HarvestEngine has an AI assistant. It can explain a proposed trade, simulate a year of harvesting, look up a security's metadata, suggest replacement candidates when you sell something at a loss, and walk you through portfolio drift. We named it Clark. It's pretty useful.
What it cannot do — and what we have intentionally designed it not to do — is tell you what to buy or sell. The line between "explain a decision" and "make a decision for you" is a legal one as much as a product one, and we've built our entire AI system around staying on the right side of it.
This article is the engineering and compliance argument. If you care about how an AI can be helpful without becoming a robo-advisor, read on.
The legal line: investment advice vs investment information
The Investment Advisers Act of 1940 (15 U.S.C. § 80b) defines an "investment adviser" as anyone who "for compensation, engages in the business of advising others... as to the value of securities or as to the advisability of investing in, purchasing, or selling securities." Anyone meeting that definition has to register with the SEC (or state regulators), maintain fiduciary obligations, comply with custody rules, file Form ADV, etc.
The exclusion that lets us be software, not an RIA: providing information, analysis, or tools without rendering specific recommendations as to securities. The SEC has consistently treated "explain how this works" as not-advice and "you should buy this" as advice. The line is the act of recommendation.
How the AI is constrained
Clark has access to a registered set of "tools" — read-only queries against your portfolio, market data, tax calculations, etc. The tool catalog is carefully scoped. There are three tiers of tools and the model only sees the tier it's allowed to act on:
- Tier 0 (read-only): portfolio inspection, metadata lookups, tax bank queries, drift reports, simulator runs. No state changes anywhere.
- Tier 1 (proposes drafts): creates proposal drafts that land in PENDING_REVIEW. No trades happen until you explicitly approve them on the proposal page. The AI never executes.
- Tier 2 (executes — confirm + PIN required): approves, rejects, executes, or activates the kill switch. Every Tier 2 call returns a "pending action" that requires your PIN to go through. The AI cannot bypass the confirmation step.
The system prompt instructs Clark to explain mechanics, surface information, and execute the rules you set — never to recommend specific securities to buy or sell. Asked "should I sell NVDA?" Clark surfaces what it can about NVDA (cost basis, current gain/loss, wash status) and refers the decision back to you. It does not opine.
The product surface that reflects this
Three places where the boundary shows up:
1. Proposal generation
When the daily pipeline finds harvestable losses, it generates proposals — but those proposals land in PENDING_REVIEW. They are recommendations generated by software you signed up for, following rules you turned on, against a portfolio you designed. The legal posture is "tools generated this, review and approve if you'd like" — not "your advisor recommends this trade."
Compare to a robo-advisor: in discretionary mode, the firm trades on your behalf. They've taken on fiduciary duty and SEC registration to do it. We haven't, by design.
2. Replacement suggestions
When you harvest a loss, the system suggests replacement candidates that satisfy the wash-sale rule. The legal phrasing is specifically "candidates that match your stated criteria" — not "we recommend you buy these." The user picks. The user can decline all candidates. The user can pick a write-in.
3. AI explanations
Clark's notebook on every proposal page explains what the proposal does, what risks it carries, and what the simulator estimates. The phrasing is neutral: "this proposal would..." not "you should..." The AI is presenting analysis, not advice.
Why the distinction matters to you, the user
Three reasons it actually matters, beyond legal cover:
1. You stay accountable for your decisions
If you turn on autonomous mode and approve every harvest the governor proposes, you've made a series of explicit decisions: you chose direct indexing, picked the index, set the loss threshold, configured the autonomy thresholds, approved the first run. The software is executing your strategy. If something goes wrong, the strategy is yours and you can change it.
If you hand discretion to a robo-advisor, you've delegated. That's fine when you trust the delegation, but the "I didn't make that call, the robo did" is its own kind of distance from your money.
2. The AI can be honest about uncertainty
Robo-advisors operate under fiduciary duty: they have to make specific recommendations and they have to defend those recommendations. That pressure shapes the AI's outputs — the model gives confident "buy this" answers because that's the product function.
Clark is allowed to say "I don't know" and "this depends on factors I can't see." Asked "should I harvest this loss?" the honest answer is sometimes "the math is favorable, but you also have an earnings call in 8 days and the wash-sale window from a recent buy makes timing tight — let me show you the trade-offs and you decide." That's a more truthful answer than "yes, harvest."
3. The safety properties compose better
An RIA that does TLH for you has to maintain custody, trade in discretionary accounts, and coordinate across all clients. Failure modes can affect many people simultaneously.
HarvestEngine doesn't custody anything. If our software has a bug, it generates wrong proposals — but those proposals don't execute until you approve them. The blast radius is bounded by your own approval gate. We've shipped real bugs (e.g., a notional- sizing issue that overshot a target by 12%); the right thing happened: users saw the trades on the proposal page, flagged the mismatch, we caught it, fixed it. No one's money moved without a human checking it.
The thing autonomy mode does
Some of you will set up autonomous harvesting and have us auto-approve loss harvests within rules you defined. We do that. It doesn't change the legal posture: the user has explicitly configured the rules, the user has the kill switch, the user can revoke autonomy at any time. The rule-configuration step is the meaningful human decision; the trade-execution step is software running rules.
This is the same model that automated bill-pay, automated payroll, and automated savings sweeps use. None of those are "financial advice" — they're software acting on rules you set. TLH autonomy is the same shape.
The careful framing on every page
The footer of every page says "Software, not advice. We're not a registered investment advisor. You decide what to trade. Past performance doesn't predict future results." That language isn't legal CYA we copy-pasted from a template. It's an accurate description of what we are.
If you want a registered investment advisor, hire one — there are good ones, and the fee is usually worth paying for the things they actually do (estate planning, holistic financial planning, behavioral coaching). If you want software that runs your tax-aware direct-indexing strategy on your existing brokerage, with an AI assistant that explains and executes the rules you set, that's us.
Both products are legitimate. They're different products.