It's Not the AI, It's the Last Mile.

BLOG 2 OF 5

Last week I shared the numbers. $37 billion spent [1]. 88% of enterprises using AI [2]. 6% getting value [2]. 95% of pilots delivering zero return [3]. The gap between adoption and impact has never been wider.

The natural instinct is to blame the technology. The model hallucinated. The accuracy wasn't good enough. We need a better algorithm.

Sometimes those are real problems. But every serious study from the past year converges on the same conclusion: technology is rarely the primary cause of enterprise AI failure.

The RAND Corporation found that AI projects fail at more than twice the rate of non-AI IT projects [12] — and the root causes are organisational, not technical. Across McKinsey, MIT, BCG, Informatica, and WorkOS, five failure modes appear with depressing consistency. Here they are.

1. Pilot Purgatory

S&P Global found that 46% of enterprise AI proofs of concept are abandoned before reaching production [5]. Not delayed. Abandoned. The demo was impressive. Leadership got excited. And then nobody designed a path from sandbox to real operations. Nobody addressed integration with existing systems, compliance requirements, user training, or the seventeen legacy systems that actually run the business.

The PoC sits there — impressive and useless — while the team moves on to the next demo.

2. The Bolt-On Trap

This is the big one. McKinsey tested 25 organisational attributes against actual EBIT impact from AI. The winner — above talent, technology, budget, and executive sponsorship — was workflow redesign [2].

Organisations that layer AI onto existing processes, without redesigning how work flows, consistently underperform. You can't bolt a probabilistic, generative technology onto a deterministic, rule-bound process and expect transformation. You get friction. You get workarounds. You get employees routing around the AI because it doesn't fit how they actually work. And then you get a board presentation about how "AI didn't deliver."

The AI delivered just fine. The process wasn't ready to receive it.

3. Data Unreadiness

Informatica's 2025 survey of 600 data leaders found that 43% cite data quality as the number-one obstacle to moving AI into production [6]. Two-thirds of organisations can't transition even half their pilots because the underlying data isn't trustworthy, accessible, or governed. And 97% struggle to demonstrate business value — not because the AI doesn't work, but because the data feeding it is incomplete, inconsistent, or inaccessible.

Enterprises have vast data. But it's scattered across dozens of systems, poorly governed, and formatted differently everywhere. One CIO compared it to Lake Michigan: enormous volume, none of it drinkable without serious filtration.

4. Disconnected Teams

Product builds AI features. Infrastructure manages deployment. Data curates datasets. Compliance writes policies. Legal drafts governance. And none of them share success metrics, timelines, or accountability [14]. The AI project that needs simultaneous sign-off from data, security, compliance, and the business unit gets stuck in a coordination death spiral that no engineering talent can fix.

5. The Integration Tax

The quiet killer. Analysis shows that 60% of total AI development time is consumed by integration work [25] — connecting to existing systems, managing APIs, ensuring data flows correctly, maintaining security. This is not the exciting part. This is not what anyone pitched in the board deck. But it's where projects go to die.

The pattern is clear. None of these are model problems. They're plumbing problems. The AI works in the lab. It fails where the lab meets the real world — where models have to connect to workflows, systems, compliance, and actual human beings.

That intersection — the last mile between a working model and a production business process — is the decisive gap in enterprise AI. And almost nobody is talking about it, because it's not as glamorous as the next model release.

Next week: why the "build it yourself" era is over, and what the winning architecture actually looks like.

Sources:

[1] Menlo Ventures, Dec 2025 → https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

[2] McKinsey, Nov 2025 → https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[3] MIT NANDA, Jul 2025 → https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

[5] S&P Global Market Intelligence, 2025

[6] Informatica, "CDO Insights 2025" → https://www.informatica.com/resources/articles/cdo-insights-2025.html

[12] RAND Corporation → https://www.rand.org/pubs/research_reports/RRA2680-1.html

[14] WorkOS, Jul 2025 → https://workos.com/blog/why-most-enterprise-ai-projects-fail

[25] Beam AI analysis → https://beam.ai/agentic-insights/the-great-ai-flip-why-76-of-enterprises-stopped-building-ai-in-house