The Bias You Don't See Is the One Making Your Technology Decisions
AI amplifies the biases of the people using it, on top of training data bias (HBR, 2026). In Tversky and Kahneman's data a loss weighed about 2.25 times an equal gain. The cognitive shortcuts that derail technology decisions, and the structured processes that counteract them.
Harvard Business Review published a finding in January 2026 that stopped me mid-scroll: AI amplifies the biases of the people using it, on top of whatever bias sits in its training data. The tools carry biases of their own, and the operators add more. The space where those two meet is where most technology decisions go wrong.
I studied this problem from the behavioral science side before I ever touched a line of code for a client. Five certifications in psychology, consumer behavior, and NLP later, the pattern is always the same: the decision to adopt, reject, or ignore a technology is rarely about the technology. It's about the cognitive shortcuts the decision-maker is running, and whether anyone in the room knows how to spot them.
The short version: every technology decision passes through a handful of predictable biases. Loss aversion keeps broken tools in place, confirmation bias decides the vendor before the demo starts, the bandwagon effect turns a competitor's press release into your roadmap, and status quo bias protects whatever is already running. AI adds one more, automation bias (NIST calls it automation complacency), where people stop checking a system because it sounds sure of itself. You can't remove bias from people. You can design a process that makes bias visible:
Write the problem and the success measure down before you look at any product.
Score vendors against criteria written before the first call, so confirmation bias has less to work with.
Run a premortem: assume the project failed and list why.
Give one person real authority to argue against the decision.
Keep a decision journal and reread it in 90 days.
Kahneman and Tversky's prospect theory (1979, measured in their 1992 follow-up) found that people feel a loss about twice as intensely as an equivalent gain. In technology, this means teams tolerate broken tools rather than face the perceived loss of switching.
Loss aversion is the silent budget killer
Kahneman and Tversky's prospect theory (1979) demonstrated that losses loom larger than gains, and their 1992 follow-up in the Journal of Risk and Uncertainty put a number on it: the median participant weighed a loss about 2.25 times as heavily as an equivalent gain, roughly twice as intensely. In technology decisions, this shows up as the sunk cost trap: continuing to pay for software that doesn't work because switching feels like admitting the original purchase was a mistake.
Kahneman and Lovallo's 1993 paper in Management Science spelled out what that bias does to organizations: loss aversion favors inaction over action and the status quo over any alternative. Teams will tolerate a tool that costs them hours every week in workarounds rather than face the perceived "loss" of migrating to something better. The math rarely supports this, but the psychology usually wins.
The cost hides in plain sight because nobody adds it up. Brian wrote about this in the subscription trap, where he works out what each software subscription costs per billable hour and cuts the tools that don't earn their place. Doing that arithmetic counters the bias directly, because it turns a feeling of loss into a number you can compare with the gain.
This is why the problem usually isn't you: much of the weight comes from decisions made before you arrived. The software was chosen by someone who's no longer there, and nobody wants to be the person who says the emperor has no clothes.
The Behavioural Insights Team has observed that business leaders often respond to cognitive bias concepts with nodding and passive agreement, but no follow-up. The bandwagon effect operates the same way in technology adoption.
Confirmation bias turns vendor demos into theater
Once a decision-maker has a preference (even a subconscious one), they filter every piece of incoming information to confirm it. Columbia University's risk management research (November 2025) found that traditional enterprise risk frameworks "assume rational decision-making, even though real-world outcomes are shaped by cognitive heuristics and behavioral tendencies."
In practice, this means the vendor demo is already decided before it starts. The buyer liked the sales rep. The logo looked enterprise-grade. A competitor uses it. Everything after that is confirmation theater: the team watches a curated walkthrough and nods because the comparison already happened in their head, not on a spreadsheet.
More demos won't fix this. Structured evaluation criteria written before the first call will, something I covered in depth in how to actually evaluate a technology vendor. When the rubric exists before the pitch, confirmation bias has less room to operate.
