AI’s Biggest Opportunity Lies In Paying For Outcomes, Not Tokens

Enterprises are rationing AI because they are paying for effort instead of outcomes. The startups that build the outcome-pricing layer, particularly in India’s inference stack, can own the next shift in AI adoption.

AI today is sold and priced by the units of effort the model expends (tokens) rather than the outcome the user receives. This category error is now producing visible dislocations of expectations at enterprises across the world.

If the largest AI companies, on the path to their IPOs, are claiming that their agents will replace the future human workforce, this antinomy will need to be solved sooner than later.

Speaking at the intelligence at work event in late May,  OpenAI CEO Sam Altman conceded that token cost has become “a huge issue” within a single quarter. Customers are telling him:

“My company spent my entire 2026 budget in Q1. Can you make this more efficient?” Altman noted that cost concerns are now the second-most common complaint he hears from enterprise customers, behind only workflow simplification.

Goldman Sachs projects that agentic AI could push token consumption 24X by 2030, to 120 quadrillion tokens per month. The meter is not slowing down.

Below the headlines, the pattern repeats at a smaller scale. An operator running a dozen AI automations recently described budgeting for tooling and integration but never modelling token usage. Three weeks in, the question had moved from “is this saving us time” to “is this actually cheaper than hiring someone.”

Humans Are Paid For Outcomes. AI Is Still Paid For Effort.

As a founder, when you hire a contract lawyer to draft a sales agreement, you pay a flat fee for a clean, enforceable document. You do not pay extra because the lawyer re-read the Companies Act three times before drafting. The effort is the lawyer’s problem. The outcome is what you bought.

This is the structure of the most mature professional services market. Architects price by project. Surgeons price by procedure. Investment bankers price by transaction. The seller absorbs effort variance because the seller is closest to it and can manage it.

Token billing inverts this. The buyer pays for the AI’s deliberation, retries, verbosity, memory reconstructions, and routing inefficiencies. Every “thinking” token is on your invoice whether or not it produced anything useful.

The seller has no incentive to be efficient. In some cases, the seller has an incentive to be verbose. Instead of checking bugs and dependencies on just new code, the AI will read broadly across your entire codebase, pulling in files, dependencies, and historical context until restricted.

Effort-based pricing is what every immature services market starts with, because it is the easiest thing to meter. Outcome pricing arrives when the seller has enough confidence in delivery to absorb variance. AI is not there yet. The market is now forcing it there.

The Major Differences In Token Billing

Token pricing has three structural differences that SaaS and cloud never had.

• First, no predictable unit. SaaS had seats. Cloud had compute hours. Tokens have neither a stable cost curve nor a predictable consumption pattern. A single prompt change can triple a bill overnight.
• Second, no departmental ownership. AI usage crosses every function. Finance cannot allocate the cost to a single P&L line. The 21-year-old in FP&A who could explain every line item in 2024 struggles to explain the Anthropic invoice in 2026.
• Third, no shared language of value. A sales team has qualified leads. A support team has resolved tickets. A token tells you the meter ran. It does not tell you whether the work was worth doing or that a goal was achieved.

Uber is a working demonstration of this — 95% of Uber engineers use AI tools monthly and  70% of committed code is now AI-generated. Uber’s COO, Andrew Macdonald, has publicly admitted he cannot draw a clear line between this spend and measurable improvements in consumer products. High internal adoption, but no attribution to outcomes.

The ramp is counter-proof. High AI-intensity firms have doubled revenue since 2023. Low-intensity firms went flat. The answer is not to stop spending. The answer is to spend with the same rigour companies apply to headcount and vendors. That rigour requires outcome metrics. Token billing does not provide them.

Enterprises Are Choosing Control Over Choice

The rationing era has started, and the names involved are not small.

In April, Uber CTO Praveen Neppalli Naga publicly disclosed that the company had burned through its entire 2026 AI coding budget in four months. Claude Code adoption jumped from 32% to 84% of Uber’s 5,000-engineer organisation in three months.

Individual engineer bills ranged from $500 to $2,000 per month. Naga himself burned $1,200 in tokens during a two-hour internal demo. By May, Uber imposed a $1,500 monthly cap per engineer on agentic coding tools.

In May, Microsoft cancelled most Claude Code licenses across its experiences and devices division, the team behind Windows, Microsoft 365, Outlook, Teams, and Surface. Thousands of engineers were redirected to GitHub Copilot CLI by June 30.

