ADP Broke the Wage Signal. 1 in 172 Engineers Is Why.

On September 2, ADP printed 38,000 August jobs, the weakest month since January, and chief economist Nela Richardson said the quiet part out loud: AI has overtaken demographics and inflation as the force distorting wage patterns. The same release showed job-changers earning 7.3% more year over year against stayers at 4.4%. If you are still pricing offers off Radford medians or a Levels.fyi screenshot, you are guessing.

What ADP actually said, and why comp teams should care

ADP's chief economist publicly conceded that the wage patterns the Federal Reserve and every enterprise comp team rely on have been scrambled by AI, which means the benchmark inputs feeding your offer letters are calibrated on a market that no longer exists.

Richardson's exact framing, from the September 2 release: "Once predictable wage growth has been overtaken by complexities of demographic change, persistent inflation, and AI's effects on jobs." That is the first time ADP has named AI as a top-tier wage distorter in a headline release. The dataset behind the claim is not small: ADP tracks pay for more than 14.8 million individual workers, comparing each one's pay over rolling 12-month intervals.

The specifics of the August print:

• 38,000 private jobs added, missing the Dow Jones consensus of 47,000
• Job-stayer base pay up 3.0% YoY, job-changer base up 4.7%
• Job-stayer gross pay up 4.4%, job-changer gross up 7.3%
• Job-changer gross growth decelerated from 7.5% to 7.3%, meaning the switching premium peaked earlier and is still elevated
• Manufacturing, professional services, and information categories all shed jobs

1.66x

Why the benchmark stack quietly went stale

Every major comp benchmark is a lagged, aggregated snapshot of a market that is now repricing weekly at the top and shedding jobs in the middle, which mechanically produces medians that are wrong in both directions at once.

Here is the mechanism, benchmark by benchmark:

• Radford, Mercer, Willis Towers Watson. Survey cycles run 6 to 12 months. When job-changer gross pay is running at 7.3% and stayers at 4.4%, a benchmark refreshed even one quarter ago understates market-clearing offer levels by 300 to 700 basis points depending on role heat.
• Levels.fyi. Self-reported and public-reporting-based. OpenAI does not publish comp bands; Levels.fyi's own Competitive Intelligence page notes it is inferring numbers from press coverage of retention packages.
• LinkedIn Salary. Aggregates member-reported base and bonus. It cannot see clawback-structured signing grants, PPUs, or acquihire earnouts.

The blended result is a single "software engineer" median that averages a shrinking cohort of stayers against an exploding frontier-lab tail. The median lies about both populations. Richardson called wage growth "once predictable," and predictability was a data assumption, not a market fact.

The 1-in-172 problem your median is hiding

The mechanical driver of the wage-signal breakdown is scarcity: LLM-skilled engineers are roughly 1 in 172 U.S. software engineers, so a handful of outlier packages at the top of the distribution drag the mean without moving the count of people who can actually clear the bar.

In Refolk's index of professional profiles, the U.S. software engineering pool breaks down like this:

Two things fall out of this table. First, the "AI engineer" label is not one market. 16,997 people is a real pool with functional benchmarks in the $170K to $245K enterprise range, per pin.com's 2026 guide. 2,053 people is a bidding war. Second, the outliers set the anchor for every counteroffer conversation the specialist tier has, even though 99.42% of software engineers will never see one.

The median lies about both populations, and every comp committee in the country is still quoting it as if it did not.

The outlier packages that broke the anchor

Named comp outliers are not statistical noise anymore; they are the anchoring price that every frontier-lab candidate walks into a negotiation carrying, and benchmarks that filter them as outliers are systematically wrong about what "market" means for the top 2,053 people.

The receipts:

• Ruoming Pang, ex-Apple foundation models lead, took a Meta package worth $200M over several years, exceeding Apple's pay scale for anyone except Tim Cook.
• OpenAI's August 2025 retention program, per Levels.fyi's own Competitive Intelligence page: $300K new-hire grants vesting over 2 years, existing-staff bonuses ranging up to $1.5M, staff-level TC confirmed up to $700K.
• Meta Superintelligence Labs bands: E5 at $525K to $700K, E6 at $750K to $1M, E7 typically $1M to $1.5M on standard offers.
• Scale AI: Meta's $14.3B investment doubled as an acquihire of Alexandr Wang and team. Benchmark data cannot see it as compensation at all.
• Signing bonuses as the new base: a $2M signing paid over two years with clawback through year three means the recipient must stay three years to realize the full bonus. Standard practice across frontier labs, invisible in "base + equity" schemas.

If your comp philosophy caps at the 75th percentile of a Radford cut, you are not competing in the tier where these anchors set expectations. That is fine. Just say so out loud, and target the 16,997 not the 2,053.

