The GenAI Divide.
MIT Project NANDA's *State of AI in Business 2025* is the cleanest read of why most enterprise GenAI investment is stuck - and what the 5% who broke through are doing differently. This page is our reading of the study, organised around the questions our clients keep asking us. Numbers and direct claims are the report's; framing and emphasis are ours.
Last updated:
95% of enterprises are getting zero return on their GenAI investment.
Despite $30-40B in enterprise GenAI investment, 95% of organisations are getting zero return. The split between the 5% extracting millions and the 95% stuck in pilot purgatory is what the authors call the GenAI Divide - and it is driven by approach, not model quality, regulation, or talent.
That last sentence is the whole study, compressed.
Eight numbers that frame the divide.
80%+ piloted, ~40% deployed
Of organisations have explored or piloted ChatGPT or Copilot. Roughly 40% have deployed at least one general LLM tool company-wide.
60% → 20% → 5%
For task-specific or custom enterprise GenAI - the kind that handles real workflows - 60% evaluated, 20% piloted, only 5% reached production. General LLMs reach ~83% pilot-to-implementation; custom internal tools collapse far earlier.
90% / 40% - the shadow AI economy
90% of employees use personal AI tools at work. Only 40% of companies pay for an official LLM subscription.
External partners win 2-to-1
External vendor partnerships reach production about twice as often as internal builds - 67% vs 33%.
$1.2M ARR in 6-12 months
Top GenAI startups crossing the divide hit $1.2M ARR in 6-12 months. Speed of value delivery is correlated with the gap.
9-to-1 vs 70/30
On complex, long-term tasks, humans beat AI 9-to-1. On quick, well-bounded tasks, AI wins 70/30. The implication is the architecture, not the model.
50-70% of budgets, wrong place
About 50-70% of GenAI budgets go to Sales & Marketing - visible KPIs, easy reporting. The biggest documented ROI lives elsewhere, in the back office.
State of AI in Business 2025 - "The GenAI Divide"
Aditya Challapally · Chris Pease · Ramesh Raskar · Pradyumna Chari (MIT Project NANDA). Methodology: 300+ public AI initiatives reviewed · 52 structured organisation interviews · 153 senior-leader surveys (Jan-Jun 2025).
The #1 blocker is not what most teams think it is.
It is not infrastructure, regulation, or talent. It is that GenAI tools do not remember, do not adapt, and do not learn. Users love ChatGPT for ad-hoc work; they reject it (and custom internal tools) for mission-critical work because they reset every session.
- Model quality - "GPT-4 is not smart enough for our domain"
- Regulation - "we cannot ship until legal signs off"
- Talent - "we cannot hire AI engineers"
- Data - "our data is too messy"
- Tooling - "the platform we picked is wrong"
- GenAI tools that do not remember Tuesday on Wednesday
- Architectures that reset every session - no accumulating context
- No fit to the workflow the team actually does
- Internal builds without the iteration loops vendors run by default
- Investment in demos, not in deep workflow integration
Five moves the 5% have in common.
Buy, do not build
External vendor partnerships succeed at twice the rate of internal builds. Custom internal tools collapse on the pilot-to-production gap.
Treat vendors like BPOs, not SaaS
Demand customisation. Hold to business outcomes, not feature checklists. Co-evolve through real failures.
Source ideas from the frontline
From the prosumers already using ChatGPT for their own work - not from a central AI lab three layers up.
Decentralise implementation, centralise accountability
Line managers own delivery; one accountable owner sets the bar.
Land in narrow, high-value workflows first
Voice and call summarisation. Document automation. Code generation. Then expand. The teams that try to "transform the company" go nowhere.
The study describes exactly the pattern our agents are built around.
Memory. Our agents read the same context every call, write to a daily log, and keep a vector index over past conversations. Tuesday's agent remembers Monday. Continuous improvement. Every recurring task accumulates a living lessons file; consecutive runs are materially better than the previous. Deep workflow integration. Typed, role-aware tool wrappers around your real systems - not a chat layer bolted on top. Vendor partnership. Discovery is paid because we put our best person on it; we co-evolve through real failures, with weekly checkpoints and an open kill-switch. If your team is inside a pilot that has gone quiet, the diagnostic call walks through what the study says you are missing - and how we close that gap.
What the data actually shows.
Is AI about to replace most jobs soon?
Is GenAI transforming business?
Are enterprises slow adopters?
Is the blocker model quality, legal, or data?
Should the best enterprises build their own?
This is what we build.
Our agents have memory, learn from each run, and integrate at the system level. If you are inside a pilot that has gone quiet, the diagnostic call walks through what the study says you are missing - and how we close that gap. Free; written summary either way.