7 more thinking traps — in AI, strategy and marketing.

What this is about

In Part 1 we discussed the six classics in SME digital business — overconfidence, sunk-cost traps, confirmation bias, authority bias (classic and AI-related), the fascination with winner stories and the confusion of cause and effect. These are cognitive biases we know from everyday life that regularly hit hard during digital transformation.

Part 2 goes one level deeper. Here come the more specific traps that appear particularly often in strategy, marketing and AI contexts — and one new bias that only became a mass reality with Generative AI in 2023–2026. These exact patterns accompany us in almost every one of our transformation mandates because they rarely sit in the technology — they sit in the thinking.

Seven cognitive biases, same four-block structure as in Part 1: What it is → Why it happens to us → Examples from business & SME practice → How to prevent it.

7. Incentive sensitivity (Incentive-Superresponse Tendency) — when the reward system swallows the strategy

What it is

As soon as a clear incentive is defined in an organisation — bonus, commission, KPI target, OKR — behaviour adjusts unconsciously. What gets incentivised gets done, often at the expense of things that would actually matter more in the background. Incentives act more strongly than they should at first glance.

Why it happens to us

Incentives speak directly to the reward system, not to the conscious, strategic mind. People do what they are rewarded for — even if the strategy slide says something else. This gap between stated and rewarded priority causes many strategies to fail without anyone noticing. A sober framework like the market-product matrix helps to align goals and incentives cleanly.

Examples from business & SME practice

• Banking bonuses for loan closings: When bonus payments are tied to the number of loans closed, employees grant more loans — including ones that are not sustainable long term. There was never bad intent behind the incentive. Still, the incentive system helped trigger the 2008 subprime crisis. That is incentive-superresponse tendency in its purest form.
• Delivery date as the sole incentive in software projects: When a software team is primarily judged on hitting deadlines, other measures suffer silently — test coverage, user experience, maintainability, attention to detail. The product ships on time but accumulates technical debt over the following months that costs three times the time saved.
• Galaxy Note 7 (2016): Samsung launched the new flagship as usual in autumn — release plan held, all incentives met. What followed shortly after: exploding batteries, worldwide recalls, billions in damages. It remains an open question whether the uncompromising adherence to the release plan was part of the cause. It seems plausible.
• OKRs without a counter-metric: A management team sets "growth of the user base" as the central metric. What they do not set: an accompanying retention metric. Result: the team optimises for acquisition, ignores churn, and after twelve months the user base is larger than ever — and at the same time less profitable.

How to prevent it

• Always think incentives in pairs. One quality metric per growth metric. One stability KPI per speed KPI. One "depth of X" per "amount of X"
• Derive team and personal goals from company goals — not the other way around
• Incentive audit twice a year in the leadership team: which incentives have we set formally and informally? What behaviours actually emerge from them? What unintended side effects do we see?
• Critically question bonus structures that are based on a single metric. That is almost always a trap

8. Outcome bias — when the result distorts the evaluation of the process

What it is

We judge the quality of a decision by its result, not by the process that produced it. A decision that was competently made with the information available at the time but turned out unfortunate looks like a mistake in hindsight. A decision made recklessly but with a lucky outcome looks like genius in hindsight. Both views are distorted.

Why it happens to us

Our brain loves clean retrospective stories. Success is declared in hindsight to be the logical consequence of "right" decisions — even though the same decisions could just as easily have led to failure. There is also an entire market of success biographies that caters exactly to this bias.

Examples from business & SME practice

• The 10-apps thought experiment: Imagine ten teams each get one million francs to develop successful apps. After two years, five have failed, two are limping along, two have reached break-even and one is an international hit. Immediately media and analysts begin dissecting the success recipe of the one app. Yet we know from probability theory: with ten attempts it is statistically likely that some will succeed, even if all ten teams had the exact same decision-making process. The decision process therefore cannot be judged by looking at the success.
• Dating apps: One of the most crowded app categories of all. Millions of users worldwide, but only a handful of platforms are actually successful. Hundreds of similarly designed apps "rot" away in the app stores. Anyone who analyses Tinder's success without looking at those that failed is doing outcome bias combined with survivorship bias from Part 1.
• AI startup successes 2025/2026: A few visible winners (OpenAI, Anthropic, Cursor, Lovable) spawn an entire consulting industry selling the AI success recipe. No one really knows how much of it was talent, how much market timing, how much luck. What gets sold is still a "recipe".

