clean, structured, contextually rich data to function. The gap between the two is the debt.
AI Cost vs ROI in Enterprise: Why Value Is Harder to Measure Than Cost
AI productivity gains are real, but it is also heavily obscured by the fog of exaggeration and increased human cognitive workload caused by AI.
Stanford research from Erik Brynjolfsson confirms that double-digit productivity gains exist at the micro-economic, task level. High performers who aggressively redesign their workflows are seeing returns. The KPMG 2026 Global Tech Report adds a further dimension, that high performers are achieving an average of 4.5 times ROI, compared with the industry average of approximately 2 times. But only 24% of organisations are scaling AI successfully across use cases. (Source: KPMG Global Tech Report, 2026)
The gap between the leaders and the rest of the field is widening, not closing and the productivity benefit is neither uniform nor universal.
The HBR and Axios research on “workslop”, an AI-generated output that is deceptively plausible but lacks the substance to move work forward, quantifies a productivity dynamic that standard ROI models do not capture. Across 1,150 US desk workers surveyed, 40% reported encountering AI-generated workslop in the preceding month. The estimated economic loss runs to approximately $186 per worker per month from the human verification overhead required to distinguish credible AI output from output that needs rework. (Source: HBR/Axios, 2025) . There is also new work. Deepak Subramani, a practitioner post on X, also captured this in operational terms: AI coding tools generating ten times the volume of pull requests, with human review work increasing twenty-five-fold over six months, turning initial productivity gains into exhaustion.
Even granting these gains, there is still the distance problem between task-level gains and enterprise-level ROI.
The Deloitte 2026 State of AI in the Enterprise report, surveying 3,235 senior leaders across organisations sets out the gap precisely. 66% of organisations report productivity and efficiency gains. Only 20% are realising revenue uplift. Agentic AI, now in use at 57% of organisations, is delivering significant ROI at just 10% of those deployments.
Gartner warns that many “Agentic” claims are hyped and true autonomous capability is actually concentrated in a small subset of vendors. They project that more than 40% of Agentic AI projects will be canceled by the end of 2027 due to escalating costs and unclear business value. True autonomous capability with agents that can actually complete complex multi-step business processes without human intervention is concentrated in a very small subset of vendors.
In an interesting insight from Deloitte State of AI in the Enterprise, 2026, the typical AI investment payback period has stretched to two to four years, compared with seven to twelve months for prior technology cycles.
Enterprise AI Cost Structure: CapEx vs OpEx and the Economics of AI Architecture
The cloud transition converted capital expenditure into operational expenditure. AI is complicating the equation again in ways that CIOs and CFOs are working to resolve.
Under FASB ASU 2025-06, which modernises internal-use software accounting by removing outdated project-stage language, certain Fabric components built as internal-use software can be capitalised as depreciable assets, provided the organisation can demonstrate a “probable-to-complete” assessment and committed development governance. (Source: FASB ASU 2025-06; Deloitte DART; KPMG, 2025). However, exploratory Fabric work and volatile token consumption remain strict OpEx. Because the “probable-to-complete” assessment is rigorous, organisations that lack strong governance will find their AI development flowing entirely through OpEx, eroding operating margins. Token consumption and API calls remain operational expenditure by nature. Cloud infrastructure supporting the Fabric , compute, storage, network , typically remains OpEx.
The strategic implication is significant. Organisations that architect their Fabric as owned internal software orchestration layers, agent policy engines, context retrieval services, internal evaluation harnesses can treat this as a depreciable capital asset, typically over three to five years. Those that rely entirely on vendor-hosted, consumption-billed infrastructure remain in an OpEx-only position, with no balance sheet recognition of the strategic value they are building.
FinOps, the cloud financial management discipline that brought governance to cloud spending, is evolving its frameworks explicitly for AI workloads. Showback and chargeback models that make AI run-costs visible to the teams creating them, unit economics KPIs connecting consumption to outcomes, and spend volatility controls are the emerging operating model. The organisations managing AI as a financial variable, not just as a technology initiative are the ones building durable ROI capability. (Source: FinOps Foundation, 2025/2026)
In a constrained business climate where finding budgets, public valuations, and private fundraising depend on strict P&L management, accounting rules have become a critical survival mechanism. The distinction between Capital Expenditures (CapEx) and Operational Expenditures (OpEx) is actively dictating how enterprises budget their AI initiatives.
Why Enterprise AI Pricing Is Complex: Risk, Liability & Hidden ROI Costs
Accounting rules determine how AI investments appear on the balance sheet. But there is a cost that does not appear on any balance sheet at all.
When an AI agent makes a costly autonomous error, misquotes a price, mis-routes a shipment, triggers an unintended workflow, the financial hit lands on the enterprise. Standard AI service agreements are structured to ensure this.
The OpenAI Services Agreement, effective January 2026, is illustrative. Vendor liability is capped at the trailing twelve months of payments. Indemnification is limited to intellectual property infringement. Operational misuse, autonomous failures, and output errors are explicitly excluded from coverage. (Source: OpenAI Services Agreement, January 2026) The Clifford Chance analysis of Agentic AI contracts identifies the structural gap: agreements written for deterministic software tools do not allocate risk appropriately for autonomous agents that take actions rather than execute instructions. When an Agentforce agent autonomously misquotes a price, standard indemnities exclude this scenario. The enterprise absorbs the cost of remediation, the discounts, credits, customer recovery, unless it can demonstrate a specific vendor breach. (Source: Clifford Chance, February 2026)
Every autonomous workflow deployment also introduces a liability exposure that does not appear in typical standard ROI calculation.
