No Layoffs Is Not a Finding

First in a series on occupational substitution in records and archival work.

Last week, I wrote about Jinfang Niu’s interviews with fifteen archivists — the first dedicated empirical study of what AI adoption has actually done to archival work rather than what it might do. The headline finding was that none of the interviewees knew of an archivist who had lost a job to AI, and that they generally expected staffing levels to hold steady.

I said then that counting jobs was the wrong test. I want to make that case properly now, because the citation pattern is likely taking shape and will be sloppy in a specific, predictable way. Within a year, “the empirical evidence shows no displacement in the archival profession” will appear in a conference paper, then a professional association statement, then a budget justification. It will be sourced from a study of fifteen people, ten of them in leadership roles, none of them working in a for-profit archival organization, several of whom described implementations that were still transitioning from experiment to production at the time they were interviewed.

That is not a criticism of Niu. Her limitations section says all of this plainly, and she is right about her own data. It is a criticism of what the field is about to do with it.

Here is the underlying problem, and it is not about sample size.

Think about the causal chain that has to be completed before a records job disappears:

AI adoption → task automation → productivity change → organizational response → occupational restructuring → headcount effects

Six links. The records-specific empirical literature is strongest for the first three and progressively weaker for the last three. Archivists adopt tools. Tasks get automated. Efficiency changes — though Niu’s most useful finding is that it changes late, only at full production maturity, and considerably less than anyone promised in the demo.

That is a claim about our field’s evidence, not about labor economics generally. Economists have ample evidence on organizational restructuring and its employment effects. What we don’t have is any of it about records work.

And headcount is the terminal variable in that chain. It is also the loudest one, which is why we keep measuring it. A layoff is legible. It has a date, a memo, an affected person who can be interviewed. It is the one form of labor substitution that announces itself.

It is also the form least likely to be implemented first in a cost-center function within a resource-constrained institution. Layoffs require a decision, a budget process, sometimes a union, usually a manager willing to own it. Not refilling a vacancy requires none of those things. It requires a director to not write a requisition, in a fiscal environment where not writing requisitions is already the default. The org chart loses a row, and nobody sends an email.

So the field has built its monitoring apparatus around detecting the one outcome the theory doesn’t predict, and reported the absence of that outcome as reassurance.

I want to be more specific than “restructuring,” because the vagueness is part of why this is hard to research. There are at least four distinct mechanisms by which records work could be substantially reorganized while total employment holds steady. They are separable, separately measurable, and none of them appears in a headcount:

Non-replacement. Positions vacated through attrition are not refilled, or are refilled at a lower level, or converted to contract. No separation event, no announcement, no data.

Ladder compression. The distribution of positions shifts toward seniority. Coordinators and junior analysts thin out; managers and directors persist. Total employment can be flat while the entry pipeline closes.

Occupational substitution. Technical staff — data engineers, AI engineers, information architects — take on work the records function previously owned, whether or not they sit inside it on the org chart.

Occupational migration. Records work continues, but under different titles and reporting lines. Data governance, privacy, compliance, eDiscovery. The function persists; the occupation dissolves.

Four mechanisms, four articles in this series. What makes them worth treating as one theory rather than four complaints is that they are not alternatives. They can all be running at once, and their combined effect on total headcount could be approximately zero.

Let me be precise about the state of the evidence, because I don’t want to overclaim in the other direction.

On employment effects in corporate records management and information governance, there is essentially no direct evidence. Niu’s organizations include no for-profit archives, and she says so explicitly, noting that corporate archives may operate under stronger efficiency and cost-reduction pressures. The adjacent archival literature is overwhelmingly academic, governmental, and nonprofit. There may be scattered corporate implementation cases in vendor literature, digital asset management writing, eDiscovery, and pharmaceutical records practice — I intend to go looking. But there is no systematic evidence on the workforce.

On for-profit archival adoption, unmeasured. We do not know whether corporate archives in the pharmaceutical, financial services, technology, or media sectors are adopting faster, slower, or differently than universities. There are strong theoretical reasons to expect all three answers, but no data to choose among them.