How to prevent it

• Never judge an already-taken decision by its result alone. Ask instead: "Did we have a good decision-making process with the information available?"
• Keep a decisions journal: For important decisions, write down — what we knew, which assumptions, which alternatives, which expectation. Process versus result can only be separated in hindsight if the process was documented
• Look at success cases critically: For every success example studied, study at least three comparable failures. Anyone who only sees the winners calibrates wrong

9. Illusion of control — when we believe we control the uncontrollable

What it is

We systematically overestimate our influence on outcomes that are essentially determined by factors outside our reach. A well-known test: people throw a die harder when they want a high number and softer for a low number — as if throwing strength would influence the result.

Why it happens to us

The experience of control reduces anxiety and gives psychological safety. If the world were too random, it would be cognitively almost unbearable. So we attribute more influence to ourselves than we have. In digital projects and marketing this regularly leads to forecasts and plans that suggest more certainty than they can deliver. It gets more honest when you base forecasts on predictive data analysis instead of gut feeling — including openly communicated uncertainty.

Examples from business & SME practice

• "Will my idea be successful?": One of the most frequent enquiries we get at AHEAD OF TIME. Clear answer: in a fifteen-minute initial call you cannot judge that, and you cannot after three hours either. What we can deliver is a subjective assessment of the idea. What we cannot deliver is a guarantee. Anyone who offers that is practising the illusion of control at the customer's expense.
• Quarterly forecasts with three decimal places: A management team presents a quarterly revenue forecast to the board with the precision "12,437,500 CHF" — in a market that fluctuates by double-digit percentages monthly. The precision is theatre. It suggests a control that does not exist.
• A/B-test interpretation: Marketing teams announcing a "clear winner" after four days of A/B testing without checking statistical significance. What they interpret as insight is mostly chance. Real insight needs sufficient sample volume and Bonferroni correction — both rarely given.

How to prevent it

• Let it flow. What we cannot influence we should not artificially try to control. Energy instead goes into what is actually influenceable
• Test hypotheses early. Validate assumptions with prototypes, user research, small pilot setups before large investments flow
• Use iterative user-centered design processes. Co-creation with the target audience from the concept phase prevents building for months on a solution that nobody wants
• Understand the difference between influenceable (e.g. own communication, product quality) and not-influenceable (e.g. market movements, competitor actions). Direct energy according to influenceability

10. Cognitive dissonance — when a wrong decision suddenly becomes a learning success

What it is

We psychologically struggle to admit mistakes. Instead, we spin reality until it fits our self-image. A clearly failed project becomes a "valuable learning experience", an obviously wrong tool purchase becomes a "strategically important intermediate step". Self-deception protects against pain but blocks real learning.

Why it happens to us

Aesop's Greek fable nailed it 2,500 years ago: a fox tries several times in vain to reach juicy grapes and finally decides that the grapes are "not ripe yet anyway" and therefore sour. The brain protects the ego by reinterpreting reality. In modern psychology this is called cognitive dissonance.

Examples from business & SME practice

• The "ignored gut feeling" project: It happens to me too on a regular basis that I take on a project despite a bad gut feeling. Amazingly often the gut feeling is not wrong. Friction arises, misunderstandings, frustration in the team. Then comes the classic debrief — "But we really learned a lot." When the actual insight was: "We should not have taken on this project." The more honest version would have helped next time.
• AI pilot projects 2024–2025: "Our AI proof-of-concept didn't scale, but we learned a lot about AI." Sometimes that is true. Very often it is whitewashing — the project failed, and without honest analysis of why, the pattern repeats in the next PoC.
• Wrong tool choice: "HubSpot was actually the wrong choice, Salesforce would have been better — but HubSpot does the job too." This cognitive dissonance prevents both honest reflection and timely correction