The emerging per-resolution pricing models from Sierra, Decagon, Intercom at $0.80 to $0.99 per successful resolution, offer a partial response to this. Vendors paid only on successful outcomes have a financial stake in reducing failures. This acts as a commercial alignment mechanism. However, it is not an indemnification transfer. The liability gap below the waterline remains. But outcome-based pricing at least aligns vendor incentives with enterprise outcomes in a way that pure seat or consumption pricing does not. (Source: Sierra.ai, Decagon, Clifford Chance)
Enterprise AI Business Models: Building Sustainable ROI Beyond AI Adoption
Yet there are captains who are both transformatively and incrementally, navigating the journey even when fog reduces visibility and icebergs reduce manoeuvrability.
JPMorgan Chase (featured in Mindvista Edition 56 for Enterprise wide AI ) spends a substantial $19.8 B in IT, an increase of $2B in 2026 It’s internal LLM used by 150,000 employees to save 4 hours daily per user, though benefits embed into operations rather than direct cost savings.
Jamie Dimon had this to say on JP Morgan’s Investor Day on February 24, 2026,
“We have a LLM and 150K people use it every week. They think they’re saving 4hrs a day. That’s not in a net present value. We don’t see the 4hrs a day in terms of reduced headcount like that. We look at all of it, and it’s deeply embedded in what we do, and that’s true for all technology projects. I think the hardest thing to measure has always been technology projects. That’s been true my whole life. It’s also been true my whole life that technology is what changes everything. Going to mainframes, going to servers, going to speed.”
There are also high performers in the KPMG survey, that represent the 24% of organisations who successfully scale AI across use cases to achieve an average 4.5x ROI, vastly outpacing the industry average of 2x.
As per Deloitte 2026 AI report, what distinguishes these leaders is that they do not simply buy standalone tools or rely on vague, top-down narratives of “10% efficiency gains” across the workforce.
Instead, they actively redesign workflows and scale best practices to prioritise bottom-up unit economics and tangible revenue uplift. How they accomplish this is by applying strict FinOps-style governance to their AI workloads and investing early in data infrastructure to avoid implementation debt. They define KPIs that connect raw compute consumption directly to business value, tracking specific unit-cost-per-activity metrics such as the cost per resolved case, the cost per automated workflow completion, or the cost per successful escalation avoided. Ultimately, these high performers recognise that if they cannot express their AI investments in these specific unit outcomes, they have a spend report rather than a true ROI model.
How Enterprise AI Pricing Models Create Sustainable ROI
The mandate for both buyers and vendors is the pursuit of Sustainable AI. It requires deep vigilance, robust architecture and clarity.
For enterprise leaders the first the bifurcation of AI investment by time horizon is the first discipline.
- Generative AI delivers short-cycle productivity gains that can be measured at the task level and managed within standard fixed and variable cost ROI frameworks.
- Agentic AI is a two-to-four-year process redesign investment, evaluated against operational transformation outcomes rather than monthly efficiency metrics. Running a twelve-month SaaS payback calculation against an architectural transformation produces wrong answers. The model needs to match the investment type.
The second discipline is cost visibility at the unit level. A spending report is not an ROI model. The FinOps framework, cost per resolved case, cost per automated workflow, cost per escalation, makes AI spending governable in a way that aggregate consumption data does not. Organisations that can’t express their AI deployment in unit economic terms can’t manage it.
The third is explicit liability accounting. Every autonomous workflow carries an unmodelled downside exposure. Building that exposure into the ROI calculation, as a statistical risk, not as a vague caveat, changes which deployments get approved and which governance structures get built around them. Vendor per-resolution pricing is a useful signal of where liability is being implicitly priced. It is not a substitute for the enterprise’s own risk modelling.
For vendors
The pricing signal from Veeva is not a curiosity. It is a strategic challenge to every incumbent charging for data access and API consumption. Open, fast, performant data infrastructure, reducing buyer implementation debt is a competitive posture that builds trust. Tollbooth pricing at the data layer creates friction that accumulates into implementation debt for the buyer. The vendors who win the next three years will be those who understand that reducing buyer cost of adoption is their own growth strategy.
Outcome-based pricing, adopted selectively and designed carefully, shifts vendor incentives toward sustained performance rather than initial deployment. It does not resolve the liability gap. But it aligns commercial interests in a way that pure consumption billing does not.
When enterprises ask vendors directly: where are you willing to be paid on outcomes and what happens when it fails, their answer needs to be definitive and accretive.
The Future of Enterprise AI Economics: Sustainable AI Requires Better Pricing Models
To navigate the iceberg of costs and the fog of value, enterprise leaders must return to disciplined seamanship.
They use incrementalism as a Lead Line to sound the depth of each deployment, ensuring initiatives rest on measurable unit economics rather than optimistic projections.
When visibility drops, they sound the Foghorn of transparency, making costs, usage, and outcomes explicit and listening carefully for the echo between planned productivity and realised business impact.
They rely on the Sextant of core business values and long-term strategy, aligning AI investments with durable competitive advantage rather than short term hype cycles.
Through Dead Reckoning, they track their starting baselines of cost structures, process cycle times, error rates, so that every claimed gain can be measured against a known origin. And they steer by True North of leadership commitment.
Sustainable AI demonstrating ROI in 2026 is that crossing. The fog is real. The icebergs are real. The journey is there to be taken.
- What is your experience navigating the fog of AI value and the icebergs of enterprise AI costs ?
- Are you taking the transformative top down view of AI like JPMorgan or ground up build as seen by AI leaders in Deloitte survey?
- Do you think the AI investment cycle by buyers and vendors will sustain on ROI or run on optimism and FOMO?
Love to hear your comments, thoughts and ideas. Explore and stay tuned for more!
Best wishes,
References
Mindvista Editions
External Sources & Key Takeaways
