# Pints AI · AI your regulators can trust

> Enterprise agentic AI for regulated industries.
> Production agents for banks, insurers and asset managers. Every decision
> governed, source-attributed and defensible by design, running inside the
> institution's own perimeter.

- Domains: KYC, claims adjudication, underwriting, document intelligence
- Pages: / (the argument) · /platform.html (Autothought, the export,
  delivery) · /solutions.html (workflows by industry) · /trust.html
  (security, compliance, drift, research) · /proof.html (production figures,
  testimonial) · /about.html (company, milestones, leadership) ·
  /blog.html (news & press, insights) · /blog-pre-series-a.html (the
  Pre-Series A announcement, published 15 Jun 2026) · /faq.html (38
  questions across seven categories, served static with FAQPage JSON-LD)
  · /events.html (FOUNDATION, a closed forum with Profinch, Singapore,
  24-26 Sep 2026, by invitation) · /contact.html (talk to us)
- CTA: Talk to us · https://pints.ai/contact
- In production with: three institutions in insurance and diversified financial services.
  Named where written permission to publish is on file, and nowhere else: one of
  the three is named against the testimonial below; the other two are described
  by sector until their permission lands.

## The problem

You are not short of AI solutions. You are short of AI you can defend.

1. **Pilot purgatory**: proofs of concept that never clear risk review,
   because nobody can explain how the model reached its answer.
2. **Black-box output**: decisions with no line back to the source document;
   indefensible the day a regulator or auditor asks.
3. **Data leaving the perimeter**: vendor APIs that move regulated records
   across jurisdictions your licence does not cover.

## Autothought · the platform

Pipeline: Source -> Agent -> Decision -> Audit trail

| Stage | What happens |
| --- | --- |
| Source | Documents ingested; every field indexed against its origin |
| Agent | Multi-step reasoning (extraction, rules, fraud checks), logged as it executes |
| Decision | Approve/deny, confidence-scored, and handed to a named human role: the agent recommends, your reviewer decides |
| Audit trail | Complete run exports as one trail; attribution down to page and field |

Example run · Claims Adjudication Agent (09:23:15 → 09:23:27):

```
09:23:15  document_intake    claim_form_2847.pdf, medical_records_batch_7.pdf
09:23:18  data_extraction    injury="Lumbar strain"; charges=$15,400  [Page 2 · Field 8]
09:23:22  rule_application   coverage=active; within_limits=true      [Policy §4.2]
09:23:25  fraud_check        provider=in-network; anomalies=none      [Registry match]
09:23:27  routed             recommend=APPROVE; confidence=91.2%; threshold=95.0%;
                             below_threshold=true; assignee=senior_claims_assessor
```

The run is illustrative: file names, values, scores and the threshold describe a
fictional case. The threshold is the institution's policy, not a Pints default,
and the confidence figure in *Proof in production* below is an aggregate across
production runs, not the score on this or any single demo case.

## Solutions by industry

Different workflows, the same burden of proof. Autothought runs the
compliance-heavy work of insurance, asset and wealth management, and
commercial banking.

- **Insurance**: claims adjudication (FNOL, triage, fraud screen,
  settlement) · underwriting file review (in production, figures under
  *Proof in production*) · medical record review · policy issuance and
  servicing · regulatory and statutory reporting (RBC, statutory returns)
- **Asset & wealth**: investment research digests with citations ·
  portfolio risk monitoring · client reporting · compliance monitoring
  (personal trading, mandates, marketing review) · RFP and due diligence
  responses from an approved answer library
- **Commercial banking**: loan application processing (spreading, credit
  memo) · covenant monitoring · KYC and onboarding (CDD, gap list) · AML and
  sanctions alert review · annual credit reviews

Every card on the page keeps one shape: what the workflow does in plain
language, then how it stays defensible. The featured workflow band repeats
the underwriting review-time and error-rate figures from *Proof in
production* below, same denominators, no new numbers.

## Deployment

Three modes, one property: the work runs inside the institution's own
perimeter. VPC (your cloud tenancy, your region, your keys; no traffic to
Pints AI) · On-prem (your data centre; no external dependency in the serving
path) · Air-gapped (an isolated network with no route out; models and updates
ship into the gap).

