The AI gold rush has entered its accountability era

Enterprise AI is no longer a single software purchase. It is a continuously expanding consumption ecosystem spanning embedded assistants, standalone licences, APIs and tokens, application-specific models, retrieval services, cloud and GPU capacity, and autonomous agents—each acquired, metered and governed differently, often by different teams.

Gartner forecasts worldwide AI spending across services, software, models, data, cybersecurity and infrastructure to total $2.59 trillion in 2026, a 47% increase from 2025. It separately forecasts worldwide end-user spending on AI models and platforms to reach about $64 billion in 2026, up 63.4%, with specialised and domain-specific generative AI models growing 210%. Enterprise AI budgets are under greater scrutiny, with increased focus on usage efficiency, cost control and measurable outcomes.

The challenge for organisations is therefore no longer simply how to provide access to AI. It is how to provide the right AI capability to the right user and workload, at the right level of cost and control.

Access has outpaced accountability

The market is moving from “How do we encourage the use of AI?” to “Can we explain what we are paying for, who is using it, what it is doing, and whether it is creating value?” Access is an entitlement. Value is an outcome.

1. Spend visibility gap: Licences, APIs, tokens, models and agents often lack one consumption and cost baseline.

2. Fragmented AI portfolio: Embedded features and standalone tools overlap, while inactive entitlements and shadow AI remain hard to detect.

3. Architecture cost drag: Premium models handle routine work that capable models, small task-specific models or deterministic rules could serve.

4. Demand and governance gap: Requests are framed around vendors and SKUs rather than business need; lifecycle ownership and agent controls are inconsistent.

The cheapest option can be the most expensive decision

The lowest-priced licence, model, API or infrastructure option is not necessarily the most economical. Apparent technology savings can be offset by inaccurate or lower-quality outputs, longer response times, additional human review and rework, poor adoption, security or privacy exposure, regulatory risk, reduced scalability, interruption to important processes, or failure to achieve the intended business outcome.

Organisations therefore need to assess the total value of each workload across cost and consumption, quality and accuracy, performance and latency, scalability, security and privacy, risk and regulatory requirements, user adoption, productivity improvement and business outcomes. The right unit of economics is not cost per token. It is cost per trusted business outcome.

Six avenues for AI consumption rationalisation

1. Baseline AI consumption: Build a joined inventory of AI applications, embedded features, licences, models, APIs, tokens, infrastructure and agents. Link each item to its owner, users, workload, commercial model, data, risk and intended outcome.

2. Provision by persona and lifecycle: Allocate licences and credits by role, use case, frequency, context need and data sensitivity. Use telemetry, inactivity warnings and reclaim rules so premium access does not become a permanent entitlement.

3. Right-size models and route workloads: Use deterministic rules for fixed logic, capable models for repeatable high-volume work and frontier reasoning for complex or high-risk tasks. Route by complexity, risk, required accuracy and business criticality.

4. Rationalise prompts, context and compute: Retrieve only relevant information, compress repeated context, standardize outputs, cache reusable instructions, batch non-urgent work and right-size runtime or GPU capacity.

5. Rationalise the portfolio and commercial model: Identify duplicate tools and fragmented use cases. Compare per-user, consumption-based, enterprise commitment and hosted options using total workload economics—not headline price.

6. Govern agents and measure outcomes: Maintain an agent registry; control data, tool and API access; set budgets and thresholds; monitor traces; preserve human approval and rollback where needed; measure cost per accepted output or completed process.

The goal is not less AI. It is better-matched AI

That means providing the right AI capability to the right user and workload, through the right commercial and architectural model, with the right controls.

· Monitor: Inventory AI tools, applications, models, licenses, APIs, token flows, infrastructure and agents. Establish baselines and telemetry, create dashboards, and identify unused entitlements, duplicate tools, anomalous usage and cost hotspots.

· Rationalise: Align access to personas and use cases; assess model and architecture fit; rationalize prompts, context and compute; consolidate overlapping capabilities; and evaluate commercial options against workload value.

· Govern: Centralise demand intake, apply allocation and lifecycle rules, establish agent registries and risk-based controls, set usage and budget guardrails, strengthen cost attribution, and conduct recurring value reviews.

Together, this approach would create visibility, cost efficiency, performance, scalability, governance at scale.

(Ramchandran is Risk Markets Leader, EY India; Kochhar is Director – Digital Risk Consulting, EY India. Views are personal.)