Artificial intelligence agents are moving beyond answering questions and analysing business information. They’re beginning to take action, manage workflows and make decisions that previously required human approval.
But giving these systems access to more data doesn’t always make them more effective.
In an article published on 3 November 2026, David Edelman, executive advisor, Harvard Business School faculty member and author, challenges the assumption that more information automatically produces better AI decisions.
Writing in The Personalize Community, Edelman examines how poor data quality can create problems when businesses move from human-led analytics to autonomous AI systems.
His argument centres on a straightforward concern. A human employee may recognise an incorrect record before acting on it. An AI agent connected to business software could process the same information and take action before anyone notices the mistake.
For companies deploying AI across customer service, marketing and operational processes, this raises questions about data management and the controls placed around automated decisions.
Why More Data Doesn’t Always Mean Better AI Decisions
Businesses have spent years collecting information about customers, transactions, products and operational performance.
That information is often spread across customer relationship management platforms, financial databases, analytics tools and other enterprise applications.
AI agents can connect these systems, retrieve information and use it to determine what action to take next.
The problem begins when the information itself is inaccurate.
A duplicated customer profile might create an incorrect account history. An outdated payment record could suggest that a customer owes money. Two databases may contain conflicting information about the same transaction.
These problems are familiar to data management teams. What changes with autonomous AI is the speed at which an error can affect a business process.
When a person reviews a report, there’s usually an opportunity to question an unexpected result.
An autonomous agent may move directly from interpreting data to executing an action.
Edelman’s analysis suggests that companies need to reconsider how information is validated before allowing AI systems to operate independently.
How One Incorrect Billing Record Could Damage Customer Relationships
To illustrate the risks, Edelman presents a hypothetical example involving a business-to-business company and its automated billing process.
Imagine that a billing system incorrectly marks several major customers as overdue.
If an employee were reviewing the accounts, the unusual status might prompt an investigation before any collection activity began.
But an autonomous collections agent could interpret the records as accurate.
It might send payment reminders, escalate the language of its communications or threaten account suspension.
The customers would receive demands based on an error that originated elsewhere in the organisation.
And the consequences could extend beyond the original data problem.
A company might face complaints from long-standing customers, additional work for its support teams and damage to relationships built over several years.
The example is hypothetical, but it demonstrates the problem Edelman wants businesses to address.
An AI agent can follow its instructions correctly and still produce a damaging result when the information behind those instructions is wrong.
Data Validation Must Happen Before AI Agents Take Action
Edelman argues that businesses need stronger controls over the information entering autonomous workflows.
This involves checking data before it influences decisions, rather than relying entirely on monitoring after an action has been completed.
For instance, a collections agent could be required to verify that an outstanding balance agrees with the accounting system before sending a reminder.
If records conflict, the system could stop the workflow and request human review.
The same approach can apply to other business activities.
A marketing agent managing campaign budgets should work with verified spending limits. A customer service agent changing account details should confirm that the requested action is authorised.
These controls require more than carefully written instructions to an AI model.
Edelman recommends deterministic checks, meaning software rules that produce defined outcomes and can prevent an action even when an AI model recommends proceeding.
This approach is consistent with guidance from the Open Worldwide Application Security Project (OWASP), which recommends limiting agent permissions and requiring independent validation for actions with serious financial or operational consequences.
Why AI Prompts Alone Cannot Guarantee Business Safety
AI agents commonly receive instructions describing their responsibilities, restrictions and expected behaviour.
These instructions can help guide their responses, but they aren’t a substitute for controls enforced by the surrounding software.
An agent may misunderstand a request, encounter conflicting information or use an authorised tool in an unintended way.
OWASP identifies excessive agency and tool misuse among the risks associated with AI systems that can take action through connected applications.
Consider an agent authorised to issue refunds.
A written instruction might tell the model to avoid refunds above a certain amount. But if the underlying payment system doesn’t enforce that limit, the business remains dependent on the model following the instruction correctly.
A stronger design would place the refund limit within the payment workflow itself.
The model could suggest the refund, but the transaction would still need to pass a separate check.
For higher-value transactions, the company could require approval from an authorised employee.
That separation allows businesses to use AI for analysis and recommendations while retaining control over actions that carry financial consequences.
Turning AI Outcomes Into Feedback
Another principle in Edelman’s article concerns the need to learn from what autonomous systems actually do.
An agent may perform well during initial testing but encounter situations that weren’t anticipated when it enters daily business operations.
Customer behaviour changes. Records are updated. New products and internal processes introduce different conditions.
Edelman recommends collecting the results of agent actions and feeding those outcomes into future evaluation and improvement.
For a marketing team, that could mean comparing an agent’s campaign decisions with actual spending, customer responses and approved performance measures.
For a customer service department, it could involve reviewing complaints linked to automated communications.
