97 percent invest, 5 percent are ready: why enterprise AI data readiness is a budget allocation problem

Dun and Bradstreet published the findings of its 2026 AI Momentum Survey in May 2026. The survey covers 10,000 businesses across 32 countries, conducted on a quarterly cadence. The headline finding: 97 percent of organisations worldwide report active AI initiatives. Five percent say their data is adequately ready to support them (Dun and Bradstreet, AI Adoption Surges, Data Readiness Limits ROI, 2026).

That 92-point gap is the central fact of enterprise AI adoption in 2026. It tells you more about why the Gartner survey finding that 40 percent of CIOs cannot point to the value they get from AI is structurally predictable than any number of individual implementation post-mortems.

The gap is not a technology problem. It is a budget-allocation problem.

What data readiness actually requires

Data readiness for AI is not the same as having data. Most enterprises have extensive data assets accumulated across decades of transactional systems, CRM platforms, ERP databases, and operational logs.

Data readiness for AI means having data that is structured, labelled, deduplicated, and accessible in the formats and at the latency that autonomous AI systems require to operate reliably across mission-critical processes. A financial reconciliation agent that checks invoice data against purchase orders requires both datasets to be consistently structured, fully deduplicated, and queryable within the agent’s processing window. A customer service agent that makes real-time decisions based on account history requires that history to be accessible within the agent’s response time and permissioned so the agent can reach it.

Most enterprise data environments were designed for human workflows. A knowledge worker can tolerate inconsistent field labelling, navigate between legacy and modern data models, fill in gaps from institutional memory, and apply contextual judgement when data is incomplete. AI systems cannot. They pattern-match against what they can access, and inconsistent, siloed, or missing data produces inconsistent outputs.

The architecture required for reliable AI performance at scale is different from the architecture required for human-readable reporting. The gap between them is not filled by changing the model. It is filled by capital investment in the infrastructure layer.

Why the budget allocation gets it wrong

Enterprise AI budgets in 2025 and 2026 were structured around model licensing, API access, and pilot projects. The logic was sequential: demonstrate value in a pilot, secure funding to scale, fund the infrastructure required to scale.

The problem with the sequential logic is that it confuses pilots with scale. Pilots can be run on purpose-built datasets cleaned specifically for the pilot, on infrastructure managed manually for the duration of the test, and with evaluation criteria defined to show the model in its best operating conditions. Scale requires production data, production infrastructure, and production conditions. The pilot succeeds. The scale attempt fails. The failure is attributed to the model or to AI in general rather than to the data environment the scale attempt ran on.

Sixty percent of businesses in the D&B survey report at least some measurable ROI from AI, including 24 percent reporting broad or strong returns. The 40 percent reporting no measurable ROI are not predominantly running failed models. They are predominantly running models on data infrastructure that cannot support reliable AI output at scale.

The Gartner observation that 40 percent of CIOs cannot point to the value they get from AI is a downstream symptom. The upstream cause is that the data-infrastructure investment required to make the model-layer investment productive was not funded at the time the model-layer investment was approved.

What the OpenAI Deployment Company’s service model reveals

OpenAI launched the OpenAI Deployment Company on 11 May 2026, structured as a separate entity with more than 4 billion dollars in initial investment and approximately 150 Forward Deployed Engineers (OpenAI, OpenAI launches the OpenAI Deployment Company, 11 May 2026). The acquisition of Tomoro, an applied AI consulting firm, accompanied the launch.

The service model is the signal. The Deployment Company’s stated mandate is not model selection or API access. It is workflow redesign and infrastructure integration: identifying where AI can make the biggest impact, redesigning organisational infrastructure and critical workflows around it, and turning those gains into durable systems.

OpenAI already sells model access. The rationale for a separate 4-billion-dollar entity is that model access is not the constraint. Integration and infrastructure are the constraint. The Deployment Company is, structurally, a data-readiness and workflow-integration service operated at enterprise consulting rates by the same organisation that sells the model.

That structure is a reliable indicator of where the bottleneck is.

The four prerequisites for scale that most pilot budgets do not fund

Pilots require model access and a defined use case. Meaningful scale across production workflows requires four investments that most enterprise pilot programmes do not budget:

Consistent data pipelines that route structured, clean data to AI systems in production, not to a one-time pilot dataset. Building and maintaining those pipelines is an engineering investment in the data infrastructure layer, not in the model layer.

Permissions and governance architecture that allows AI systems to access the data they need without creating security or compliance exposure, and that produces an audit trail. This is a governance infrastructure investment with its own capital requirements.

Integration with existing operational systems at the latency the AI task requires. Most enterprise data infrastructure runs on batch-processing cadences appropriate for human-readable reporting, not on the sub-second response times that production AI agents require. Closing the latency gap requires infrastructure change.

A monitoring and correction layer that identifies when AI outputs degrade because of data drift or model-behaviour change, and that routes corrections back into the system. This is the infrastructure equivalent of quality assurance for human workflows, and it does not come bundled with a model subscription.

None of these appear in a typical proof-of-concept budget. All of them are required before a pilot produces the reliable, scalable outcomes that justify continued investment.

The CFO conversation

The data-readiness gap does not require a new budget line. It requires a reframe of the existing one.

The enterprise has already approved model-layer spend: subscription costs, API access, and whatever in-house or outsourced development supports the pilot portfolio. That approved spend implies a return, and the return is contingent on the data infrastructure the models run on.

The CFO conversation is not about adding a data-infrastructure budget. It is about protecting the return on the model-layer budget already committed. A model subscription that the enterprise cannot use at scale because its data environment is not ready is generating below-plan ROI. The infrastructure investment that closes the gap is the prerequisite for the model investment to perform, not a separate initiative.

A four-line budget addendum that separates model-layer spend from infrastructure-layer spend and frames the latter as the prerequisite for the former is the mechanism that converts the 5-percent data-readiness figure from an industry statistic into a capital-allocation conversation the CFO can act on.

Claim AM-165 is registered in the Holding-up ledger. 90-day review: 20 Aug 2026.