It is late, you are tired, and you have just asked a coding agent to clean up your database migrations. There is no plan and no preview; the run simply starts. Shell commands scroll past faster than you can read them, and somewhere in the blur you catch the word DROP. You are now parsing output you cannot keep up with, reconstructing what has already changed, and deciding whether to kill the run. That decision needs your full attention, and the anxiety the tool just created is eating into it, so you make the call with less focus than you had two minutes ago.
The tool did not merely fail to help at a risky moment. It created the risk and then taxed the person facing it. Our essay The Human Factors recommends lowering the stakes at exactly these moments, and this chapter shows you how.
Anxiety and attention draw from the same pool
Cognitive scientists describe working memory capacity as executive attention, the ability to keep what matters right now active and quickly retrievable.
Your attention is one fixed pool that the task and your emotions both spend from, so the worry a tense moment creates leaves less attention for the decision that moment demands.
Everything competes for that pool: the task in front of you spends from it, and so does emotion.
The research on how anxiety raids the pool is specific:
- Worry rises as confidence falls. The less you expect to do well, the more of your attention worry occupies.
- Anxiety peaks at maximum uncertainty, when the outcome feels like a coin flip.
- Under enough pressure, the emotion itself consumes the attention the task needed, and performance drops.
Now put an unpredictable system into a consequential moment. When the stakes are high and you cannot predict what the system will do next, uncertainty is at its highest, and that is exactly when anxiety peaks and steals the most attention, so you have the least focus available right when the decision needs the most.
A caution is worth raising before the fix. A reasonable amount of anxiety is useful because it supplies an appropriate degree of caution, and a person approving a production deploy should feel a little of it. The goal is to tune that anxiety down to a workable level, not to switch the caution off entirely.
Lower the stakes with preview, confirm, and undo
One of the graduate research papers behind this part reviewed an online stock-trading interface, a product where anxiety and money meet daily. It recommended a simulated trading mode with nothing at risk, because practice builds skill, skill raises confidence, and confidence shrinks worry. It also recommended a journal where the trader records sell conditions while calm and ties them to alerts, so that when the market moves, the trader executes a decision already made instead of improvising mid-panic. Both ideas move the thinking to a calm moment and shrink the consequences of the stressful one.
For an agentic product, one that takes actions on your behalf rather than just answering, the risky moments are easy to list: just before it sends the email, runs the command, moves the money, deletes the record, publishes the post. At each of those moments, the same principle becomes a ladder of moves.
Preview moves the decision to a calm moment. Show the whole result, or the whole plan, before anything is committed. The user weighs it with their full attention available, then watches an execution they already approved.
Confirm keeps the consequential act in human hands. A deliberate yes at the genuinely risky step gives the user one clear moment of control instead of a blur of motion they can only react to.
Undo lowers the stakes themselves. When any mistake is one step from being reversed, the worst outcome stops being permanent, so the moment never becomes high-stakes enough for anxiety to peak in the first place.
How shipped products lower the stakes
-
Adobe Photoshop's Generative Fill previews everything and commits nothing. Each generation produces variations on a separate layer, shown as thumbnails, and the original pixels are never written to. You can compare options, regenerate, or discard the whole layer at any time. The design removes the fear of the AI ruining hours of work on a real file, which is exactly the anxiety that would otherwise tax every creative judgment in the session.
-
Gmail's Help me write treats sending as the risky act. Gemini produces the email as a proposal that you must explicitly insert into the compose window, with one-tap refinements available before you commit, and nothing is ever sent on your behalf. The genuinely consequential act, an email going out in your name, stays entirely in your hands.
-
ChatGPT agent requires confirmation before it acts in the world. It prompts for an explicit yes before consequential actions such as purchases or outgoing email, and on sensitive sites such as financial services it requires the user to actively watch while it works, with the option to interrupt or take over at any time.
Put it to work this week
-
Show the plan before anything runs. Any action that touches more than one thing gets a preview the user approves first. The decision happens while they are calm, and the execution happens on terms they chose.
-
Reserve confirmation for the irreversible. Tier your actions by reversibility and prompt only at the top tier. Pair this with make the warning impossible to miss so the prompt that does fire registers.
-
Make undo one step, and say so at the moment of action. An undo nobody knows about lowers no anxiety. Put "you can reverse this" next to the commit button, where it changes the felt stakes before the click. Gmail's Undo Send shows the pattern: the moment a message goes out, the sent notice offers Undo for a short window you can configure, so the classic irreversible act stays recoverable right when regret arrives.
-
Give beginners a consequence-free practice mode. Offer a dry run, a sandbox, or a simulated mode. Intercom applies this at the organizational level: teams can question its Fin support bot in a preview panel and run batch tests against real customer questions before the bot ever answers a paying customer.
-
Keep long runs visible. Stream what the system is doing in plain language while it works, because a window into the run holds uncertainty below its peak. The memory half of that problem is covered in keep the session on the screen.
These moves belong to the same operating discipline as your guardrails, and the verify-first rule in the Playbook applies the same posture to your own AI sessions.
Try this today: inventory your irreversible actions
Set aside fifteen minutes and list every action your product can take on a user's behalf: send, write, delete, publish, spend, message, deploy. Mark each one reversible or irreversible, and be honest about it, because a sent email is irreversible and a draft is not. Then check every irreversible action for a preview, a confirmation, and an undo, and write down the reason for anything that is missing.
Most teams who run this find at least one unprotected irreversible action, usually one that shipped early and that nobody has looked at since. Fix that one this week. The full version of this drill is part of the human factors audit at the end of this part.
Chapter Summary
- Attention is one fixed pool, and both the task and the user's emotions spend from it.
- Worry peaks when the outcome feels like a coin flip, and that is when it drains the most attention.
- An unpredictable system at a high-stakes moment creates that peak, leaving the user with the least focus right when the decision needs the most.
- A little anxiety keeps people careful, so the goal is to calm the user down, not to switch the caution off.
- Preview moves the decision to a calm moment by showing the full result or plan before anything is committed.
- Confirmation keeps the one hard-to-reverse act in human hands, so save it for those steps and do not put a dialog on everything.
- Undo lowers the stakes themselves: when a mistake is one step from being reversed, the moment never gets tense enough for anxiety to peak.
- Run the fifteen-minute drill to list every irreversible action your product can take, then check each one for a preview, a confirmation, and an undo.
- Next, show people what the system can do, because a user who cannot predict what the system is capable of carries uncertainty before anything even runs.
Sources
- Atkinson, J. W., & Feather, N. T. (Eds.). (1966). A Theory of Achievement Motivation. Wiley.
- Engle, R. W. (2002). Working memory capacity as executive attention. Current Directions in Psychological Science, 11(1).
- Liebert, R. M., & Morris, L. W. (1967). Cognitive and emotional components of test anxiety. Psychological Reports, 20(3).
- Luu, P., Tucker, D. M., & Derryberry, D. (1998). Anxiety and the motivational basis of working memory. Cognitive Therapy and Research, 22(6).
- Pessoa, L. (2009). How do emotion and motivation direct executive control? Trends in Cognitive Sciences, 13(4).
- Graduate research paper on working memory, emotion, and online stock trading (2013), one of the research papers behind this part.
- Product documentation: Adobe Photoshop help on Generative Fill; Google Gmail Help and Workspace Updates on Help me write and Undo Send; OpenAI Help Center on ChatGPT agent; Intercom Help Center on Fin previews and batch tests.