Why We're Building an Advisor, Not Just Better Advice

The best conversation still ends

Suppose the advice you got was genuinely good. It fit you rather than people in general. It was specific to a tee. It arrived at a moment you could hear it. Our whole first year went into making that happen in a single conversation, and the demo on this site is that year, compressed to a few minutes.

Then the conversation ends. You close the tab, and the advice goes off to meet the only judge that matters, which is your actual week. Whatever happens next, happens off-screen: the walk taken or not taken, the conversation with your manager had or postponed, the decision made or quietly re-litigated at midnight. The interesting failures of advice are not in the moment. They are on a random day next week.

The gap between agreeing and doing

This gap has a name in the research, and it is one of the better-documented effects in psychology. Reviewing decades of studies on intentions, Sheeran and Webb put the figure bluntly: roughly half of the things people sincerely intend to do never happen. That is mainly because intending is a different act from doing, and the gap between them is where real life comes into play.

The same literature says the gap is not fixed. In a 2006 meta-analysis covering 94 studies, Gollwitzer and Sheeran found that giving an intention a concrete shape, a when and a where and a what-if, closed a substantial part of the distance, with an effect size around d = .65. The gap responds to structure and to follow-through.

Notice what this means for advice as it is usually consumed. The conversation produces an intention and then stops, which is to say it stops at exactly the point where the evidence says the real work begins.

Advice that works has four parts, and every one of them is hungry for time

We have argued before that advice which actually lands has four parts: it is built on a real model of you, it is specific enough that there is a risk of being wrong about you, its confidence matches its evidence, and it arrives when you can take it in. You do not need the full argument here. What matters is what happens to each part when the advisor gets more than one sitting.

The model of you deepens. Some things about a person only surface when a choice forces them, and no single conversation contains enough choices.

The “so specific it risks being wrong” claim gets a scorekeeper. A specific prediction about you is only worth something if someone is still around when it resolves. "You talk yourself out of things by the second day" is a claim with a verdict date, and an advisor who is present on the second day gets to collect it.

Calibration stops being a performance and becomes a record. An advisor who has watched three of their own calls about you play out knows exactly what their confidence is worth on the fourth.

And receivability is mostly timing. One session gets one shot at your state. Someone present across days can hold a true thing until the hour you can hear it.

A single sitting caps all four parts at whatever it happens to catch. Time compounds all four. That is the entire argument, and it is why better answers alone were never going to be enough.

The relationship carries weight of its own

There is evidence that continuity is an active ingredient rather than ambience. In psychotherapy research, the working alliance, the bond and agreement between the two people, is the most robust predictor of outcomes that the field has found. A 2018 meta-analytic synthesis by Flückiger and colleagues, pooling 295 studies, put the alliance-outcome correlation around r = .28, steadier across conditions than differences between techniques.

Therapy is not advice, and nothing about an AI advisor inherits this automatically. We hold the finding for what it shows: the relationship itself does work in the outcome, over and above what gets said inside it.

What we are engineering differently

Building for a relationship instead of an answer inverts the architecture. Four of the decisions, stated as decisions, because we are iterating on them daily with the cohort and the details will keep moving:

One conversation that never starts over. Knowing your problem and knowing you are both slow accruals, and they need one continuous place to happen. The app opens into the same space it opened into yesterday.

Memory that graduates. Conversational memory blurs by design; every compression loses texture. So anything that matters beyond the flow of talk must be caught and structured before the blur takes it: a placement on the map, a fact about your life, a plan, a promise.

The only dated thing in the system is a kept promise. "If I didn’t do the thing by Thursday, bring it back" creates the one row that carries a date. Time delivers what was promised and nothing else; it never manufactures a reason to contact you.

A step closes only when reality closed it. "Go on the walk" completes when the walk happened, not when the conversation moved on. This is the intention-behavior gap, built into the data model.

The counterfeit version

There is a cheap way to simulate an advisor who comes back: streaks, guilt, subject lines that withhold. Products measured on whether you returned reach for these naturally, and we have written about why that scoreboard produces a different kind of product. In the document that governs this build, these mechanics are banned by name. If the relationship is worth keeping, it will be kept for what it does, and a relationship engineered to be hard to leave is measuring its own failure.

What we watch instead is whether life moved: the steps only reality could close, closing. A trend we steer by, never a quota anyone is pushed toward.

The journey, restated

Generating good advice turned out to be the easy half. Modeling the receiver was the missing half, and we spent a year on it. What the cohort is teaching us now is that even that pair is a moment, and a moment needs a keeper: someone still there when the prediction resolves, when Thursday comes, when the walk does or does not happen.

The advisor is being built in the open with a small group, and it is not the public door yet. The demo at getinward.com is: a few minutes, no signup, and the first half of everything described here.

If you’d like to build with us, or reach out, send a note to hello@getinward.com