Odau Report 03 · 2026

The 95% figure, read properly.

AI usually fails at the business-system level, not the model level. The tool can work while ownership, integration, adoption and measurement do not.

Seven minutes · Every number sourced · Nothing gated
Sixty seconds
78%
Use AI somewhere
↑ from 55% YoY
Stanford HAI, 2025
1 in 20
Reached rapid P&L impact
Pilot-to-value finding
MIT NANDA, 2025
69%
Say governance will take over a year
Controls do not arrive instantly
Deloitte, 2025
26%
Explore autonomous agents extensively
The risk boundary is moving
Deloitte, 2025
The correction

The 95% figure is not a verdict on AI.

It describes pilots that did not reach rapid, measurable P&L impact. It does not say the models were incapable or the work produced nothing.

The gap is pilot to production, then adoption to value.

What it means

Most pilots never crossed into a measured business result.

What it does not mean

Ninety-five percent of AI technology is unusable.

Source: MIT NANDA, State of AI in Business 2025.
Twelve failure patterns

The first two come from Odau experience—not a survey.

Our team repeatedly encountered these patterns in real operating environments and heard them echoed in conversations with industry leaders. No invented sample size. Just the same operating failures, repeatedly.

What we repeatedly saw in the field

This comes from the Odau team’s direct operating experience and industry conversations—not a survey or third-party benchmark. We label it separately because that distinction matters.

From the Odau team

No real owner

A revenue system gets handed to someone who knows the technology or the function—but rarely both. Nobody owns implementation through its full lifecycle.

The system was purchased. The work around it was never designed.Check your exposure
From the Odau team

Everyone wants AI. Nobody defines the work.

Teams say they want to “implement AI” without naming the workflow, decision or measure. Many already had AI in Salesforce Einstein without recognizing it—because the system, people, process and data were never treated as one implementation.

If nobody can say what changes in the work, there is no implementation—and nothing to prove worked.Check your exposure

AI looking for a problem

Many demos, no baseline. Start with one constrained workflow.

Pilot purgatory

Repeated trials, no production design. Fund the full path first.

Poor data readiness

Inconsistent output and workarounds. Establish ownership and quality tests.

Weak integration

People copy and paste. Put AI inside the system of work.

No adoption design

Good model, low repeat use. Redesign roles, incentives and escalation.

Vague ROI claims

Usage rises; operations do not. Tie value to measurable change.

Inadequate governance

Unknown use and no audit trail. Apply controls by risk.

One vendor or model

Cost and performance become rigid. Preserve portability where justified.

Premature autonomy

Small errors create irreversible actions. Start read-only, then add approval.

Operating cost underestimated

Usage grows without attribution. Add budgets, unit economics and limits.

Patterns 01–02: recurring Odau field experience. Patterns 03–12: synthesis of attributed published research.
What works differently

A technology purchase is not a working business change.

The tool can function perfectly while the business result never arrives. That gap is where most implementation stories end.

Technology purchase

Proves the company acquired a capability.

Business change

Proves the company created a result.

Measurement discipline

Time saved is not value until the business can point to the result.

A credible result exists outside the usage chart and can be explained without relying on the technology’s own metrics.

A usage chart is not a business case.

Claim

The tool saved time.

Result

The business can show what changed.

No gate

Want more like this?

We’ll send you select resources.

We hate spam more than you do. Unsubscribe anytime.