AI agent economics to shape next phase of enterprise GenAI adoption; 60% of agentic AI costs go to response refinement: McKinsey
New Delhi, July 17
As generative artificial intelligence moves beyond experimentation into enterprise-scale deployment, business leaders are increasingly grappling with the economics of AI agents rather than the technology itself, according to a new McKinsey report.
The report said that while the first two years of Gen AI adoption were largely centred on access, experimentation and deployment, the next phase will be defined by financial sustainability and return on investment as organisations scale agentic AI systems.
"For the first two years of gen AI adoption, most enterprises focused on access, experimentation, and deployment. As agents move into production, a different set of questions is emerging... The decision to scale an agent is increasingly becoming a complex and fast-changing economics decision, not a technical one," the report said.
According to the report, enterprises are now shifting their attention from reducing AI costs to demonstrating measurable business value, with chief financial officers (CFOs) and chief information officers (CIOs) increasingly demanding evidence that AI investments are delivering tangible returns.
It says, despite the growing adoption of agentic AI, many companies lack systems to track the business impact of AI-driven decisions, while nearly 60 per cent of the technology's operating costs are spent on verifying and refining responses.
The report identified six major factors that drive operating expenditure in agentic AI systems. It noted that long-lived context is one of the biggest contributors to costs.
Noting Agentic AI tasks can use nearly 1,000 times more tokens than conventional code reasoning or chat-based AI tasks, the report said, "Per-token pricing has stopped being a useful measure for what enterprises actually pay for gen AI."
Furthermore, context accounts for a significant portion of agentic costs, it said.
Secondly, modifying the AI-generated responses is even more expensive than generating the initial answer. "About 60 per cent an agentic task's costs, in fact, are tied to refining answers," it noted.
The report also highlighted that autonomous AI systems create cost variability, as the same task can incur different expenses depending on the tools used, the reasoning path taken and the number of retries required.
Another major cost driver is the use of advanced reasoning for simple tasks. While extended reasoning delivers value for complex problems, it adds unnecessary computing costs when applied to routine or straightforward tasks.
McKinsey further noted that agent orchestration--how AI agents coordinate with tools, models and one another--can significantly influence costs. Efficient task allocation and coordination can reduce expenses substantially without affecting business outcomes.
Finally, the report said information structure plays a crucial role in AI operating costs. Factors such as prompt design, context length, formatting, language and data structure influence token consumption. "Non-English text, for example, gets fragmented into more tokens per meaning, so the same conversation costs more in some languages than others," it added.
— ANI
Reader Comments
As someone working in a Bengaluru startup deploying AI agents for customer support, this hits home. We've noticed that 60% cost going to refinement is accurate—our teams spend more time testing and fine-tuning responses than building models. The cost variability from autonomous systems is another headache. Indian companies need to plan for these hidden costs.
Finally someone is talking about actual economics instead of just generating ChatGPT demos. The point about "advanced reasoning for simple tasks" is crucial—in many Indian enterprises, we're seeing over-engineered AI solutions that burn money unnecessarily. Need more focus on efficient orchestration and right-sizing models for tasks.
As a CIO in a Mumbai-based financial firm, I appreciate this report. We've been struggling to get our CFO to sign off on more AI investments because of unclear ROI tracking. The report's mention of "long-lived context" costing more is something we've experienced—our customer service agents waste tokens on irrelevant context. Important wake-up call for Indian enterprises.
The bit about non-English text costing more is interesting for India. With multilingual AI agents, companies using Hindi, Tamil or other Indian languages will face higher costs due to token fragmentation. This is a barrier for inclusive AI adoption in our country. Need more efficient tokenization methods for Indian languages.
Good reality check from McKinsey. The fact that 60% costs go to verification/refinement shows we're still in early days of making AI truly autonomous. I'd add that in Indian context, the cost of data labeling and human-in-the-loop systems are also significant hidden costs not discussed here
We welcome thoughtful discussions from our readers. Please keep comments respectful and on-topic.