How to price offers from first principles in 2026

When the benchmarks lie in both directions, the answer is not a better benchmark, it is a pricing method built on the individual candidate's realized alternatives rather than an aggregated median. Here is the working recipe.

1. Segment the role before you price it. Decide whether you are hiring from the 353,811 pool, the 16,997 pool, or the 2,053 pool. Different populations, different anchors, different offers. Do not let one requisition drift up the ladder.
2. Price against realized alternatives, not medians. For every finalist, ask what their next-best offer actually looks like today. Job-changer gross pay is up 7.3% YoY; if your offer only clears the incumbent's base by 8%, the candidate is roughly indifferent.
3. Break out clawback-adjusted realized comp. Model year-one, year-two, and year-three take-home separately. Frontier-lab candidates already do this. The recruiter who does not will lose to the one who does.
4. Refresh comp bands quarterly for AI roles, annually for the rest. The switching premium moved from 7.5% to 7.3% in a single month. An annual band refresh for AI-specialist roles is not a policy, it is a delay.
5. Track your own win/loss data as a benchmark of one. Every declined offer is a market data point. Roll them into a rolling 90-day view. It is the only signal that is not lagged.

The upstream problem is finding the right candidates to price in the first place. When the specialist pool is 2,053 people nationally, keyword-Boolean sourcing across LinkedIn returns thousands of false positives who list "GPT-4" in a skills section and have never shipped an eval harness. That is the exact gap Refolk closes: describe the person in plain English ("engineers who have shipped a production RAG pipeline at a Series B or later company, currently based in the US") and get a ranked shortlist pulled across GitHub, LinkedIn, and the open web.

Where the enterprise AI tier still has functional benchmarks

The enterprise AI engineer tier ($170K to $245K) still has usable benchmarks because the employer set is large enough that no single outlier moves the median, and this is where most comp teams should actually be hunting rather than chasing frontier-lab price signals they cannot match.

Employers hiring at this tier that show up in Refolk's index outside the frontier-lab headline names include Distyl, Sandgarden, Develop Health, RAI Institute, and PostEra. They are hiring the 16,997, not the 2,053. Their offers clear market by paying at the 65th to 75th percentile of a functional benchmark and closing quickly, not by outbidding Meta.

The tactical implication for a sourcing team: if your comp band tops out at $245K, do not waste cycles pitching candidates whose LinkedIn headline already reads "MSL" or "Frontier." Filter for the exact experience shape you can afford - shipped LLM features in production, not currently at OpenAI, Anthropic, or Meta MSL - which is the difference between a 3% and a 30% reply rate on outbound.

What the latest Pulse changed for Q4 planning

The NER Pulse showed the switching premium is decelerating from a higher peak but is still nearly 2x stayer growth, which means Q4 is the quarter where comp bands either catch up or lose every contested candidate to whoever moved first.

For the four weeks ending August 29, private employers added an average of 16,250 jobs per week. That is a hiring market that looks soft in aggregate and molten at the top. Two operational reads:

• The soft aggregate is why your CFO thinks you have leverage. You do not, not in the roles you actually need to fill.
• The molten top is why every AI-adjacent requisition open past 60 days is going to close 15 to 25% above where it was posted.

Plan the band revision now. Do it as a documented policy tied to the ADP September 2 release so it survives the next comp-committee meeting. And staff your sourcing pipeline against the 16,997, not the 2,053, unless you have Meta's balance sheet.

FAQ

Are Radford and Mercer still worth the subscription?

For non-AI roles, yes. Their methodology is sound and the survey depth in general engineering, sales, and G&A is unmatched. For AI-specialist roles the honest answer is that the benchmark cycle of 6 to 12 months is longer than the repricing cycle of the market itself. Use them as a floor, not a ceiling, and build a rolling win/loss log alongside them.

How do I explain "the benchmark is wrong" to a comp committee?

Anchor on ADP directly. The chief economist of the largest private payroll dataset in the country, tracking 14.8 million workers, named AI as a top-tier wage distorter in a public September 2 release. The job-changer premium is 1.66x stayer growth. Both facts are attributable to a Fed-grade source, which is the framing that gets a comp band revised.

What is the job-changer wage premium in 2026 and why does it matter for offers?

It is the pay differential between people who switched jobs in the last 12 months and people who stayed. As of ADP's August print, job-changers earned 7.3% more YoY on gross pay against 4.4% for stayers, a 2.9-point gap or roughly 1.66x. It matters because your candidates are the job-changer population by definition. Pricing them against the stayer median, which is what most benchmarks blend toward, underprices them by roughly the size of that gap.

Where does Refolk fit in a comp-first sourcing workflow?

Refolk handles the "find the person" half so you can spend the comp conversation on the person who actually clears your bar rather than sorting through hundreds of skills-section false positives. Describe the candidate in plain English, get a ranked shortlist across GitHub, LinkedIn, and the open web, then price each finalist against their realized alternatives rather than a broken median.