The objective is to identify repeated errors and update the system’s controls before the same problem affects more customers.
Such feedback needs careful handling. Poorly measured outcomes could reinforce the wrong behaviour, while feedback data may contain errors of its own.
Businesses need to define what constitutes a successful action and who is responsible for reviewing unexpected results.
Treating Business Data as Evidence Rather Than Truth
One of the more useful ideas in Edelman’s argument is what he calls input-to-judgment thinking.
Instead of treating every database entry as a fact, an AI system should assess whether the information is reliable enough to support the action being considered.
A customer record may be current but incomplete. A transaction might be correctly recorded in one system and outdated in another.
Some sources may carry greater authority than others.
An AI agent responsible for updating a customer account should be able to identify disagreements between records rather than selecting whichever result appears first.
If the available evidence is insufficient, it may need to pause.
This approach also raises questions about data ownership within organisations.
Who is responsible for correcting inaccurate records? Which database should take precedence when systems disagree? How quickly can an error be traced back to its source?
Without clear answers, companies risk giving AI agents access to data environments that employees already struggle to manage.
The Business Risks Go Beyond Customer Communications
Although Edelman’s example focuses on billing, the underlying problem can affect many forms of enterprise automation.
AI agents are being explored for advertising, procurement, financial operations and account administration.
Each application introduces a different level of risk.
An incorrect product recommendation may inconvenience a customer. An unauthorised financial transfer can have far more serious consequences.
Businesses should consider how much authority an agent needs for each task and what could happen if its information is wrong.
The National Institute of Standards and Technology (NIST) provides an AI Risk Management Framework designed to help organisations assess and manage risks associated with AI systems.
Its approach supports continuous evaluation of how systems operate and the effects they may have on individuals and organisations.
For companies deploying AI agents, this means examining both the accuracy of the model and the business process surrounding it.
A system that generates sensible recommendations during testing may still be unsuitable for autonomous deployment if it relies on unreliable records or lacks effective approval controls.
What Businesses Should Check Before Deploying AI Agents
Edelman encourages organisations to examine the safeguards they already have in place.
Three areas deserve attention: whether software-enforced rules prevent unacceptable actions, whether the results of automated decisions are reviewed, and how often corrupted or duplicated records reach operational systems.
Companies can use these questions to examine individual workflows before expanding AI automation.
For example, a business introducing an autonomous customer service agent could begin by allowing it to retrieve account information and prepare suggested responses.
Once the data sources have been tested, the organisation could permit selected low-risk actions under defined conditions.
More consequential decisions might continue to require human approval.
This staged approach creates opportunities to observe system behaviour while limiting the damage that an incorrect decision could cause.
It also makes it easier to identify whether failures originate in the AI model, the information supplied to it or the software connecting the different systems.
Why Trustworthy Data Will Shape Enterprise AI Adoption
Edelman’s article highlights a problem that can be overlooked during the excitement surrounding autonomous AI.
Businesses often focus on what a new model can accomplish, how quickly it can complete a task and how much work it might automate.
Those capabilities matter, but they depend on information that may have been collected over many years through systems with different standards and responsibilities.
An AI agent cannot automatically resolve every inconsistency simply because it can access more records.
Sometimes it will need better information. Sometimes it should ask for human judgement. And sometimes the safest response is to take no action.
As companies give AI systems greater authority over customer interactions and business decisions, the quality of their existing data will become harder to ignore.
The questions raised by Edelman provide a practical starting point for organisations planning more autonomous operations.
Before allowing an AI agent to act, businesses need to establish what information it can trust, which decisions it can make and how mistakes will be detected.
A sophisticated model cannot compensate for every weakness in the systems around it.
About David Edelman
David Edelman is an executive advisor, public speaker, author and faculty member at Harvard Business School. His work focuses on marketing, customer experience, business growth and the use of artificial intelligence in organisational strategy.
Through The Personalize Community, Edelman shares perspectives on AI adoption, personalisation and the changes affecting marketing and customer relationships.
His article The Fallacy of “All Data Is Good Data” for AI Agents was published on 3 November 2026.
LinkedIn: https://www.linkedin.com/in/daveedelman/
Sources
1. David Edelman, The Fallacy of “All Data Is Good Data” for AI Agents, 3 November 2026
Published in The Personalize Community on LinkedIn. https://www.linkedin.com/newsletters/the-personalize-community-7281665178532810752/
2. David Edelman, LinkedIn Post and Video https://www.linkedin.com/posts/daveedelman_ai-data-marketing-ugcPost-7523008194114260992-Ur7i/
3. OWASP, AI Agent Security Cheat Sheet https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html
4. OWASP, Excessive Agency Risk Guidance https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
5. NIST, AI Risk Management Framework https://www.nist.gov/itl/ai-risk-management-framework
