In short
Our published salary dataset set its floor at 5 disclosed listings. Checking three monthly snapshots, medians from samples of 12 or fewer swung by 50% to 89% month over month, while samples of 23 or more moved under 5%. We have raised the floor to 30 and now name every segment we withhold, with its sample size.
Until this week, the open dataset behind our State of Agentic Hiring report published a salary median for any segment with at least five disclosed listings. One consequence sat near the top of the file: a top-paying-employer median of $402,500, drawn from five listings.
That number was not wrong in the sense of being miscalculated. It was the true median of those five listings. It was wrong in the sense that it would not survive contact with next month's data, and anyone who quoted it would have been left holding a figure we no longer reported.
What we did about it
We keep monthly snapshots, so the reliability of a published median is a measurable question rather than a matter of taste. We took every salary segment that appears in all three of our July, August and September 2026 snapshots and measured how far its median moved between them.
Sorted by sample size, the pattern is not subtle:
- Junior, 4 disclosed listings: swung 62% ($140K, then $95K, then $86K)
- Intern, 5 listings: swung 51%
- Executive, 12 listings: swung 89% ($150K, then $262K, then $284K)
- Generative AI, 19 listings: swung 29% ($211K, then $178K, then $164K)
- LLM Engineer, 15 listings: swung 21%
- AI Infrastructure, 23 listings: swung 2.6%
- Data Scientist, 29 listings: swung 2.6%
- ML Engineer, 33 listings: swung 4.1%
- Senior, 217 listings: swung 1.9%
Below about 20 listings, a median is a coin flip dressed as a finding. Above about 23, it is stable to within a few percent. There is no gentle slope between the two, which is what makes it a usable cutoff rather than a judgement call.
The exception worth naming
One segment breaks the pattern. AI Agent Engineer has 47 disclosed listings, comfortably above the line, and still swung 11.7%. That is not the estimator failing. That category grew by 73 roles in the last 30 days, the second fastest on the board, and a category whose composition is changing that quickly will move its own median honestly.
Sample size tells you whether a number is noisy. It does not tell you whether the underlying thing is moving. Those are different questions and it is worth keeping them apart.
What changed in the dataset
The floor for any published salary median is now 30 disclosed listings, applied to roles, seniority levels and employers alike. That cost us twelve of sixteen role medians and three of seven seniority medians. The $402,500 employer figure is gone, replaced by five employers whose smallest sample is 32.
What survives is worth more than what we lost. The seniority table now runs Staff at $263K from 163 listings, Senior at $216K from 217, Lead at $211K from 78 and Mid at $205K from 249, all of which moved less than 4% across three months. The Staff-over-Lead inversion we wrote about in August holds up precisely because those samples are large.
Segments under the floor are not deleted. They are named, with their disclosed-listing counts, in the report and in the machine-readable dataset under salary.withheld. If you want to know why there is no Prompt Engineer salary this month, the dataset tells you: four listings disclosed pay.
What to ask of any AI salary report
Only 728 of our 1,737 listings disclose pay at all, which is 41.9%. Every salary figure published about this market, ours included, rests on that minority. Segment it by role and by seniority and the samples get small quickly. This is a structural feature of AI compensation reporting, not a quirk of our corpus.
So three questions are worth asking of any salary figure you are about to rely on, including ours:
- What is n? If a report does not publish sample sizes per segment, it is asking you to trust a number it has not shown you the basis for.
- What is the spread? A median with no interquartile range hides whether the segment is tight or bimodal. Ours now publishes p25 and p75 on every row.
- What is missing, and why? A report with a number for every category has either a very large corpus or a very low floor.
We would rather publish four role medians we will stand behind than sixteen we will quietly revise. The full report, the CC BY 4.0 dataset and the withheld list are on the State of Agentic Hiring page, and the per-role salary pages are under the data section.
Frequently asked
How many listings do you need for a reliable salary median?
Based on three monthly snapshots of our own corpus, about 30. Segments with 12 or fewer disclosed listings swung 50% to 89% month over month, segments in the teens swung 8% to 29%, and segments of 23 or more moved under 5%.
Why did you remove the $402,500 employer salary median?
It was drawn from five disclosed listings. Our published floor was five, which was far too low to be stable. The floor is now 30 disclosed listings, and the five employers we do publish have a smallest sample of 32.
What is the median salary for agentic AI roles?
The overall median is $217,500 USD annual, from 728 disclosed listings out of 1,737 tracked. By seniority: Staff $263K from 163 listings, Senior $216K from 217, Lead $211K from 78 and Mid $205K from 249.
What share of AI job listings disclose salary?
In our corpus, 728 of 1,737 listings disclose pay, which is 41.9%. Every published salary figure about this market rests on that minority, which is why per-segment sample sizes matter so much.
Where can I see which salary segments you withhold?
They are listed on the State of Agentic Hiring page and in the open dataset under salary.withheld, each with its disclosed-listing count. This month that includes Data Scientist at 29, AI Infrastructure at 23 and Prompt Engineer at 4.