How to prevent it

• Don't allow yourself small lies. Name mistakes openly — to yourself and in the team. It hurts at first, but it builds trust and depth of learning
• Clear success definitions before project start. So the goal window doesn't get shifted afterwards to reframe failure as success
• Establish a real error culture. Pre-mortems, post-mortems, blame-free reviews. Mistakes become learning material, not subjects of shame
• "Yeah, now I've learned something!" mentality. A mistake is fuel for growth — but only if it can be called a mistake

11. Anchoring effect — when the first number distorts everything

What it is

We let ourselves be guided unconsciously in decisions by an anchor — usually a number, a first value, a first piece of information heard. Even when we know the anchor was random or arbitrary, it measurably influences our later judgement.

Why it happens to us

The brain uses mental shortcuts to reach decisions quickly. The first value mentioned becomes the reference point, and our later judgement deviates from it only in small steps. Even with explicit explanation of the effect, it does not go away — we cannot switch it off, we can only work around it.

Examples from business & SME practice

• Negotiations in oriental markets: Anyone who has haggled in a bazaar in Istanbul or a souk in Marrakech knows it: the first price named by the seller is virtually never the real value, but a deliberately high anchor. Even those who haggle down 50% probably still end up above the fair price — because the anchor shifted the entire playing field.
• Estimates in software projects: When a team learns before the estimation that the customer has "around 50,000 CHF budget", the estimate very likely lands exactly there. Had the team estimated unburdened, the value would probably have been different. Same when iOS developers see the Android hours already booked in the estimation sheet — their own estimate is unconsciously pulled toward the Android value.
• AI tool pricing 2025/2026: OpenAI set a market anchor with ChatGPT Plus at USD 20/month. Competitors price around it, even if their value proposition deviates significantly. Anyone who would need to charge USD 50/month doesn't dare — the anchor distorts the entire market.

How to prevent it

• Never communicate budgets before effort estimates. The estimation team needs unburdened numbers
• Let disciplines estimate independently. iOS and Android estimates should not see each other
• Multiple estimators in parallel, independently of each other. For small differences: use the average. For large differences: discuss together where the discrepancy comes from
• Actively seek second opinions without communicating your own position in advance. A second opinion with a pre-briefing is not a real second opinion

12. Scarcity fallacy — when scarce appears valuable

What it is

Rara sunt cara — rare things are precious. This ancient heuristic is deeply anchored in us: what is scarce seems to have more value, even when objectively it does not. Marketing professionals know the effect and use it systematically. Understanding when we fall for it — and how we can use it ourselves — is mandatory reading for every marketing manager.

Why it happens to us

Evolutionarily, scarcity was a reliable indicator of value. Little food, few partners, little safe territory — what was scarce was worth securing. We continue to use this heuristic in modern consumer contexts, although the conditions today are completely different. Modern marketing uses this systematically and mostly successfully. In our SME marketing mandates we make sure to use such effects honestly — as real information, not as manipulation.

Examples from business & SME practice

• Artificial beta scarcity in a game launch: A new mobile game is to be launched successfully. Instead of giving all interested parties beta access, marketing limits the beta to 50 exclusive slots. Plus a live counter on the landing page: "Only 17 slots available". The scarcity creates sign-up pressure that would not arise with generous beta access.
• Hotel booking platforms: "Only 1 room left" and "2 other people are looking at this offer right now" are classics. Inner stress, a feeling of "now or never" — and the finger clicks faster than the mind can ask "do I really need this?"
• Black Friday, Cyber Monday, Singles' Day: Time-limited special offers that suggest scarcity by printing a counter ticker or hours display. In fact, many offers are available longer and better than the promotion promises
• Tech launch patterns: "Early access programmes", "invite-only beta", "first 1000 subscribers only" — Anthropic, OpenAI, Lovable and many others deliberately use the pattern to maximise initial attention

How to prevent it

• Before any purchase impulse ask: "Would I buy this if it were available without limit?" If no — scarcity fallacy detected
• Define your own purchase criteria and apply them consistently. Scarcity is not a purchase criterion, except for collectibles or investments
• Add a time buffer before larger purchases. For non-perishable goods, wait 24–48 hours. Artificial pressure dissipates surprisingly quickly
• Learn to actively recognise scarcity triggers: countdown timers, "last 3 in stock", "3 people are viewing this" — register consciously and decide whether it is manipulation or real information
• Conversely: use deliberately and honestly. Anyone selling something can use this effect constructively (beta limits, FOMO marketing for workshops). But consciously and not manipulatively — otherwise trust suffers in the medium term

