Pair every insight with its action

Tags: insights, actions, attention

Models made noticing nearly free. Acting still costs a context switch, a hunt for the right screen, and the reassembly of everything the product already knew. So the product generates a hundred cheap observations and hands each one back as five minutes of work — which is the wrong way round, and it is why so many AI features are quietly ignored.

Insights are not the deliverable

"Spend is up 23% in this category." "This account looks at risk." "Three contracts renew inside 40 days."

Each of those is true, generated in milliseconds, and useless on its own. The user reads it and then has to leave — find the category, open the account, locate the contracts — reconstructing by hand a context the product had fully assembled one second earlier and then discarded.

The ratio is what makes this fatal rather than merely annoying. When observations were expensive, one of them arriving was worth the trip. When they are free, a product can produce more of them per day than anyone can act on, and the scarce thing becomes the acting, not the noticing. Design that optimises for producing more insights is optimising the wrong side of the equation.

There is a second cost, and it compounds. Every observation a user cannot act on teaches them that this stream is not worth stopping for. By the time the one that mattered arrives, they have learned to dismiss.

Where the action lives

Three of those branches are the interesting ones.

A fact is not an insight. If nothing follows from it, it does not earn an interruption — it earns a place on a dashboard someone visits deliberately. This is the same discipline as work without being asked: the tiers are interrupt, collect, record, and most things belong in the bottom two.

Handing off means carrying the context, not pointing at it. A "Learn more" link is not a handoff. A draft email already written, a filter already applied, an export already shaped for the tool it is going to — those are. The product knows everything needed; the only question is whether it packages that or throws it away at the boundary.

Sometimes you must not offer the button. Where the judgment has to be a person's — a clinical call, a lending decision, anything with a regulator on the other side — pairing means assembling the decision rather than proposing to take it. Put the evidence, the precedent and the contrary case in one place. That is still pairing; it just resolves to a briefing instead of a control.

Grounded in

Nakatomi's treasury surface, where the forecast is the product and every reading arrives with its remedy attached. A projected deficit offers Borrow. A surplus offers Invest. A payment that would breach minimum need is intercepted before it clears, rather than reported afterwards as an insight about something that already happened.

The interception is the part worth stealing. It inverts the usual order — most products notice, then tell you, then let you go and fix it. Catching it at the moment of the action means the observation and the remedy are the same event, and the controller never leaves the screen that told them.

Anti-patterns

• The "Learn more" link. The product had the context and chose to hand over a search problem instead.
• An action that makes the user re-enter what the product already knew. Selecting the account again, re-picking the date range, retyping the filter. Every field they refill is one the insight failed to carry.
• Offering an action the product cannot actually complete. Worse than no action, because it spends the trust that the pairing was supposed to build.
• Every action is Dismiss. A stream whose only affordance is making it go away has correctly identified what it is for.
• Pairing a one-click action to a judgment nobody can check. This is the dangerous one. An unverifiable observation plus a frictionless button is automation of an unexamined decision, and it will be used exactly as fast as it is offered. If the reasoning is not inspectable, the action needs friction rather than less.
• No way back. Acting on an insight is acting on the model's word. See undo over confirm — the staging point is what makes a one-click remedy safe to offer.

The smallest version worth building

Take the insights the product already produces and, for each one, write the sentence the user would have to say next. Then ask whether the product could just do that.

Where it can, attach it. Where it cannot, carry the context to wherever it happens. Where nothing follows, stop pushing it and put it somewhere people can go and look — which will also make the remaining stream worth reading again.

Related patterns

• Work without being asked — what earns an interruption in the first place; this is what has to be true once something has.
• Undo over confirm — the mechanism that makes an attached action safe to press.
• Mark what the product vouches for — an insight the interface will not stand behind should not arrive with a one-click remedy.