How it works

Watch one document become a verified entry.

Follow it from your queue, through the AI read, into your software and back into your records.

See it work Book a demo

Sumvant sits between your software and your team. It types, clicks and files in your software and reads the screen back; inside, every item goes read, decide, act, read back; it sends exceptions to your team with the reason, and your team gives it rules in plain sentences.

Sumvant sits where a person would, between the software and the people who own the rules, and nothing about your systems has to change.

The data flow

Nine beats, one document.

The film plays on its own and loops, and any dot on the rail jumps to that beat.

An illustration of the path with sample data, not a screen from any client’s system.

  1. Beat 1 of 9 · Illustration
    State
    SESSION: SVC_AGENT · 2FA ROUTED · WORKSTATION LOCKED
    Mechanism
    Dedicated service identity on a dedicated workstation
    Metric
    adapters loaded 3 · health check ok

    The agent works on your workstation, signed in as itself.

  2. Beat 2 of 9 · Illustration
    State
    QUEUE: EMAIL, FAX, PORTAL · 3 NEW
    Mechanism
    Queue intake opens each item as a governed work order
    Metric
    work orders 3 · priority by rule

    Work arrives the way it always does.

  3. Beat 3 of 9 · Illustration
    State
    VISION: DOCUMENT INTELLIGENCE
    Mechanism
    Layout aware AI vision extraction, field by field
    Metric
    14 fields · min confidence 0.95

    The agent finds every field your rules ask for, each with a confidence value.

  4. Beat 4 of 9 · Illustration
    State
    TRANSPORT: TLS 1.2+ · AWS BEDROCK · BAA
    Mechanism
    Zero retention inference, fields returned as structured JSON
    Metric
    round trip 1.8 s · retained none

    One encrypted trip out for the AI read. Fields come back as data.

  5. Beat 5 of 9 · Illustration
    State
    RULES: RULE SET v12 · DECISION TRACE
    Mechanism
    Deterministic rules engine evaluates your SOPs in order
    Metric
    9 rules evaluated · 1 below threshold

    Your versioned rules decide what happens next, and one of them routes.

  6. Beat 6 of 9 · Illustration
    State
    SCREEN STATE: ENTRY_FORM · ADAPTER: YOUR_SYSTEM
    Mechanism
    System adapter drives the entry screen through its own login
    Metric
    14 fields typed · 0 retries

    It types into your software like a trained person.

  7. Beat 7 of 9 · Illustration
    State
    READ BACK: 14 OF 14 MATCH
    Mechanism
    Each committed field re read from the screen and compared with source
    Metric
    mismatches 0 · commit verified

    Nothing counts until it is read back from the screen.

  8. Beat 8 of 9 · Illustration
    State
    ROUTE: EXCEPTION · REASON ATTACHED
    Mechanism
    Governed exception routing below the confidence threshold
    Metric
    1 item · rule R12 · to your team

    Anything uncertain goes to a person, with the reason.

  9. Beat 9 of 9 · Illustration
    State
    LEDGER: RUN 0217 · 9 CAPTURES · HASH 7f3a…
    Mechanism
    Run ledger and evidence capture on the workstation; AWS cleared
    Metric
    filed 1 · routed 1 · unverified 0

    The receipt stays with you, with its run ID and hash. AWS keeps nothing.

Beat 1 of 9 · Illustration
State
SESSION: SVC_AGENT · 2FA ROUTED · WORKSTATION LOCKED
Mechanism
Dedicated service identity on a dedicated workstation
Metric
adapters loaded 3 · health check ok

The agent works on your workstation, signed in as itself.

Arrow keys step

For technical readers

For your IT team

The agent runs on one computer you own, inside your network, and signs in to each system with its own account, which you can turn off at any time. It reads documents and screens with AI that runs on AWS Bedrock; what it reads is not kept there and is not used to train anything. Your rules are written down as a dated set, so you can see which version made which decision. Every entry is typed, then checked on the screen before it is saved, and every step is kept in a record you can open. Anything the agent is not sure about stops and goes to a person with the reason.

The launch

Six weeks, kickoff to schedule.

  1. Week 1Setup

    Workstation, remote access and the agent’s logins go in, with your guideline drafted live.

  2. Week 2Map

    Every screen mapped, your guideline encoded as rules, the exception route wired.

  3. Weeks 3 to 4Build

    The agent is built and tested against your real screens.

  4. Week 5Supervised runs

    The agent works while your team watches, and every run is reviewed.

  5. Week 6Handover

    Run reports, your sign off, and the agent moves to its schedule.

What you provide, answered once in the FAQ →

See it on your screens.

Book a demo or write to [email protected]
Book a demo