Six Modes of Decision-Making in the Age of AI
Discussion on how AI can be used in different ways within decision making, offering implications in each case, which require consideration before deploying.
AIEVIDENCE BASED MANAGEMENT
7/12/20269 min read
Introduction
There is a moment in a lot of working days right now where someone has a recommendation or decision to make. The data looks mixed. Time is short. The impulse is to use AI. What people are asking about less, though — and this is a very consequential question for any professional — is: what exactly is the job I've given to AI?
Was it retrieving facts? Making connections across a dataset? Weighing up options with its own training-derived perspective? Or was it simply pattern-matching to whatever the user asked, in whatever direction the question pointed? These are not equivalent activities, and they have meaningfully different consequences for the decision that follows.
The evidence on AI's strategic capabilities is genuinely mixed. A recent HBR article found that AI offered what researchers called "trendslop" when asked for strategic advice — confident-sounding output that recycled generic frameworks rather than reasoning carefully about the specific situation at hand (Romasanta et al., 2026). At the same time, Anthropic's own labour market research suggests AI is already beginning to reshape the work of knowledge professions in ways that are hard to ignore (Anthropic, 2026). Both things can be true simultaneously. What makes the difference, in large part, is how deliberately AI is being deployed.
In this article, I want to suggest a framework of six modes of AI in decision-making — a way of making that deployment more deliberate, and more honest about what is actually being asked of AI at each stage.
A Framework First
Before we get to AI, it helps to have a framework for what good evidence-based decision-making actually looks like. Most organisations don't have one, although they might have lots of policies, processes and frameworks that have been created, to different extents, with evidence in mind.
The worst-case scenario for bad decisions was outlined by Jeffrey Pfeffer and Robert Sutton in 2006, in a piece in the Harvard Business Review that opened with a memorable accusation: executives dose their organisations with "strategic snake oil" — fads, management theories, and confident assertions that have almost no basis in evidence. The problem, they argued, was not that managers were incapable of critical thought. It was that the environment and incentives didn't reward it.
Since around that time, a relatively small but serious field called Evidence-Based Management (EBMgt) has developed a more rigorous response. The Centre for Evidence-Based Management, working from the foundations laid by Barends and colleagues, propose a modest but challenging idea: good decisions should draw conscientiously on the best available evidence from four distinct sources.
Scientific evidence: peer-reviewed studies, meta-analyses, and systematic reviews that bear on the question at hand. Not all research is equal, and knowing how to evaluate it critically is a real skill, requiring some pre-existing knowledge.
Local or organisational evidence: your own data, internal metrics, customer feedback, and operational records. This provides the local reality of where the decision will be implemented.
Experiential evidence: the practitioner's judgement accumulated over years of working on similar problems, synthesising theories with real-world observations and considering how people actually behave when different decisions are put in motion.
Stakeholder evidence: the values, concerns, and preferences of the people who will be affected most by the decision. In a commercial setting, it is often the user or customer. In a medical setting, it is the patient and their family.
We see this play out when getting care at a hospital. Doctors consult the medical evidence on treatment efficacy. Hospital data provides available care options. Practitioners know when to escalate to specialists. And the patient and their family inform doctors about medical history and current coping challenges. Each source does a different job; none is sufficient alone.
How do you work through these sources? One useful process was set out as the Six As (Barends et al.):
Ask a clear, specific, researchable and useful question. Acquire the relevant evidence. Appraise its quality and relevance. Aggregate across the different sources. Apply it to the decision. Assess the outcome to learn for next time.
This is a continuous learning loop. A learning organisation sees each iteration not as a journey toward a definitive answer, but as an opportunity to improve the quality of the next decision.
None of this is revolutionary. What is remarkable is how rarely it is done, even partially, given the value it can offer.
Enter AI
AI has arrived into this process with immense capability, but introducing attendant risks. Sometimes the authority with which it makes recommendations or assertions is justifiable, and at other times it is not. There is no widely agreed map of where AI is supposed to sit in a process of evidence-based decisions.
Consider six distinct roles AI can play in a decision. To make this more tangible, imagine a UK insurance company considering whether to launch a new product for the over-50s home insurance market. A leadership team member opens their AI tool. What, exactly, are they about to use it for?