The official reason was tooling integration. The timing, aligned with Microsoft’s fiscal year end, suggests cost was equally load-bearing. Around the same time, GitHub Copilot itself switched from a flat monthly fee to token-based billing. Users reportedly ran out of credits the same morning.

This is the same arc every enterprise IT category has followed. Shadow IT becomes governed IT. SaaS sprawl becomes vendor consolidation. AI is now entering its consolidation phase, three years in rather than ten. 

Founders should expect their enterprise customers, within the next four quarters, to stop asking “what does your AI do” and start asking “what does it cost per outcome delivered.” The companies that can answer that question will close deals. The companies that quote per-token rates may not.

The Outcome-Pricing Layer Is Already Being Built

This is a present-tense market shift. The largest enterprise vendors are already moving.

Salesforce Agentforce now prices AI-driven actions through agentic work units, consumption credits tied to workflow execution and case resolution. Zendesk has explicitly committed to outcome-based pricing for its AI agents, with CEO Tom Eggemeier framing it directly: as AI agents take on more service workload, customers should expect pricing to reflect outcomes delivered. 

Sierra, founded by Bret Taylor, charges per resolved customer support conversation, with no charge if the conversation is unresolved. Fini charges $0.69 per AI-resolved ticket. 

ServiceNow, UiPath, and Adobe are all moving to hybrid models combining subscription, consumption, and outcome components.

The shift is not trivial. It requires three pieces of infrastructure that mostly do not exist yet. 

• Outcome definition, which is the canonical “filed return” equivalent for a sales agent, support agent, or code agent. 
• Outcome verification, which determines whether the agent actually resolved the ticket or just closed it.
• Risk absorption, which means the seller takes on token variance, requiring either intelligent model routing or margin buffers thick enough to absorb the worst-case run.

The Inference Stack Is Where Outcome Economics Are Won

Outcome pricing requires control over the cost-per-outcome math. That control sits in the inference layer, not the application layer. Indian infrastructure companies are now operating at this frontier.

According to an Indian AI cloud provider, the company is currently processing 40 Bn tokens daily on its inference platform. Indian enterprises bleeding on token costs are doing so because they are running production workloads on shared, frontier-model infrastructure designed for experimentation.

By moving a production workload from frontier APIs to a specialised inference stack running open-weights models, an Indian enterprise can compress its monthly AI bill from $75,000 to $9,000. That is an 88% cost reduction without a corresponding drop in capability for the specific task.

The directional reading is straightforward. Not every task needs a frontier model. A meeting summary does not need GPT-5.5 Pro. Every code review does not need Claude Opus 4.8. Intelligent model routing alone can compress AI bills by an order of magnitude, and that compression is the precondition for outcome-based pricing to work at the seller’s margin. 

The global inference market is projected to overtake training spend by 2027. India is already the second-largest inference market in the world. There are wedges of opportunities opening as this shift becomes mainstream.

India’s Innovation Opportunities

The Indian SaaS playbook of the last decade was built on better unit economics under competitive pricing. Lower cost of engineering, sharper capital discipline, and an operator culture that defaults to profitability earlier in the lifecycle. The same advantage can apply to the outcome-pricing transition. 

An Indian vertical AI company that can price by outcome, route intelligently across models, and absorb token variance will out-compete a global peer still quoting per-token USD rates. The Indian founder is trained to think this way. The Indian customer expects it.

Reporting from May 2026 confirms Indian AI startups are actively shifting from seat-based to pay-as-you-go and outcome-linked models, with customers pushing founders to adapt. The GCC base provides scale. India now hosts 2,117 Global Capability Centres employing 2.36 Mn people and generating $98.4 Bn in revenue. More than 1,200 GCCs have already embedded AI and ML capabilities, supported by an AI talent pool of approximately 250,000 professionals. 

These are the buyers, and they are facing the same token bill problem as Uber and Microsoft, with the added pressure of data privacy and compliances that makes every cross-border prompt an emerging risk.

Specific category evolutions are worth exploring: vertical agents in legal, accounting, customer support, performance marketing automation, asset generation, system design, design verification, and code review. Each has a clean outcome unit and a buyer already accustomed to outcome-based vendor pricing.

Tokens are now an emerging pillar of business cost, alongside people and vendors. They are not going away. But the meter being used to bill them was borrowed from the cloud era and is broken by AI economics. 

The winners of the next cycle will not be the companies with the cheapest tokens or the biggest models. They will be the companies, on both the buy side and the sell side, that figure out how to price outcomes.

The post AI’s Biggest Opportunity Lies In Paying For Outcomes, Not Tokens appeared first on Inc42 Media.