AI Invest · Research × Tool
A body of research on AI in investing — and the tool it produced.
Over five parts we mapped what AI is actually doing inside wealth management: the ladder of autonomy, what the filings say versus the marketing, where the return really shows up, the execution controls nobody has built, and why most performance numbers cannot be trusted. Then we turned the framework into something you can run — AI Tool Diligence, a Claude plugin that evaluates AI investment tools, the strategies they generate, and the portfolios they actually produce. Both halves are below. Both are free.
The research
AI in wealth management — a five-part series
Each part stands alone; together they build the evaluation framework the tool implements. Read in order for the argument, or jump to the one you need.
From 60/40 to autonomous agents
Three regimes of portfolio construction, one technology with three very different buyers, and the autonomy ladder — a four-rung taxonomy from copilot to autonomous agent that the rest of the series runs on.
Read the pieceRead the filings first
Nine retail AI advisors scored against their own Form ADVs, Form CRSs and terms of service. The gap between the marketing and the paperwork turned out to be the finding — and the evaluation protocol nobody outside the firms can legally run.
Read the pieceTwenty minutes a week
The advisor stack, audience by audience. Do the division first: a reported 5,000 hours saved a year is twenty minutes per advisor per week. The data layer as a gate, the obligations that attach on deployment, and vendor risk reframed.
Read the pieceThe shield
Agentic execution architecture. The broker with the deepest pre-trade controls will not let an agent execute; the rails that will document none. The deterministic enforcement layer that belongs between them, on the Market Access Rule template.
Read the pieceThe number you cannot trust
The validation playbook. Forty-five variants manufacture a Sharpe of 1.0 from pure noise; the mathematics of establishing skill is settled and ignored — and in April 2026 the framework everyone cited was rescinded. Deflation, evaluation design, and drift.
Read the pieceDownload
Take the tool, and the guide
AI Tool Diligence is a Claude Code / Cowork plugin — it runs inside Claude, not as a standalone app. Install the bundle, and the six skills trigger in natural language. MIT-licensed.
The plugin
ai-tool-diligence.plugin — six skills, reference library, and the deterministic decomposition scripts. Install in Claude Code or Cowork.
The feature guide
A three-page PDF: what each skill does, the measurement that justifies the portfolio engine, the feedback cycle, and how to install and run it.
Install via your plugin settings (or /plugin in Claude Code), pointing at the downloaded file. The portfolio engine needs daily adjusted price history, which the caller supplies — the scripts take no network access and require Python 3 with numpy.
The tool
Six skills, one diligence loop
Each skill maps to a stage of the research. The first three evaluate what a tool claims; the next two evaluate what it actually did to a portfolio; the last closes the loop by checking whether the verdicts held up and recalibrating the scoring.
screen-offerFast triage of an advertised offer
Paste an ad, a URL, or a Telegram invite. Routes on the custody question first, runs the registration lookups that matter for that class of product, scores tiered red flags, and returns a risk verdict with what is unknowable stated plainly. Target: five minutes.
evaluate-platformDocumentary teardown of a named platform
Pulls Form ADV, Form CRS, BrokerCheck, terms of service and disclosure pages; places the product on the five-level autonomy ladder from Part I; scores six dimensions of documented investor protection; and tests marketing claims against filings.
deflate-claimStress-test a performance number
Applies the deflated Sharpe ratio, minimum backtest length and break-even trial count, then works a presentation-defect checklist — the Part V arithmetic, made repeatable. Extends into implementability when there is an actual strategy, not just a statistic.
evaluate-portfolioDecompose an actual book
Upload holdings as CSV or paste a position list; the engine regresses every holding on a factor set, rebuilds the covariance, and returns factor exposure, concentration and independent bets. A position is not an atom — and positions that look independent are often the same bet.
monitor-portfolioRe-run the decomposition on a cadence
Keyed to the strategy's rebalancing frequency; reports what drifted against both the original anchor and the prior run. Alerts are tiered and deterministic. Where the tool made a specific commitment, its own promise becomes the threshold.
review-registryResolve past verdicts and recalibrate
Revisits past evaluations, marks each falsifiable prediction confirmed or refuted, and proposes concrete changes to the scoring. This is the skill that makes the others improve — the calibration record is the only honest basis for trusting the verdicts.
The measurement that justifies the portfolio engine
AI tools produce thematically coherent baskets — thirty tickers that look like thirty decisions and behave like one bet. Standard diversification measures cannot see this. On a drift test where a diversified book was rebalanced into a single theme, effective N by risk actually rose, from 7.56 to 7.60, while the book went from 5.74 independent bets to 2.23. The two numbers that caught it — independent bets and the largest single risk source (95% of portfolio risk versus 44%) — lead every portfolio output, both computed from the trace-normalised eigen-spectrum of the risk-scaled position matrix.
What it will not do
It will not tell you a product is safe, recommend an investment, or estimate returns. It will not recommend trades or securities from a decomposition — that output describes exposure over a stated window; it does not prescribe or forecast. And it will not claim to have determined whether a product’s “AI” is real: no external observer can settle that from public information, and where that is the question the honest answer is “unknowable, and here is what would have to be true.”
AI Tool Diligence is a research instrument, distributed free under the MIT license. It produces risk classifications about structure and disclosure, not investment advice, and it is not a substitute for professional diligence, legal review, or a registered adviser. The decomposition scripts are deterministic and take no network access. Research and education, not investment advice.