On long-run headcount effects, the evidence is close to nonexistent. Many of Niu’s interviewees explicitly declined to project ten years out on the grounds that the technology was moving too quickly, and were comfortable discussing only the foreseeable future. That is intellectually honest, and it is also the whole problem. We are debating decade-scale occupational change based on what practitioners can see from where they currently stand.

On entry-level and paraprofessional hiring, nothing. And this is the gap that matters most, for a reason I’ll come to.

On whether engineering hires complement or substitute for archival positions, we have two numbers from Niu and no way to interpret either. Ten of the fifteen interviewees described projects that require digital archivists and software engineers to work together. Five interviewees reported that data or AI engineers had actually been hired. Five shops adding engineers alongside stable records staff and five shops adding engineers as records staff attrit are the same observation and opposite futures.

Except in one case, where we can see which it was. One of Niu’s organizations automated summary-writing work, kept the incumbent abstract writer until she retired, and then hired a data engineer into the vacated position. Nobody was laid off. The position was not eliminated. The occupation changed hands. That is, as far as I can find, the closest thing in the records-specific literature to an observed occupational substitution event — and it is a single anecdote in a study of fifteen people, which is precisely the problem.

Five gaps. But they are not five isolated holes in a literature. They are the same hole — nobody is measuring occupational structure — and they only look like separate problems because we keep approaching them one study at a time.

Here is the measurement problem in miniature.

The Bureau of Labor Statistics projects 6 percent employment growth for archivists, curators, and museum workers from 2024 to 2034, faster than average, with about 4,800 annual openings. Niu cites a roughly 4 percent figure specific to archivists, sourced from CareerOneStop.

Both numbers are fine. Neither can tell you anything about AI and archival labor, and the reason is instructive: the official category bundles archivists with curators, museum technicians, and conservators — three occupations with almost entirely different exposure profiles. A decline in archival processing positions offset by growth in museum work would show up as growth.

Corporate records management is fragmented rather than absent, which is a problem in itself. There is no standalone occupational series for records and information management as a profession. BLS explicitly recognizes records and information managers within Administrative Services and Facilities Managers — the handbook splits that occupation’s duties into facility management and records and information management, and even notes that demand is expected to be particularly strong for those working in records and information management as cloud computing expands. Other records work is distributed across information and records clerks, compliance occupations, health-information occupations, and management analysis, depending on duties.

So there is an official projection saying records work is growing. Notice what it cannot do: it cannot tell you whether that growth is in senior managers while coordinators disappear, whether it reflects records duties migrating into administrative-services roles, or whether the people doing it are records professionals at all. There is no single official occupational series capable of tracking restructuring of records and information management as a function.

This isn’t only a gap in the data. It’s a gap in the classification that's harder to fill because it can’t be fixed by collecting more of the same thing.

I want to import a finding from outside the field — and then immediately import the objection to it, because the objection is doing as much work as the finding.

Brynjolfsson, Chandar, and Chen’s Canaries in the Coal Mine used payroll data from the largest U.S. payroll processor to ask where AI-related employment effects appeared first. Their answer was not “in aggregate employment.” It was: in workers aged 22 to 25, in occupations where AI automates rather than augments, with employment for experienced workers in the same occupations holding steady. The original paper put the relative employment decline for that group at 13 percent. The August 2026 revision widened it to 19 percent — while still reporting no widespread displacement overall, and finding that the divergence operates primarily through reduced hiring rather than separations.

The authors make it clear that these are descriptive indicators, not causal estimates. That care is warranted. A January 2026 Financial Times analysis of millions of Lightcast job ads found that the broad decline in junior hiring was also consistent with interest-rate and macroeconomic effects, and did not find clear evidence that AI uniquely explains it.

