Strategic Acquisition Brief · Confidential
In July, the firm published research showing eight AI agents beating a 60/40 allocation across two decades of historical data — and was careful, correctly, to note that a backtest is not proof that AI can consistently outperform. That caution is the whole problem with agentic advice at consumer scale, and FINRA has now named it: multi-step agent reasoning is difficult to trace or explain. MaxiFi is the part that does not have that problem: computationally exact, economics-based planning — for a household’s facts and assumptions, it solves, not guesses, the lifetime plan, every dollar of taxes and benefits computed under current law. Deterministic, reproducible, auditable.
On 9 July, strategists led by Thomas Salopek published results from an array of AI-powered investing agents that shift between stocks and bonds as market conditions change. Every one of the eight beat both a standard 60/40 allocation and the bank’s own rules-based regime model. The note was explicit that these were historical simulations and should not be read as proof of consistent outperformance.
That combination — genuine capability, honestly qualified — is the state of the art, and it is the position the whole industry now occupies. The firm runs more than 450 AI use cases in production against a plan for a thousand, on a technology budget above $18 billion, with both frontier labs as partners.
An agent that allocates capital, or tells a customer how to fund a retirement, is making a recommendation in substance whatever it is called in the product. FINRA’s 2026 report names the risk precisely: complicated, multi-step agent reasoning can make outcomes difficult to trace or explain, which complicates auditability — alongside domain-knowledge gaps and autonomy without human validation.
Backtested is not auditable. Confident is not correct. At the scale of a consumer franchise, the difference between those pairs is the entire risk position.
MaxiFi does not compete with the agent stack. It is the computation service the agent calls when the question has a dollar answer and a forty-year consequence — the same architectural discipline the firm already applies wherever a wrong number is unacceptable.
Wealth Plan, J.P. Morgan Personal Advisors, the Chase app. Same surfaces, same conversational interface, same advisors.
The orchestration, the models, the internal platform. Powerful, improving, and correctly qualified — and never asked to do the arithmetic.
The rules, the solver, the audit trail. Same inputs, same answer, every time, traceable line by line to the law tables in force on the plan date.
A planning output that can be reconstructed and defended years later — the thing a backtest, by construction, cannot provide.
This is not a novel doctrine; it is the one the industry already applies to tax, to pricing, to risk. Lifetime financial planning is exactly such a function, and getting it wrong is its own kind of disaster: the retiree who runs out of money at 82, the family under-insured by a million dollars.
It is also precisely where a model, left alone, fails — because it reaches for the same rules of thumb the planning incumbents use, and in this domain approximation is not close enough. It is wrong, in ways that compound every year.
MaxiFi is the financial-planning platform of Economic Security Planning, Inc., built over more than three decades by Professor Laurence Kotlikoff of Boston University. It uses consumption smoothing and dynamic programming to compute the single, mathematically optimal lifetime plan — solving simultaneously across Social Security strategy, federal and state taxes, Roth-conversion sequencing, withdrawal order, life-insurance need, estate planning, and upside investing.
Goals-based tools and rule-of-thumb calculators answer “What is the chance you hit your number?” MaxiFi answers “What is the optimal path, and how much can I spend today without jeopardizing tomorrow?” It is not a better simulator. It is a different class of engine.
Prof. Laurence Kotlikoff — William Fairfield Warren Professor at Boston University; Harvard Ph.D.; former Senior Economist on the President’s Council of Economic Advisers; named by The Economist among the 25 most influential economists. He intends to stay on with the acquirer in whatever capacity best serves the product. The more important fact for an acquirer is that the engine’s currency does not rest on it: rule maintenance is routine engineering, not founder work, and runs without his involvement.
MaxiFi’s economics build on Nobel-laureate work, and Nobel laureate Robert Merton teaches with MaxiFi at MIT Sloan as an “outstanding science-based lifecycle and retirement management platform.” Featured in Bankrate’s “Best financial planning software of 2025” roundup, cited as best for near- and long-term tax planning and the decumulation phase.
The moat is the rulebase as much as the solver: thirty years of encoded, continuously maintained federal and state tax, Social Security and benefit rules, carried under a regression suite re-run against every law change, plus patent-winning optimization algorithms built from economic theory rather than scraped text. The maintained surface is concrete: federal, Social Security, Medicare Part B and 42 state income tax codes, updated as provisions are released. Stated plainly, because it will be checked: the solver is the replicable half — the mathematics is published, much of it by Kotlikoff himself. The rulebase is not, because encoding thirty years of law correctly is the decade.
Planning tools die on data entry. Inside a bank that already holds the deposit, card, mortgage, investment and retirement relationship, the inputs problem largely disappears — and the same computed plan can be surfaced through a self-directed app, an advisor-led relationship, or an agent, with one auditable answer behind all three rather than three different approximations.
A firm building its own agents has good reason to be sceptical of buying technology. The distinction that matters is between the half of this asset a capable team could rebuild and the half it could not.
The solver is the replicable half. The mathematics of lifecycle consumption smoothing is published, much of it by Kotlikoff himself, and a strong quantitative team with a frontier model could write one.
