AI Gives Great Answers. But Does It Have the Whole Story?
A polished answer can still be based on one side of a story, missing records or an outdated policy. For business AI, the next question is not only whether the answer sounds right—it is what the system actually knows.
Imagine it is late, an argument is still on your mind, and you ask an AI chatbot: “My partner did this. Am I being unreasonable?” The reply arrives quickly. It is patient, articulate and seems to understand. But the system has heard one account. It has not spoken to the other person, witnessed what happened or necessarily been told the surrounding history.
A SuperAging article republishes an essay by Maha Khawaja in The Conversation that uses this kind of moment to discuss the appeal and limits of AI advice. It is an illustration of incomplete context, not a survey measuring how many people use AI for relationship advice or evidence about the effect on every user.
The same gap appears in customer service, HR, procurement and operations: a system can recommend from supplied information while missing facts it needs.
Confidence is not context
A detailed answer can feel like an informed answer. Those are different things. For AI decision making, the key question is what information the system actually has. It cannot use a relevant fact that was never provided or retrieved, and its sources may be partial, stale, unclear or in conflict.
Consider these ordinary AI decision-support questions:
- “Should we give this customer a refund?” It may lack complaint history, current policy, contract terms or promises already made.
- “Should we employ this candidate?” It may see incomplete applications or unverified notes, not consistent job-relevant evidence.
- “Is this supplier safe to use?” It may summarise public claims without checking security evidence, contract terms or system access.
A well-written answer may rest on an unchecked assumption. Scattered policy, email and CRM records can carry that gap into an AI workflow.
Why a partial answer can be so persuasive
Several different risks can produce the same impression: an answer that sounds more certain and complete than its evidence warrants.
The user’s framing can shape the response
People naturally describe a problem from their own point of view and may include the facts that support their current interpretation. In a 2023 study, Anthropic researchers found sycophantic responses across five AI assistants on four text-generation tasks; their analysis also found that people were more likely to prefer responses matching the user’s views, including some persuasive but incorrect answers. That study describes particular models and tasks at that time—it does not show that every current assistant always agrees with every user. Read the Anthropic research.
A plausible answer may include invented or mistaken content
NIST’s Generative AI Profile uses the term confabulation for erroneous or false content that a system can present confidently. Fluent language is not proof that the model has checked the underlying facts, and a source citation is only useful when the source exists and actually supports the claim. See NIST AI 600-1.
A human reviewer can still over-rely
Human review helps only when the reviewer has enough context, time and authority to challenge the output. In a 2023 preprint, Vasconcelos and colleagues report five studies with 731 participants: task difficulty, explanation difficulty and incentives affected over-reliance in simulated maze tasks. An explanation does not guarantee a person will check. A “human in the loop” is not automatically an effective check. Read the open-access study.
These risks differ: a system may lack context, flatter a user’s view or apply a correct answer to the wrong case. Ask what evidence it used and what is missing.
For another practical treatment of checking outputs, see The More Authority You Give AI, the More Verification Matters.
The problem gets bigger when agents can act
A chatbot might recommend a refund. An AI agent may be able to issue one, change the customer record or send a message. Agents can use tools and connected systems to carry out multiple steps, so an incomplete assumption can become a real business action before a person notices it.
“Fix this problem” is not a permissions policy. Can the agent contact a customer, change a configuration, delete a record or commit money? Give it an identity and only the access its job needs. Gate consequential changes, log its actions, and provide a way to stop or reverse them.
The UK National Cyber Security Centre’s Guidelines for Secure AI System Development recommend secure defaults and least privilege, and call for riskier capabilities to be explained and opted into. These are useful design principles for business agents too—not a substitute for assessing the risks of a particular workflow.
For a concrete agent boundary checklist, read Your AI Agent Needs a Job Description — and Boundaries.
Human in the loop does not mean checking everything
Requiring manual approval for every AI step can remove the value of automation. The better approach is risk-based autonomy: let routine, low-impact work proceed within clear limits; involve a person when the consequence, uncertainty or exception warrants it.
Low impact: let routine help run
An assistant can organise meeting notes or categorise internal requests within an approved workflow. Use sensible quality checks, but do not make a person approve every harmless formatting step.
Moderate impact: allow bounded action
An agent may draft a customer reply or update a record when the source and policy are clear. Keep unusual requests, complaints, refunds or conflicting records out of automatic handling.
High impact: require accountable approval
Financial commitments, employment decisions, security changes, contract decisions, sensitive customer actions and deleting data need stronger evidence, named human review and a way to challenge or correct the result.
Unsuitable for independent action: narrow the job
Where a decision could seriously affect a person or the business and the evidence is incomplete, do not treat a confident answer as authority. Use AI to prepare information or options while an accountable person decides.
Set escalation triggers before launch: missing evidence, conflicting records, unusual requests, policy exceptions or sensitive information. Reviewers need to see the sources and be able to decline or change the proposed action.
Employment, financial, contractual and data-protection decisions can carry specific legal obligations. This is general operational guidance, not legal advice; check current requirements for your organisation and use case.
Give AI better context—without handing over everything
The answer is not “do not use AI”. Decision support needs relevant, current, approved information: the right policy, process or customer record. Retrieval can improve on general model memory, but a retrieved source can still be incomplete or out of date.
Nor should “give it more context” become “connect every system”. Additional data and permissions increase what the service can see and what could be exposed. Decide which sources are appropriate, confirm access rights, prefer the minimum information needed, identify which source takes precedence when records disagree, and tell the system to stop and ask when key facts are missing.
For personal data, verify the service, purpose, data categories, contracts and current requirements before connecting information. The ICO’s AI and data-protection guidance is marked as under review following the Data (Use and Access) Act, so check the latest ICO material rather than relying on a saved copy. See also Where Your Business Data Goes When Staff Use AI.
Ask five questions before letting AI decide
Use this five-part review when an AI system moves from generating ideas to supporting a business decision or taking action. Record the answer for each question; unanswered questions identify where the workflow needs more context, tighter permissions or human involvement.
- What does the AI know—and which source supplied it? List the prompt, documents, records and system results it actually used. Check their source, date and relevance.
- What relevant facts or perspectives might be missing? Identify missing people, events, policy, history, constraints or conflicting facts. Ask what would change the recommendation.
- What systems and information can it access? Name the data and systems in scope. Access should be limited to what the job needs, not everything the employee can see.
- What actions may it take without approval? Separate reading, summarising and drafting from sending, changing, approving, deleting or committing the business.
- Who can stop, review or reverse the outcome? Set clear reasons to escalate, name the person who can intervene, and decide how actions are logged or reversed.
Apply these questions to a standalone assistant, Microsoft Copilot, an automated service desk or an agent linked to business systems.
The relationship-advice lesson for business
AI can organise information, challenge assumptions, generate options and prepare a decision. In a business, it can help a team review evidence or prepare a recommendation.
But a chatbot cannot reliably fill in the facts it was never given, and an agent should not turn a missing fact into permission to act. More context can improve relevance; carefully limited access, risk-based approval and useful escalation protect the organisation while it does.
The question is not just “Can AI answer this?” Ask what it knows, what it is allowed to do, and who steps in when the story is incomplete.