Trust Tax

Open your wealth app. Somewhere on the home screen, there's a small AI-generated number telling you something about your portfolio — your projected returns, your “wellness score,” your suggested next action. Look at it for half a second. Notice the tiny instinctive question your brain runs:

Should I trust this?

That half-second of friction is the most important UX moment in financial products today. Nobody puts it in a PRD. Nobody A/B tests it. It doesn't show up on the dashboard. But it is the entire reason most AI features in wealth apps quietly fail to move retention, even when the underlying model is technically impressive.

I think of it as the Trust Tax. Every AI feature shipped into a financial product collects one — paid not in dollars, but in customer confidence drawn down over time. Most product teams pay zero on the way in and then get a bill ten times larger at the worst possible moment.

This essay is an argument that the Trust Tax is the central design constraint in AI-powered wealth — more important than model accuracy, more important than feature velocity, more important than the agentic-AI roadmap your competitors just announced. And it's an attempt to define what that tax actually consists of, why product managers in this space tend to underprice it, and what to do about it.

The Indian wealthtech industry is in the middle of a coordinated land-grab. INDmoney is deploying AI across onboarding, portfolio insights, and customer support.[1] Groww's CEO has said publicly that AI-driven tooling let them ship seven new products in two years without proportional hiring. Upstox's CTO talks about AI as the reason feature rollouts accelerated.[2] The country's AI-in-finance market is projected to grow at a 22.7% CAGR to $14.2 billion by 2031. Robo-advisory revenues are expected to scale from $286 million to $2.1 billion by 2030.[3]

The shared narrative is roughly: AI makes wealth platforms cheaper to operate, faster to ship, and more personalized at scale. That story is broadly true. It is also incomplete in a way that matters.

What gets celebrated in industry posts and pitch decks is automation share — what percentage of operational work can be done by an AI, what fraction of user queries get resolved without a human, how many features shipped without headcount growth. These are real numbers. They're also the numbers easiest to put on a slide.

What rarely makes it onto the slide is what the same AI features cost when they're wrong. And in wealth products, that cost is structurally larger than the gain.

Every time you ship an AI feature into a financial product, you are taking on three obligations you didn't have before. Together they form the Trust Tax.

The explainability cost. When a user sees an AI-generated number — a tax projection, a portfolio rebalancing nudge, an “optimal” SIP amount — they will eventually ask why. Not always out loud. Often only in a support ticket, six months after onboarding. But they ask. If your answer is “the model said so,” you have produced a worse outcome than if the number had never existed. You have built a product that confidently claims authority without earning it. That's a position users punish hard.

The reversibility cost. Some AI errors are cheap. A movie recommender suggests something boring; you scroll past. Some AI errors are catastrophic. A wealth product mis-classifies a high-risk product as conservative for a 65-year-old, and by the time anyone notices, the damage is two layers deep — financial loss plus the loss of the user's belief that the platform is on their side. Most PMs treat these errors the same way at the design stage. They are not the same.

The escalation cost. When the AI is wrong, who notices, how fast, and what happens next? Most teams ship AI features with elaborate logy on the happy path and a “we'll figure it out” plan for failure. In wealth, failure is the path. A 2023 Deloitte study found that 68% of robo-advisors failed to properly verify client risk profiles through their digital assessments.[5] Industry estimates put current LLM hallucination rates between 3% and 27% depending on the task.[4] The question isn't whether your AI will be wrong; it's how often, in which scenarios, and what your platform does in the next thirty seconds when it is.

The Trust Tax is the sum of these three obligations, paid silently in user confidence drawn down with each interaction. Most product teams I've talked to don't have any of these on a metric dashboard.

There's a body of research worth knowing here, dating back to Berkeley Dietvorst's work on what he called algorithm aversion[6]. The finding, replicated across many follow-up studies: people abandon AI advisors after a single visible error, even when those AI advisors objectively outperform the human baseline. One mistake from an algorithm produces more lost trust than five mistakes from a human.

