The New Data Bottleneck Is Decision Accountability

The New Data Bottleneck Is Decision Accountability

For years, companies tried to democratise data.

They built data warehouses, hired analysts and distributed dashboards. Yet when a manager wanted to understand why sales had fallen, which customers were at risk or where costs were increasing, the request often joined an analytics queue.

That constraint is beginning to disappear.

Natural-language data agents can now investigate business questions, generate analyses and build dashboards without requiring the user to write SQL. OpenAI’s recently introduced Data agent, for example, connects to approved enterprise sources, applies existing table-, row- and column-level permissions, and uses an organisation’s metric definitions and semantic layers to interpret questions. That is a vendor announcement, not independent evidence of universal performance, but it illustrates the direction of travel: querying organisational data is becoming dramatically easier. OpenAI

This sounds like data democratisation finally arriving.

It also creates a new problem.

When anyone can produce a plausible analysis, the bottleneck moves from obtaining an answer to deciding:

The next data strategy is therefore not simply about access.

It is about decision accountability.

More answers do not automatically produce better decisions

Imagine that revenue growth slows.

The chief revenue officer asks an AI data agent to find the cause. Finance does the same. So do marketing, product and regional sales.

Within minutes, each team has a convincing answer:

These findings could all be correct. They may be looking at different time periods, customer cohorts, revenue definitions or causal assumptions.

The old problem was waiting two weeks for the analysis.

The new problem is receiving five analyses in two minutes—with no agreed mechanism for turning them into one decision.

AI reduces the cost of producing evidence. It does not eliminate organisational disagreement about what that evidence means.

In fact, it can industrialise disagreement by giving every function the ability to generate a polished, data-backed case for its preferred interpretation.

Separate the right to ask from the right to decide

Most companies treat data access as a technical permission: can this employee see this table, document or customer record?

AI requires a broader set of decision rights.

A person - or agent - may have permission to access data without having authority to:

Leaders should distinguish four different rights:

1. The right to ask

Employees should have broad freedom to explore data relevant to their work. This is where democratisation produces learning and speed.

2. The right to publish

An analysis presented as an official company position should meet a higher standard. Its definitions, sources, assumptions and material limitations should be visible.

3. The right to decide

Every consequential decision needs a named owner. AI may recommend an action, but accountability cannot be assigned to “the model.”

4. The right to execute

Allowing an agent to change a price, contact a customer, move inventory or approve a transaction is a separate delegation of authority. It should be bounded by explicit thresholds, monitoring and escalation rules.

This distinction becomes particularly important as companies move from AI that explains information to AI that acts on it. McKinsey describes agency as a transfer of decision rights, while BCG’s research on decision agents observes that their introduction can expose unclear accountability and unresolved data or process gaps. McKinsey, BCG

Not every decision needs the same controls

The answer is not to force every AI-generated analysis through a central committee. That would recreate the reporting queue under the banner of governance.

Controls should match the consequence of the decision.

The central principle is simple:

This is not primarily a data-platform decision. It is an operating-model decision.

Centralise meaning, not every analysis

Stanford’s 2026 Enterprise AI Playbook, based on 51 deployments across 41 organisations, found that data was scattered across multiple systems in 59% of the examined cases, while only 16% had fully centralised it. Yet successful deployment did not always require complete centralisation; access, integration and documented organisational knowledge were often more important. The sample covers selected successful deployments and relies partly on reported organisational experience, so it should not be treated as a market-wide success rate. Stanford Digital Economy Lab

The lesson is not that data architecture no longer matters.

It is that companies should avoid waiting for a mythical moment when every dataset is perfect and centralised. Instead, they should centralise the elements that allow decentralised analysis to remain coherent:

McKinsey similarly argues that AI systems need governed information with clear versioning, context and links between unstructured material and the structured records defining customers, products, contracts and transactions. McKinsey

A company does not need one team answering every question.

It needs one organisational understanding of what “revenue,” “active customer,” “churn,” “margin” and “qualified opportunity” mean—or a transparent record of where definitions legitimately differ.

Accountability must sit with the business

Data teams can maintain pipelines, semantic definitions and analytical standards. Risk teams can establish controls. Technology leaders can govern access and system behaviour.

But none of them can own every business decision produced from the data.

The executive responsible for the outcome must remain accountable for the decision process.

NIST’s AI Risk Management Framework makes this organisational requirement explicit: roles, responsibilities and lines of communication for managing AI risks should be documented and clear. Its governance model also emphasises that accountability must be supported by continuing documentation and review rather than treated as a one-time approval. NIST AI Risk Management Framework

For leaders, this means every important AI-supported decision should answer five questions:

If the fifth answer is “the data team,” “the AI committee” or “the model,” the decision design is incomplete.

What to do on Monday morning

Choose one recurring decision: pricing, inventory allocation, renewal intervention, hiring approval or sales forecasting.

Map this chain:

Question → Data → Definition → Analysis → Decision → Owner → Outcome

Then look for three gaps:

Do not begin by buying another governance platform.

Begin by naming the person who owns the decision—and documenting the evidence that person is expected to inspect before acting.

The strategic conclusion

The first phase of data democratisation gave more employees dashboards.

The next phase will give them analysts on demand.

That will be valuable—but it will also make it cheap to create persuasive, conflicting interpretations of the business.

The winning companies will not be those that generate the most answers. They will be those that preserve shared meaning while allowing questions to proliferate.

Decentralise interrogation. Centralise definitions. Make decision ownership unmistakable.

Sources and why they matter