HR data access recertification now has to cover AI agents
Employee-built AI agents already read HR data in many organizations, and most access reviews still assume a human user with a badge and a termination date.
Most HR data access recertification processes were designed for a world where every user had a badge, a manager and a termination date. That world is gone. Vendor research covered by HRD America in September 2026 found that employee-built AI agents already hold access to HR data inside surveyed organizations, which means those agents were created outside the sightlines of both IT and HR. Separate coverage in the same cycle concluded that agentic AI adoption is running ahead of the governance frameworks meant to contain it.
Agents do not offboard
A standard access review walks a list of named employees, checks their role, and revokes what no longer fits. The control works because people leave, and leaving generates a signal. An agent built by an analyst in a low-code tool generates no such signal. It holds a service connection, an API token, or a scheduled extract, and it keeps running after the person who built it has handed back the laptop.
The practical failure modes are mundane. A departing analyst's custom agent keeps pulling from the HRIS. A copy of a compensation extract sits in a workspace nobody owns. A listening-survey feed continues to refresh into a dashboard that an executive still opens on Monday mornings. None of this requires bad intent. It requires only that no one ever wrote down who owns the connection.
This is why the recertification question is now phrased differently. The old question was: should this person still have access? The new one is: does every identity that can read people data, human or not, have a named owner who will answer for it?
The old question was whether a person should still have access. The new one is whether every identity reading people data has a named owner.
Market consolidation is quietly making the problem worse
RedThread Research's 2026 People Analytics Technology Report put the people analytics technology market at 12.2B USD in 2025, with growth cooling, contracts unbundling, and vendors reconsidering the level of customer support they offer. SHRM, in August 2026, framed the same shift as a market maturing rather than expanding, with buyers consolidating and re-scoping their stacks.
Maturing markets are where orphaned licenses and stale integrations accumulate. Unbundled contracts mean more connection points between more narrowly scoped tools, and fewer vendor hands available to help a buyer map them. The integration and data-quality burden shifts to the HR team that signed the contract. When a module is swapped out, the pipe feeding it often is not switched off.
Treat this as a data-integrity problem, not a legal one
It is tempting to route this work to legal or privacy and consider it handled. That misreads the risk. The sharper exposure is to the numbers themselves. An unmapped agent feeding stale headcount, an outdated cost centre mapping, or a superseded comp extract into a dashboard does not announce itself. It quietly corrupts the reporting executives act on.
Josh Bersin's September 2026 piece on the changing role of data in HR placed ownership of people data squarely back on the HR function rather than on IT. Gartner's 2026 Future of Work trends for CHROs continue to centre AI governance and workforce data readiness. Both point the same direction: the function that produces the number is the function accountable for its provenance.
Three artifacts to produce this quarter
The work does not require a platform purchase. It requires documentation that currently does not exist in most HR functions. First, an inventory of every identity, human and non-human, that can read people data. That includes service accounts, integration credentials, scheduled exports, BI connections and any agent or script an employee built against an HR system.
Second, an owner name against each line. Not a team, not a function, a person who will confirm annually that the connection is still needed. Ownership that is distributed is ownership that is absent. Third, a recertification cadence placed on the same calendar as existing system access reviews, so the people data review is not a separate initiative competing for attention.
Start with the systems carrying the most sensitive fields: compensation, performance ratings, exit and listening data. If the inventory for those three is accurate by the end of the quarter, the rest can follow in the next cycle.
