Recording the decision, not the outcome

Ask a reward function what someone is paid and the answer arrives in seconds. Ask why they are paid it, who decided, on what evidence or under which version of the rules. The answer takes weeks. It usually arrives as a reconstruction, assembled from email threads, old spreadsheets, a consultant report and the recollection of people who may no longer work there.

That gap is a design flaw rather than an administrative one. The systems hold the result of a decision and almost nothing about the decision itself. The reasoning, the evidence in front of the panel, the exception someone authorised and the name of the person who exercised judgement sit in correspondence or in memory.

Where a system does capture change, it captures a transaction. An effective-dated record shows a salary moving on a given date under a reason code marked promotion or market adjustment. A reason code is a category. It is not a reason. The change is stored as a new state rather than as an event with an author, a rule and a basis behind it.

This was survivable while the only question being asked was what the number is. The questions now being asked are different.

The questions have changed

Directive (EU) 2023/970 required member states to transpose pay transparency rules by 7 June 2026. Employers must make the criteria used to set pay levels and pay progression accessible to workers. Those criteria must be objective and gender neutral. Where an employer has not met its transparency obligations, the burden of proof shifts to the employer to demonstrate that no discrimination occurred. Workers can also request their own pay level alongside average pay levels for those doing equal work, broken down by sex, with two months allowed to answer.

In Great Britain the direction of travel is the same. The government opened a consultation on equal pay and pay discrimination on 14 July 2026, closing on 27 October 2026. Its own diagnosis states that an initial equal pay breach is often not an isolated incident. It frequently signals wider vulnerabilities in a pay structure. The document then lists three: a lack of pay transparency, opaque or overly complex pay systems and inadequate record-keeping.

Record-keeping is named by government as a structural cause of equal pay liability rather than as an administrative afterthought.

The proposals follow from that framing. A tribunal finding a breach would have to order implementation of a non-discriminatory job evaluation scheme where an up-to-date one is not in place. A new Equal Pay Regulatory and Enforcement Unit could require a job evaluation scheme or an equal pay audit as part of an investigation. A statutory questionnaire would return, with tribunals able to draw inferences from evasive answers or from silence. Judicial discretion to extend time limits on a just and equitable basis would replace the fixed six-month limit.

Taken together, those proposals increase the amount of evidence an employer may have to produce about decisions already taken. They also increase the pressure to have better decision-making infrastructure in place before a dispute arises. The current system delivers little of either. On the government’s own account, only one equal pay audit has ever been ordered by an employment tribunal.

Documentation is not a record

Most organisations believe they have this covered because they produce documentation. Documentation is written after the event by someone reconstructing what happened, usually because a claim, a grievance or an audit has prompted it. A record is created at the moment of decision as a by-product of taking it. The difference is evidential, not semantic. A contemporaneous record fixes the inputs, rules and judgement before the outcome is known. A reconstruction selects them afterwards. The Financial Conduct Authority has been putting the same proposition to senior managers for years, usually summarised as "if you cannot evidence it, you did not do it".

Nobody decided it

Reward may be the last high-consequence decision domain with no named decision maker.

In financial services, every senior management function holder has a statement of responsibilities, backed by a duty of responsibility that turns on whether they took reasonable steps. In clinical practice, notes are signed. In job evaluation, outcomes emerge from a panel. A panel is a collective noun. The sentence "the panel agreed" is not attribution.

That becomes material when a material factor defence is tested. The employer has to prove the reason for a pay difference was genuine and neither directly nor indirectly discriminatory. Proving it means producing the person who applied the factor or a contemporaneous record of them applying it.

The evidential life of a pay decision

A pay decision outlives the people and the systems that produced it, usually by a wide margin. In the Next case, claims brought in 2018 reached a finding of equal value in June 2023 and a material factor hearing in May 2024, involving more than 3,500 claimants and reported back pay in excess of £30m. Arrears run up to six years in England and Wales and five in Scotland. A decision taken this month may need to be explained twelve or fifteen years from now.

Set that against how long anything is actually kept. Retention periods follow operational convenience. The EU AI Act gives deployers of high-risk systems a minimum six-month retention obligation for automatically generated logs, subject to longer periods required by the intended purpose or by other law. Six months is reasonable for monitoring how a system behaves in service. It is nowhere near adequate for defending a pay decision.

Silence is not a safe alternative. The Supreme Court confirmed in Royal Mail Group v Efobi that a tribunal is entitled to draw an adverse inference where a respondent does not call evidence from the people who actually took the decision, with the significance attached to that absence depending on the circumstances. Missing witnesses and missing documents leave the same hole.

The methodology is not the defence

A job evaluation scheme is routinely described as a defence to an equal value claim. Read precisely, it is not. What defends the employer is a valid analytical scheme applied consistently, without bias and evidenced as such. The guidance and the case law behind it are concerned with how a scheme was constructed and how it was applied, rather than with the fact that one exists.

Consistency is a claim about hundreds or thousands of separate judgements taken by different people over several years. There is no way to demonstrate it without a record of each one. An organisation running the same methodology across many sites, with independent panels and local practice, does not have one scheme. It has as many schemes as it has panels, with no means of proving otherwise.

The inputs to a pay decision are examinable as well as the outputs. In Next the tribunal did not confine itself to whether market rates had been used. It examined the basis of the comparison, taking the view that benchmarking warehouse pay against a market that was itself predominantly male risked perpetuating the inequality the legislation exists to prevent. An employer relying on market data needs to record the source, the cut taken from it and the date, alongside the judgement it made about the result.

