Only 23% of Enterprises See Sustained Value From AI. Accenture's Fix Is Treating Tokens Like a P&L Line.

Only 23% of C-suite leaders report sustained business value from their AI investments, according to an Accenture Generative AI Insights piece authored by Chief Strategy and Services Officer Manish Sharma, Chief AI and Data Officer Lan Guan, and Principal Director Surya Mukherjee. Their argument: as Goldman Sachs’ $800 billion 2026 AI-spending projection plays out, token consumption is becoming the primary cost and governance line item, and most enterprises aren’t managing it as one. Their framing, in their own words: “The question isn’t how much AI costs. It’s whether it’s returning enough to matter.”

The waste is structural, not incidental. Accenture estimates only 10-20% of enterprise tasks are complex enough to justify a frontier model — implying the bulk of enterprise AI work could run on cheaper models — and cites one telecom operator that cut annual AI spend 68% purely through routing optimization. Enterprise token consumption is projected to reach 120 quadrillion tokens a month by 2030, a 24x increase from today, and only 15% of organizations say they’d bank the savings if token prices fell 25% — most would just reinvest it into more usage, meaning cost discipline has to be designed in now rather than assumed later. Accenture’s four disciplines: make AI economics visible to leadership, govern before scaling entrenches bad habits, route work to the appropriate intelligence level, and treat token economics as continuous management rather than a one-time cost review.

BCG’s earlier “Return on AI” research supplies the sharper mechanism for why this compounds. Per-token cost between a simple model and a frontier model can run 5 to 25 times higher, and because agentic systems resubmit growing context on every loop, cumulative token cost rises roughly with the square of session length — a session twice as long can cost four times as much. BCG also cites IDC’s estimate that the top 1,000 global companies will underestimate AI infrastructure costs by up to 30% through 2027.

Put together, the two pieces make the same case from different angles: the AI spending conversation is shifting from “how much” to “return on how much,” and the enterprises that win won’t be the ones with the smallest or largest bill — they’ll be the ones who can show the math either way.