The live causal twin of your AI-run business

Ask your business why.

Your business now runs on AI agents. Cohom gives it a twin that knows why.

Become a design partnerSee how it works

In development. Designed to run in your own cloud and start in shadow mode, read-only.

The shift

The org chart now includes software that decides.

Businesses are handing decisions to fleets of AI agents, tools and data feeds that act together. An outcome is now the product of many parts, not one.

HUMANBusiness ownerHUMANFraud operationsHUMANAI platform teamHUMANAnalystsAGENTScreening agentAGENTTriage agentMODELKYC modelAGENTInvestigation agent

Swipe to see the whole chart.

The blind spot

When outcomes change, today’s tools can’t say why.

Reports, observability and evals show what happened, not why.

changed · Tue 14:02 Auto-approval rate · illustrative why?

After the fact

They explain a failure once the damage is done, not while decisions are drifting.

One trace at a time

They see single runs, not how agents, shared data and human handoffs affect each other.

Correlation, not cause

An LLM judge or a cluster of similar errors stands in for proof of what caused the change.

Causal AI exists for IT. Twins exist for processes. In our October 2026 review, we found none that covers AI-run decisions.

Cohom

The live causal twin of your AI-run business.

One model of how your agents, systems, data and people drive outcomes. Scroll to watch it work.

01 · Assemble

It learns your ecosystem.

Qualia is designed to build the twin from your telemetry: agents, models, feeds, rules and people, and how they connect.

02 · Change

Something changes. An outcome moves.

A model update lands on Tuesday. Auto-approvals drop 8%. The twin starts from what changed, not from a guess.

03 · Cause

It names the cause, and checks it.

Qualia traces the drop to the KYC model update and confirms it by counterfactual replay. The watchlist feed stays labelled as a weaker, observational lead.

04 · Ask

Then it tells you why, in plain language.

Drishya, the visibility layer, answers with the cause and its evidence, forecasts the what-if, and keeps watch.

Causal graph of a fraud triage flow (illustrative) Eight nodes: KYC model, watchlist feed and velocity rules feed the screening, triage and investigation agents, which lead to the auto-approvals outcome and analyst review. Auto-approvals moved down 8% after the KYC model changed on Tuesday at 14:02. The cause path runs from the KYC model through the triage agent to auto-approvals, replay-confirmed. The watchlist feed is an observational lead. Illustrative example, not customer data. MODELKYC model DATA FEEDWatchlist feed RULEVelocity rules AGENTScreening agent AGENTTriage agent AGENTInvestigation agent OUTCOMEAuto-approvals HUMANAnalyst review changed · Tue 14:02 ▼ 8%
CAUSE PATH Causal graph of a fraud triage flow (illustrative)Same graph as the desktop version: the cause path runs from the KYC model through the triage agent to auto-approvals, replay-confirmed; the watchlist feed is an observational lead. Illustrative example, not customer data. MODELKYC model DATA FEEDWatchlist RULEVelocity rules AGENTScreening AGENTTriage AGENTInvestigation OUTCOMEAuto-approvals HUMANAnalyst review changed · Tue 14:02 ▼ 8%
CAUSE PATH
From tile: Auto-approval rate · 7 days
Why did auto-approvals drop 8% since Tuesday?
Why · cause ranked 1 of 2

The KYC model update on ◆ Tue 14:02 shifted the risk scores the triage agent relies on.

Replaying the flow with the previous model reverses the drop.

The watchlist feed is a weaker lead, labelled observational.

What-if
→

Rolling back is forecast to recover approvals in about 6 h [4–9 h]. Not yet checked.

Illustrative example, not customer data.

Qualia, the causal engine

From telemetry to cause to foresight.

Qualia builds the twin, finds the cause of a change, checks it, and forecasts what happens next.

01

Assembles the twin

Built to learn the graph of agents, tools, data, models and people from your telemetry, not from hand modelling.

02

Attributes cause

Names the component behind a change in outcomes as it shifts, not in a post-mortem.

03

Confirms it

Counterfactual replay, with the suspect change reverted, confirms the cause before you act.

04

Forecasts

Simulates what-ifs and predicts what happens next, with a range, labelled as a forecast.

Every answer carries its evidence.

Labels run strongest first. When the evidence isn’t there, Qualia says “can’t tell”.

Counterfactual replay, with the suspect change reverted, confirmed the effect. The strongest label.

A before-and-after test around a logged change, such as an interrupted time series.

A correlation. A lead worth checking, not a cause.

Rests on a stated prior or a link someone entered, not on data from your ecosystem.

?

The evidence isn’t there, so Qualia says so. Shown, never hidden.

Not evidence
→

A what-if prediction with a range. It becomes evidence only once it’s checked.

Drishya, the visibility layer

See your twin. Ask it why.

Live views show what moved across your agents, models, data and people. Any tile opens a why-question in chat, and the answer comes with its evidence. Watches call out shifts before they pile up.

Live views

See what moved.

Live views of your twin, built from the twin rather than raw metrics. Every tile keeps why one click away.

Chat surface

Ask why and what if.