The bandwagon effect scales with your board deck
The Behavioural Insights Team (BIT) published a 2024 piece by Scott Young observing that business leaders often respond to behavioral science concepts with "nodding and passive agreement (and a lack of follow-up)." The same pattern applies to technology trends. A founder reads that their competitor adopted a specific platform. The board asks about AI strategy. Suddenly the urgency is artificial, and the decision is about optics rather than operations.
This is the bandwagon effect at enterprise scale. It's why AI suggestions sometimes make products worse: the technology isn't bad, but the adoption was driven by social proof rather than operational need. The question that cuts through it is simple: "What specific problem does this solve that we measured before we started looking for solutions?"
Brian's piece on the agentic hype ladder is a useful check on bandwagon bias. He argues that most companies announcing "agentic AI" are on the first two rungs, and that the value sits in well-defined workflows. When the board deck says everyone is doing it, ask which rung "everyone" is actually on.
An NBER 2026 working paper found that LLMs exhibit systematic behavioral biases in economic decisions, and that prompting them to 'make rational decisions' measurably reduces those biases. The same principle applies to human operators.
Status quo bias protects bad architecture
Samuelson and Zeckhauser's 1988 research on default bias demonstrated that people overwhelmingly accept the default option, even when alternatives are objectively better. In technology, the default is whatever's already running. And the longer it runs, the harder it becomes to change, regardless of whether it's the right choice.
I've seen this with tech stack decisions dozens of times. In most companies the stack was inherited rather than chosen, and inherited architecture carries the biases of every person who touched it: the developer who preferred a specific framework, the manager who liked a particular vendor's pricing model, the CTO who left two years ago but whose preferences are still running in production.
Status quo bias is the reason technology fragmentation compounds silently. Nobody chose to have four overlapping tools. They just never chose to consolidate, and the default won.
Columbia University researchers recommend premortems, assigned devil's advocates, and scenario analysis as practical tools against cognitive bias in enterprise risk management.
Which biases most affect digital decision-making?
No study ranks these biases, and I'd be wary of anyone who claims one does. What the research does give us is a map. The National Institute of Standards and Technology's SP 1270 sorts AI bias into three categories (systemic, statistical and computational, and human) and lists the human ones that matter most when people build or use these systems: anchoring, the availability heuristic, confirmation bias, automation complacency, groupthink and the sunk cost fallacy, among others.
From what I see in client work, these are the six biases that do the most damage to digital decisions, and where each one tends to show up:
Loss aversion: renewals. The tool stays because leaving feels like a loss.
Confirmation bias: evaluations. The favorite vendor's weak answers get excused and the others' get scrutinized.
Bandwagon effect: strategy. A competitor's announcement becomes the reason to buy.
Status quo bias: architecture. The inherited stack is treated as a given.
Planning fallacy (optimism bias): budgets and timelines. Estimates assume the best case.
Automation complacency (automation bias): AI output. People stop checking because the answer sounds confident.
NIST makes one point every business owner should hear: biases in human decision making "are usually implicit and unconscious," and "any assumption that biases can be remedied by human control or awareness is not a recipe for success." Knowing the names helps you talk about bias. Process is what changes the outcome.
Automation complacency: the bias AI brings into the room
NIST SP 1270 defines automation complacency as what happens "when humans over-rely on automated systems or have their skills attenuated by such over-reliance," and gives spellcheck as the everyday example. With AI tools in the loop, this bias stacks on top of the others. A team that already favors a decision asks an assistant to evaluate it, the assistant agrees in fluent prose, and the agreement feels like independent confirmation. That's confirmation bias and automation bias working together. That's the amplification the HBR authors, Grace Chang and Heidi Grant, describe: the way people frame questions and read answers shapes what the system gives back.
The models carry biases of their own, too. The NBER working paper by Bini, Cong, Huang and Jin (January 2026) ran classic behavioral economics experiments on large language models and found that prompting them to make rational decisions reduced their biases. The most effective fix was one sentence at the start of the prompt asking the model to think of itself as a rational investor. That is a form of persona prompting, and it's worth knowing that a role changes a model's framing more than its factual accuracy.
The human-side fix for automation bias is a checkpoint people can't skip. For AI agents that send, post, spend or delete, that means a named person approves the action, and the approval request shows enough detail that it can't become a rubber stamp. The guide to human approval for AI agents covers how to set that up so approvals stay meaningful.
