Skip to main content

Generate / Review / Controller

Generate, Review, Controller hero

If LLMs cannot write ledgers directly, then the system must answer a new question:

How should AI participation be structured
without collapsing responsibility and authority?

In BookiAI, the answer is a deliberate separation of roles and powers:

Generate → Review → Controller

This is not three prompts. It is an architectural boundary.

From actions to roles​

Most AI agent designs treat the model as an actor:

  • It reasons
  • It decides
  • It acts

This works only when mistakes are acceptable.

In a business system, we instead ask:

  • Who proposes?
  • Who validates?
  • Who authorizes execution?

These are distinct responsibilities, and they must not be blended.

Generate: proposing, not deciding​

The Generate role belongs to the LLM.

Its responsibility is to:

  • Interpret input and context
  • Infer possible actions
  • Produce structured proposals
  • State assumptions and uncertainty

A generated result is never an instruction. It is a proposal object.

For example:

  • Suggested journal entries
  • Inferred accounts
  • Estimated classifications

At this stage, nothing is written. Nothing is executed.

Review: validating recommendations​

The Review role exists to answer a single question:

Is this recommendation acceptable
under known rules and constraints?

Review may include:

  • Schema validation
  • Business rule checks
  • Confidence assessment
  • Secondary AI review (e.g. CPA-style reviewer)
  • Human review

Review produces an explicit Review Report:

  • What passed
  • What failed
  • What is uncertain
  • Why the recommendation should or should not proceed

Importantly, review does not execute anything.

Controller: owning authority​

The Controller is not another LLM agent.

It represents the system itself.

Its responsibility is to:

  • Interpret review outcomes
  • Apply policy and thresholds
  • Decide whether and how to proceed
  • Require confirmation when needed
  • Block unsafe actions

The Controller owns authority. It may be implemented as:

  • Deterministic logic
  • Decision tables
  • Policy engines
  • Workflow orchestration

LLMs may inform the Controller, but must not replace it.

Why this separation matters​

By separating Generate, Review, and Controller, the system gains:

  • Clear responsibility boundaries
  • Auditable decision paths
  • Safe failure modes
  • Gradual automation capability

Most importantly, no single component can unilaterally change system state.

Objects, not conversations​

This model works only if outputs are treated as objects.

Instead of:

  • Chat messages
  • Free-text explanations

We use:

  • Proposal objects
  • Review reports
  • Explicit decisions

These objects can be:

  • Logged
  • Tested
  • Replayed
  • Audited
  • Evolved

This turns AI behavior into something the system can reason about.

Manual-first is intentional​

In early stages, the Controller may:

  • Require explicit human confirmation
  • Execute nothing automatically

This is not a limitation. It is a design choice.

Automation is not removed — it is earned through evidence.

What comes next​

With roles clearly separated, the next challenge becomes quality:

How do we assess whether a recommendation
is good enough to proceed?

👉 Recommendation Assurance and validation

Written by ChatGPT, reviewed by the BookiAI team.