The number that should have stopped everyone was not 57%. It was five.
Saturn, a UK fintech, asked 121 personal finance questions and repeated each one five times. Roughly 10,000 answers. A sample large enough to say something about consistency, not just accuracy. The published report led with failure rates. Nobody led with variance.
That omission is the tell.
A model that is wrong 57% of the time is a text problem. A model that answers the same question five different ways across five attempts is a solvency problem, because it means the output is not a measurement of anything. It is a sample from a distribution the user cannot see.
I have spent the last eighteen months building filters that separate human wallets from machine wallets on Ethereum and Solana. The lesson from that work is blunt. The dangerous failure mode of a financial AI is not being wrong. It is being confidently wrong in a shape the user cannot detect. Saturn measured the first. It never reported the second.
Saturn's test is simple in construction and unpleasant in result. 121 questions spanning tax, pensions, savings, and investment. Free-tier models failed 63% of the time. Paid models โ the ones with reasoning modes and a monthly invoice โ failed 49%. On the hardest subset, the numbers degraded to 88% and 93%. The best-performing model in the set, which the report names, still failed 39% of its answers.
Somewhere in that data sits a pension question whose wrong answer carries a ยฃ17,500 HMRC penalty. And "failed" in this methodology is not only "gave a wrong number." It also includes "failed to disclose risk." That second criterion matters more than the first, and I will come back to it.
The adoption data sitting next to the accuracy data is where the story turns structural. EY found 49% of surveyed users already lean on AI for savings and investment decisions. PensionBee found nearly six in ten would act on an AI recommendation without independently verifying it. FCA research puts trust in AI financial information at 56%, above the 47% who trust television and radio. And 44% of the same respondents believe AI financial information is regulated.
It is not. Saturn's CEO said so directly: AI financial advice sits outside the regulatory perimeter, and a consumer who loses money to it has none of the redress rights they would have with a human adviser.
Read those four numbers together and a triangle falls out. High trust. Low awareness. Zero recourse. The failure rate is a footnote. The trust-to-recourse gap is the headline.
One methodological caveat before going further. The report's model naming is not reliable enough to anchor a decision on. One label in circulation does not correspond to any released version I can verify. If the naming is imprecise, the per-model numbers are imprecise too. Hold the 39% figure loosely. Hold the structure of the finding tightly.
And one more piece of context, because it shapes how much weight the numbers deserve. The test was commissioned and run by an interested party. Saturn sells financial tooling. Independent replication โ a second lab running the same 121 questions against the same model versions, with a published scoring rubric โ does not yet exist. In my own work I have learned to treat any single-source metric as a hypothesis until a second dataset agrees. The direction of Saturn's finding is concordant with everything I observe on-chain. The magnitude is unverified.
Here is where the on-chain data diverges from the chatbot narrative, and where I think most of the commentary has been misdirected.
For seven months I have run a classification pipeline across DeFi lending activity. The method is unglamorous. I filter transaction metadata for behavioral signatures associated with autonomous agents โ fixed gas-price bidding, sub-second inter-transaction latency, deterministic call ordering, session-key authorizations, and wallet age-to-volume ratios no human can produce. On that filtered set, roughly 40% of lending-sector activity in my sample originates from wallets I classify as non-human. Not retail. Not funds. Bots. Arbitrage loops, liquidation hunters, vault rebalancers, yield routers executing on schedules no person maintains.
That single number should reframe the whole AI-and-money conversation. While researchers measure whether a chatbot can be trusted to describe a pension, autonomous agents are already executing against collateral. The advisory question is downstream of the execution question. A chatbot can take your money by convincing you. An agent can take your money by signing.
Be concrete about the difference. A chatbot that hallucinates an ISA allowance wastes your afternoon. An agent that hallucinates a liquidation threshold wastes your position. One error has a cost function denominated in embarrassment. The other is denominated in collateral. The chat interface hides error. The chain exposes it.
That asymmetry is why I keep working on a metric I call the Agent-to-Human Interaction Ratio. It is a crude instrument, but it tracks something nothing else tracks: the share of value-moving transactions in a sector that begin and end without a human decision point. In lending, it has been climbing. In DEX routing, agents are the majority. In stablecoin settlement on a handful of chains, agent-initiated transfers clear a meaningful share of daily volume during high-volatility windows.
Volume spikes don't tell you who moved the money. They tell you that money moved. The interesting question is always authorship.
The code doesn't lie about that.
What the ratio also exposes is a structural blind spot in how AI risk is being debated. Regulators measure AI harm in categories inherited from consumer protection law โ disclosure, suitability, redress. Agents do not fail in those categories. They fail in collateral ratios and gas auctions. An agent that underestimates a health factor does not mis-sell anything. It simply liquidates, and the contract executes the consequence without a complaint procedure. The mismatch between the vocabulary of AI regulation and the vocabulary of on-chain execution is not academic. It means the harms regulators are drafting rules for are not the harms the agent economy is generating.
This is the part of the argument crypto gets right and consumer fintech gets wrong. On-chain, every agent action is attributable, timestamped, and consequential โ a failed transaction reverts, gas burns, state stays untouched. Off-chain, a wrong pension answer leaves no trace. No revert. No burn. No receipt. The user acts on it and discovers the error months later, if ever.