The planning fallacy and the fat tail in IT budgets
Optimism bias shows up in every project plan. Bent Flyvbjerg and Alexander Budzier studied 1,471 IT projects for Harvard Business Review (2011) and found that the average cost overrun was 27%. The average hid the real risk: one in six projects was a "Black Swan," with a cost overrun of 200% on average and a schedule overrun of almost 70%.
Optimism bias hides in the average: a 27% typical overrun sat beside one project in six that overran its budget by 200%.
IT project cost overruns: the average and the fat tail (Flyvbjerg and Budzier, Why Your IT Project May Be Riskier Than You Think, Harvard Business Review, September 2011)
Project group
Cost overrun (%)
All 1,471 projects (average)
27%
Black Swan projects, 1 in 6 (average)
200%
Those numbers are why I don't trust a single-point estimate for anything that touches core systems. A budget built on the best case is optimism bias written into a spreadsheet, and it leaves no room for the one-in-six outcome. The practical counter to this bias is small: ask for a range, ask what the estimate looks like if the project lands in the worst sixth, and decide in advance what you'll do if it does. That last question is a premortem in miniature.
Risk compensation: when a new tool makes people careless
Risk compensation is a bias toward taking more risk when you feel better protected. The idea is usually traced to economist Sam Peltzman's 1975 study of automobile safety regulation in the Journal of Political Economy, which argued that drivers offset the safety devices by driving less carefully. Later research disputed how large the effect was on the road, but the bias is easy to recognize in technology.
A company buys a security tool and stops patching as carefully. A team adopts an AI writing assistant and stops reading the drafts closely. A backup system goes in and nobody tests a restore. The tool was supposed to add a layer, and it quietly replaced one. To counter the bias, when you add a safeguard, write down which existing habit it does not replace, and check that habit at the next review.
How to evaluate a vendor demo without confirmation bias
A demo is designed to show the best path through the product. Getting beyond surface-level reasoning means taking control of the path. This is the checklist I use to keep bias out of the score:
Bring your own scenario. Send a real, messy task from your business a week ahead and ask to see that, not the standard script.
Score blind first. Each evaluator scores against the written criteria before the group discusses, so the most senior opinion doesn't anchor everyone else (anchoring bias at work).
Ask for the failure case. What happens when the data is wrong, the integration drops, or the person who configured it leaves?
Ask how the recommendation reached you. If a consultant, review site or "best tools" list pointed you here, ask whether there's a referral fee or partnership. A neutral recommendation can say why the tool is wrong for some buyers.
Get the exit terms in writing. How you leave, what data you get back, and in what format.
How to design decisions that resist bias
Columbia's risk management researchers recommend premortems, assigned devil's advocates and scenario analysis. Here are the three practical interventions I've adapted for technology decisions at Kief Studio:
Premortems. Before committing to a technology decision, ask the team: "It's six months from now and this failed. What went wrong?" This surfaces objections that status quo bias and groupthink would normally suppress, and it gives the team a structured way to disagree.
Gary Klein, who named the method in Harvard Business Review in 2007, built it on 1989 research on "prospective hindsight": imagining that an event has already happened helps people come up with more reasons for how it could happen. Klein reports the gain as 30%. The method costs about half an hour, and it gives quieter people permission to say what they've been thinking.
Devil's advocates with authority. Assign someone the explicit role of arguing against the decision, and give that person actual authority to delay or block it. BIT's Scott Young writes, from his own experience, that Operations and IT people are more open to receiving help and thinking differently than leadership often is, which makes them ideal candidates. When nobody inside has the distance, an outside reviewer can take the seat; that's a large part of what technology advisory work is for.
Decision journals. Write down what you decided, why, and what you expected to happen. Review it in 90 days. Auditing what AI is actually doing in your business follows the same principle: you can't improve what you don't measure, and you can't measure what you didn't write down.
