The agent that settles it before the deadline.
Setliva is an AI agent for revenue recovery. It reads the evidence, decides the next action inside your policy, and produces the artifact that action requires — a network-compliant representment, or the next step in a receivables sequence. Every claim it makes traces to a record.
- +10AVS full matchEV-4417-02
- +3.2CVV2 matchEV-4417-02
- +28Signed delivery to the billing addressEV-4417-03
- +7Terms accepted at checkoutEV-4417-06
- +18Receipt acknowledged in writingEV-4417-05
5 exhibits attached · self-check passed
Open this case →Read from the engine when this page rendered — the gate count is the length of the gate catalogue, the fact count the size of the rule tables.
Five stages the agent runs on every case.
Extract
Read the systems of record — carrier documents, processor responses, ledgers, CRMs, chat threads — and derive facts, each one carrying the record it came from. If a record is absent the fact is absent, and that silence is scored.
Weigh
Score those facts against your thresholds, not ours. Corroborating evidence of the same kind discounts rather than stacking — three proofs the buyer was real are not three times one.
Decide
Aggregate into an outcome, with 15 hard gates that bypass the score entirely. A weighted model that can average away "we already refunded this" is the wrong model, so those facts never enter the arithmetic.
Express
Assemble the artifact by binding every sentence to a signal it can prove. The agent drafts from the evidence rather than composing freely, so it cannot assert what the evidence does not contain.
Self-check
Verify every citation resolves to a signal, every signal to a record, and every monetary figure to one or the other, before anything is released.
The models read the evidence, score it and draft the filing. What they never do is decide alone. Every outcome passes through a policy core where a disqualifying fact overrides the score outright — because a model that can average away “we already refunded this” is the difference between a recovery and a pre-arbitration loss.
That boundary is what makes an agent deployable in a regulated operation. Generated text is bound to extracted signals and re-checked before release, and the decision path is reproducible — the same case a year from now produces the same digest, verified across processes rather than merely across calls.
One representment. Spend it on the right case.
Merchants lose chargebacks by default because assembling the evidence inside the network window costs more than the transaction. The usual answer is to fight everything, which is worse: a lost representment forfeits the only attempt and moves the case to pre-arbitration, where the arbitration fee alone exceeds most tickets.
Setliva reads delivery confirmations, session and device continuity, AVS and CVV2 responses, 3-D Secure results and their liability implications, terms acceptance, access logs and prior transaction history; checks Compelling Evidence 3.0 qualification against the real age and matching-element rules; and files only what clears your evidence floor, 58 of 100.
Most disputes should not be fought, and it concedes them — including cases it would probably win on the merits but which fall under your $25.00 net value floor. Deciding what not to spend the representment on is most of the work.
- G-REFUNDED
A refund already settled. Representing claims money that was returned.
- G-DUPLICATE
Two identical charges, second unrefunded. The cardholder is right.
- G-WINDOW-CLOSED
The network window has passed. Recorded so the intake delay is visible.
- G-DOCUMENTED-DEFECT
The cardholder photographed the defect and your policy provides a remedy.
- G-HIGH-VALUE
Above your ceiling. The packet is ready; a person signs it.
- G-NO-EVIDENCE
Nothing to file. A representment with no exhibits is not a representment.
Days past due is a property of the invoice, not the payer.
A customer who is reliably thirty days late is a different problem from one who is thirty days late for the first time. Treating them alike is what makes conventional dunning ladders so ineffective, and so corrosive to the relationship.
Setliva scores propensity to settle unaided from the payment record itself — median pay lag against terms, the spread around it, promise-keeping ratio, write-offs, responsiveness, part-payment pattern and tenure. The score is shown band by band, with the contribution of each input on the page, because a propensity score nobody can argue with is worthless to whoever has to sign the escalation.
Above your leave-alone ceiling of 82, a recently-due account is left alone. Doing nothing is a first-class outcome here, and a ladder that cannot express it will burn a good customer to collect an invoice they were going to pay.
B2B receivables where the terms are informal and the channel is WhatsApp.
Collections software assumes an email address, a signed contract and a days-past-due counter. Across much of the world none of those hold: terms are agreed verbally, the working channel is WhatsApp, settlement runs over PIX, UPI or M-Pesa, and part-payment is normal rather than a default event.
That is precisely the shape a conversational agent fits and a rules engine does not. Setliva works those accounts in channel: it reads the reply, recognises a promise to pay and the terms attached to it, and escalates only once the thread stops moving. Consent governance still binds it — an account with no recorded consent for WhatsApp does not get a WhatsApp message. The agent falls back to email rather than sending, and records why.
Judge it on the working, not on the pitch.
It is built for two shapes of book. Card-not-present merchants in digital goods, furniture and B2B software, where disputes arrive faster than an analyst can assemble the evidence for them. And B2B suppliers running informal terms across emerging markets, where the ledger is real but the contract is a WhatsApp thread and part-payment is normal rather than a default event.
Open any case and you can see the records it read, the facts it drew from them with the weight applied to each and why, the gate that fired if one did, the artifact it produced, and the self-check that had to pass before release. The scope and limitations page sets out where it stops — worth reading before the demo, not after.
A share of what it recovers. Nothing otherwise.
Between 5% and 20% of recovered value, banded by volume and mix. No seat licence, no platform fee, no minimum. If a month recovers nothing, that month costs nothing — which is the only structure under which a vendor and a collections team want the same thing.
Digital goods and subscription merchants above 5,000 disputes a month.
Mixed dispute and receivables books. The band most teams land in.
B2B receivables and high-value disputes where each case carries real assembly cost.
What it displaces: a dispute analyst at roughly 22 minutes a case, or an agency taking 25–35% of face value and the relationship with it.
The arithmetic. 400 disputes × 21% contested = 86 filed. × 90% mean win probability = 78 recovered × $120 = $9,360. Your current 12% recovers $5,760. The fee applies only to the difference.
The 21% contest rate and 90% mean win probability are modelled from the engine’s own decisioning, and both move with your evidence mix — a book with reliable delivery data contests more and wins more. Around 54% of cases never reach scoring at all, because a hard gate settles them. The 314 cases not contested still consume roughly 115 analyst hours a month today, and that time is the part most teams underestimate.
Recover what you are already writing off.
Open the queue and read a case end to end — the evidence, the working, the gate that fired, and the artifact it produced. Nothing is behind a form.