The System of Record Has No Object for the Act

HR systems were built to record what people do to other people's employment. Agents now do some of that work, and the record has no type for it.

In an April 2026 essay for HR Executive, Ramprasad Reddy Mittana names the gap. When an agent screens a candidate, approves a time-off request, or routes an HR service case, "the system of record typically captures the outcome but not the actor." 1 The candidate moves to the next stage. The request is approved. The case is closed. Nothing on the file says a non-human process produced any of it.

That would be a design quirk if the record were optional. It isn't. Federal rules require employers to preserve any personnel or employment record having to do with hiring, promotion, termination, pay, or selection for training for at least one year. 3 California's Civil Rights Council went further in rules that took effect on October 1, 2025: employers must keep employment records, including automated-decision data, for a minimum of four years. 2

Vendors see the problem. Workday made its Agent System of Record generally available in February 2026, with agent interactions "recorded and tracked." 4 That gives the agent an entry of its own. It doesn't show that the candidate's file can say which agent acted on it.

So the question is narrower and harder than whether agents are governed. Does the system of record have a record type that can hold the agent's act on the person it touched? Mittana's answer is that most platforms were never built to. 1

1. 1. The record keeps the outcome and drops the actor. Screening decisions, leave approvals, case routing, and performance flags land in the HR system as results, with nothing in the record to show that an agent produced them. 1
2. 2. The duty to keep the record already exists. Federal rules require one year of retention for records tied to hiring, promotion, termination, pay, and selection for training. California requires four years for employment records, and it names automated-decision data. 23
3. 3. A registry for the agent is a different object from a record of the act. Workday's Agent System of Record lists agents alongside employees and tracks their interactions. That answers which agents exist, not whose file they changed. 14
4. 4. Autonoma synthesis: the gap is a schema problem before it is a policy problem. An enterprise can write an AI governance policy in a week. It can't produce a record of an act its system has no field for. Score HR agent readiness on whether the act lands on the person's record. 1234

When an employment decision is challenged, the reconstruction runs through records. Mittana notes that regulators, courts, and plaintiffs' attorneys even read the org chart as a document that defines the authority behind a challenged decision. 1 The question here is whether those records can still say who decided once an agent did part of the deciding.

Earlier Briefs cover the neighbors. 007 asked whether the agent belongs on the org chart. 012 covered agents acting on stale workforce data. 021 asked who assigns, completes, and writes a learning record. 022 asked whether an agent's approval and its tool call join into one record inside the agent's own trail. This Brief moves to the far end of that trail: the employment record the act changed.

The file keeps the result and loses the actor

Mittana starts from the design assumption. Every HRIS built in the last four decades, he writes, was organized around the employment relationship: a legal start date, pay, a reporting line, and an audit trail HR must keep. 1 Agents carry "no W-2, no 1099, no position in your job architecture, no manager of record," yet they now do work that people with all of those things used to do. 1

He then describes three gaps. In the visibility gap, no single place in most HR platforms shows which agents exist or who owns them, because agents arrive as service accounts IT provisions. In the authority gap, an agent that screens out a candidate "isn't acting under a delegation HR approved," so HR is accountable for a decision that traveled a path it wasn't part of. 1 The audit gap sits between them, and it's the one this Brief is about. The record shows a result with no sign that a non-human process produced it. "This is the gap that creates legal exposure," he writes. 1

Picture how that reads six months later. A candidate was moved out of a requisition and a concern was flagged on an employee's record. Each outcome sits in the right table with the right date. If the record names anyone, it names the service account IT set up. Ask who made the call, and the file answers with what happened. (That scene is our illustration, not a case from his essay.)

Mittana's second step for CHROs spells out the object that is missing. Every AI action in an HR process should leave a record in the system of record showing which agent acted, on which record, using what inputs, with what outcome, and whether a human reviewed or changed the result. 1 He adds that HR has to be able to produce that trail from the HR system itself, "not from a vendor dashboard or a server log." 1

The law already asks for the record

The federal rule is short and broad. Under 29 CFR 1602.14, any personnel or employment record an employer makes or keeps, including application forms and "other records having to do with hiring, promotion, demotion, transfer, lay-off or termination, rates of pay or other terms of compensation, and selection for training or apprenticeship," must be preserved for one year from the date of the record or the personnel action, whichever is later. 3 Once a charge of discrimination is filed, the employer has to keep every relevant personnel record until the charge or the lawsuit is finally resolved. 3

The rule doesn't care who produced the record. For L&D readers, the last clause matters most: an agent that picks who gets into a leadership cohort, or who is assigned required learning, is producing a record about selection for training.

California's rules are more pointed. The Civil Rights Council's regulations on automated-decision systems were approved on June 27, 2025, and took effect on October 1, 2025. 2 The state's Civil Rights Department says they make clear that an automated-decision system can violate California law if it harms applicants or employees based on protected characteristics, and they require employers to keep employment records, including automated-decision data, for at least four years. They also define "agent." 2

Neither rule says the HR system must carry an actor field, only that the records must exist and be kept. An employer that keeps the outcome on the person's file for four years, and the automated-decision data somewhere in a vendor's logs, has two stores to reconcile after a charge arrives. The retention duty reaches both. The schema covers only one. That reading is ours; neither rule says anything about database design.

