FORGE
Book a Demo

Governed delegation

What is governed delegation?

Governed delegation is letting software act on an organisation's behalf under the same conditions a person would be given: a defined authority, limits on what it may do, someone to approve the exceptions, and a record of what it did and why.

The term names the problem that appears once AI can make the routine judgment calls that fixed rules never could. The question stops being whether software can carry out a step, and becomes whether we can let it decide. That is not an intelligence problem. It is a delegation problem.

An assistant drafts. A delegate acts.

An AI assistant produces something for a person, who then decides what to do with it: a draft email, a summary, an answer. The person is still the one who acts.

A delegate acts itself. It records the payment, approves the routine case or sends the reminder. The difference is not how capable the software is but whether it holds authority, and authority needs someone to have granted it, limits on it, and a way to see afterwards how it was used.

Why application permissions are not enough

Business software already has permissions, approvals and audit logs. Three things change when software does the work.

  • Who acts. Permissions were built for people pressing buttons. They decide who may open a screen, not what software may decide when it reads an email and interprets it.
  • Where the work happens. Each application governs only itself. Nothing governs an invoice's journey from email to ERP to bank, and that gap between systems is where people do the chasing today.
  • What the record shows. An audit log says that a user approved at 10:42. When software decides, you also need to know why: which document it read, which rule it applied, and who authorised the exception.

The three parts

Governance has to move from inside each application to a layer above them, and that layer has three parts.

  • Policy governs actions. The system knows which actions it may take, which need approval and which are prohibited. Privacy rules apply before sensitive information reaches a model, not after.
  • Deterministic systems execute them. AI interprets the document; code matches the amounts and moves the money. Judgment and arithmetic are kept apart, so the parts that must be exact are exact.
  • Evidence establishes what happened. Actions, approvals, policy decisions and the documents behind them are recorded, so any outcome can be reconstructed and challenged.

Deciding what may run on its own

Start where a mistake is cheap and visible, and widen the limit as the evidence shows the routine cases are handled well. These questions decide where an action sits:

  • Can it be undone? An action that can be reversed can be delegated sooner than one that cannot.
  • How much is at stake? A limit, such as a payment amount, is the simplest boundary to state and to check.
  • Can the rule be written down? Where it can, code should apply it. Where it cannot, a person should decide.
  • How often are there exceptions? Frequent exceptions mean the process is not ready to delegate yet.
  • Will there be a record? If you could not show afterwards what was done and why, do not delegate it.

In the loop, on the loop, and out of it

Human in the loop means a person approves each action before it happens. Human on the loop, sometimes called over the loop, means the software acts and a person watches and can step in. Out of the loop means the software acts alone within its limits, and people review the record afterwards.

Governed delegation is how an action moves between these. A new kind of action starts in the loop. As the record shows the routine cases handled well, it moves on the loop and then out of it, while exceptions and anything above the limit stay with a person.

What the record has to hold

An auditor asking about one decision made by software needs to see:

  • which documents it read, and who introduced them;
  • what was sent to a model, and what was masked before it went;
  • which rules were applied, and what the AI produced against what the rules decided;
  • who approved an exception, when, and with what note;
  • what was finally done, and what it cost.

Where Forge fits

Forge, by GradTensor, is our implementation of governed delegation, built as a layer above existing systems of record rather than another one. Today it masks personal details before they reach a model under rules members cannot weaken, runs an organisation's own programs without network access, moves records through defined workflow stages, and records every access and AI call.

The argument in full is in the essay Governed delegation: the future of business software, on our blog.

An invoice, delegated

  1. 1An invoice arrives by email. The system reads it, matches it against the purchase order, and checks the policy.
  2. 2The amounts match and the payment is within its authorised limit, so payment is initiated and the evidence recorded.
  3. 3The next invoice is 4% higher than the purchase order. The system stops and presents both documents and the discrepancy to the finance controller.
  4. 4She approves the exception with a note, and the work continues.
  5. 5Later, finance reviews the history of exceptions and decides whether to raise the limit. The AI did not get smarter. The organisation delegated more, because it could see the evidence.

What governed delegation is not

  • It is not removing people from decisions. The hard calls stay with people; what goes is the chasing.
  • It is not a model-quality claim. A more capable model does not make delegation safe; limits and evidence do.
  • It is not the whole of AI governance. Fairness, bias and model risk are real concerns that arise mostly in training and model selection. Governed delegation is about what software may do on your behalf.
  • It is not solved across every system. It applies most clearly today to document-driven processes, and extends from there to work across applications and across organisations.

Questions

What is governed delegation in AI?

Letting AI act on an organisation's behalf with a defined authority, limits on what it may do, a person to approve exceptions, and a record of what it did and why.

How is it different from AI governance?

AI governance is broad: policy, ethics, model risk and compliance. Governed delegation is the operational part of it that decides what software may do for you, under which conditions, and with what evidence.

Does every AI action need human approval?

No. The point is to delegate more over time. Routine actions within a set limit run on their own and are recorded; exceptions and actions above the limit go to a person.

What is human in the loop in agentic AI?

A person approves an AI agent's action before it takes effect, such as a payment, a message sent outside the company or a change to a record. In governed delegation it is where a new kind of action starts, and where exceptions always go.

What is the difference between human in the loop and human on the loop?

In the loop, a person approves each action before it happens. On the loop, the software acts and a person monitors it and can intervene. The first suits actions that are new, costly or hard to undo; the second suits routine actions with a good record.

What are examples of human in the loop AI workflows?

An invoice that differs from its purchase order going to a finance controller before payment. A new supplier's documents checked automatically, then approved by someone in finance before the supplier goes live. A contract summary reviewed by a lawyer before it is sent to a client.

Why keep deterministic rules if AI can reason?

Because some parts of a process must be exact every time. AI is good at reading and interpreting; code is better at matching amounts, checking formats and applying a limit. Keeping them apart means the exact parts are exact and can be checked.

Do I need this if my staff only use chat assistants?

Less so. A chat assistant drafts and a person acts, so existing controls on the person still apply. Governed delegation matters from the moment software takes an action itself.

Read next

Book a Democontact@gradtensor.com