AnantState
Industry · Financial services

An account is a relationship with a history.

AnantState tracks where each account, merchant or case is heading, with the evidence and the cost of acting attached, so a decision reads as a reviewable record.

In plain termsSee how each account is moving, beyond the flag it has today.

One case, start to finish

An entity here is usually a relationship with a history: an account, a member, a merchant, a card, a device, a dispute. The case below is ranked with its evidence, and where sources disagree it is marked contentious, because that is the case an experienced analyst should see and the one a single score hides.

One case, start to finish

  1. 1. Account activity to AnantState: Behavior breaks from its history
  2. 2. AnantState to An analyst: A ranked case, sources in conflict
  3. 3. An analyst to AnantState: Approves a priced hold
  4. 4. AnantState to Audit trail: Evidence, decision, actor, timea reply or asynchronous
A case is ranked with its evidence, including where sources disagree, an analyst approves a priced action, and the evidence, decision, actor and time go on the audit trail.

The false-positive question

Declining a legitimate customer is a real loss that rarely appears as a line item, so the rule that causes it is rarely challenged. Step-up authentication, a hold and a manual review each carry a cost here, which turns the false-positive trade-off into an explicit, reviewable decision rather than a rule someone set two years ago.

  • Alert volume is managed by headcount when it could be managed by value: the queue is ranked by money at stake.
  • Model risk review runs behind the pattern, so the validation function gets a record it can read, including the horizons where the model does not beat the simple guess.
One OgMart entity opened: ranked candidate actions with cost and return, a notice that the evidence is conflicting so autonomous action is held, the top drivers, and the trajectory against the events that triggered it
One entity in depth: what to do, what it costs, why the platform thinks so, and a flag where the evidence disagrees.

What moves in an account

StateBaseline driversUnexpected drivers
Fraud postureSeasonality, channel mix, product behaviorNew attack patterns, account takeover, mule activity
Credit postureTenor, payment cycle, utilizationEmployment change, hidden leverage, behavior break
Dispute pressureVolume, product, partnerMerchant failure, systemic event
Investigation loadCase arrival rateA single organized pattern creating correlated cases

A worked example

Five thousand open cases, and a team that can review one in seven.

Round, illustrative figures for open cases: they show the arithmetic, leave out the subscription, and are for you to change in the calculator.

InputValueWhat it stands for
Open cases in scope5,000Everything the team is responsible for
Value at stake per one$1,800Loss on one case if nothing is done
Cost of the action per one$45A manual review or step-up check
Share acted on15%The cases at the top of the queue: 750 of 5,000
Modeled, USDPessimisticExpectedOptimistic
Share of value at stake avoided5%15%30%
Cost to act, in total$33,750$33,750$33,750
Avoided loss$67,500$202,500$405,000
Net position$33,750$168,750$371,250

Break-even is 2.5%: each action costs $45 against $1,800 at stake, so it has to avoid that share of the loss to pay for itself.

Regulatory notes

  • Explainability is a regulatory expectation in most jurisdictions, and an engineering property here: every conclusion carries its evidence.
  • Model governance maps to model risk management expectations: documented purpose, measured performance against a defined baseline, a recorded promotion decision, and drift monitoring.
  • Data residency and tenancy are deployment questions. See governance and control.

What exists today

Financial services exists as a catalog template: loadable, and the starting point for an ontology workshop.

Price the false positive

Bring a rule whose false-positive cost nobody has measured. We will price it against the cost of the loss it prevents.