Field Report·MIT NANDA “GenAI Divide”·State of AI in Business 2025

Everyone adopted AI. Almost nobody got paid.

MIT’s Project NANDA studied 300+ enterprise AI deployments and interviewed 150 leaders. The verdict: adoption is everywhere, returns are almost nowhere, and the money that did come back came from the least glamorous place in the building. The back office.

$30–40B Poured into enterprise GenAI initiatives
95% Of organizations report no measurable P&L impact*
5% Of custom AI tools ever reach production
90% Of firms have employees using unofficial “shadow AI”

01The money went to the wrong floor.

Asked to budget for AI, leaders steered 50–70% of spend toward sales and marketing, the demos that look good in board decks. The provable returns showed up in operations, where the savings are already a line item someone signs every month.

50–70% BUDGET → SALES & MKTG THE REST → BACK OFFICE (WHERE ROI LIVES) 33% BUILT IN-HOUSE: DEPLOYED 66% BUILT WITH A PARTNER: DEPLOYED
Left: where budgets go vs. where returns came from. Right: partnered tools reached deployment at twice the rate of internal builds. Source: MIT NANDA, State of AI in Business 2025.

02Where the real ROI lived.

The winners didn’t fire their people. They fired their invoices. The returns came almost entirely from cutting external spend on work AI now does in-house.

Customer service · Document processing

$2–10M / yrBPO contracts eliminated

Voice agents and document automation replaced outsourced intake, logging, and processing of contracts, invoices, and forms.

External spend, deleted
Creative · Content

−30%External agency costs

Content and creative work moved in-house on AI tooling. The agency retainer became the savings line.

External spend, deleted
Financial services · Compliance

~$1M / yrOutsourced risk checks

Risk and compliance reviews previously farmed out to consultants, now run internally with AI assistance.

External spend, deleted

03Why the other 95% stalled.

Not the models. Not regulation. Not talent. MIT’s diagnosis is a learning gap: tools that can’t remember, deployed into workflows they don’t fit, by buyers shopping for demos.

No memory

Most tools retain no feedback and no context. Great in a demo, useless by week three when the workflow demands what it learned last month.

Learning gap

Brittle fit

Generic chatbots hit 83% adoption on trivial tasks, then shatter on the exceptions and edge cases that make up real operations.

Workflow mismatch

Wrong buyer

Central AI labs shopping like SaaS customers. The 5% acted like BPO clients instead, demanding customization and operational metrics.

Approach failure

Too big, too slow

Mid-market firms went pilot-to-production in ~90 days. Enterprises took nine months or more, and converted the fewest pilots.

90 days vs 9 months
  • Quick tasks: 70% prefer AI
  • High-stakes work: 90% still want humans
  • The dividing line: memory, not intelligence

The winners picked one pain point and started in the back office.

Narrow scope. A partner instead of a lab. Tools that learn the workflow and run where the work lives. Savings you can point to on an invoice, not a dashboard. That’s the 5%, and none of it requires an enterprise budget to copy.

*Source: “The GenAI Divide: State of AI in Business 2025,” MIT Project NANDA (July 2025): 300+ deployment reviews, 150 leader interviews, 350 employee surveys. Released as preliminary findings; the 95% figure has been debated, the back-office ROI pattern is corroborated across independent studies.

The 5% playbook is the only one we run: one pain point, back office first, tools that fit the workflow, savings you can point to on an invoice.

Start with a Recon