Why most enterprise AI stalls — and it's not the model

McKinsey publishes a chart every year showing that around 70% of enterprise AI initiatives stall before they reach production. BCG says something similar. Deloitte's number moves around but hovers in the same range.

The response from most consultancies is: “You need our AI strategy engagement.”

That's the wrong response. The AI initiatives aren't stalling on strategy. They're stalling on data.

The pattern

Every stalled AI project we've been asked to review over the last three years follows the same shape:

• Someone senior gets excited about a use case.
• A proof-of-concept ships in six weeks and looks impressive in the demo.
• Someone asks “can we scale this to production?”
• Six months later, nothing has moved.

The stall is always in the last step. And in every case, when we trace the root cause, it's the same problem: the data foundation the AI model needs to consume in production is broken, undocumented, or too expensive to feed the model at the required cadence.

What “data readiness” actually means

Everyone talks about “data readiness” but the definition drifts. Here's the concrete version:

• The source-of-truth question is answered. For every field the AI needs, someone can point to a single system that owns it. Not two systems. Not “we're working on that.”
• The freshness question is answered. The data the model needs at inference time is available within a latency budget that makes sense for the use case.
• The cost question is answered. The compute cost of feeding the model at production volume is understood and defensible.
• The security question is answered. PII exposure, access controls, and audit trails are in place before the model touches customer data.

Most stalled AI projects can't answer any of these. That's why they stall.

The Snowflake symptom

Snowflake is where this shows up most visibly in the mid-market. Not because Snowflake is the problem, but because Snowflake is the exposed surface — the place where your data-operations problems become measurable and painful.

The signs are consistent:

• Credits climbing month-on-month without a clear driver.
• Multiple teams building competing versions of the same table.
• Pipelines running for reports nobody reads any more.
• Development warehouses running 24×7 for no defensible reason.

If any of these describe your Snowflake bill, your AI initiative is going to stall on the same underlying issue. The AI model wasn't the problem. The data foundation was.

The order matters

We run two connected practices at Blue Thread. One does Snowflake environment audits and data-platform engineering. The other does AI strategy through production. The same senior team works both. And we insist on the sequencing: data foundation first, AI second.

Not because we're pedants about process. Because we've watched too many teams try to do them in the other order and burn out on the second step.

If you're going to invest in AI, invest in the substrate that makes AI work. If the substrate is Snowflake — and increasingly it is — audit it before you build on it.

What “fix the pipes” actually costs

A fixed-fee Snowflake environment audit runs £10–30k depending on scale, takes 5–15 days, and gives you a board-ready PDF with quantified findings. It's the cheapest, fastest way to answer the four data-readiness questions above.

Compare that to the cost of a six-month AI initiative that stalls in month five. The maths is straightforward.

Fix the warehouse first. Then build the AI on top.