Registering the agent does not record the act

The strongest vendor answer so far is Workday's. Its February 18, 2026 announcement frames the Agent System of Record around questions organizations struggle to answer: which agents are running, who owns them, and what their role is. 4 With it, Workday says, agents are "managed as part of the overall organizational structure, alongside employees," and "AI agent interactions are recorded and tracked," with data access backed by Workday's security model. 4 Mittana counts the product as movement in the right direction. 1

What the announcement describes is an object for the agent and a log of what the agent touched. That goes a long way toward closing the visibility gap. It doesn't show whether the candidate's record or the performance record carries the agent's identity as part of the act, and it says nothing about how customers configure the product.

Mittana's own renewal checklist keeps the two apart. One item asks whether a platform can register agents as distinct workforce entities alongside employees. The next asks whether an agent's action is captured "in the record itself, where HR and legal can access it on demand," rather than in a separate analytics tool. 1 A registry answers the first question. Only a record type for the act answers the second.

Autonoma forecast: Over the next 12 months, HR suites will ship agent registries and agent analytics faster than they add an actor to the employee and candidate records those agents change. Buyers will accept the registry as the audit trail because the registry is the part that demos well. The timing is Autonoma synthesis. 14

The first four signs are documented in the sources. The last four are signals to watch for in your own systems.

• HR records show screening, leave, case, and performance outcomes with no sign that a non-human process produced them. 1
• Agents enter HR processes as IT service accounts, outside any inventory HR owns. 1
• Retention rules cover employment records for one year federally and four years in California, including automated-decision data. 23
• Suite vendors sell a registry that lists agents alongside employees and tracks their interactions. 4
• A request to show who screened a specific candidate gets answered from a vendor log, not from the candidate's record.
• Agents are registered, but the leave, case, or performance record has no field that names the agent.
• An agent selects people for a training cohort or assigns required learning, and the result is recorded as a plain enrollment.
• Contract renewals add agent registration and leave out action logging on the record itself.

For CHROs and HR operations: Treat the act record as an HR requirement and put it in writing: which agent acted, on which record, with what inputs and outcome, and whether a person reviewed it. 1 If HR can't produce that from its own system, HR doesn't hold the record of its own decisions.

For HRIT and suite owners: A registry tells you what exists. Ask separately where the act lands, then test it on one candidate an agent screened, using the candidate record alone.

For employment counsel and compliance: The retention clocks are already running. 23 Map where automated-decision data lives, and how long each vendor keeps it, before a charge arrives.

For CLOs and L&D: Selection for training is named in the federal rule. 3 If an agent picks who enters a program or assigns required learning, the enrollment is an employment record with the actor missing.

For buyers in renewals: Ask two questions. Can the platform register agents as workforce entities, and can it write the agent into the record of each act? 1 A yes to the first is not a yes to the second.

Weight: Moderate. The strongest objection is that this is already solved, or close to it. Workday's registry records and tracks agent interactions. 4 The argument goes that a legal team needing to know who screened a candidate can pull the agent log, match the time, and produce an answer. On that view, an actor field on every employee record is a data-model preference dressed up as a governance gap.

Parts of that are right. Logs are real evidence, and neither rule says where records must live. 23 The Workday product does address the visibility gap Mittana names. 14

What the objection doesn't answer is the join. A log entry showing an agent acted at 3:40 p.m. and a candidate record showing a rejection at 3:40 p.m. are two records that agree on a time. Proving they are the same act means rebuilding it, one request at a time, from a system HR doesn't control. That's Mittana's authority gap in another form: HR carries the accountability for a decision it can't see in its own record. 1 The objection also assumes the vendor log lasts as long as the employment record. California's four-year floor covers automated-decision data 2, and whether a vendor keeps interaction data that long is a contract term someone has to ask for.

A second objection is about the evidence. The mechanism rests on one practitioner's essay, and Mittana's "most" is his judgment, not a survey. That's fair. This Brief uses his essay for the mechanism, not for a count. No source here documents a named employer whose record failed a legal challenge for this reason, and none measures how many HR systems lack the field. A vendor document showing that major HR platforms already write the acting agent onto the person's record would weaken the argument.

HR has always been able to answer one question about any decision in its records: who did this? The answer was a person, because only people did the work. That assumption went into the schema so long ago that nobody thinks of it as a design choice.

Agents break it without breaking anything visible. The records stay clean. They're just missing the one fact a regulator, a court, or the employee will ask for first.

A registry of agents is worth having, and HR should own it rather than leave it with IT. But a list of who might have acted is not a record of who did. The test takes an afternoon. Pick one file an agent touched and try to prove, from that file alone, that an agent touched it.

If you can't, you have a decision to make now, not after the first charge. Either add the actor to the record, or stop letting agents write to records you would have to defend.