Decision Accounting

The Ledger for Decisions Made by AI.

Decision Accounting is the management discipline that treats the decision, not the worker, the hour or the output, as the resource a firm manages and accounts for. A decision is produced under declared authority, recorded as it is made, costed per unit, and answerable to a named person. NATARAJA created the discipline and builds the platform that runs it.

a measuring instrument precedes every management discipline

Why Now

Every management revolution followed a measuring instrument. Decisions never had one, until the governed run.

The Ledger Made the Firm

Double-entry bookkeeping let capital and obligation be managed. The stopwatch gave us scientific management; control charts gave us quality. Management follows measurement.

Decisions Were Never Measured

Reasoning stayed in heads and meetings, so firms managed proxies instead: headcount, hours, budgets, projects. Nobody managed the decision itself.

AI Industrialised Decisions

An executive's agents can produce a thousand analyses a day. The constraint is no longer making decisions. It is knowing which ones you may act on, and who answers for each.

An Unbooked Liability

An ungoverned machine decision appears on no statement until it appears in an incident, a court or a regulator's record-keeping review. The EU AI Act's high-risk duties land exactly here. Singapore's MAS now makes the board own AI risk appetite and keep an inventory of AI use, and the UK's FCA found firms registering every agent.

four terms make a decision bookable

The Practice

Four things turn a machine decision into something you can account for. Each exists on the platform today, and a board can ask for each by name.

Authority Envelope

What the agent may decide and commit, set in machine-readable form before it runs. A limit the agent cannot argue with, because it is enforced outside the agent.

Decision Record

What was asked, what was read, what each step produced and who filled it, hash-chained as it happens. Written by the runtime, never by the agent, so it cannot be rewritten by the party it records.

Cost per Decision

What the decision consumed, including the human oversight it needed, priced in decision credits when the run seals. The number a board can put beside the decision's importance.

The Seal

A named person approves, amends or rejects at the gates you demand, with a short minute in their name. The record closes as Autonomous or Human-approved, and keeps that word.

These are the controls our articles on the agentic AI governance framework and agentic AI risk describe from the practitioner's side. The discipline gives them names.

the discipline has a maturity ladder

The Ladder: Where Your Decisions Are Kept

Each level is verifiable from artefacts, not from self-assessment. Level 1 shows documents, Level 2 shows traces, Level 3 shows a close.

L0

Oral

Decisions live in heads and meetings and are reconstructed after the fact. No books.

L1

Recorded

Memos, decision journals, architecture Decision Records. A single-entry diary: something is written down, nothing is checked against it.

L2

Governed

Declared protocols, gated runs, named principals. The journal has controls. This is where the platform starts.

L3

Accounted

Decisions are costed, matched to the obligations they discharge, and closed in periods, the way the books are.

L4

Optimised

The portfolio of decisions is managed from its own data: which protocols earn more autonomy, which need revision, where oversight costs more than it protects.

A firm at Level 0 or 1 running agents at scale is operating a decision factory with no books. The exposure is real, current, and off the balance sheet.

episteme runs the discipline

How NATARAJA Runs It

Decisions run on Episteme as governed protocols. This is what ships today.

Protocols

A decision method as a declared sequence of steps, versioned and hashed. Nine are built for advisory work, two for review and multi-disciplinary decisions, and you can register your own.

Human Gates

A step a person must approve, amend or reject, with a minute in their name. Comments from colleagues go on the record beside the decision.

The Seal and the Credits

A completed run seals as Autonomous or Human-approved and is priced in decision credits. Your first runs draw on an opening allocation, granted once.

Decision Record Reading

Each run read as its record at one address, from its start to its seal: what was asked, each step, the gates and who signed. Read it as a document or as a graph.

Decision Books

Every run you own or take part in, open or sealed, grouped by period, with search and freshness. Each row opens its record, and each month's credits close in a sealed statement.

A sealed decision does not stay true forever. When the inputs it rested on change, its record reads Stale instead of Checked or As sealed, and the decision can be re-checked against what changed, the way an asset is revalued.

Think of an agent as a member of staff and Episteme as the office it works in. The agent brings its own job description, training and tools. The office supplies the authority it may act under, the file every decision is written into, the cost centre the work is charged to, and the person who signs. Any vendor's agent can work here, and a better one is welcome. The procedures are the same for all of them, and they are not a binder on a shelf, because an agent has no conscience of its own to keep them. The runtime checks the envelope before the act, writes the record as the work happens, and will not let a gate be skipped.

Connect an AI assistant over MCP →  ·  Run it yourself, as one person →

the ladder starts with an assessment

Where to Start: The AI Value Review

The review is a Decision Accounting maturity assessment. It tells you which level your decisions are kept at today, your Cost per Decision, and what staying there costs. Its prescription is Levels 2 and 3, reached one decision at a time.

Score Your Portfolio

Five minutes, four dimensions, a level and a next step.

Take the self-assessment →

Scope a Review

A focused engagement that ends with a decision package your board can defend.

How the review works →

Book One Decision

You do not convert the organisation. You convert one decision, prove it, and repeat.

Request a governed pilot →
four questions come up first

Questions

What Is Decision Accounting?

Decision Accounting is the management discipline that treats the decision, not the worker, the hour or the output, as the resource a firm manages and accounts for. Each decision is produced under a declared Authority Envelope, recorded as it is made, costed per unit, and sealed in the name of an accountable person. NATARAJA created the discipline and builds the platform that runs it.

How Is Decision Accounting Different From AI Governance?

AI governance sets what is allowed. Decision Accounting books what was decided: the record, the cost and the named person behind each AI decision, so that governance has evidence to act on. Governance is the policy; accounting is the record. The two meet at the Authority Envelope, which governance sets and the record proves was kept.

Is a Sealed Decision an Audit?

No. A seal attests that a declared process was followed and who signed off. It is not a statutory audit and carries no auditor liability. What it gives you is a record that can be read, checked and relied on, by your own board, your auditors or a regulator, long after the run.

How Is Decision Accounting Different From an AI Inventory?

An AI inventory lists which AI systems and agents exist, who owns them and what they can reach. Singapore's MAS now expects one in its final AI risk guidelines, and the FCA found firms keeping one for their agents. On the ladder that is Level 1: something is written down. Decision Accounting books what each of those agents decided: under which Authority Envelope, with what Decision Record, at what Cost per Decision, and sealed by whom. The inventory tells you who could act. The books tell you what was done in your name.

booked decisions deliver outcomes
1 record
per decision, written by the runtime, kept after the run
2 outcomes
Autonomous or Human-approved, never blurred
5 levels
from decisions in heads to decisions on a ledger

Book the First Decision.

Sign in to run one decision as a governed protocol on the opening allocation, or start with the review and find out which level you are at today.