KPMG AI Leader: Lessons from Accounting for Measuring AI’s Business Value

TLDR

• Companies need a more disciplined way to connect AI costs and consumption with business value, taking a cue from accounting ledgers.
• Outcome-based pricing could gain ground as companies focus less on token prices and more on results.
• Companies are now going past productivity metrics to other measures that tie AI investments more directly to business value.

Enterprises have spent the past several years counting AI users, prompts and tokens. KPMG’s Swami Chandrasekaran says the next challenge is more crucial: quantifying what companies are actually getting for the money.

Chandrasekaran, global head of AI and Data Labs at KPMG, said companies need a more disciplined way to connect AI consumption with business outcomes. His idea is to create what he calls an AI economics ledger.

The idea borrows from accounting systems that track financial resources, but adds something companies have not traditionally had to measure as a separate resource: intelligence.

“We need something similar for the AI economics,” Chandrasekaran told The AI Innovator on the sidelines of the Reuters Momentum AI conference in Austin, Tex. “We can borrow a lot of constructs from accounting and finance, but there are also net new things because intelligence is a new resource now.”

That resource can include tokens purchased from model providers as well as human expertise and other assets involved in producing an outcome, he said. The goal is to account for those inputs and ultimately connect them with business value.

“We need to have something more disciplined so you have a system of record for your value,” Chandrasekaran said.

The problem is becoming more pressing as companies move from experimenting with generative AI toward deploying it in specific business functions. KPMG’s own research found active AI use across finance functions more than doubled from 30% in 2024 to 75% in 2026, while 71% of leaders said AI was meeting or exceeding their ROI expectations in finance.

Chandrasekaran said the ROI conversation inside companies has already shifted.

Earlier, a company might justify an AI investment by noting that employees are completing their tasks faster. He said that productivity metric alone is no longer enough. Companies increasingly want to connect AI to actual business outcomes: Are customers happier? Is revenue or profit increasing? Is AI detecting fraud or financial leakage the company previously missed?

From AI costs to business outcomes

One obstacle is that the bills companies receive from AI suppliers don’t answer those questions.

Cloud and model providers can tell customers how many tokens they consumed and what those tokens cost. Chandrasekaran said enterprises shouldn’t expect providers to determine the business value produced by those tokens.

He compared it with buying ingredients at a retailer. A receipt can show what the cheese, bread and sauce cost, but the retailer can’t tell the customer the value of the meal ultimately prepared with them.

AI providers similarly supply the raw ingredients.

“Somebody needs to translate those tokens into, ‘Where am I using it? For what am I using it? And what is the value I’m going to get?” Chandrasekaran said. “That is going to be the responsibility of the enterprises.”

Chandrasekaran said traditional business goals don’t necessarily change because AI is involved. Companies still care about revenue growth, customer satisfaction, speed and regulatory compliance.

What changes is how AI’s contribution to those outcomes is calculated.

Consider the month-end financial close. Chandrasekaran said AI could move companies closer to what he called a “zero-day close,” in which books could theoretically be closed every day rather than periodically.

The value isn’t merely the reduction in time required for the close, Chandrasekaran said. More frequent financial information could improve forecasting, which could allow a company to spot rising demand in a region sooner, reallocate salespeople and potentially increase sales.

Don’t use a sledgehammer for a nail

Measuring value also requires companies to understand which AI models they actually need.

“The number one literacy (requirement) in the phase we are in right now is to understand what models to use for what workloads,” Chandrasekaran said. That means not using a sledgehammer to drive a small nail.

Enterprises should develop evaluation tests for their use cases and compare different models against them, he said. In many cases, companies don’t need the biggest model available.

That can lower costs, but model selection involves more than price. Companies also need to consider resilience, response time, capacity limits and whether a model behaves consistently enough for a particular business application.

“I would highly recommend everybody gets that literacy (lesson) right away,” Chandrasekaran said.

The same calculation applies to open-weight models.

Companies can run those models on their own infrastructure, which can be useful when they face data-sovereignty requirements, need AI operating at the edge or want greater control over customization. But they then have to account for the compute infrastructure, expertise and operating costs required to run them.

Chandrasekaran also distinguished open-weight models from fully open-source models. Open-weight models provide model weights but typically don’t provide all of the training data and training pipeline. Open-weight models are more popular, especially for enterprise uses.

Token prices aren’t the whole cost story

AI model prices have been declining, but Chandrasekaran said enterprises increasingly look beyond the headline price per token.

A newer model, for example, might use substantially more tokens to perform the same task because it conducts more reasoning. Even if the unit price falls, greater consumption can offset some of those savings.

Clients are therefore asking not only what each token costs but why a task requires a particular number of tokens and whether they can control reasoning or “thinking” levels to manage consumption.

“They’re double-clicking into how many tokens you’re consuming, how long some of these things are taking to process,” Chandrasekaran said. “I don’t want to end up paying (more) because the multipliers add up.”

That scrutiny could eventually change how AI itself is sold.

Chandrasekaran said several pricing models already coexist. Some AI tools charge per user per month. Model providers sell token consumption directly. Other products combine subscriptions with usage-based overages.

A fourth approach charges for outcomes. Some customer-service providers, for example, charge for a successfully completed call rather than the tokens or computing resources consumed to produce it.

Chandrasekaran expects that approach to gain ground.

Paying for outcomes is “getting more traction,” he said. “If I have to place a bet, I’m going to say clients are looking to buy outcomes.”

Outcome pricing effectively shifts part of the AI economics problem from the buyer to the vendor. Instead of the enterprise trying to calculate the value of every token consumed, a supplier could agree to deliver a specified business result at an agreed price.

That doesn’t mean consumption-based pricing will disappear. Chandrasekaran expects multiple models to coexist and said dynamic pricing could emerge as the market develops.

But the direction is toward making AI spending easier to connect with what businesses actually care about.

“Pricing is ripe for disruption,” Chandrasekaran said.