IWhat this is

A company now has three kinds of deciders: people, algorithms, and AI assistants and agents. Most companies have written rules only for the first kind.

The Simon G. Method is a short, practical way to put all three under the same control. It tells you what to write down, what to record each time, and what to measure afterwards. It needs no particular software. You can start with one page and a spreadsheet.

It gives an executive team four things:

PromiseWhat it means
ControlData and models run where the company decides
AccountabilityA named person answers for every class of decision
EvidenceEvery decision leaves a record an auditor can read
ReturnThe money effect of each decision is measured, not assumed

IIWho it is for

Any company where software proposes or takes decisions about stock, prices, credit, payments, people or customers. It was written for companies of 50 to 2,000 employees, where there is no separate governance department and the same few executives carry the consequences.

IIIThe seven rules

1. Every class of decision has a named owner

List the decisions that matter: a dozen or two, not hundreds. Put one person's name beside each. A committee is not a name.

2. Every freedom given to a machine has a signed limit

For each class, write what the machine may do alone and where it must stop and ask. The limit is a number or a condition, signed by the owner, with a review date. A vendor's default setting that nobody changed is a limit nobody signed.

3. Every proposal carries its evidence

A proposal states what is true now, what is proposed, the reasoning, the expected effect in money, and which inputs were missing. A proposal that cannot show its reasoning is not approved.

Missing information is stated as missing. It is never replaced silently with a zero or a default.

4. A person signs

Inside the limit, the machine may act and the record says so. Outside it, a named person approves or rejects, and gives a reason when rejecting. Above a second threshold, the person who proposed cannot be the person who approves.

5. The record cannot be rewritten

Decisions are added to the record and never edited or deleted. A correction is a new entry that points to the old one. This holds for administrators too.

6. The effect is measured against a baseline fixed beforehand

Before a decision is applied, write down the figure it is meant to change and its current value. Afterwards, measure the same figure the same way. A decision is then marked verified, underperforming or not measurable.

A result without a baseline proves nothing, in either direction.

7. Data and models stay inside a perimeter the company controls

Decide where data may be processed and which models may see it. Anything that crosses the perimeter is written down: what, to whom, and why. Local processing is the default, and sending data out is the exception that needs a signature.

IVThe three documents

DocumentWhat it holdsSize
Decision charterWho may decide what, within which limitsOne page
Decision recordOne entry per decision, with evidence, signature and outcomeOne row each
Perimeter statementWhere data and models run, and what may leaveHalf a page

The decision charter

Six columns, one row per class of decision:

ColumnContent
ClassThe kind of decision
OwnerOne named person
Machine may act aloneThe conditions and the ceiling
Escalates whenThe trigger that sends it to a person
Health signalsWhat is watched to see whether the gate works
Review dateWhen the row is looked at again

A template ready to fill in: decision-charter-template.csv.

The decision record

FieldMeaning
Class and charter versionWhich row of the charter applied
SubjectWhat the decision is about
ProposerA person, an algorithm, an AI model or an agent, with its version
InputsThe data used, its date, and what was missing
Current and proposedThe value now and the value proposed
EvidenceThe reasoning and the expected effect in money
Limit checkWhether the proposal was inside the signed limit
DecisionsWho approved or rejected, in which role, when, and why
BaselineThe figure to be changed, fixed before the decision
OutcomeThe figure afterwards, the measured effect, and the verdict
PerimeterWhether any data left the company, and where to
Previous recordThe entry this one corrects or replaces

A record moves through these states: proposed, approved or rejected, applied, then verified, underperforming or rolled back. A proposal nobody acts on expires.

The full definition: decision-record-0.1.schema.json.

VFour signals that the gates work

SignalWhat it reveals
Rejection rateA gate that never says no is a formality. A gate that rejects most proposals means the limit or the machine is wrong.
ReversalsHow often an approved decision is undone shows whether the limit sits in the right place.
Time to decisionA slow gate teaches people to go around it.
Review triggersReview on a calendar, after an incident, and after any model or vendor update.