Mode 1: Researcher. In its most basic function, AI retrieves and summarises information. "What is the approximate size of the UK over-50s home insurance market?" This is a useful and relatively low-risk application, provided the user understands that AI-generated facts require verification — which is possible because the data has a source and a date.
Mode 2: Commentator. Here, AI moves from retrieval to perspective. "What are the main risks and opportunities in launching a home insurance product for over-50s?" Now AI is drawing on its training to offer its own assessment, blending facts with interpretation. This can be valuable, but it is the AI's view, shaped by everything it has been trained on, not a neutral read of the specific situation.
Mode 3: Analyst. Now the user uploads their own data — customer data, claims rates, renewal patterns, competitor pricing, survey research — and asks AI to work across it. "Given this data and the broader market context, what does the data show are the opportunities and risks for launching a new product?" AI is now combining its general knowledge with specific organisational evidence. This is the first mode where the quality of your inputs begins to drive the quality of the outputs as much as the capabilities of the AI itself.
Mode 4: Partner. In this mode, AI shifts from answering questions to shaping the direction of the enquiry. It coaches the process, returning prompts such as: "Have you considered how the regulatory environment might affect this? Is your question actually the right question?" The AI here is functioning more like an intelligent colleague who has read everything and is challenging your assumptions. Some emerging AI platforms are beginning to offer this as a distinct mode of interaction, separate from standard chat. This is arguably where AI can add the most distinctive value to evidence-based decision-making — not by having the answers, but by improving the quality of the questions.
Mode 5: Consultant. Asked to go further, AI synthesises everything and offers a recommendation. It churns through data, requests more, coaches the refinement of questions, and then says: "Based on the available evidence, I would recommend launching in Q1 with a focus on water damage cover and an initial focus on a digital-only journey for 50-60 year olds who are already confident with online applications." This is decision-support you might expect from a marketing consultant. It is where the stakes in getting it wrong are highest, and where delegation to AI requires trust. The decision is still a human's to take, but the direction of travel has been recommended.
Mode 6: Decision-taker. This is the frontier, and it is arriving faster than most organisations are ready for. Here, AI executes: pricing decisions made algorithmically, claims assessed and settled by automated systems, product withdrawals triggered by performance thresholds. AI writes the marketing copy and automates the CRM campaign. Human oversight doesn't disappear, but it becomes supervisory rather than active. This mode requires a governance framework — with a human in the loop — that most organisations have not yet fully built.
The point of this spectrum is not that higher modes are better. From Mode 1 to Mode 6, we benefit more from AI's speed and processing power, at the cost of losing transparency and control. Different jobs, at different moments, will tend to lead toward different modes. Pressure and culture can adversely affect the choice. Deploying the wrong mode carries real risk.
Crucially: if the choice is not explicit, AI will make the choice for you
Choosing Your Mode
So what mode should we deploy? A useful way to think about this is through five questions.
Is the decision repeatable rather than novel? AI tends to be on safe ground for repeatable, lower-stakes actions where we know what outcomes we want, or at least the characteristics of a good outcome. AI can correct our writing, translate it, generate lists of ideas, extract themes from longer texts, and accelerate information retrieval of the kind we might previously have done on a search engine. We can effectively set rules for these decisions: if a customer seems distressed on web chat, escalate their call to an agent automatically. Modes 5 and 6 are often applicable here. Where the decision is genuinely novel — involving conditions, combinations, or consequences that don't map cleanly to historical patterns — human judgement becomes harder to substitute and the upper limit on the appropriate mode drops accordingly.
How high are the strategic stakes? Decisions with major consequences for customers, employees, or organisational direction need more human engagement. Mode 4 is often the upper limit, and there needs to be not only oversight but immersion in the loop. Complex new considerations may enter the frame. Human values and ethics are as important as transaction data to an organisation's strategy. It is not that we expect AI to get all the answers wrong in high-stakes situations. The concern is that if it does get things wrong, the costs are too high to treat it as merely a system error.
How much does this decision affect real people? Some decisions fall directly on individuals: employees facing restructuring, customers being declined for cover, or being redirected to communication channels that present barriers to accessing services. These present ethical dimensions that still require human judgement and human empathy to manage and resolve. We might operate at Mode 2, while using Mode 4 to augment our thinking.