## How we deliver

1. **Workshop** (half a day): your documents, rules and infrastructure
   constraints; one workflow mapped end-to-end.
2. **Pilot** (weeks, not quarters): a governed agent on that workflow, inside
   your perimeter, scored by AutoEval against your own settled cases.
3. **Production** (owned by you): your models, your infrastructure,
   continuous evaluation, full audit trail per run.

## Proof in production

- Underwriting file review: 60 min -> 5 min median per file, and field error
  rate 4.1% -> 0.9% on the same files
  (n = 1,240 files, 12 weeks to 31 Jul 2026, measured 4 Aug 2026)
- Decision confidence: 98.3%, mean per-run score from AutoAssure, the
  confidence scorer that grades every run and fires the review gate
  (n = 18,412 runs, 12 weeks to 31 Jul 2026, measured 4 Aug 2026)
- Source-attributed output: 100%, every value traces to page and field
  (n = 214,880 extracted values, 12 weeks to 31 Jul 2026, measured 4 Aug 2026)

Scope: review time and error rate are one insurer's production underwriting
queue; confidence and attribution are every production run in the window,
across all three institutions. Re-measured quarterly; next 4 Nov 2026.

> "The first AI system our risk committee approved without a caveat. When they
> asked how a decision was made, we showed them the trail."
> · Adeline Koh, Chief Operating Officer, Havenmark Life Assurance, Singapore
> (published with Havenmark's written permission)

## Why Pints

1. **Own the AI pipeline**: model-agnostic, infrastructure-agnostic, no lock-in.
2. **Governed by design**: the audit trail is the architecture, not a report.
3. **Built for regulated work**: KYC, claims, underwriting, document intelligence.
4. **Proven in production**: deployed inside the institution's own perimeter,
   on its own models and infrastructure, in Asia, India and the Middle East.

## Trust & compliance

- Deploy inside your perimeter: on-premise, private cloud or VPC
- Privacy: agents read records in place; nothing is retained by Pints AI and
  nothing enters a training set (stated in the DPA)
- Model-agnostic: commercial, open-weight or fine-tuned, swappable
  without rebuilding the pipeline
- Token cost: each step runs on the smallest model that clears the workflow's
  pass mark; routing, caching and batching tuned per workflow; spend reported
  per run
- Human-in-the-loop: confidence gates route edge cases to your reviewers
- Continuous evaluation: AutoEval, Autothought's evaluation service, scores
  every agent against the institution's own settled cases, before deployment
  and continuously after it
- Compliance posture (Aug 2026): SOC 2 Type II and ISO/IEC 27001:2022
  certified; MAS TRM mapped; ABS OSPAR attestation audit booked Oct 2026;
  PDPA and GDPR compliant; CREST pen test Jun 2026; four subprocessors,
  all listed. In-perimeter deployment means your certifications govern the
  data path: the records never reach us
- Access control: role-based, every read and override is an audit event

## Research

- Report: "AI Sovereignty: why regulated institutions are bringing the
  pipeline in-house" (PDF, email-gated, no sales follow-up unless you ask)

## Company

- Pints AI Pte. Ltd. · Singapore · Asia · India · Middle East
- Founded 2021 in Singapore; Autothought in production with insurance
  clients since 2024; US$5.6M pre-Series A (2026) led by Tin Men Capital,
  co-led by SBI Ven Capital, to scale across Asia Pacific and the Gulf
- Recognition: AI Singapore AI in Finance Global Challenge grant (2024),
  SFF Global FinTech Award (2023); programmes and partnerships with
  AI Singapore, MAS, AWS and Google (not customers)
- Milestones (aggregate across client deployments; per-workflow figures on
  /proof.html): $13.2M saved for clients in under two years · 12 live
  production deployments · 1.5 Pints, a compact domain-trained language
  model for insurance
- Leadership: Partha Rao (Co-Founder & CEO) · Calvin Tan (Co-Founder & CTO)
- Pre-Series A announcement (published 15 Jun 2026, /blog-pre-series-a.html):
  participation from SEEDS Capital, NTUitive, SUTD Venture Fund and Tenity
  alongside the leads; as at that date, 12 institutions across four
  countries had saved more than US$10 million (the later $13.2M figure on
  /about.html is the same series, measured later)
- Research partners: SUTD, AI Singapore
- Contact: Talk to us · https://pints.ai/contact