Indeed’s Hiring Lab has documented a related pattern in postings. As of May 2026, senior-level postings were up 14.7 percent year over year while entry-level postings were down 7.5 percent. Measured against January 2025, senior postings were up 13.5 percent while entry-level and mid-level postings were each down roughly 6 to 7 percent. Indeed does not attribute this to AI either.

Put those three together, and you get a genuinely useful epistemic position:

Something unusual is happening at the junior end of the labor market. Attribution to AI remains contested. And that is precisely the argument for measuring the records profession directly, rather than importing a causal story from the aggregate.

What I take from it is narrower than a prediction and, I think, hard to argue with: the workforce layer where Stanford’s data show the clearest emerging employment divergence is the exact layer no study of records or archival work has examined.

Niu names this herself, and it’s buried in her limitations where it belongs in the abstract. AI currently affects technical and operational work often performed by practicing archivists, student workers, interns, and paraprofessionals. Her sample disproportionately captures leadership. Directors are well positioned to observe whether a line disappears from an org chart. They are poorly positioned to report on what happened to the contingent, student, and entry-level labor working under them.

AI currently affects technical and operational work often performed by practicing archivists, student workers, interns, and paraprofessionals.

We pointed the instrument at the wrong altitude.

There is a second-order version of the entry-level question that I think is more serious than the employment number, and almost nobody is asking it.

Many of the records tasks now being automated or AI-assisted — transcription, metadata remediation, routine description, bulk information processing, first-pass file review — have historically been performed in part by junior, paraprofessional, temporary, and student workers. Niu documents exactly this: student workers and staff assistants who had manually updated archival descriptions were redirected to rehousing physical materials.

Whether that produces entry-level compression is an empirical question, and I want to keep it open. But there’s a consequence that follows even if headcount is unaffected.

Those tasks are not merely junior work. They are the mechanism by which a new person learns how a specific organization’s records actually behave — where the naming conventions break down, which department has never once followed the retention schedule, what the 2011 migration did to the email archive, which “final” folder is the real one. That knowledge is not in the policy manual. It is acquired by handling several thousand records badly and slowly.

If AI performs that work, the profession may not have merely automated some tasks. It may have automated its own apprenticeship. And if that learning pathway is removed rather than redesigned, the bill may not arrive as a layoff. It may arrive years later as difficulty producing experienced records professionals.

I want to be clear that this is my inference, not Niu’s finding. She establishes that AI is affecting technical and operational work performed by practicing archivists, student workers, interns, and paraprofessionals. She does not establish that this work constitutes the profession’s training mechanism, nor has anyone tested it. But no employment survey would detect it if it were true — it is a question about the reproduction of professional capability, not headcount — and the field has a limited window to look, while the cohort that learned the old way is still working.

Before previewing the rest of the series, one methodological point that changes everything.

Niu’s central operational finding is that substantial efficiency gains emerged only once systems reached mature production — not during experimentation, not during pilot, not during partial production. If that’s right, then any study comparing “organizations using AI” against “organizations not using AI” will dilute the effect it’s trying to detect, because most of the treatment group hasn’t reached the point where the effect exists.

Building on Niu’s distinction between experimental use and transition, partial, and full production, I’ll code organizational AI maturity across this series on a five-point scale: no use, individual experimentation, pilot, partial production, mature production. The extension is mine, not hers.

The employment question then stops being "Does this organization use AI?" I propose measuring what happens in roughly the 18 months after a system reaches mature production — an interval chosen as a design decision, not one dictated by the literature.

Over the coming months, I’m going to work through what a research program on this would look like. Roughly, mapped to the four mechanisms above:

The vacancy that was never posted. Vacancy refill rates — positions refilled ÷ departures — with replacement disposition coded separately: same occupation, junior, senior, contractor, technical, redistributed, eliminated. Stanford’s finding that adjustment ran through hiring rather than separations is the reason to look here first.

Hollowing out the ladder. A longitudinal corpus of records and IG postings, measuring the junior/senior ratio rather than the total. Indeed’s seniority tilt is the external benchmark; the open question is whether records hiring shows the same pattern, a stronger one, or none at all.

