Who is "The Evidence Base: Why Most AI Initiatives Fail" for?

This Expert module in AI Strategy & Implementation is for operators, founders and growth teams who run AI in real work and want production-grade judgement. It takes about 9 minutes and ends with a 4-question quiz.

No. This module is about decisions, economics and implementation planning, so no coding is involved.

What comes before and after this module?

"The Evidence Base: Why Most AI Initiatives Fail" is the first module of AI Strategy & Implementation, so it has no prerequisites. Afterwards, continue with "Competing in the Age of AI".

The Evidence Base: Why Most AI Initiatives Fail

The failure rate is not a technology story. Every major study lands on the same cause: organizations that don't learn.

By Ronan Liedmeier, fractional head of growth and course author · Reviewed 13 October 2026

The big idea

MIT's NANDA report gave it a name: 'adoption without transformation.'

Generic AI assistants do boost individual productivity — that part works. What stalls is the enterprise-grade stuff: the big sanctioned deployments, the flagship initiatives, the transformation programmes.

The root cause isn't the technology. It's the learning gap. Organizations deploy tools that don't retain feedback, don't adapt to context, and don't improve over time — into workflows that were never redesigned to take advantage of them.

Static tool. Static org. No compounding. That's the 95% failure story in one sentence.

Numbers that matter

95% of enterprise GenAI pilots delivered no measurable P&L impact, despite $30–40B invested (MIT NANDA, 2025)
~10% of companies achieve significant financial benefits from AI (MIT SMR × BCG)
47% of organizations say they have no strategy for AI (MIT SMR × BCG, 2025 (n=2,102))
~67% vs ~33% deployment rate: external vendor partnerships vs internal builds (MIT NANDA)
~90% of workers use personal AI tools regularly while only ~40% of firms have enterprise subscriptions (MIT NANDA — the shadow-AI economy)

Deep dive

The four structural findings worth memorizing

Four findings from the research that are worth having in your head before every client conversation.

**1. Buy beats build, roughly 2:1.** External partnerships with specialised vendors reached deployment about 67% of the time. Internal builds: about 33%. 'We have engineers' is not, by itself, a business case for building.

**2. Budgets and ROI live in different departments.** AI spend concentrates in sales and marketing — the visible, glamorous use cases. Measured returns concentrate in back-office operations: document processing, compliance, finance. Money follows visibility; value follows friction. That gap is your biggest opportunity when scoping.

**3. The shadow-AI economy is real.** Around 90% of workers use personal AI tools regularly. Around 40% of firms have enterprise subscriptions. Employees are extracting value bottom-up while the sanctioned pilot stalls in procurement. That's not a compliance problem to lock down — it's validated demand routing around a broken official channel.

**4. Empowered line managers beat central AI labs.** The organisations that succeeded pushed use-case selection and iteration to the people who actually own the process — not to a central AI team picking from the top.

Consulting move

Open every discovery conversation with these four findings — buy vs build, front vs back office, shadow AI, line-manager empowerment.

This does something important: it reframes the entire conversation. Instead of 'what AI should we buy?' (a shopping question), you're now asking 'which of these four failure patterns is this organisation currently running?' (a diagnostic question).

That's the difference between a vendor and a consultant.

How to run it

Cite the headline figure with its caveat. The 95% 'zero return' figure rests on a small executive-interview base the report itself calls directionally accurate — cite it that way, not as a precise statistic.
Check budget-vs-value mismatch first. Where is AI spend concentrated (sales/marketing) versus where measured returns actually cluster (back-office operations)? Name the gap explicitly.
Look for the shadow-AI signal. ~90% personal AI usage vs ~40% enterprise subscriptions is validated demand, not a compliance problem — check for it before assuming low appetite.
Test the build-vs-buy justification. Is 'we have engineers' doing the entire justification for an internal build? That alone is a red flag given the ~2:1 deployment gap favoring vendor partnerships.
Identify who picks use cases. A central AI lab choosing use cases top-down underperforms empowered line managers and process owners — note who currently holds that decision.

In the field

A regional bank's flagship 'AI transformation' launched a chatbot for wealth clients — glamorous, budgeted, publicised. Twelve months in: no measurable P&L impact. Meanwhile a compliance officer had wired ChatGPT into her draft-review process on personal subscription and cut cycle time by ~40%. The 5%-vs-95% picture in one org: budget followed visibility; value followed friction — and the sanctioned channel missed the demand routing around it.

When not to use it

Cite the 95% figure honestly — the 'zero return' finding rests on a small executive-interview base the report itself describes as directionally accurate, and the synthesis method for the 300 deployments has been criticized. The direction of the finding replicates across other studies; the exact number is one signal among several. Cite it with the caveat and you look serious rather than credulous.

Pitfalls and takeaways

Chasing the shiny front-office use case while measurable ROI clusters in back-office friction workflows.
Building where the data says buy — internal builds fail at roughly twice the rate of vendor partnerships.
Blocking the shadow-AI economy instead of channelling its demonstrated demand.
Central AI lab picking use cases the process owners never asked for; empowered line managers outperform.

95% failure has structural causes, not model causes.
The winners run a learning loop; the losers ship static tools.
Budget flows to visibility; value hides in back-office friction.
Cite research with caveats — credibility depends on it.

Do the work

Goal: Diagnose a real or hypothetical organization against the four MIT NANDA failure patterns and produce a discovery brief that names which pattern applies, with cited evidence. (60 min)

Pick a real or realistic target organization. Choose a company you know well enough to describe its current AI initiatives, budget allocation, and where employees already use AI tools unofficially.
Map their current AI spend by department. List where AI budget is currently allocated (sales/marketing vs back-office) and note whether this matches the 'glamour bias' pattern from the research.
Identify evidence of the shadow-AI economy. Find or infer where employees are already using personal AI tools unofficially to solve problems the sanctioned pilot hasn't addressed.
Check the build-vs-buy posture. Determine whether the organization is building internally or buying from a specialized vendor, and whether that choice is justified by the 2:1 buy/build evidence.
Score ownership: central AI lab vs empowered process owners. Determine who is picking AI use cases — a central team, or the people who actually run the workflow.
Write the one-page discovery brief. Name which of the four failure patterns (glamour bias, build-over-buy, shadow-AI suppression, central-lab ownership) is most active, cite the relevant MIT NANDA statistic, and propose the first diagnostic question you'd ask in discovery.

Deliverable: A one-page discovery brief naming which of the four MIT NANDA failure patterns is most active in the target organization, with evidence cited and one opening discovery question proposed.

Sources

MIT NANDA, 'The GenAI Divide: State of AI in Business 2025'
MIT Sloan Management Review × BCG, ninth annual global executive study (2025)
Duke Fuqua / Federal Reserve CFO Survey