The settings you never touch decide more about your AI assistant than the ones you do.
Every AI product ships with defaults. Most teams never question them. The assistant that acts on its own, the one that remembers everything, the one that never asks — those are choices someone made for you, usually on a Tuesday, usually under a deadline, usually with the assumption that you would change them later. You almost never do.
Defaults are not a technical leftover. They are a product decision, and for business AI they are one of the most important decisions a team can make. Not the model choice. Not the prompt. The defaults.
What defaults actually are
A default is the behaviour an assistant falls back to when no one is watching and no one has told it otherwise. It is the answer to a question your customer never asked but lives with every day.
Should the assistant act, or should it ask first? Should it remember this conversation, or forget it? Should it surface the uncertainty, or smooth it over? Should it trust the model’s confidence, or verify before it moves? Should it speak up, or stay quiet?
Every one of those questions has a default answer baked into the product. And the baked-in answer is what your customers actually experience — not the settings page they will never open.
The one that matters most: act or ask
If you take one default seriously, take this one. Should your assistant do the thing, or should it check with a human first?
Default-to-act feels impressive in a demo. The assistant just handles it. The workflow just runs. Nobody has to think. Then a customer sees it do something sensible at the wrong moment — send the email, change the record, refund the order, post the reply — and the trust cost is enormous. Not because the action was wrong in isolation, but because nobody expected it to move on its own.
Default-to-ask feels slower. It is slower. It is also the default that keeps small teams in control of tools they cannot afford to babysit. When the assistant asks before it acts on anything consequential, the human stays the decision-maker and the assistant stays a tool. When it does not, the assistant quietly becomes a colleague you cannot manage.
The useful products do not pick one side forever. They default to ask on the actions that cost something to undo, and default to act on the ones that are trivial to reverse. That single line — reversible versus irreversible — is more useful than any confidence threshold you will ever tune.
Remember or forget
Memory defaults are where privacy becomes a product decision instead of a policy page.
An assistant that remembers everything about every customer is convenient until a customer asks what it remembers, or until you need to explain where the data lives, or until someone leaves and the assistant still knows their voice. An assistant that forgets everything is safe and useless — it re-introduces itself every morning.
The default worth fighting for is scoped memory. Remember what is needed to be useful inside this context. Forget what is not. Keep a customer’s preference for a follow-up call in the morning. Do not keep a record of every question they ever asked at 2am. The line is not hard to draw once you stop treating memory as a feature and start treating it as a decision about what a relationship should hold.
This is also the line that separates a privacy-respecting product from one that merely says it respects privacy. A privacy policy tells people what could happen. A default tells them what does happen. Customers can feel the difference, even when they cannot quote either one.
Surface or bury
Every assistant encounters moments it is not sure about. A question it cannot fully answer. A request that sits between two valid interpretations. A customer who is clearly frustrated and a reply that could be read two ways.
The default is to bury it — to pick the most plausible answer and move on. That default keeps the conversation flowing. It also quietly trains customers to stop trusting the assistant, because they learn that the confident response and the correct response are not the same thing.
The better default is to surface the uncertainty where it matters. Not everywhere — nobody wants an assistant that second-guesses every sentence. But at the moments where a wrong answer costs the customer something: their time, their trust, their money, their next step. There, the assistant that says “I am not certain, here are two options, which one do you mean?” is the one people come back to.
Surfacing uncertainty is not a weakness. It is the behaviour that lets a small team run an AI assistant without staffing a safety desk. The assistant handles the easy path on its own and makes the hard path visible. That is the division of labour that actually works in a real business.
Why small teams feel defaults the most
Enterprises can survive bad defaults. They have review boards, legal teams, and enough people to watch the assistant and correct it in real time. A small team has none of that. A founder, an operator, maybe one other person. The defaults are the whole safety model.
This is why the defaults that look like minor product choices — ask before acting, forget what is not needed, surface what matters — are not minor for the businesses we build for. They are the difference between an assistant that compounds trust and one that slowly erodes it.
A small team running a website assistant does not have time to audit every conversation. They are relying on the assistant to ask before it sends, to forget what it should, and to hand off the moments it cannot own. If those defaults are wrong, no amount of prompt engineering fixes it later. The team just learns to distrust the tool and eventually turns it off.
The defaults worth changing
Most defaults are invisible until you look for them. A few are worth questioning the moment you do.
- Act or ask. Default to ask on anything that costs something to undo. Default to act only where the action is trivial to reverse.
- Remember or forget. Default to scoped memory — enough to be useful, not enough to become a liability. Let the customer’s context stay the customer’s.
- Surface or bury. Default to surfacing uncertainty at the moments a wrong answer costs the customer something. Let the easy path run and make the hard path visible.
- Speak or stay silent. Default to quiet confidence, not chatter. An assistant that talks less and means more is the one people keep using.
Defaults are the product
It is easy to talk about AI as if the model is the product. It is not, for most businesses. The model is raw material. The defaults are the product — the decisions that turn a capable model into an assistant people trust, keep using, and recommend.
The next time you evaluate an AI tool, do not start with the feature list. Start with the defaults. What does it do when you have not told it anything? What does it do when no one is watching? What does it ask before, and what does it do on its own? Those answers will tell you more about the product than any roadmap ever will.
For us, the defaults are the whole point. Ask before it acts. Remember only what helps. Surface what matters. Keep the human in the decision. That is not a settings page. That is how we build.
You’ve got this.
