An HRIS is built around the employee record. Everything it can tell you is a property of a person: headcount, compensation, tenure, reporting line, job architecture, engagement score, performance rating. The AI value question has a different atomic unit. It asks which work changed, by how much, at what cost, and whether the money showed up anywhere finance recognizes.
You cannot get from the first to the second by adding features, because the system never captured the task. It captured the person who performed the task.
Two things are true at once here. Systems of record are very good at the job they were built for, and that job is not this one.
What a system of record is built to hold
The clearest statement of this boundary in 2026 came from a vendor describing its own product, approvingly, in a product release post.
Lattice, writing in September 2026 about turning performance data into workforce intelligence: "That's workforce intelligence: knowing which top performers need to be retained, spotting rating patterns across a review cycle, and making business decisions with the data that already lives in Lattice."
That is an accurate description and it is not a criticism to quote it. Retention risk on a named person, rating patterns across a review cycle, and decisions from the data already in the system are all real capabilities and they are all genuinely useful.
The phrase doing the work is "already lives in"
Every system of record scopes the questions it can answer to the object it was designed to store. A CRM can tell you everything about an opportunity and nothing about the engineering effort that shipped the feature that closed it. A ticketing system knows the ticket, not the revenue. A general ledger knows the dollar, not the hour that produced it.
An HRIS knows the person. So when the question becomes "what did our AI investment change about the work," the honest answer from inside a system of record is that it can tell you what happened to the people, measured the way it has always measured people.
This is a property of the category, not of any one product
It applies to the HRIS you run, the one you evaluated and did not buy, and the one that has not shipped yet. A roadmap does not change the unit of analysis. It changes what is reported about the same unit. That is the whole reason a Human Capital Operating System is a layer above the stack rather than another system inside it.
Why the AI value question has a different unit
The most useful public illustration of this comes from another vendor in the category, and it is worth reading carefully because the arithmetic is where most AI business cases break.
Visier calls it the Fractional Headcount Fallacy. In their example, a company deploys a new AI tool and finds that "20 employees each save two hours a week. On paper, this easily equals 40 hours of work a week, and that effectively saves an organization a full headcount." The problem is that "those time savings are distributed across 20 different roles."
Nothing is recoverable. There is no person to remove, no budget line that changes, and no number a CFO can bank. The savings are real at the level of the individual and imaginary at the level of the organization.
Look at what the two halves of that example are made of. "2 hours saved per employee" is a fact about the employee record. "One recoverable headcount" is a fact about how work is distributed across roles. They are different units, and the fallacy is the addition that pretends they are the same one.
Visier's own prescription in the same piece points where you would expect: "it is time to bring back the process-improvement mentality: structured, disciplined, task-level reengineering of work." Task level. That is a vendor in the workforce analytics category saying the unit has to move. The boundary between that category and a measurement layer is covered in workforce intelligence versus people analytics.
What each kind of fact actually looks like
The distinction stops being abstract as soon as you write two sentences side by side about the same team.
The shapes below are illustrations of form, not results. They carry no figures deliberately, because the point is the structure of the sentence rather than any particular answer.
An employee-record fact has this shape: this person holds this title at this band, was hired on this date, was rated this way in the last cycle, holds this license, and reports to this manager. Every clause is true, auditable, and about a person.
A task-level fact has a different shape: this category of work moved by this much on an adopting team, against a comparable team that did not adopt, over the same window, controlling for tenure and role mix, with the difference reported alongside its confidence.
The second shape contains no people at all. It contains work, a comparison, a window and a stated limit. Nothing in the first can be rearranged into the second, no matter how many fields you add, because the first describes who was present and the second measures what changed.
That is also why the honest version of the second sentence is longer and less quotable than the first. Comparison and confidence take words.
What practitioners describe when you ask them
In 2024 our founder ran a set of structured conversations with people who own workforce systems, at organizations ranging from a few hundred to well over 100,000 employees. We are describing them by rough size and sector rather than by name, because the conversations were private.
The pattern was consistent and it was not about software quality.
| Organization | Stack | What they described |
|---|---|---|
| Education group, several hundred employees, PE owned | SuccessFactors, PeopleSoft, ADP, applicant tracking | Manual KPI creation, no integration between systems |
| Global technology firm, over 100,000 employees | Workday, internal performance tool, Salesforce | Review inputs entered by hand, data not used in the decision |
| Pharmaceutical company, tens of thousands of employees | Workday, Teams, internal tracking | Goals entered manually, no automation in tracking |
| Defense manufacturer, tens of thousands of employees | Teams, Oracle-based HR | Difficulty tracking goals, poor interface, lack of automation |
| Financial services firm, over 20,000 employees | ADP, Workday, in-house review, deal and CRM systems | Review process disjointed across business groups |
Every one of these organizations runs competent software. None of them described a system that connected a goal to the work that satisfied it, and the two who raised goal tracking directly described it as manual. The systems hold the person and the intent. The work happened somewhere else.
What this is not
This is not an argument for replacing anything. A system of record should be boring, authoritative and correct, and you want exactly one of them. Levos reads from the systems you already run rather than competing with them, which is why the stance is "and," not "or."
It is also not a claim that vendors are being misleading. The Lattice sentence quoted above is accurate, volunteered and unforced. Vendors describe their scope correctly far more often than they are given credit for. The boundary is structural, and structural boundaries are the reason an intelligence layer above the systems exists at all.
If the practical question is how you would actually run the measurement, how to measure AI ROI across your workforce is the operational version of this argument, and the Human Capital Operating System is the category it belongs to.
What you would need instead
Three things the employee record does not contain.
- Observed work, in the systems where work happens. The unit has to be the task, which means reading from the tools where the task leaves a trace rather than from a summary of the person who did it.
- Cost joined to that work. A change in effort is not a number until it has a price attached. SaaS Stack Intelligence covers the tool-level layer here; the rest of the cost picture still comes from finance.
- A comparison group. Without one, a change is indistinguishable from a trend. That is what controlled cohort analysis with confidence scoring is for, comparing adopting teams to non-adopting teams while controlling for tenure, role and tool stack, and disclosing the limits rather than reporting one headline number.
Skills are the sharpest version of the same problem. A skills field on an employee record is a claim someone typed. A skill derived from demonstrated work is an observation. Those are different objects that happen to share a name, which is why skills are derived from demonstrated work in connected tools rather than collected by survey.
Frequently asked questions
Why can a system of record not answer the AI value question?
Because of unit of analysis, not feature coverage. The HRIS captured the person who performed the task, never the task. No amount of additional reporting on the employee record produces a fact about the work.
Does this mean you should replace your HRIS?
No. A system of record is very good at holding an authoritative record of employment, and you want exactly one of those. The argument is about boundaries, not quality.
What is the Fractional Headcount Fallacy?
Visier's name for the error where 20 people saving 2 hours each is read as 1 recoverable headcount. The savings sit across 20 roles, so nothing is recoverable. Time saved per employee and value realized are different units.
What would you actually need to measure to answer the AI value question?
Observed work in the systems where work happens, cost joined to that work, and a comparison group so a change can be told apart from a trend.
Is this a criticism of HR software vendors?
No. Every system of record reports on the object it was designed to store. That is as true of a CRM or a general ledger as it is of an HRIS.
Where this leaves you
If you are being asked what your AI rollout returned, notice which system the question is being routed to. If the answer is going to be assembled from the employee record, the answer will be about people, because that is the only kind of fact in there.
The question was about the work.
Request a Demo to see how observed work, cost and comparison groups come together, or read how the AI Impact signal family derives adoption and outcome signals from the tools where the work happens.
Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.