How complex, ambiguous, and stable is the information? These are three separate considerations, but all concern the risk that the data AI uses can mislead when applied to our decision. Where data is contradictory, incomplete, or rapidly changing, AI's confidence can be at its most dangerous. Organisations' survival strategies changed rapidly during COVID as new rules emerged and consumer demand shifted radically. Human appraisal was irreplaceable in that environment. Today, a volatile geopolitical and financial landscape similarly means that AI trained on a decade of trend data is not necessarily calibrated for conditions that represent genuine outliers.
What are the hidden effects of AI on the humans working alongside it? This may be the least obvious consideration, but research suggests it is among the most important. Work published in the Proceedings of the National Academy of Sciences in 2024 (Treiman et al.) found that when humans know they are labelling decisions to train an AI system, they change their behaviour — subtly enough that neither they nor the organisations relying on the resulting AI tend to notice. Imagine a team of experienced credit analysts labelling historical loan decisions to train an underwriting model. In their normal working lives, these analysts exercise nuanced judgement: approving borderline cases, declining applications that look fine on paper but trigger professional instinct. When they know they are training an AI, something shifts. They gravitate toward safe and defensible options. Hard cases — where experienced judgement was previously critical — get resolved more conservatively, because nobody wants to teach an AI to take risks they cannot explain later. The AI, trained on this modified behaviour, learns a subtly different version of human decision-making. Its output will match the data it was trained on, and so will look correct and human-like — until compared against decisions made by humans working in non-training conditions. Many organisations will never make that comparison. Similar dynamics have been observed in medical imaging and hiring. The act of being observed and recorded nudges human behaviour toward safety, even when the best real-world decisions are a function of risk and reward, not of defensibility alone. A related point comes from Steyvers and Kumar's 2024 work on human-AI complementarity: the conditions under which AI genuinely improves human decisions are not automatic. They depend on whether the human has an accurate mental model of what AI is and isn't good at — a mental model that takes time, effort, and deliberate calibration to develop, and that most users do not yet have. Both findings point in the same direction: the risks of AI in decision-making are not only in what AI does, but in what it does to us.
In Sum
AI can genuinely improve the quality of evidence-based decisions, but only if the human has specified its role. One of the conditions for it adding value is that the mode of application is chosen deliberately. Only if users understand which mode they are in, what the AI is doing with their question, and where its limitations lie, can they optimise AI's contribution across both risk and reward.
That understanding does not come simply from becoming proficient with any one tool, though proficiency helps. It requires the same discipline that Evidence-Based Management has always called for: Ask. Acquire. Appraise. Aggregate. Apply. Assess.
The question is not whether to use AI in your decisions. It is how.
Sources and Further Reading
Anthropic (2026). Labour market impacts of AI: A new measure and early evidence. https://www.anthropic.com/research/labor-market-impacts
Barends, E., Rousseau, D., & Briner, R. (n.d.). Evidence-Based Management: The Basic Principles. Centre for Evidence-Based Management.
Booyse, C. & Scheepers, C. (2024). Barriers to Adopting Organisational Decision Making Through the Use of AI. Management Research Review.
Cheong, M. (2024). Transparency and Accountability in AI Systems. Frontiers in Human Dynamics.
Pfeffer, J. & Sutton, R. (2006). Evidence-Based Management. Harvard Business Review, January 2006.
Romasanta, A. et al (2026). Researchers Asked LLMs for Strategic Advice. They Got "Trendslop" in Return. Harvard Business Review. https://hbr.org/2026/03/researchers-asked-llms-for-strategic-advice-they-got-trendslop-in-return
Rynes, S. & Bartunek, J. (2017). Evidence-Based Management: Foundations, Development, Controversies and Future. Annual Review of Organizational Psychology and Organizational Behavior, 4.
Steyvers, M. & Kumar, A. (2024). Three Challenges for AI-Assisted Decision-Making. Perspectives on Psychological Science.
Treiman et al. (2024). The Consequences of AI Training on Human Decision Making. Proceedings of the National Academy of Sciences.
These are the author's personal thoughts and views, and do not necessarily reflect those of any organisation by which the author is employed. This piece is offered as a set of considerations, not recommendations.