VIFour levels

A company does not need to do everything at once.

LevelNameWhat existsEffort to reach
1WrittenThe charter and the perimeter statementAn afternoon
2RecordedDecision records are kept, even in a spreadsheetTwo to four weeks
3MeasuredBaselines are fixed and outcomes verifiedOne quarter
4EnforcedThe system refuses what the charter does not allow, and the record cannot be alteredSoftware

The effort figures are estimates. Levels 1 to 3 need discipline and no purchase. Level 4 is what Simon G. does.

VIIStarter decision classes

These are the classes most companies share. Owners and limits are examples to be replaced with your own.

FunctionClassTypical ownerMachine may act aloneEscalates when
Supply chainReplenishment orderHead of supply chainLow-value, steady items under a value ceilingHigh-value item, value above the ceiling, unusual demand
Supply chainChange of safety stock or reorder pointPlannerNeverAlways a proposal
Supply chainTransfer between locationsPlannerUnder a quantity ceilingAbove it, or when the sending location falls below its own minimum
Supply chainDisposal or markdown of excess stockCommercial directorNeverAlways a proposal
SalesPrice changeCommercial directorInside a signed corridor around the list priceOutside the corridor, or for key accounts
SalesDiscount above list termsSales managerNeverAlways a proposal
FinanceCustomer credit limitCredit controllerSmall increases for customers with a clean payment historyAny decrease, any increase above a threshold
FinanceRelease of a payment runHead of financeNeverAlways two signatures
FinanceWrite-off of a receivableCFONeverAlways a proposal
ManagementOverride of a forecastPlannerNeverAlways, with a stated reason
Customer serviceRefund or compensationService managerRoutine cases under a value ceilingAbove it, or when the customer is in distress
PeopleScreening of job applicantsHead of HRSearch, sort and summariseEvery rejection belongs to a person
ITApproval of a new AI toolCIONeverAlways a proposal
ITData leaving the perimeterCIONeverAlways a proposal
ITUpdate of a model or vendor versionCIONeverAlways, and it triggers a review of the classes that use it

VIIIHow the method relates to regulation

The method does not make a company compliant by itself. It produces the evidence that several rules ask for.

RuleWhat it asksWhat the method provides
EU AI Act, Article 14Human oversight of high-risk systemsRules 1, 2 and 4
EU AI Act, Article 26Users of high-risk systems keep logs and name the people who oversee themRules 4 and 5
EU AI Act, Article 4Staff who use AI understand it well enoughThe charter, as training material
GDPR, Article 22No decision with legal effect on a person by machine aloneRule 4 for classes that touch people
NIS2Risk management and documented security measuresRule 7
ISO/IEC 42001A management system for AIThe three documents, as its operating core

Most stock, price and payment decisions are not high-risk under the AI Act. For them the reason to use the method is commercial: knowing what each decision earned, and who stands behind it.

IXHow it relates to existing standards

StandardWhat it coversRelation
OMG DMNHow decision logic is modelledDescribes the rule; the decision record describes each use of it
W3C PROVProvenance of data in generalThe record's inputs can be expressed in it
Open Policy Agent decision logsPolicy decisions in software systemsA technical log one level below the business record
OpenTelemetry for AI agentsTraces of what an agent didA source of evidence for the record

None of these records a business decision with its approver, its limit and its measured effect in money. That is the gap the decision record fills.

XWhat the method does not do

XIStart this week

  1. List your decision classes, using the starter table.
  2. Put a name beside each.
  3. Write one limit per class, and sign it.
  4. Write the perimeter statement.
  5. Record the next ten decisions in a spreadsheet, using the fields above.

The Simon G. Method is published by Groma. The reasoning behind it is set out in "The machine may decide. A human signs where.". To see part of it at work, try the free Simon G. Stock Check.