An employee uses AI to summarize a public industry report. Another connects an AI assistant to the company’s document repository. A third gives an AI agent permission to respond to customers. All three activities may be described as “using AI,” but they create very different business risks.
That distinction belongs at the center of every adoption decision. The question is not simply whether AI can perform a task. It is whether the business can trust the process, protect the information involved, and intervene when something goes wrong.
AI encompasses technologies ranging from forecasting systems to tools that generate text, images, and code. For leaders adopting generative AI and increasingly capable agents, the practical starting point is straightforward: choose a specific business problem, establish boundaries around data and authority, and expand only when testing supports the decision.
1. Start With a Task, Not a Tool
A subscription is not an AI strategy. Neither is an instruction to employees to “find ways to be more productive.” Both leave the hardest questions unanswered: What should improve? What mistakes are acceptable? Who owns the result?
Choose a narrow, reversible workflow for the first pilot. Drafting an internal briefing from public reports is a better starting point than autonomously changing customer accounts. The briefing can be checked and discarded. An account change may create contractual, financial, or operational consequences.
Before testing, document:
- The problem: What takes too long, costs too much, or produces inconsistent results?
- The baseline: How is the work completed today, including review and correction?
- The success criteria: What improvement would justify continued use?
- The owner: Who is accountable for evaluating results and approving expansion?
Keep the pilot specific enough to evaluate honestly. “Help the operations team” is not a testable objective. “Prepare a first draft of the weekly industry briefing using approved public sources” is.
2. Decide What Information May Enter
The most important AI decision may happen before anyone submits a prompt: deciding what information the system may receive.
An employee copying a document into an AI service is making a data-handling decision, whether or not they recognize it. Connections to email, shared drives, and customer systems can extend that decision far beyond a single document.
Create a Usable Approved-Use Policy
A practical policy should identify approved tools, permitted uses, prohibited information, accountable owners, and a clear route for questions or exceptions. Employees should not have to interpret a lengthy security document to determine whether they can summarize meeting notes.
Make distinctions concrete. Public marketing copy may be suitable for an approved tool. Customer records, confidential contracts, employee information, passwords, and API keys require different treatment. Credentials should not be pasted into prompts.
Before approving sensitive information, review the specific product, plan, configuration, and contractual terms. Examine retention, deletion, use of submissions for model training, access controls, connected applications, and available audit records. Do not assume every AI service trains on every submission—or that an enterprise label resolves every concern.
Check Where Connected Information Comes From
Data security also concerns integrity. Who can change the documents an assistant retrieves? Are those documents current? Could an unauthorized contributor alter material the system treats as reliable?
Joint guidance released by NSA and partners addresses data security throughout AI development, testing, and operation, including provenance, poisoning, and drift. For business leaders, the implication is practical: assign owners to important information sources and establish how unauthorized changes or declining performance will be detected.
3. Keep People Responsible for Consequential Outputs
Generative AI can produce clear, confident language without producing a correct answer. A polished summary may omit a crucial exception. A plausible citation may not support the claim—or may not exist.
NIST’s Generative AI Profile addresses risks including confidently incorrect output, privacy exposure, information-security threats, and problematic human reliance. It is a voluntary companion to NIST’s AI Risk Management Framework, not a certification or guarantee of legal compliance.
The appropriate review depends on the consequence of an error. Brainstorming internal presentation titles is different from interpreting contract terms, recommending a hiring decision, or preparing financial information.
- Verify material facts: Check names, dates, calculations, and substantive claims against authoritative sources.
- Open the citations: Confirm that the source exists and supports the statement being made.
- Require qualified review: Use people with relevant expertise for consequential decisions and external publication.
- Check omissions and unfair assumptions: A technically accurate passage can still exclude important context or treat people inappropriately.
For the industry-briefing pilot, the reviewer should compare the draft with the original reports—not ask another AI response whether the first response looks correct.
AI can assist with the work. Accountability for accepting and using that work must remain clear.
4. Separate Assistance From Authority
An assistant that drafts a customer message is not equivalent to an agent that sends it. The second system has authority to affect the business, not merely generate content.
That difference should determine its controls. Giving an agent broad access because it might need additional capabilities later reverses the principle of least privilege.
Limit Permissions and Require Meaningful Approval
Start with read-only access where possible. Restrict the agent to the records, systems, and actions its task requires. Use temporary credentials where feasible, and separate testing from production.
Require explicit approval before consequential actions such as issuing payments, changing account details, deleting records, or publishing externally. Reviewers need enough context to understand what they are approving; a generic confirmation button is not meaningful oversight.
The UK National Cyber Security Centre’s guidance on agentic AI recommends limited scope, least privilege, temporary credentials where possible, monitoring, and incident planning. These measures reduce exposure; none makes an autonomous workflow inherently safe.
Treat Outside Content as Untrusted
AI systems may encounter malicious instructions embedded in websites, documents, or retrieved material. As NCSC explains in its AI and cybersecurity guidance, prompt injection can lead to unexpected behavior, confidential-information disclosure, or other unintended consequences.
For example, a document being summarized could contain instructions to transmit internal information elsewhere. That document is task input, not an authorized manager. Design the workflow accordingly: restrict available tools and destinations, separate untrusted material from instructions where possible, and require approval for sensitive actions. Telling the model to ignore malicious instructions is not a sufficient defense.
5. Measure the Complete Workflow Before Expanding
A fast first draft can conceal a slow overall process. If employees spend substantial time checking citations, fixing errors, and rewriting the result, the apparent productivity gain may disappear.
Evaluate completed work, not impressive demonstrations. Useful pilot measures include:
- Net time saved: Include preparation, review, correction, and exception handling.
- Error and rework rates: Distinguish cosmetic issues from mistakes with business consequences.
- Cost per completed task: Include subscriptions, integration, oversight, and maintenance.
- Security and privacy events: Track inappropriate access, disclosure, and attempted unauthorized actions.
- Operational resilience: Confirm that employees can stop the workflow and complete essential work manually.
These are proposed evaluation measures, not universal benchmarks. Set acceptance criteria that reflect the task’s risk and the organization’s obligations.
Test ordinary cases, ambiguous requests, missing information, and suspicious inputs. Reassess when models, integrations, permissions, or source material change. Approval should apply to a defined workflow—not permanently to everything a product may eventually do.
Assign someone authority to suspend the workflow. The response plan should explain how to stop actions, revoke access, preserve relevant records, investigate, and involve security, legal, or other responsible teams when necessary.
The Five Questions Every AI Approval Should Answer
Before approving a pilot or expanding access, insist on clear answers to this checklist:
- What problem does this solve? Identify the task and measurable benefit.
- What information can it access? Include prompts, connected systems, and retained records.
- What actions can it take? Distinguish recommendations from execution.
- Who checks the result? Name the accountable owner and required reviewer.
- How do we stop it when something goes wrong? Identify the shutdown process and manual fallback.
Responsible AI adoption does not mean avoiding experimentation. It means making experiments bounded, observable, and accountable.
Choose one workflow, assign an owner, and establish these boundaries before the pilot begins. Then measure the complete result. Expand because the evidence supports it—not because the demonstration was impressive.