Crypto has the better safety harness precisely because it charges for being wrong at the moment of being wrong.
I have seen that property matter before. When I traced the 2017 Parity wallet drain, the forensic work was possible only because the attackers left a permanent, queryable footprint across fourteen wallet clusters. The trail existed because the medium recorded everything, including the dusting attacks used to obscure it. Financial advice has no such property. A chatbot's suggestion to move a pension into a riskier wrapper is not an event on a ledger. It is a sentence in a scrollback window, and scrollback windows get closed.
We don't audit the interactions that leave no state change. That is the entire problem.
Now map Saturn's findings onto the agent layer. The failure modes are not exotic. Saturn's three categories โ arithmetic error, stale knowledge, invented rules โ translate directly. Arithmetic error becomes a miscalibrated health-factor threshold. Stale knowledge becomes a liquidation bot that does not know collateral parameters were changed by governance four days ago. Invented rules become an agent calling a function signature that never existed, reverting, and retrying with identical wrong inputs because nobody wrote a fallback path.
The price gradient is shallow, too. Saturn found 63% failure on free models and 49% on paid. Fourteen points for a subscription. In agent terms, that is the gap between a cheap bot and an expensive bot, and both hit the same wall on hard cases. Compute does not buy reliability in a domain whose failure mode is epistemic rather than computational. I watched the 2022 algorithmic stablecoin unwind from inside that logic. The mechanism did not fail because it ran out of processing power. It failed because its belief about its own solvency was wrong, and no amount of throughput corrects a wrong belief.
Then there are session keys, which almost nobody is discussing. An autonomous agent in a lending protocol usually does not hold your seed phrase. It holds a scoped authorization โ a session key, an allowance, a delegation. That is sold as a safety feature. It is also the precise surface where an AI error converts from advisory to financial. An agent with a session key able to move 15% of a position, executing on a hallucinated rule, is a loss event with a smart contract for a witness and no human in the loop to say no.
There is a middle path the report barely acknowledges, and it is probably where the industry lands. Not replacement, and not refusal. Hybrid deployment: a model drafts, a human verifies, a licensed entity carries the liability. I have built versions of this in the on-chain context โ a monitoring agent that flags an anomalous wallet cluster, and an analyst who decides whether the flag is a signal or noise. The agent does the volume. The human does the judgment. The liability sits with the entity that can actually pay for being wrong.
That architecture solves the Saturn problem in principle. It also explains why the report's conclusion is convenient for its author, and why the FCA's eventual position will determine which firms survive the transition.
That is the bridge Saturn's report never crosses. It stops at the text. The text is not where the money is.
Now the part the discourse will not like.
Saturn is not a neutral observer. It is a UK fintech with a compliance posture, and its CEO publicly called for FCA action. That is a legible commercial position. If AI financial advice gets pulled inside the regulatory perimeter, general-purpose model providers absorb the compliance cost, while firms that already hold licenses and redress frameworks gain a moat. The test is real. The data is not fabricated. The framing is interested.
I have written enough governance analyses to know that the entity commissioning a benchmark is never downstream of the benchmark's conclusion. That does not invalidate the finding. It does mean you should read "57% failure" as a problem-existence signal, not a problem-size measurement.
The methodological gap is real and under-discussed. Saturn never reported the five-attempt consistency rate. For financial decisions, inconsistency is worse than inaccuracy. A model that is reliably wrong teaches a user to distrust it. A model that is randomly right teaches a user nothing, and the user keeps asking. Non-determinism dressed as helpfulness is the most dangerous output shape in consumer finance. I would trade a 57% failure rate for a 40% failure rate with perfect repeatability without hesitating. I would not trade in the other direction at any price.
There is a regional bias nobody has priced. If the 121 questions skew toward UK tax and pension rules, models trained on predominantly US corpora are being penalized for a knowledge-coverage gap rather than a reasoning gap. That inflates the headline and distorts the ranking. The fix is architectural, not parametric โ retrieval against authoritative sources, tool calls into a real calculator, structured output schemas. None of those configurations appear in the published material. The report measures the failure rate of a naive deployment. It does not measure the failure rate of the technology.
Which brings me to the contrarian turn. Everyone is asking whether AI should be permitted to advise humans. Wrong question. Ask whether anyone is auditing the agents that already sign. Between the hash and the human, there is a silence, and in that silence sits the agent economy โ executing, reverting, liquidating, with no consent screen anywhere in the frame.
The forward signal I am watching is not the FCA's next statement, though that will matter. It is the Agent-to-Human Interaction Ratio in lending. If it keeps climbing while model failure rates on financial reasoning stay in the double digits, the industry is scaling a capability gap into a solvency surface. The first serious loss will not come from a chatbot's bad advice. It will come from a well-authorized agent doing exactly what it was programmed to do, correctly, in a world whose parameters changed after the last retrain.
Watch the ratio. Watch session-key revocations. And ask the question nobody is asking: if we would not let a 39%-failure model give pension advice without a human signature, why are we comfortable letting a similar model hold a key?