More recent research has refined this. Algorithm appreciation exists too — in tasks where users have no expertise, no immediate feedback, and high quantitative complexity, they often prefer the algorithm at the start.[7] Financial advisory sits squarely in that zone, which is why AI onboarding into wealth has been so easy in the first wave. But the same research shows that initial appreciation collapses faster than human-advisor trust does, when an error becomes visible. The trust curve is steeper in both directions.

Translate this into product terms. In a normal consumer app, your trust function looks roughly like: ship → users use → some errors happen → users tolerate them if value is high → trust slowly accumulates. Standard SaaS playbook.

In a wealth product, the function looks more like: ship → users use → one visible error → trust collapses to zero in this user's mind → they tell three friends → cohort retention damage three quarters later. This is what an asymmetric loss function looks like in practice. The upside of a working AI feature is “users feel slightly more confident.” The downside of a broken AI feature is “users feel betrayed by their money platform.” These are not the same magnitude.

Most product frameworks for AI feature evaluation come from the B2B SaaS world. They optimize for time saved, NPS, completion rates. None of them have a trust capital term. In wealth, trust capital is more or less the only moat you have. Burning it to ship faster is an unusual bargain.

If the trust argument feels soft, here's the hard version. In February 2025, SEBI notified Regulation 16C through the SEBI (Intermediaries) (Amendment) Regulations.[11] The text is short and worth reading in full at least once. Its substance is this: every SEBI-regulated entity that uses an AI tool — whether built in-house or procured from a vendor, regardless of scale — is solely responsible for the privacy and security of investor data, for the output of that AI, and for compliance with all applicable laws.[9][10]

Sole responsibility. Not shared. Not limited. The vendor you bought the model from is, from SEBI's perspective, not your liability shield. If the LLM hallucinates an instrument suitability assessment and an investor loses money, the case isn't against OpenAI or Anthropic. It's against you.

For anyone paying attention, this is the Indian equivalent of the Air Canada precedent. In Moffatt v. Air Canada (2024), the British Columbia Civil Resolution Tribunal ruled the airline liable when its customer-facing chatbot gave incorrect information about bereavement fares.[12] Air Canada argued the chatbot was a separate entity responsible for its own actions. The tribunal called this argument what it was — and ruled that the company was responsible for everything its website said, whether the words came from a static page or an AI. The damages were trivial ($812). The precedent was not.

SEBI 16C is the same idea, codified before the lawsuits. The regulator has essentially told every wealth platform in India: if you ship it, you own it. You can't disclaim it in a terms-of-service footer. You can't push the blame to your model provider. The “but the AI said it” defense doesn't exist.

This is not a compliance burden. It's an architectural directive. It changes how you should build, what features you should ship, and which AI use cases are worth the candle.

Here's the part most AI product narratives get wrong. The instinct in 2026 is to make AI features more visible, more autonomous, more “agentic.” Bigger surface area. More chrome. More tasks completed without a human in sight.

In wealth, the inverse often earns more trust. Some examples worth studying:

Robinhood's Strategies product has roughly 250,000 paying customers at around $250 a year — a real, ARPU-generating AI advisory product. But Robinhood doesn't pitch it as autonomous AI advice. It's an AI tool guided by human advisors.[13] The model is doing most of the cognitive work; the framing puts a human in the loop. That framing is a feature, not a marketing flourish. It changes how users interpret the same recommendations.

The most thoughtful Indian wealthtechs are putting their best AI in back-office reconciliation, exception handling, ops drafting, document parsing — places where the model can be wrong 5% of the time and a human catches it before the user ever sees it. The flashier customer-facing applications get reserved for high-confidence narrow tasks: classification, summarization, draft generation that a human approves. The most expensive AI sits behind the wall, not in front of it.

This goes against every “agentic AI is eating the enterprise” narrative being written right now. It's also the right answer for products where trust is the entire game.

There's a behavioral reason this works. Research on explainable AI consistently finds that visible human oversight increases trust and increases tolerance for errors when they happen.[7] Users forgive an AI-assisted-human more readily than they forgive an autonomous AI. Same underlying performance. Different perceived contract.