The rulebase is not. Thirty years of encoded federal and state tax, Social Security and Medicare provisions — versioned, continuously maintained, and carried under a regression suite re-run against every law change. Not because the rules are secret, but because encoding them correctly and proving they are encoded correctly is the decade. That is the half AI has made scarcer, not cheaper: a model has no correct reference point, so no error in its output is decidable.
MaxiFi does not approximate. It computes — iteratively, multivariately and simultaneously across taxes, benefits, longevity and cash flow, year by year for a whole life. It is provable, not merely confident: the answer that holds up when someone with an adverse interest checks the math. That claim is about the computation — the optimization and the tax and benefit math are exact and inspectable — not about predicting markets.
And the clock is real. A build arrives in years; the agents, the liability and the competitive window run in quarters. The engine — and its economist — exist now, once.
The report identifies, as explicit risks of agentic AI: auditability and transparency — complicated, multi-step agent reasoning can make outcomes difficult to trace or explain; domain knowledge — general-purpose agents may lack what complex, industry-specific tasks require; and autonomy — agents acting beyond intended scope or without human validation. FINRA and the U.S. Treasury have since published an AI Lexicon and a Financial Services AI Risk Management Framework.
Being “AI-generated” is not a liability shield, and the substance of a recommendation is governed regardless of the interface delivering it. For a consumer franchise measured in tens of millions of relationships, the exposure is not theoretical — it is the reason the agent programme is carefully qualified in the first place.
A correct-by-construction engine addresses the exposure directly: if the math is right, reproducible and auditable, the answer holds up to scrutiny on its own terms — including the scrutiny of an examiner reconstructing a recommendation made three years earlier under a different tax regime.
And because the engine is deterministic, the assurance can be underwritten: a bounded accuracy guarantee no probabilistic rival can offer, because their output has no correct reference point to warrant.
CBS MoneyWatch (May 7, 2026) ran an identical retirement question — a 50-year-old single woman retiring at 65 — through two leading AI models. The verdicts diverged. MIT’s Andrew Lo was quoted on the underlying structural point: today’s consumer AI carries no best-interest duty. Kotlikoff was quoted describing the risk that AI “may do more harm than good” when it mishandles claims like Social Security timing or substitutes an average for a maximum life expectancy.
A concrete, checkable example: AI engines trained before the One Big Beautiful Bill Act (enacted July 2025) told users the federal estate-tax exemption would “sunset” on January 1, 2026 — reverting to roughly half its level. In fact, the Act permanently raised the exemption to $15 million per person starting in 2026.
A model repeating pre-2025 training data would confidently tell a household to rush an irrevocable estate move it no longer needs — a costly, hard-to-reverse error delivered with total confidence. A computed engine, fed current law, does not carry stale assumptions forward as fact.
Neither example is about any single company’s brand. It is the same structural point twice: confidence is not correctness, and an answer’s value depends on the currency and correctness of the computation behind it — not the fluency of the sentence delivering it.
Larry’s Economics Matters Substack — 137,000+ subscribers — has run a six-post sequence testing named frontier engines against MaxiFi on dollar-specific household problems, including the labs the firm partners with. The variance across engines on identical, checkable prompts is the proof: the correctness cannot come from the model layer.
Acquiring MaxiFi acquires the megaphone these pieces ship from — pointed, with credibility no one in the category can match, at the exact question the firm’s own research raised. The CBS finding is the named, neutral proof; the Substack series is the dated, dollar-specific record behind it.
Durable value accrues to whoever owns the deterministic engine under the trusted interface — not to the interface, and not to the model. The firm partnered with both frontier labs precisely because neither differentiates it. The one un-owned layer left in consumer finance is the deterministic planning engine.
The constraint on putting planning guidance in front of tens of millions of customers is not model quality — it is what the firm is willing to stand behind. A computed, auditable answer moves planning from a qualified experience to a defensible product, across self-directed, advisor-led and agentic channels at once.
MaxiFi’s determinism makes a planning-side accuracy guarantee offerable for the first time: a computational error is objectively decidable, so the warranty prices at a rounding error and is insurable. Bank of America, Wells and Schwab cannot answer it, because a goals-based output has no correct reference point to warrant.
The firm is further into agentic deployment than its peers, which means it carries more of this exposure than its peers. A correct-by-construction engine retires the largest part of it. We are not selling an insurance policy; the insurance is included.
There is exactly one MaxiFi and it will sit somewhere. In the model layer it reaches every customer of those models through the same API the firm rents — competitors included. Owned, it is deployed under your brand and available to others only on your terms.
The July research answered whether an agent can allocate well. It could not answer whether any particular recommendation to any particular household is right, because a backtest cannot. A deterministic engine answers exactly that question, for every household, every time, and shows its work.
MaxiFi is being offered through a focused strategic process — the engine, its IP, and thirty years of R&D. The preference is an acquisition; that is where the strategic value sits. Continuity de-risks it: Larry Kotlikoff intends to stay on with the acquirer in whatever capacity best serves the product — architect, spokesperson, advisor. The next step is a 30-minute live demonstration: MaxiFi solves a real household’s plan while the leading models are asked to match it. The gap is the thesis. Evidence deepens with commitment — nothing is deployed, nothing left behind, and the full case is provable in an acquirer’s first quarter of ownership.