Every important decision balances competing priorities — and most AI tools blend them into a single answer behind the scenes. UtilityMax lets you set the priorities yourself, in the prompt.
Validated across three frontier models · arXiv:2603.11583
Tell us the decision you want made — rank candidates, optimise a page, generate an email — and the context that matters. The role spec, the brand guidelines, the regulatory regime, the customer profile. The kind of background you would give a thoughtful colleague taking on the work.
The framework breaks your decision into the components that actually matter — and scores each one independently. Cultural alignment is its own score, separate from technical fit. Compliance is its own check, separate from brand voice. Each component carries its own evidence, traceable back to the source.
A ranked shortlist, an optimised draft, or a generated artefact — with the reasoning behind it. Per-component scores, per-component evidence, and the configuration that produced the answer. If your priorities change, the configuration changes — and the same input produces a new answer the framework can justify.
The task's competing objectives are decomposed into an influence diagram. The LLM's answer is the sole decision node. Each objective becomes a precisely defined chance node — a random variable whose outcome can be observed and measured.
A utility function is defined over the chance nodes. Multiplicative when all objectives must jointly hold. Additive when each contributes independently. The ambiguity inherent in natural language is eliminated entirely through formal mathematical specification.
The LLM generates candidate answers, estimates each utility component individually using its domain knowledge, computes the objective score for each candidate, and returns the answer that maximises expected utility.
"Rather than restructuring how the model reasons, UtilityMax replaces the natural language objective itself with a formal mathematical specification — constraining the LLM to reason explicitly about each component of the objective."
— arXiv:2603.11583The following analyses were conducted on real content across independent client engagements.
Artist Biography
Maximise profile follows and stream conversions from first-time listeners encountering the artist for the first time.
Press Release
Reinforce platform credibility as action-driving rather than awareness-only, and build confidence among enterprise prospects in the product roadmap.
Job Specification
Attract passively employed senior candidates with domain-specific expertise while filtering out applicants with only general experience.
Property Listing
Convert sceptical out-of-town buyers on an unfamiliar short-term rental investment model by demonstrating honest, verifiable ROI.
Each percentage reflects a composite utility score across five formally defined conditions, estimated independently against the stated objective. Each analysis completed in under one minute.
The disagreement in the room becomes a configuration on the page.
UtilityMax is available for consulting engagements across strategy, content, recruitment, product, and any domain where a decision objective can be formalised.
If you have a problem with competing objectives and want to know what the optimal answer looks like, get in touch.