Clarify intent
The LLM structures business context, surfaces ambiguity and proposes a formulation using approved financial definitions.
The decision engine for finance
Decide with rigor.Act with control.
AGX brings business context, financial models and mathematical optimization together to help finance teams decide how to fund obligations, deploy cash and manage risk within explicit limits.
In development · Demo in preparation
01 / A decision in practice
Open full example What changes when your priorities do?
The numbers live in systems.
The decision spans them.
The decision spans them.
Cash balances, payment schedules and forecasts rarely answer the question on their own: what should we do next? AGX is being built to connect those inputs to a decision that finance can inspect, challenge and authorize.
02 / Building the decision
The model starts with
your business definitions.
“Preserve liquidity” is an objective. Turning it into a model requires agreement on what liquidity means, over which horizon, and under which conditions.
- 01
Define the decision
Owner: finance leadershipSpecify the objective, available actions and non-negotiable limits. Make acceptable trade-offs explicit, including which targets may flex and who may approve exceptions.
- 02
Ground it in business meaning
Owners: finance and data teamsMap entities, facilities, obligations and cash flows to their sources. Agree definitions, currencies, dates and data freshness. Missing or conflicting facts remain unresolved until verified.
- 03
Approve the assumptions
Owners: finance and model reviewersSelect the forecast, stress scenarios and treatment of uncertainty. Review dependencies between cash flows. Risk appetite, probabilities and objective weights require an accountable owner.
- 04
Formulate, test and compare
Owners: modeling team and decision ownerTranslate approved definitions into variables, constraints and an objective. Check units, feasibility, known cases and sensitivity before reviewing the solver’s proposed action and alternatives.
An optimal solution is conditional on its model. Validation must also establish whether that model is a useful representation of the business.
03 / Technical architecture
A distinct job for
each part of the system.
AGX’s design separates interpretation, financial modeling, optimization and authorization. The model, proposed action and approval each have their own checks.
Quantify consequences
Reviewed financial and risk models supply cash-flow effects, scenarios and assumptions about uncertainty.
Evaluate the options
Mathematical solvers search within the formulation. The result must state feasibility, solve status and any remaining optimality gap.
Authorize action
Policy checks and delegated authority determine whether a proposed action may proceed through connected execution systems.
The integration objective: carry one versioned decision from source data and model assumptions through approval and execution evidence.
Your existing stack: ERP, treasury and banking systems remain the sources and execution channels. Connection and deployment requirements are scoped for each use case.
04 / Control & evidence
Know when to proceed.
Know when to stop.
The intended control model keeps authority explicit and gives reviewers enough context to challenge a decision, including the assumptions behind it.
- Unresolved inputs stop the decision.
- Stale data, conflicting balances or an unapproved assumption require a named owner to resolve them. A plausible estimate cannot silently become a verified fact.
- Infeasible means escalate.
- If no action satisfies the hard constraints, report the conflict. Any change to a limit or objective requires explicit approval.
- Authority is limited and revocable.
- Begin with recommendations and human review. Any later execution authority is restricted by decision type, amount and conditions, with an owner who can pause it.
- Execution needs its own checks.
- Recheck data and authorization before acting. If an external system gives an uncertain result, reconcile its status before retrying a financial instruction.
05 / Where to start
One recurring decision.
A measurable test.
A useful starting point is a treasury or finance operations decision with material consequences, several feasible actions and limits you can state clearly.
Choose the decision.
Cash positioning, short-term funding or payment timing. Define its owner, cadence, source systems and the approach used today.
Agree what good looks like.
Set the baseline and evaluation criteria before testing: funding cost, liquidity headroom, review time and exceptions. Include historical cases and adverse scenarios.
Compare before acting.
Evaluate recommendations alongside the current process. Review disagreements and model failures before considering any execution authority.
Broader applications, including working capital and capital allocation, require their own models, evidence and approval. Each new decision has to stand on its own evaluation.
Before we talk
A few practical
questions.
What can I see today?
AGX is in development and a demo is being prepared. The example on this page illustrates the approach. Customer deployments, measured outcomes and product performance results are not yet available.
How does this fit with our existing tools?
Treasury platforms, ERP systems, analytics and internal models already support parts of the process. AGX’s focus is the connection between the decision model, its controls and the resulting action. A first assessment should establish whether that connection solves a gap in your current setup.
Why combine an LLM with a solver?
Language models can help express and structure a problem; a solver evaluates the formal model. Research such as MIT’s work on LLM-based formalized programming explores this separation. Applying it to finance also requires approved definitions, model validation and authorization controls.
What about security and deployment?
Data access, identity, segregation of duties, retention, hosting and system connections need to be agreed for the specific environment before a pilot. Supported integrations and security controls must be demonstrated during that evaluation.
Start with your decision
What decision would
you like to make better?
Start with one recurring finance decision. The assessment looks for a meaningful opportunity, interacting choices and hard constraints, then identifies what we would need to examine in a research interview.
Assess your decision
Six short steps. A preliminary research-fit result, with no file upload required.
Have a question? Contact Agentronics