Beyond the AI Echo Chamber: Why Different Perspectives Make Better Decisions

Introduction

Ask an AI assistant for an opinion, and it will often give you a very convincing answer. The problem is that a convincing answer is not necessarily the right one. When we use AI to test an idea, are we genuinely seeking a different perspective, or are we hoping it will confirm what we already think?

A useful response differs from a reassuring one. A confident explanation, neatly organised into headings and action points, can make an uncertain conclusion look settled. If we accept the presentation as proof, we risk moving straight to implementation without checking the premise.

For example, a business owner might ask why email revenue has dropped and suggest that subscribers are bored. An agreeable answer could recommend a new content calendar. Yet the real issue might be deliverability, broken tracking or a change in the mix of people receiving the messages. Before rewriting the newsletter, we should check what has actually changed.

Research into agreeable AI makes this more than a theoretical concern. A study published in Science tested 11 AI models and found that they affirmed users' actions 49% more often than human respondents in the scenarios examined. Experiments involving 2,405 participants also linked sycophantic responses to stronger convictions that they were right. These findings concern interpersonal judgements rather than marketing decisions, but they show why agreement shouldn't be confused with independent assessment.

Source: Science study

Why agreement isn't always helpful

Agreement feels productive. It can also be misleading. If our starting assumptions are weak, an assistant that follows them too readily may produce a polished recommendation built on shaky foundations. That is particularly risky when we are making decisions about customers, budgets or business priorities.

Take a campaign proposal that recommends doubling the paid media budget. It might sound sensible when the brief prioritises growth. But what if the sales team cannot respond to more enquiries quickly enough? What if the conversion rate has declined because the offer no longer matches customer expectations? What if the reporting counts duplicate leads?

An assistant may not raise these issues unless we permit it to question the brief. Try asking: 'What assumptions have I made? Which three would be most damaging if they were wrong? What information would change your recommendation?' You will not always like the answers. That is rather the point.

Research provides a useful warning here. In a 2025 Microsoft Research survey of 319 knowledge workers, greater confidence in generative AI was associated with less reported critical thinking. Higher confidence in their own ability to carry out a task was associated with more. This does not establish that AI causes weaker judgement, but it raises an important question: are we checking the work as carefully as we think?

Source: Microsoft Research study

Different perspectives, not just more voices

Asking five assistants the same question adds little value if all five work from the same, identical assumptions. A more useful approach is to give each perspective a clear purpose. One might examine the data, another the customer experience and a third the practical implications of implementation. Someone must then challenge the combined recommendation before treating it as a decision.

The distinction matters because a role label alone does not create expertise. Asking an assistant to 'be a finance director' does not give it access to your accounts. Asking it to 'be a customer' does not make its invented reactions genuine research. Specialist perspectives help structure questions, but they don't replace qualified people, reliable data, or direct customer feedback.

A better instruction is: 'Review this proposal from the perspective of someone responsible for delivery. Identify dependencies, likely points of failure and questions that remain unanswered. Do not invent evidence.' The response should reveal gaps that a human can investigate, not manufacture certainty.

The UK's AI Security Institute has also studied how question wording influences AI agreeableness, sometimes called sycophancy. Its experiments found that confidently asserting a belief could encourage agreement, whereas turning statements into questions helped reduce it. For example, instead of 'I know our prices are too high. Tell me how to cut them', try 'What evidence would help us establish whether price is the main reason customers do not buy?'

Source: UK AI Security Institute research

What constructive challenge looks like

Suppose website enquiries are falling. Marketing may want to invest in advertising, sales may question lead quality, and operations may point to follow-up delays. The answer likely won't come from a single dashboard. We need to understand which explanation the evidence supports, where information is missing, and what to test before committing money.

Imagine a business with 20% fewer enquiries than last quarter. One explanation is weaker search visibility. Another is a website change that makes the enquiry form harder to complete. A third is that the business is attracting fewer suitable visitors because its messaging has shifted.

A data-focused review could compare traffic sources, conversion rates and changes by landing page. A customer-focused review might test the enquiry journey on a mobile phone and examine questions raised in sales conversations. An implementation review could check recent website releases, tracking changes and response times.

Rather than asking AI to choose its favourite explanation, ask: 'What evidence would distinguish these possibilities, and what is the lowest-cost check we can make first?' You might discover that visitor numbers are stable while form completions have fallen after a redesign. That points towards a different response from increasing advertising spend. The figures here are illustrative, not a real case study.

The human still makes the decision.

AI can organise evidence, identify contradictions and suggest alternative explanations. It cannot take responsibility for the commercial consequences. People must decide which risks are acceptable, what matters to customers and whether a recommendation fits the organisation's priorities.

That does not mean people need to verify every sentence with equal effort. A draft headline and a recommendation to change pricing carry different levels of risk. The more consequential the decision, the stronger the evidence and review should be. If an assistant cites research, check that the source exists and actually supports the claim. Where customer or employee information is involved, use appropriate approved tools and follow your organisation's data protection rules.

It also helps to name the decision owner. Otherwise, an apparently thorough discussion can end with nobody accountable for acting on it or measuring what happens next.

The productivity argument is real, but it needs context. Harvard Business School's research with Boston Consulting Group involved 758 consultants and found that AI improved speed and work quality on tasks within its capabilities. The researchers also described a 'jagged technological frontier': AI could help with one task yet hinder performance on another. It is a useful reminder that AI's presence doesn't automatically make a recommendation sound.

Source: Harvard Business School research

A practical exercise

Try taking a decision you are considering and asking your AI assistant to approach it three ways: as a customer advocate, as a commercial analyst and as a sceptical reviewer. Ask each perspective what evidence it would need before agreeing. Then compare the points of disagreement rather than simply counting recommendations.

Here is a prompt you can adapt: 'I am considering [decision] because [reason]. Act first as a customer advocate and identify who might be disadvantaged. Next, assess the commercial case, including costs, benefits and assumptions. Finally, act as a sceptical reviewer. Challenge unsupported claims and list the evidence needed before proceeding. Keep facts separate from hypotheses.'

For a small business, the decision might be whether to stop offering a service that attracts enquiries but generates little profit. The commercial view may favour stopping it. A customer perspective might reveal that the service is an important first step towards more valuable work. A sceptical review would ask whether customer journey data and profitability data support either claim.

For agencies, in-house teams and independent marketers

Agencies can use structured challenge to test recommendations before presenting them to clients. In-house teams can examine assumptions across departments instead of optimising a single channel. Independent marketers can use AI to ask additional questions when a colleague or specialist isn't available. In every case, the benefit comes from better scrutiny rather than greater output volume.

A useful habit is to record the original recommendation, the objections raised and the final decision. Revisit the outcome after a suitable period. Did the challenge expose a genuine weakness? Was a concern overstated? What would you ask differently next time? This turns an interesting AI conversation into a repeatable improvement process.

Conclusion

The most useful question may not be 'What does AI think?' but 'What have I not considered?' Better decisions come from bringing evidence, different perspectives and human judgement together. AI should help us think further, not remove the need to think.