13. AI hallucination — the cognitive bias of AI itself

What it is

In Part 1, bias 4, we covered how we blindly trust AI tools (authority bias for AI). Now the other pole: the bias AI itself makes. Generative AI systems like ChatGPT, Claude or Gemini produce content with disturbing self-confidence that sounds plausible but is factually wrong — hallucinations. Invented quotes, non-existent studies, wrong legal paragraphs, wrong product features.

Hallucinations are not a bug that can be "fixed". They are a structural feature of generative language models.

Why does this happen? (From the AI's perspective)

• Probabilistic architecture: language models predict the next token based on probabilities. When the probability is too uncertain, the token still falls somewhere — and the sentence becomes a hallucination
• Training pressure to answer: models were predominantly trained to answer, not to stay silent or communicate uncertainty. "I don't know" was significantly rarer in the training corpus than plausible (but wrong) answers
• Knowledge cutoff: what happened after the training date the model does not know — but cannot reliably recognise this itself and sometimes mixes old and current knowledge

Examples from business & SME practice

• Invented studies in AI research: ChatGPT, Claude and Gemini regularly invent studies — with correctly formatted authors, plausible journals and even DOI numbers that exist nowhere. In scientific texts, media articles and strategy consultations, this leads to embarrassing corrections when the output is taken over unchecked.
• Air Canada chatbot ruling 2024: The airline's customer service chatbot invented a bereavement refund policy that never existed. A customer sued successfully. The court decided: if your chatbot promises something, it counts. Air Canada had to pay. A cautionary tale for anyone considering a chatbot in customer communication.
• Code hallucinations: AI coding tools invent library functions, methods or API parameters that do not exist. The code looks "right" but crashes immediately on execution. Extremely hard for inexperienced developers to detect
• Invented people and sources: "As Prof. Schmidt of ETH Zurich explains in his book 'Digital transformation for SMEs' (2019)..." — Prof. Schmidt doesn't exist, the book doesn't exist. Sounds convincing. Still ends up in strategy decks and sales pitches

How to prevent it

• Hallucinations as default assumption. Don't assume AI output is correct — assume it could be correct, and verify
• RAG (Retrieval Augmented Generation) for use cases with factual claims. The model is forced to answer from a real, vetted knowledge base — not from inside the model. Reduces hallucinations drastically
• Check sources instead of trusting blindly. Every link, every study, every quote — click, read, reconcile
• Keep temperature low for factual tasks (0.1–0.3 instead of 0.7). Low temperature = less creativity, but also fewer hallucinations
• Multi-model cross-check for critical statements. Ask the same question of Claude + ChatGPT + Gemini. Significantly different answers are worth deeper checking
• Use web-search-capable models deliberately. Perplexity, ChatGPT with Search, Claude with web tool hallucinate less on current topics because they pull real sources. But beware: the AI's source selection can also be unreliable
• Clarify chatbot responsibility in the company. If your chatbot hallucinates, your company is liable (Air Canada case). Guardrails, escalation paths, regular spot-checks are not optional — they are mandatory

Conclusion — the 13 biases at a glance

Here again all 13 traps from both parts — compact, with one self-reflection question per bias that helps recognise it in your own work day.

Part 1 — the six classics

1. Overconfidence bias — Did I estimate realistically, or does my plan reflect the wish scenario?
2. Sunk cost fallacy — Am I evaluating this decision really from a future perspective — or do past investments have a say?
3. Confirmation bias — Am I actively looking for evidence against my thesis — or just collecting confirmations?
4. Authority bias (classic + AI) — Am I trusting this authority (human or AI) because the argument is sound — or because it sounds convincing?
5. Survivorship bias — Have I also studied the failures in the same market — or only the winner stories?
6. Swimmer's body illusion — What was cause here, what was effect? Have I checked both directions?

