Most conversations about scaling AI agents focus on models, prompts, and tools. A report published this summer points to a more basic limit: how much of the business the agents can actually see.
On 12 August 2026, MIT Technology Review Insights published Scaling AI agents with trustworthy data, produced in partnership with Google Cloud. It is based on a survey of 300 data and technology executives.
What the report found
- AI has access to an average of only 45% of company data.
- “Data leaders” give AI access to more than 70% of their data. “Laggards” provide 30% or less.
- Only about half of organizations trust their agents’ decisions to be accurate and relevant. All of the data leaders report that trust.
- Two-thirds of laggards say legacy systems limit their ability to scale agents (66%) and to make decisions quickly (68%), compared with 8% of leaders.
- The top priorities are improving agents’ access to structured and unstructured data and strengthening data and AI governance with business context.
The report is sponsored content, and Google Cloud sells data and AI platforms. The direction of the findings matches other 2026 research, including a September Harris Poll survey in which 72% of AI decision-makers traced AI shortfalls to the data foundation.
What partial data does to an agent
An agent that sees less than half of the business still answers with confidence. It just answers from part of the picture: the contract without the amendment, the customer record without the open support ticket, the policy without the regional exception. The errors that follow look like model mistakes, but they are data gaps.
The fix is not to give agents access to everything. Organizations that win expand access deliberately, with permissions, definitions, and owners attached, so more data also means more trusted data.
Lessons learned
- Map what each use case needs. List the systems and documents an agent must read to make a decision correctly, then close those gaps first.
- Unlock unstructured content with permissions intact. Contracts, emails, and documents hold much of the context agents miss. Prepare them for retrieval without flattening who is allowed to see what.
- Add business context. Definitions, metadata, and relationships turn raw records into something an agent can reason about correctly.
- Plan around legacy systems. Use APIs, replication, or a governed data layer so old platforms stop being the ceiling on what agents can do.
- Measure trust. Track accuracy against an evaluation set, and widen access or autonomy only when the numbers support it.
What this means for your AI program
More capable models will not fix an agent that cannot see the data it needs. Our AI-Ready Data Foundation package prepares one priority data domain, including documents, for AI in six weeks, and our Knowledge Agent Launchpad connects assistants to your content with permissions and citations.