This is something I'd want to coin: the Blast Radius Framework. Borrowed from systems engineering — where blast radius describes how much of a system breaks when one component fails — and applied to AI feature design in financial products.

The premise is simple. When evaluating any AI feature, the most important question isn't “how good is the model” or “what does this automate.” It's: if this is wrong, how far does the damage spread, and how do we contain it? The smaller the blast radius, the lower the Trust Tax.

Run any proposed AI feature through four questions, in order:

1. 01.

   Is the outcome reversible?

   If the AI is wrong, can the user undo the impact? A wrongly generated insight is fully reversible — they just ignore it. A wrongly executed trade is partially reversible (you can sell, but you've taken slippage). A wrongly filed tax submission may be effectively irreversible. The further down this scale, the higher the Trust Tax.

2. 02.

   Is the action reversible?

   Even if the outcome can be reversed, can the system catch the error in time? Some AI features have natural circuit-breakers (a human approval step, a confirmation screen, a delay). Others fire instantly. AI features that act faster than humans can catch them are expensive to ship.

3. 03.

   Is the trust reversible?

   This is the brutal one. If this feature is visibly wrong once, do users come back, or do they assume the platform is broken? For high-frequency, low-stakes interactions (a chatbot that misunderstands a query), trust recovers. For low-frequency, high-stakes ones (a portfolio recommendation that turns out badly), trust often doesn't. Treat trust like a depleting resource, not a permanent asset.

4. 04.

   Is the blame reversible?

   When something goes wrong, can your platform credibly say “here's what happened, here's what we're doing about it” — or does the architecture make root-cause impossible? This is where audit logs, model versioning, and explainability infrastructure matter. Not for the regulator. For your own ability to recover trust after the inevitable error.

A feature that scores poorly on all four has a large blast radius. It's not necessarily wrong to build. But it requires a different kind of investment — heavier human-in-the-loop, slower rollout, narrower scope, more guardrails — than the team usually budgets for.

A few practical implications for product teams shipping AI into wealth, healthtech, insurance, or any other category where being wrong is asymmetrically expensive:

• Add a Trust Tax line to your feature reviews. Before approving an AI feature for build, force the team to articulate the three obligations — what will you do when the user asks why, what will you do when the model is wrong, and what is the escalation path? If those answers are vague, the feature isn't ready.
• Prefer narrow over agentic. A model that does one well-bounded task with 99.5% accuracy will earn more trust than an agent that does ten tasks at 92%. The marketing logic and the product logic point in opposite directions here.
• Put your best AI behind the wall. Reconciliation, ops drafting, exception handling, classification — these are where the unit economics of AI compound without the trust risk compounding. The most valuable AI in your stack might be the AI your users never see.
• Design for visible reversibility. Show the user how to undo. Show them where the human is. Show them the audit trail. Friction here is not friction — it's the product earning the right to act on the user's behalf.
• Build the post-error playbook before the error. The “we'll figure out what to do when something breaks” plan is not a plan. Pre-write the in-app message, the support workflow, the comms template. Errors will happen at 11 PM on a Saturday; the response quality should not depend on someone being awake.

The shape of competition in Indian wealthtech over the next five years is becoming clearer. The first wave (build AI features fast, ship anything that automates an ops process) is largely played out. The next wave is harder and less visible: which platforms can deploy AI in a way that compounds trust instead of slowly draining it.

The platforms that win this won't be the ones with the most AI features. They'll be the ones whose AI features the user forgets are AI features — because they're correctly scoped, transparently explained, gracefully reversible, and embedded in a product surface that takes responsibility when something goes wrong.

The Trust Tax is the price of building in this category. It is paid either deliberately, in upfront design cost, or accidentally, in churn and complaints and regulatory letters. The PM's job is to choose which.

In wealth, the customer is not buying alpha. They're not buying features. They're buying the belief that a serious adult is in charge of their money. Every AI feature you ship is either an argument for that belief or against it. Ship accordingly.