Part 2 — the seven advanced

7. Incentive sensitivity — What behaviours do my incentives actually produce — and what unintended side effects?
8. Outcome bias — Am I judging this decision by the process or by the result?
9. Illusion of control — Which factors really influence the result — and which can I actually steer?
10. Cognitive dissonance — Am I honest with myself about where I was wrong? Or am I crafting a story where I was right?
11. Anchoring effect — Was I influenced by a first value without noticing it? Who set the anchor?
12. Scarcity fallacy — Would I also buy this if it were available without limit?
13. AI hallucination — Where might the AI be hallucinating right now? Have I independently verified the critical statements?

Closing thoughts

Everyone makes cognitive biases all the time. Anyone who claims otherwise has probably just committed bias 1 (overconfidence). Even after this post, mental traps will continue to catch us — the system can't be switched off. What does change through knowledge: we recognise patterns earlier, ask ourselves more critically, and set up our decision processes so that the typical distortions cause less damage. Often a sober outside view helps, as our VCS case as a pure screening mandate illustrates.

In my view, three practical levers are particularly effective: consistent self-reflection instead of quick judgements, feedback from diversely composed teams instead of echo chambers, and actively seeking counter-opinions and disproving data instead of comfortable confirmation. This kind of critical sparring is exactly what external advisory engagement brings into a team.

New since 2024: with AI tools we have not only received a reflection mirror for our biases, but also a sparring partner that brings its own distortions. Anyone who understands both — their own biases and those of the AI — has a clear advantage in SME digital business 2026.

Q&A — the most frequent questions on cognitive biases in practice

Which bias hits SMEs most frequently in 2026? Currently in consulting practice: cognitive dissonance in AI PoCs (failures get reinterpreted as learning successes instead of being honestly analysed) and the anchoring effect in effort estimates (budgets are communicated too early). Both are less spectacular than authority bias around AI, but statistically more frequent.

Will AI models soon train away hallucinations? Probably not completely. They will become less frequent, less plausible and contained in many use cases through RAG setups. Structurally, however, hallucination is part of the architecture of generative models — plan for it long term.

How do I have the cognitive-bias conversation with management? With concrete examples from your own company, not with abstract theory. "We did X last year — which bias from this list probably caught us?" works significantly better than "We should discuss our cognitive biases." How such reflection can be conveyed didactically is shown by our HSO digital case from higher education.

How do I distinguish authority bias around AI from healthy scepticism? Rule of thumb: if you simply adopt AI output without checking it — authority bias. If you treat it as a hypothesis and verify — healthy practice. The decisive point is the verification step.

Are there tools that automatically detect AI hallucinations? Sort of. Tools like Galileo, Patronus, Confident AI offer hallucination detection for AI applications. They are not yet perfect and do not replace human verification. For SMEs for now: healthy mistrust, multi-model cross-check and RAG for factual claims.

Where to read next

• Part 1: the six classics in SME digital business → in case you haven't read yet — overconfidence, sunk cost, confirmation, authority (classic + AI), survivorship and swimmer's body illusion
• AI agent practice series for SMEs → four posts on AI agent automation — directly usable despite hallucination risks
• Multi-touch attribution for SMEs → structured evaluation of marketing success, beyond outcome bias

Sources and further reading

• Dobelli, R. (2011). Die Kunst des klaren Denkens. Hanser.
• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
• Cialdini, R. (1984). Influence: The Psychology of Persuasion. Harper Business. — Klassiker zu Knappheitsirrtum und sozialer Beeinflussung.
• Mata v. Avianca, Inc. (2023). 22-cv-1461 (S.D.N.Y.). — US-Anwaltskanzlei-Case zu ChatGPT-Halluzinationen vor Gericht.
• Moffatt v. Air Canada (2024). 2024 BCCRT 149. — Air-Canada-Chatbot-Urteil zur Haftung bei KI-Halluzinationen.
• Anthropic (2024). Claude Documentation. — Technische Grundlagen zu LLM-Halluzinationen und RAG.

This list covers the primary sources for the concepts discussed in this article. It is not exhaustive but should provide an entry point for those interested in going deeper.