Until recently, much of the business use of generative AI focused on drafting, summarising and searching for information. If the result was poor, someone could correct it or decide not to use it.
Agentic AI changes that. An AI agent may not simply summarise an invoice or retrieve a customer record. It may update a system, route a case, trigger a payment, contact a customer or begin the next stage of a workflow.
That is why the quality of the information available to AI is becoming more important, not less. A July 2026 study conducted by Forrester Consulting for Boomi found that 86% of organisations have moved beyond AI agent pilots, but only 34% trust the actions their agents are taking. Adoption may be accelerating, but confidence in what those agents are doing has not kept pace.
Much of the debate about that risk starts with the model. Can it reason reliably? Has it been given the right guardrails? Who is accountable for its actions? These are important questions. But before an agent can make a reliable decision, it needs reliable information on which to base it.
The business information AI cannot see
Businesses often talk about enterprise data as though it is already sitting in clean, structured databases. A great deal of it is not.
Important information still arrives through invoices, contracts, application forms, identity documents, customer correspondence and handwritten notes. Some will arrive on paper. Some will be attached to emails or uploaded as PDFs. Even when a document is technically digital, its contents may not be searchable or connected to the correct customer, supplier, employee or case.
A 2026 survey conducted by The Harris Poll on behalf of Box shows how significant this gap has become. While 96% of organisations surveyed said it was important or very important for agents to access company-specific content, only 36% of those using or experimenting with agents had connected them to trusted internal content across multiple use cases.
This is the first mile of an AI workflow – and it is easily overlooked. A document may be perfectly legible to a person while remaining difficult for a system to interpret. Its text may be readable, but the system may not know whether it is the latest version, what process it belongs to or which details matter.
A better AI model cannot recover context that was never captured.
From a wrong answer to a wrong action
These weaknesses existed before agentic AI, but agents raise the stakes. If a chatbot produces an incomplete summary, an employee may spot the problem. If an agent uses the same incomplete record to progress an application, approve an invoice or update a customer file, the error moves directly into the operation.
People also adapt when they do not trust company systems. They keep their own spreadsheets, save personal copies of documents and use email chains to correct the official record. These workarounds are understandable, but they create information that an AI agent connected to the main repository may never see.
That leaves the agent working from only part of the picture.
For me, this puts information capture firmly within the AI governance conversation. Good capture is not simply about creating a clear scan. It means identifying the document, extracting the relevant information, applying consistent metadata and sending it to the correct system with the appropriate access and retention rules.
It must also make exceptions visible. A handwritten form may be unclear, an invoice incomplete or a contract inconsistent with the information already held. In those circumstances, the right action may be to stop and ask a person, not continue automatically.
Four checks before an agent is allowed to act
Before an agent is given authority to act, I would start with four fairly basic checks:
- Where did the information come from, and how was it captured?
- Is it complete, current and connected to the right business context?
- Which system or version is authoritative?
- What causes the agent to stop and refer the decision to a person?
These questions sound basic. In practice, answering them can expose duplicate repositories, scan-to-email habits, inconsistent naming and processes that depend on employees knowing where unofficial information is kept.
We saw the practical value of this in our work with a leading French bank. Branch teams were manually photocopying, entering and sharing customer documents, including identification, proof of address and contracts.
By standardising capture and connecting it directly with the bank’s systems, staff could check and correct information while the customer was still present. This reduced errors and repeat visits and saved more than €4 million a year in document transport and processing costs.
It was not an agentic AI project, and that is why it matters. It shows that the quality, availability and movement of information must be addressed before more autonomous technology is added.
AI readiness needs different measures
Businesses should not judge an agentic workflow only by whether the agent completed its task. They should also measure how much source information passes through correctly first time, how often it requires correction and how long it takes to become ready for use.
Human interventions and overrides should be treated as useful evidence, particularly when they reveal repeated problems with a document type or information source.
That, to me, is organisational intelligence in practice: learning how information moves through the business and making it more usable for people and technology.
Agentic AI is intended to move work forward with less human intervention. Once software is taking action, however, correcting a mistake after the event becomes harder and potentially more costly.
Getting information right at the point of capture therefore becomes part of how a business manages AI risk, not simply an administrative task that happens before the important technology takes over.
