On 26 August 2026, Axios reported that the US Department of Labor has signed memorandums of understanding with several large technology companies — OpenAI, Google, Meta and Amazon among them — to obtain data on how businesses are using artificial intelligence and where they expect adoption to go. The stated aim is to supplement official labour statistics with a faster private-sector read on AI's occupational effects.
My first reaction was not about the companies involved. It was about what the arrangement concedes. A national statistical apparatus does not go looking for vendor data unless its own numbers arrive too late to be useful. And if a government economist finds the official series too slow for policy work, an employer planning headcount two quarters out is in a considerably worse position.
That is the transferable lesson for Singapore, and it has nothing to do with American labour policy. It is that the data most of us use to plan hiring is a lagging indicator of a market that is currently moving fast.
What was reported, precisely
- Who: the US Department of Labor, with the effort led by Acting Labor Secretary Keith Sonderling.
- With whom: major technology companies including OpenAI, Google, Meta and Amazon.
- What: data sharing on how businesses use AI and their expectations of future adoption, under signed memorandums of understanding.
- Why: to supplement traditional government statistics with a faster view of AI adoption and its effects on occupations.
- When: reported 26 August 2026 by Axios. Related reporting indicates the department is also preparing an AI workforce hub.
- The open question: the companies supplying the data are the same companies selling the technology whose labour-market effect is being measured. That is a real methodological caveat, and it is worth holding.
Expert view (1/3)
The measurement gap is the story, and it exists in Singapore too. Official labour series everywhere share two structural problems when the nature of work changes quickly. They publish on a lag — typically one to several months — and they classify people into occupational categories defined long before the current wave of automation. When a role quietly changes from “produce the first draft” to “review and correct the machine's draft”, the job title does not change, the headcount does not change, and the statistics register nothing at all. Yet the skills you need to hire for have changed completely. No amount of vendor data fixes that; it only makes the lag shorter.
Why published statistics arrive after your decision
What Singapore data actually shows
It is worth separating the American policy story from local evidence, because they point in slightly different directions.
A Singapore government study reported in mid-August 2026 found that firms using AI recorded gains in both revenue and employment, with adoption concentrated in information and communications, electronics, professional services, and finance and insurance. In other words, at firm level, adoption has so far been associated with growth rather than contraction.
Alongside that, the government worked with more than twenty partners — Accenture, Microsoft and J.P. Morgan among them — to curate close to 2,000 job and training opportunities for fresh graduates and mid-career professionals moving into technology roles, announced by Minister Josephine Teo on 20 August 2026. The opportunities span AI and data, cybersecurity and software engineering.
Minister Teo framed the response as “AI bilingualism”: professional knowledge and AI skills as complements rather than substitutes, with the premium on human judgement rising as machines get better at producing code, analysis and content. That framing matches what we see in our own pipeline far better than any displacement narrative does.
Planning 2027 headcount without a reliable external signal?
We benchmark your open roles against live Singapore pipeline data — real time-to-fill by role family, SGD bands that are actually closing, and the requirement shifts we see arriving in specifications weeks before they appear in salary surveys.
Let's talkThe 3 metrics I track instead of waiting
None of these require new tooling. All three come from data sitting in your applicant tracking system and your HR records, and all three move before published statistics do.
1. Time-to-fill by role family, monthly
Not average time-to-fill across the company, which averages away everything interesting. Break it into six or eight role families and track each one month by month. A family whose time-to-fill is stretching is a family where scarcity is increasing, and it shows up well before salary data confirms it.
The inverse signal is just as useful and much less discussed: a role family whose time-to-fill is shrinking unusually fast may be one where supply is rising because the work is being automated elsewhere. Both directions are information.
2. Internal moves versus external hires
Count how many roles you filled by moving someone internally against how many you filled from outside, by role family. This measures something no external dataset can tell you: whether your own people are able to move into the roles that are growing.
If a role family is growing and every hire into it comes from outside, your internal mobility is broken and you are buying capability you could have built. That is a much more actionable finding than a national displacement figure.
3. Share of requirements that did not exist 18 months ago
Take your current open specifications, and mark each requirement that would not have appeared in the same role eighteen months ago. Express it as a percentage.
This is the closest thing to a direct measure of how fast the work itself is changing, and it is startlingly informative. In the specifications we handle, that share is presently high in AI-adjacent engineering and near zero in several long-established backend roles — which tells you exactly where retraining effort pays and where it does not.