Birmingham City Council has paid out £1.1bn settling equal pay claims since the 2012 Supreme Court ruling, having issued a section 114 notice in September 2023 with further liability then estimated by officers at up to £760m. Whatever else that history shows, it is not a failure to own a methodology.

Three processes, three failure modes

The requirement is the same across job evaluation, promotion and pay review. What is missing differs in each.

Job evaluation. The score is recorded. The reasoning behind each factor level generally is not. A total of 412 points tells nobody why one factor was rated at level four rather than level three, what evidence sat in front of the panel or whether a member dissented. When a score later changes, the record needs to show what changed in the job. Most systems show only that the number moved.

Promotion. The decision is comparative. The record is not. Systems capture the person promoted and the effective date. They rarely capture the population considered, the criteria applied, the assessments made or the reasons other candidates were not selected. A record naming only the successful candidate answers nothing about the unsuccessful ones. Challenges originate there. Promotion is also where progression criteria have to work. Criteria that exist only in a policy document have never been tested against an actual decision.

Compensation review. Highest volume, most discretion, least recorded. An annual cycle can involve several thousand individual judgements made by line managers inside a three-week window. What survives afterwards is a new salary and an increase percentage. What does not survive is what the manager was shown, what the guidance said and who approved any departure from it. Exceptions concentrate here more heavily than anywhere else in reward.

What a recorded decision looks like

Six properties turn a stored state into a recorded event. They apply equally to all three processes:

1. Named attribution. The record identifies individuals rather than bodies. Where a panel decides, each member is named and their position captured, including disagreement.

2. Contemporaneity. The record closes when the decision is committed. Nothing can be added afterwards without the addition being visible as such.

3. The evidence as it stood. Job description, questionnaire responses, benchmarking extract and performance inputs are captured as they were when used, rather than linked to a live source that will since have changed.

4. The rule and its version. The record identifies the scheme, the factor definitions, the pay range and the policy that applied, each at its version number. Decisions taken under a superseded version need to remain readable against that version.

5. Exceptions and their authoriser. Off-scale appointments, retention payments and out-of-cycle increases are where liability concentrates. The exception, the reason for it and the person who approved it are the records that matter most.

6. Append-only integrity. Corrections are appended, never overwritten. An overwritten record is indistinguishable from a concealed one.

What this does not mean

The case is for recording decisions, not recording everything people say. Capturing deliberation in full is a different exercise and a worse one. Free-text opinion about individuals creates exposure without adding evidential value. The test for inclusion is narrow: is this needed to explain the decision and to show the rule was applied as written. Informal discussion, personal remarks and running commentary on performance do not meet it. Retention needs a defined basis for the same reason.

Where to start

Rebuilding job evaluation is the wrong opening move. Three smaller ones deliver most of the value.

Start with exceptions. Off-scale appointments, market supplements, retention payments and out-of-cycle increases are low in volume and high in liability. They are usually approved by email. Moving them into a structured record with a named approver is contained work with a disproportionate return.

Start with the current cycle. The record begins where the next process begins. Reconstructing the back catalogue is the practice being replaced.

Name people in the processes that already exist. Panel composition, individual positions and dissent can be captured in any system an organisation already runs. It costs almost nothing and changes behaviour immediately.

Three objections

The first is that a full record creates discoverable material capable of being used against the employer. It does. Absence is not protection either: it removes the ability to explain while leaving the liability untouched. A record showing a defensible decision defends it. A record showing an indefensible one has surfaced a problem the organisation was carrying anyway.

The second is that structured recording invites boilerplate. Managers learn which justification passes and reproduce it. Recording structured evidence and inputs rather than free-text rationale limits that, since structured inputs can be compared across managers. A manager whose justifications never vary is an audit finding in itself.

The third is burden. Asking managers to state a basis before committing an outcome adds time at the point of decision. It removes considerably more later. The cost of a reconstruction exercise is never budgeted, because it lands on whoever is available when it is demanded.

The stronger argument is upstream

Audit trails are usually sold as protection after the event. That undersells them.

Requiring a named person to state the basis for a decision before the outcome is committed changes the decision. Judgements taken anonymously and judgements taken under attribution are not the same judgements. Reward functions know this already, since it is why panels exist at all. The panel was an attempt to solve accountability through process. Recording solves it with evidence.

There is a management case alongside the legal one. A system that stores only states can report on what people are paid. A system that stores decisions can report on how the organisation decides: where a scheme is being applied inconsistently, how often individual managers use discretion and where exceptions cluster. No current reporting produces that, because the underlying events were never captured as events.

One point on timing

For anyone whose systems touch automated decision support, the EU AI Act classifies systems used for promotion decisions, task allocation and performance evaluation as high risk. Article 12 requires high-risk systems to allow the automatic recording of events across their lifetime. Those obligations were originally scheduled to apply from 2 August 2026. The Digital Omnibus has deferred the application of the Annex III high-risk obligations to 2 December 2027.

The deferral will be read as breathing space. It should not be. Logging is the single requirement that cannot be retrofitted cheaply, because it has to be designed into how the decision is taken. It is also the one with a business case that survives without the regulation.

Mercer’s 2026 global survey of more than 1,600 organisations across 60 markets put employer preparedness for pay transparency requirements at close to 50%, up from around 32% the year before. Only 14% had fully implemented their approach. UK employers were among the less prepared.

Even organisations inside that 14% may struggle if what they have implemented is a policy rather than a record. Publishing pay criteria answers the question of what the rules are. It does not answer whether the rules were followed in any particular case. That question arrives second. It is the one that costs money.

Sources