Ask in plain language. The language model handles the words; Qualia computes every answer, so no LLM is judging an LLM.

Watches

Know when it shifts.

Set a watch in plain language. The twin calls out downstream shifts, by email or Slack, with the evidence so far.

One product, three surfaces

From a tile to a why-question, in one click.

Each surface links to the others. Ask why from a tile and the question arrives in chat with its context. Pin an answer to a view, or turn it into a watch. Drishya never sends a question for you.

Try it. Illustrative example, not customer data.

Auto-approval rate
61.2%
▼ 8% since Tue

Ask your business why.
Choose Ask why → on the tile.
Read-only. Drishya never changes a live decision.

Live views, chat and watches are planned for the first deployment. ‘Drishya’ means sight.

How it works

Connect. Learn. Ask. Act.

  1. 01

    Connect

    Deploy in your cloud, read-only. Point Cohom at decision logs, traces, analyst rulings and your change log.

  2. 02

    Learn

    Qualia assembles the twin from your telemetry: the agents, models, feeds, people and handoffs, and how they drive outcomes.

  3. 03

    Ask

    In Drishya, see what moved, ask why and what if, and get answers with their evidence.

  4. 04

    Act

    Your team decides what to change. Cohom never changes a live decision. Set a watch to see the result.

Steps describe how Cohom is designed to work. Nothing here is shipped yet.

Why, what if, what next

Why.

The twin names the component behind a change: an agent, a data feed, a model, a policy or a handoff. Then it confirms the cause by replay.

“Why did auto-approvals drop since Tuesday?”

What if.

Ask what happens before you act. Qualia simulates the change and forecasts the result, labelled as a forecast.

“What if we roll back the KYC model?”

What next.

Schedule watches in plain language. The twin calls out downstream shifts before they pile up.

“Alert me if any downstream agent drifts.”

Example questions, illustrative.

Where we start

We start where AI decisions matter most.

Fraud and AML decisioning at banks, NBFCs, fintechs and payments firms.

High stakesDecisions here carry real money.
Ground truthAnalyst rulings let us score the twin’s answers.
PullAgent stacks are growing faster than teams can explain what changed.

A beachhead, not our identity.

The same engine is built to extend to any AI-run business.

Own a fraud or AML decision flow? Become a design partner →

DATA FEEDKYCDATA FEEDWatchlistsDATA FEEDTransactionsAGENTScreening agentAGENTTriage agentAGENTInvestigation agentHUMANAnalyst reviewGROUND TRUTHAnalyst rulingsscore the twin’s answers

The modelled flow. Illustrative.

By design

Independent. In your cloud. Shadow mode first.

Cohom is designed to fit the stack you already run, and to observe before anyone acts.

Independent

A neutral twin across your vendors’ agents, not a feature of one platform.

Across models and clouds

Your observability, agent-platform and registry data become inputs to Qualia, whichever models and clouds you use.

Runs in your cloud

Designed to run in your own cloud, so your data stays with you.

Starts in shadow mode

Read-only, beside your systems. Cohom never changes a live decision.

Design commitments · not yet shipped

Product-led from day one: no billable delivery. Onboarding is unbilled, short and product-assisted.

Where it goes

One causal twin, from agents to the whole enterprise.

01

Agent decisions

Why did the AI’s decisions change?

Where we start
02

Business health

Why did the KPI move, and what if we act?

Vision
03

The enterprise

One causal view of how the AI-run company behaves.

Vision
04

Infrastructure

Predict demand from business causes, such as pre-warming GPUs before a spike.

Vision

Vision, not shipped. Each rung follows the one before it.

Team

Built by people who’ve shipped AI at banks and hyperscalers.

Satheesh Velayuthan

Co-founder

VP of AI in BFSI. Mathematics and computer science.

Photo and bio pending approval

Ganapathy Subramani

Co-founder

Amazon, Google, Adobe. JPMC, Goldman Sachs, Merrill Lynch, Bank of America, Standard Chartered, SocGen.

Photo and bio pending approval

Shruthi Chinnasamy

Co-founder

M.Tech, IIT Jodhpur. SAP Labs. Principal engineer in AI.

Photo and bio pending approval

Based in Bengaluru. Want to build the causal twin with us? Write to us careers email TBD

Design partners

Bring us one decision flow.

We’re looking for design partners who own an AI-run fraud or AML decision flow. Together we’ll run Cohom on one flow, in your cloud, in shadow mode, and measure whether it explains what moved.

  • Early access to Cohom on one decision flow, read-only.
  • A direct line to the founders, and a say in what we build.
  • No billable delivery. Onboarding is unbilled, short and product-assisted.
  • Planned: a free self-serve diagnostic on your own telemetry. Design partners get it first.

✕ Enter your name.

We’ll only use it to reply to you.

✕ Enter a valid work email, like name@company.com.

✕ Enter your company.

✕ Choose your role.

Protected by a privacy-friendly bot check. Submissions go only to the Cohom founders.

✓ Thanks. Your request is in.

A founder will reply by email.

Every AI-run business will have a twin. We’re building it.

Become a design partnerTalk to the founders