A decision journal entry can be five lines: the decision, the date, the options you rejected, the result you expect with a number attached, and the date you'll check. When the check comes, compare the result with the expectation, not with what you remember expecting. Hindsight bias rewrites memory quickly, and the journal is the only record it can't edit. If you want more on why we overvalue what we already chose, the IKEA effect and survivorship bias in best practices are close cousins of the biases above.
Frequently asked questions about bias in technology decisions
Can cognitive bias be eliminated from technology decisions?
No. Columbia University's 2025 research is explicit on this point. As Rich Lauria of Columbia's Enterprise Risk Management program put it, you can't eliminate bias, "you can only design processes that help counteract it." The realistic goal is a decision process that makes biases visible and accountable: premortems, structured evaluation criteria, and decision journals that create a paper trail you can audit.
How does AI amplify human cognitive biases?
Harvard Business Review's January 2026 analysis found that AI systems amplify the biases of their operators, in addition to the bias they inherit from training data. When a user asks leading questions, the AI confirms them. When a team uses AI to validate a decision they've already made, the output reinforces the confirmation bias. The model has biases of its own, and leading prompts add the user's.
What's the most expensive cognitive bias in technology adoption?
Loss aversion. Teams will spend more maintaining a failing system than they would switching to a better one, because the perceived "loss" of admitting the original decision was wrong outweighs the measurable gain of improving operations. Kahneman and Lovallo's 1993 research showed that loss aversion favors inaction and the status quo over the alternatives, which is exactly what a failing system needs to survive another renewal.
Does behavioral science training help with technology decisions?
Yes, though the benefit comes from recognizing biases in real time more than from knowing their names. The Behavioural Insights Team has observed that business leaders often nod at bias concepts but don't follow through. Applied behavioral science works when it's built into how decisions get made, such as the evaluation rubric, the premortem and the journal, rather than presented once in a slide deck.
Which biases most affect digital decision-making in 2026?
No study ranks them, but six come up again and again: loss aversion at renewal time, confirmation bias during vendor evaluations, the bandwagon effect in strategy, status quo bias in architecture, the planning fallacy in budgets, and automation complacency with AI output. NIST SP 1270 lists automation complacency, anchoring, confirmation bias, groupthink and the sunk cost fallacy among the human biases that affect AI systems.
What is risk compensation, and how does it show up in technology?
Risk compensation is a bias toward taking more risk when you feel better protected, an idea usually traced to Sam Peltzman's 1975 study of auto safety rules. In technology it looks like relaxing patching after buying a security tool, or reading AI drafts less carefully once an assistant is in place. Name the habits a new tool does not replace, and keep checking them.
How do you evaluate a vendor demo beyond surface-level reasoning?
Write your criteria before the first call, send the vendor a real scenario from your business to demo, have each evaluator score independently before the group talks, and ask to see what happens when things go wrong. That structure gives confirmation bias less room, because the comparison happens on paper instead of in the room.
How can you tell if a marketing tech recommendation is neutral or paid?
Ask directly whether the person or site recommending it earns a referral fee, commission or partner discount, and look for a disclosure. A neutral recommendation usually names who the tool is wrong for and what it costs to leave. If every option on a list is described only in positive terms, treat it as advertising with a built-in bias and run your own evaluation.
Are IT platform consolidation savings real, or just vendor claims?
Sometimes both. Consolidation can cut real costs when you're paying for overlapping tools, but the projected savings in a vendor's pitch are a best-case estimate, with all the optimism bias that implies. Flyvbjerg and Budzier found that one in six IT projects overran its budget by 200% on average, so price the migration as a range, include the worst case, and measure actual savings against your own baseline after 90 days.
Survivorship bias overstates the typical U.S. stock fund's reported returns by roughly 1.6 percentage points a year, per CRSP: surviving U.S. stock funds averaged 8.8% over the decade ending 2003, and 7.2% once dead funds were counted. Every best practice you copy carries the same tilt.
In a famous experiment, owners demanded twice what buyers would pay for the same mug. That's the endowment effect, and it quietly distorts the cloud versus self-hosting decision in both directions. The fix is to decide as if you owned neither option.
A company spent nearly a million dollars on failing software and chose to continue. Not because the future looked promising — because the past felt too heavy to abandon.