The Data Agent Era

Why AI Needs Trusted Business Context Before It Can Make Better Decisions.

The next stage of business AI is moving beyond generic knowledge. Increasingly, AI systems are being connected to the data, definitions, documents, tools, and permissions a company actually runs on.

OpenAI’s September 2026 introduction of Data agent in ChatGPT Work is one signal of that shift. OpenAI describes a system that can connect to company data, investigate changes, work with metric definitions and business context, and produce dashboards, reports, and actions while respecting connected-account permissions and workspace controls.

That capability exposes a hard truth:

An AI system cannot make trustworthy business decisions from context the business itself has not governed.

Data is only one layer of context

When people hear “business data,” they often think of tables: revenue, orders, customers, inventory, campaign results. Those are important, but a useful business agent also needs meaning.

It may need to know which revenue definition finance uses, which policy is current, which product description is canonical, which customer commitments are active, which repository controls implementation rules, and which user is authorized to see or change a record.

That broader layer is business context.

The Trusted Business Context System

A practical context system can be organized into seven layers:

1. Authority.
   Which system, document, or record is controlling for each important fact?
2. Semantic meaning.
   What do business terms and metrics actually mean? Definitions are part of the data.
3. Freshness.
   When was the source last updated, and how quickly can the fact become stale?
4. Provenance.
   Where did the value or statement come from, how was it transformed, and what evidence supports it?
5. Permissions.
   Who may retrieve the context, and what actions may they take with it?
6. Change control.
   How are approved rules, policies, templates, and definitions revised without silently creating conflicting versions?
7. Validation.
   What deterministic checks, reconciliations, or approval gates establish that an important conclusion is trustworthy enough to use?

Those layers turn a collection of sources into governed context.

Source authority comes first

Companies often accumulate systems faster than they accumulate governance. A spreadsheet becomes a database. A sales tool connects to billing. A second dashboard appears because the first does not answer one team’s question. Someone exports a private workbook. Different groups create slightly different definitions.

Eventually, a simple question produces several competing answers.

Take: What were Q3 sales?

• The CRM may count signed deals.
• The billing system may count invoices issued.
• The payment processor may count cash collected.
• The finance system may count recognized revenue.
• The commerce platform may count orders before refunds.

All five numbers can be technically correct and still answer different questions.

The agent needs an authority rule before it needs a bigger context window.

Definitions are operational assets

A metric name without a definition is incomplete context.

“Active customer” could mean someone who purchased in the last 30 days, someone with an active subscription, someone with an open contract, or anyone whose account has not been closed.

A governed semantic layer records the business definition, calculation, exclusions, owner, effective date, and authoritative source. That prevents an agent from silently mixing several meanings into one confident answer.

Freshness needs to be explicit

Not every source ages at the same rate. A constitution may remain authoritative for months. Inventory can change in minutes. A platform policy can change without your local documentation changing. A customer brief may supersede an older email.

Attach freshness expectations to important context:

• Static or slow-changing: operating doctrine, brand rules, approved templates.
• Periodic: pricing tables, product registries, KPI definitions.
• Near-real-time: balances, inventory, active incidents, current orders.
• External and volatile: laws, platform requirements, public pricing, product capabilities.

If the requested answer depends on volatile context, the workflow should refresh it before acting.

Provenance makes an answer inspectable

Trust improves when an important result can answer: Where did this come from?

A useful agent workflow should preserve enough provenance to trace a conclusion back to source records, transformations, assumptions, and time range. That becomes critical when the answer is surprising.

If revenue appears to drop sharply, the first question should not automatically be “What should we cut?” It should be “Is this change real?” A source may be delayed, a definition may have changed, refunds may have posted differently, or a data pipeline may have failed.

Provenance creates the path for checking.

Permissions should follow the user

An agent that can answer more questions can also expose more information if access is poorly designed.

The safer architecture carries existing authorization boundaries into retrieval, analysis, and action. A user who cannot access payroll data should not gain that access by asking an AI to summarize it. The same principle applies to customer records, confidential contracts, unpublished strategy, financial details, and personally identifiable information.

OpenAI’s business-data documentation states that organizational data for its business offerings and API is not used to train models by default. That is one vendor-level policy; organizations still need their own internal access and data-classification rules for what they connect and who may use it.

Change control prevents silent context drift

Context is not trustworthy if several active versions all claim to be current.

For high-value rules and documents, preserve a canonical source, version or effective date, change record, and superseded state. If an old rule must remain available for history, label it as historical rather than leaving it indistinguishable from current authority.

This matters for human teams and AI agents for the same reason: retrieval cannot resolve contradictions if the system never recorded which source is controlling.

Validation matters before action

Data agents become more powerful when they can move from analysis into workflows. That is also where errors become more expensive.

Before a high-impact result triggers action, use deterministic checks where possible:

• Does the total reconcile to the authoritative source?
• Is the time range correct?
• Are refunds, transfers, duplicates, and cancellations handled correctly?
• Has the metric definition changed?
• Are all required sources current?
• Does the user have authority for the proposed action?
• Is confidence high enough, or should the workflow stop for review?

The strongest systems combine flexible reasoning with hard validation.

Run context health checks

A context system should be auditable. A periodic health check can identify:

• Sources with no declared owner.
• Duplicate or conflicting definitions.
• Stale documents still marked current.
• Broken source links or disconnected integrations.
• Permissions broader than the workflow requires.
• High-value metrics with no reconciliation rule.
• Policies that changed externally but were never refreshed internally.

That audit turns context quality into an operating discipline instead of an assumption.

Governed context is the real goal

“Clean your data” is useful advice but incomplete. A perfectly formatted dataset can still have unclear authority, missing definitions, stale records, weak provenance, or inappropriate access.

The stronger target is governed context: information whose source, meaning, freshness, provenance, permissions, change history, and validation rules are understood.

That foundation makes AI systems more useful because they spend less time guessing what the organization means.

The sequence is:

Trusted context → governed reasoning → validated outputs → authorized action.

Reverse that order and automation scales confusion.

Start with The Lean AI Stack. For the governance layer around tool-using agents, read AI Agents Need an Operating Contract, Not Just a Prompt. For a governed execution view, explore Kairos/NVIDIA Nemotron Activation.

Sources

• OpenAI — Now everyone can put data to work
• OpenAI — Business data privacy, security, and compliance