Expert view (2/3)
Freezing hiring in response to displacement headlines is the most expensive available mistake. A freeze is indiscriminate: it stops you filling the growing roles just as effectively as the shrinking ones, and it preserves your existing shape at exactly the moment the shape needs to change. The Singapore firm-level evidence points to redistribution rather than contraction, and redistribution is something you manage by moving people and rewriting specifications, not by stopping. The organisations that will look well-run in 2028 are the ones that kept hiring but changed what they hired for — and they will have known what to change because they were watching their own operational data, not waiting for a published series.
Three internal signals, and what each one tells you
What Singapore employers should do this quarter
| Action | Effort | What it prevents |
|---|---|---|
| Split time-to-fill by role family | Half a day | Averaging away the only useful signal |
| Count internal moves per family | Half a day | Buying capability you could build |
| Mark new requirements in open specs | 2 hours | Training spend aimed at stable roles |
| Review the 2,000 curated national placements | 1 hour | Missing a subsidised retraining route |
| Rewrite one spec for AI bilingualism | 2 hours | Hiring for tasks machines now do |
This shift is regional rather than local. Our colleagues at HireDeveloper.ae report the same requirement churn in Dubai specifications, and the team at JapanDev sees it arriving more slowly in Tokyo, where longer tenures dampen the signal. If you are standing up a new function rather than adjusting an existing one, our guide on how to build an AI engineering team covers sequencing and the first hires.
Expert view (3/3)
Treat vendor-supplied labour data with the same scepticism you would apply to any supplier-authored benchmark. The companies handing over adoption data are the companies selling the adoption. That does not make the data worthless — it is genuinely faster and closer to the ground than a quarterly survey — but it does mean it carries a directional interest. My working rule is to use external data to generate hypotheses and internal data to test them. If a vendor dataset says a role family is contracting, go and look at your own time-to-fill for that family before you act. When the two disagree, your own numbers are describing your actual business and theirs are describing a market average that may not include you.
Frequently asked questions
What did the US Labor Department actually announce?
Axios reported on 26 August 2026 that the US Department of Labor is working with major technology companies — including OpenAI, Google, Meta and Amazon — to better understand how artificial intelligence is changing employment and hiring. The department has signed memorandums of understanding with a number of these companies, which are sharing data on how businesses use AI and where they expect adoption to go. Acting Labor Secretary Keith Sonderling is leading the effort. The stated purpose is to supplement traditional government labour statistics with private-sector data that gives a faster read on AI adoption and its occupational effects.
Why does this matter to employers outside the United States?
Because of what it concedes rather than what it creates. A government statistical agency turning to private vendors for a faster signal is an admission that conventional labour data arrives too late to guide decisions. Official series are typically published on a lag of one to several months and classify work using occupational categories designed before the current wave of automation. If that lag is too long for a national policymaker, it is certainly too long for an employer planning headcount for the next two quarters. The lesson transfers to Singapore directly: build your own leading indicators rather than waiting for published statistics.
Should Singapore employers slow down hiring because of AI displacement?
The evidence does not support a blanket slowdown. A Singapore government study reported in mid-August 2026 found that firms using AI saw gains in both revenue and employment, with adoption concentrated in information and communications, electronics, professional services, and finance and insurance. The pattern is redistribution rather than contraction: demand falls for narrowly defined execution roles and rises for roles combining domain judgement with AI fluency. The risk of freezing hiring is that you preserve exactly the roles being automated while failing to build the capability that replaces them.
Which three internal metrics should we track?
First, time-to-fill by role family, tracked monthly — a role family whose time-to-fill is lengthening is one where scarcity is rising, and this moves months before salary data does. Second, the ratio of internal moves to external hires, which tells you whether your people are actually able to shift into the roles that are growing. Third, the share of job requirements in your own specifications that did not exist eighteen months ago, which measures how fast the work itself is changing. All three come from data you already hold, and all three lead the published statistics.
The bottom line
A labour ministry asking four AI companies for their adoption data is a striking headline, but the useful content sits underneath it: the official picture of how work is changing arrives months after the change, and everyone planning headcount is working with that handicap.
You cannot fix national statistics. You can stop depending on them. Time-to-fill by role family, internal moves against external hires, and the share of requirements that are genuinely new — three numbers, all sitting in systems you already run, all moving one to two quarters ahead of anything you will read in a report. Build that dashboard this quarter and you will be making 2027 decisions with 2026 information rather than the other way round.
Benchmark your Singapore roles against live pipeline data
We give you real time-to-fill by role family, the SGD bands that are actually closing this quarter, and the requirement shifts appearing in specifications before they surface in any published survey — plus a pre-vetted shortlist in 3 to 4 weeks.
Let's talk