Businesses have always operated with imperfect data. If employees can interpret it, verify it, and follow procedures to make decisions, why should we assume AI agents cannot?
There is a recurring message in discussions about enterprise AI: before introducing artificial intelligence into business processes, organizations must first clean, standardize, and organize their data.
Data quality certainly matters. Accurate and consistent information improves efficiency, reduces uncertainty, and makes systems more reliable. But I believe the idea that clean data must always come first deserves a closer look.
Customer records contain inconsistencies, supplier information becomes outdated, documents are duplicated, and different systems sometimes provide conflicting information. Despite these imperfections, employees carry out their daily responsibilities. They interpret information, consult procedures, verify questionable records, and involve colleagues or managers when necessary.
Looking beyond traditional automation
Part of the confusion, in my view, comes from approaching AI with the same assumptions traditionally applied to software automation.
Deterministic workflows depend on predefined rules and predictable inputs. When information does not match the expected structure, additional validation, transformation, and exception-handling logic is often required.
AI agents introduce different capabilities. They can interpret information in context, consult multiple sources, recognize inconsistencies, and use available tools to investigate uncertainty.
Consider a supplier database containing three different bank account numbers for the same company. Neither an employee nor an AI agent can determine which one is correct simply by looking at those records. An employee would follow company procedures, consult additional information, or request confirmation.
An AI agent, equipped with the appropriate tools and access to company procedures, can follow a similar approach. It might also carry out additional checks that employees sometimes skip because of time constraints, workload, or unfamiliarity with specific procedures.
Agents can still misunderstand information or reach incorrect conclusions, just as employees can. But imperfect data does not automatically make them unsuitable for business operations.
Architecture makes the difference
From our experience working with enterprise AI systems, reliability depends significantly on how a solution is designed and what capabilities are available to its agents.
Employees rarely operate in complete isolation. Organizations distribute responsibilities, define approval procedures, and assign different levels of authority. AI systems can follow similar principles.
Imagine an agent responsible for preparing supplier payments. It examines purchase orders, deliveries, invoices, and supplier records to determine which payments should be prepared. Another agent, acting as a supervisor, can verify sensitive transactions, consult additional information, and approve or reject requests within its authority.
Their collaboration can follow business responsibilities without requiring every step to be predetermined in a rigid technical workflow. Different roles and access levels also make meaningful independent verification possible.
Agents should be able to consult and interpret company policies, much as employees do. At the same time, critical authorization boundaries must remain enforceable by the underlying enterprise systems.
For instance, an operational agent may recognize that a payment needs managerial approval, while the payment system independently verifies that the appropriate authorization has actually been granted, based on the identity and permissions of the actor.
AI autonomy and enterprise security can coexist when architecture preserves identity, accountability, and independent controls.
The economics of imperfect data
Clean data is generally preferable to dirty data. It can simplify operations, reduce processing time, improve consistency, and lower the cost of AI execution.
Yet cleaning an organization's entire information landscape can also be expensive and time-consuming. Whether that investment is justified depends on the processes involved and on the AI architecture.
Agents may use tools to verify conflicting information, or memory mechanisms to retain previously validated findings. Those findings can be revisited when source information or company policies change, reducing repeated work. In other situations, correcting the data at its source will clearly be simpler and more reliable.
Data preparation should be evaluated against real business needs, costs, and measurable benefits. Treating complete data cleansing as a universal prerequisite can delay opportunities that are already within reach.
Establishing trust through testing
How can an organization determine whether its available data is sufficiently reliable for AI agents to operate?
If employees successfully use that data today, there is a reasonable starting point for exploring AI adoption. This does not guarantee that an agent will perform equally well: employees may rely on implicit knowledge, informal procedures, or contextual information that has not yet been made available to the AI system.
A staging phase helps reveal these gaps. Organizations can test agents against realistic business cases, measure the accuracy of their decisions, observe how they handle ambiguity, verify policy compliance, and assess when human intervention is required.
Testing also shows where better data, additional tools, stronger controls, or more accessible business context are genuinely necessary.
In our implementations and observations, agents equipped with appropriate tools and context have often completed tasks successfully even when the underlying information was imperfect. This practical experience is not a substitute for comparative benchmarks, but it is a reason to test what an AI solution can achieve before assuming that extensive data cleansing must come first.
Introducing AI into real business environments
Organizations should continue improving data quality. Better information benefits employees, traditional applications, and AI systems alike.
At the same time, enterprise AI creates opportunities to work with information in ways that were difficult with traditional automation. Agents can interpret context, consult procedures, collaborate across responsibilities, and systematically verify information using tools connected to business systems.
With suitable architecture, governance, and validation, organizations can introduce AI progressively using the information they already have. They can measure results, identify limitations, and improve both their data and their AI capabilities as experience grows.
Data quality remains an important part of enterprise AI adoption. Its improvement can progress alongside AI adoption, guided by actual business needs and measurable results.