Talk to us
[new]We're hiring our first engineerWe're hiring( 2026-10-06 )

Routine cases, decided on arrival.

A private model for one decision your team repeats all day, like flagging a transaction or routing a claim. It is designed to decide the cases it is sure of and hand your people the rest. You set how often it may be wrong.

Replay your latest month.

On your own data, before you buy. It runs inside your network with networking off and returns metrics, not records.

A line screen seen only where light falls through a blind: tapered slats stepping from pale sodium to oxide, with a round window at the light point.

On arrival

An easy case should never wait behind a hard one.

Today the routine cases wait in the same queue as the hard ones. Tessoku is designed to decide each one it is sure of as it arrives, so your people reach the hard ones sooner. It trains on the history and corrections you already log: flagging a transaction, routing a claim, approving an expense, escalating a conversation.

Illustration, not measured data

request

How a decision is made

A probability for every option.

Including none of these. It never writes text. It picks one of your options or hands the case to a person. Nothing else.

Illustration, not measured data

A line of requests enters from the left and meets the model, which fans out into five options: fast track, send to an adjuster, refer for investigation, request documents, and none of these, with a lane under them to a person. Each request travels out along its most likely option. The ones whose top option clears the line are decided and counted at that option’s end; the rest peel off into the lane to a person, which keeps its own count. Illustration.

Below the line

A person decides.

When it is not sure, a person decides. You choose where that line sits.

An iron rule for automated decisions.

You set how often it may be wrong.

In an evaluation, the line is set on your own decisions, held back from training. Below it, a person decides.

Moving the threshold

Illustration. Raise the line and more goes to your people. Lower it and more is decided on arrival. Your error rate sets where it sits.

Illustration, not measured data

Your data

Designed to run inside your network.

  • Designed foryour GPU, cloud or air-gapped
  • Evaluationyour engineer, networking off
  • Weightsdesigned to be pinned per version
  • Fee per decisionnone

Blinds shadow on doors. Photo: 5snake5, CC0, via Wikimedia Commons. Two inks through Tessoku’s grain.

Replay your latest month.

An automation rate without its error rate is half a number.

Before you buy, it trains on your older months and replays your latest one beside what you run today, inside your network with networking off. You see how many cases could have been decided within your error rate, and how many costly errors that makes. It returns metrics, not records.

Iron rule. You set it.