The corporate blind spot. Why findings from universities and government archives shouldn’t be assumed to generalize to a cost center inside a profit-seeking company — and why corporate records management and information governance, a far larger population than corporate archives, matters more here than the archives do.

Engineers in the records office. Making Niu’s numbers interpretable by tracking both populations over time. This one comes with a serious counterweight: a June 2026 firm-level study linking corporate AI spending to workforce records across 21,559 U.S. firms found that high-intensity AI adoption was associated with roughly 10 percent greater headcount growth and 12 percent greater entry-level headcount over two years, with no statistically significant increase among low-intensity adopters. Those firms are heavily selected — larger, faster-growing, more technical, far more likely to be venture-backed — but the authors use a stronger design than a simple adopter comparison, and any substitution story has to explain the result.

Where did the records coordinators go? Following people rather than positions, because job destruction and occupational migration are indistinguishable in aggregate data and demand completely different professional responses.

Measuring the right thing. A task-level automation exposure index for records work, mapped against O*NET’s task statements rather than invented from scratch.

An observatory, not a paper. The general labor market is already building continuous AI trackers. Records management needs one more than most functions, precisely because conventional occupational statistics are so poorly suited to seeing inside it.

I am not predicting displacement. I want to be clearer about this than I usually bother to be, because the alternative is being cited as a doom piece by people who read the subtitle.

Here is everything the current evidence actually supports, stated together:

Niu finds tasks already changing, efficiency gains concentrated at mature production, workers reassigned, engineers hired, and no AI-attributed archivist layoffs reported. Stanford finds a young-worker employment gap in highly exposed occupations, operating predominantly through reduced hiring, with the authors’ own account leaving causal attribution unresolved. The Financial Times and Indeed find that junior-hiring weakness is real but that multiple non-AI explanations — rates, remote work, post-pandemic normalization — remain plausible. Ramp and Revelio find that intensive adopters can use more AI and hire more people, including junior roles, simultaneously.

That is not a picture of a profession being hollowed out. It is also not a picture of a profession that is fine. It is four partial views that do not resolve into an answer.

It is entirely possible that records and archival employment is stable, that AI is producing exactly the reallocation-and-backlog-reduction outcome that Niu’s interviewees described, and that, in 2035, this series reads as an overreaction. I’d take that outcome happily.

The claim is narrower: we do not know which of the four mechanisms is operating in records work because our occupational measurements cannot distinguish among them. The monitoring the field does have is calibrated to detect the one signal least likely to appear first. “No observed layoffs” is a true and useful statement about layoffs to date. It is not a statement about occupational structure, and it should no longer be read as one.

Niu says the field needs a larger survey in one to two years. She’s right. I’d add that a larger survey asking the same question will produce a larger version of the same non-answer. What’s needed is a different question, asked continuously, of the layer of the workforce nobody has interviewed.

Next in this series: vacancy refill rates, and why the most informative number in records employment is one no organization currently calculates.

As always — if you run a corporate records or IG program and you have quietly not refilled a position in the last three years, I would like to hear from you. That is the data.

Sources

Niu, J. “Archivists’ use of AI: practices and impacts.” Archival Science 26, 31 (2026). https://doi.org/10.1007/s10502-026-09553-w

Brynjolfsson, E., Chandar, B., and Chen, R. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Stanford Digital Economy Lab / SIEPR; August 2026 revision. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/

Indeed Hiring Lab, “The Labor Market Is Tilting Toward Seniority,” July 2026. https://www.hiringlab.org/2026/07/23/the-labor-market-is-tilting-toward-seniority/

U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, “Archivists, Curators, and Museum Workers.” https://www.bls.gov/ooh/education-training-and-library/curators-museum-technicians-and-conservators.htm

Financial Times analysis of Lightcast job-posting data, January 2026.

Ramp / Revelio Labs, firm-level AI spending and workforce study, June 2026. https://ramp.com/data/ai-jobs-impact/paper

O*NET Resource Center.

https://www.onetcenter.org/