Sourcing by Employer Reputation Reports What the Company Bought

What does an employer's AI reputation report about one candidate?

An employer's AI reputation reports a decision the company made. Adoption is recorded at the level of the firm: somebody signed a contract, issued the licenses, and told a business function to use them. Nothing in that decision names which employees learned to check what came back.

Firm adoption and personal use are measured separately, by different surveys, and the two do not agree.

In April 2026 the Federal Reserve published a note comparing the three main US adoption surveys. They disagree, and the note puts the likeliest reason in what each one counts: firms, workers, or firms weighted by the people inside them ¹.

Weighted by employment, about 78% of the US labor force sits at an employer that has adopted AI. Asked directly, about 41% of the workforce said they had used generative AI at work in November 2025 ¹.

The 78% is a fact about employers, and it is the number a sourcing filter reads.

Census figures count the firms rather than the workers inside them. The Business Trends and Outlook Survey is a biweekly, nationally representative series that asks firms about the previous two weeks. As of 3 May 2026 it found 19.8% of US businesses using AI, against 37% of firms with at least 250 employees and about 14% in retail trade ².

Sector moves the number too. AI use ran at 39.7% in Information and 33.9% in Finance and Insurance, both well above the national rate ².

A list of AI-forward employers leans toward large companies, and toward sectors like those two.

Why doesn't heavy adoption separate the people inside it?

Heavy adoption hands the tool to everybody, which is what makes it useless as a sort. At a firm that rolled licenses out to a whole function, having used the assistant is the base rate, shared by the person who reads every output and the person who forwards it unread. The variation worth hiring on sits underneath that, and no employer name reports it.

One study inside a single firm shows how unevenly the same tool lands.

Researchers at Harvard Business School ran a pre-registered field experiment in 2023, giving GPT-4 to some of 758 Boston Consulting Group consultants working across 18 realistic consulting tasks ³.

On the tasks that sat inside the model's capability, the consultants who had it completed 12.2% more of them and finished 25.1% faster ³. On those tasks, in that sitting, the tool helped.

The gain was not spread evenly. Consultants below the average performance threshold improved 43% against their own baseline, and those above it improved 17% ³.

Separately from the 18, one task was chosen to sit outside what the model could do.

There, 84.5% of the control group reached the right answer, against 60% and 70% in the two groups holding GPT-4 ³. The people with help did worse than the people without it.

The two groups holding it had not been prepared the same way. One had GPT-4 and nothing else. The other had GPT-4 and a prompt-engineering overview first.

The overview group finished 24 percentage points below the control group. The model-only group finished 13 points below it ³. On this task, being taught to prompt left people further from the right answer than being handed the tool with no instruction at all.

On the tasks inside the model's capability that ordering ran the other way, and the overview group gained more than the model-only group ³.

Everyone in that experiment worked at the same firm, so the employer explains none of those gaps. The drop on the outside task is evidence that people could not tell which side of the capability line it sat on.

GPT-4 is a 2023 model and the line has moved since, but the skill that task exposed has not: knowing when to stop believing the assistant.

The judgment transfers across tools; the tool habits do not. A candidate who learned to check a confident claim at one employer carries the habit into a stack they have never opened. A candidate who never learned it arrives from the best-equipped company on your list without it.

Which candidates disappear when the filter is an employer?

Two groups disappear, and both are groups worth wanting. The first is anyone whose employer never bought the tool, which tracks firm size and sector rather than skill. The second is the person who introduced AI somewhere nobody handed it to them, an act that is stronger evidence than being issued a license and one that no company field records.

The first cost is a proxy problem, and renaming the field does not remove it.

Screening on employer identity imports whatever that employer's workforce looks like, along with its size, its industry and its location.

Amazon's abandoned recruiting experiment carries the durable line about proxies. Built from 2014 on ten years of resumes submitted to the company, the tool was penalizing resumes containing the word "women's" by 2015, and the team behind it was disbanded by the start of 2017 ⁴.

Amazon edited the programs to be neutral to the flagged terms, and Reuters reported that the edit "was no guarantee that the machines would not devise other ways of sorting candidates that could prove discriminatory" ⁴.

The account rests on five people who spoke to the reporter anonymously, Amazon declined to comment on it, and the technology predates large language models entirely.

The principle survives all three limits. Proxies survive deletion because the record rebuilds them, and an employer name is one of the densest proxies available, since it carries a payroll, a location and an industry in a single token.

The second cost never shows up on a dashboard.

Someone who talked a sceptical finance team into a pilot, wrote the review rule, and got it past a security officer has produced a record of judgment that a colleague at an AI-forward firm may never have been asked to produce.

No number will tell you who the filter left out, because it runs before anyone applies. You cannot compute a selection rate on people who never applied, and a sourcing list that never contacted them leaves no record to compute one from.

Use the company name to read the answer, not to build the list

Keep the target list and change what it is for. A candidate's employer tells you which tools they have probably touched, which review norms they lived under, and what vocabulary will make the first call efficient. None of them measures the person, and the published validity numbers for work experience point the same way.

Prehire experience is the closest measured analogue, and it predicts almost nothing. Van Iddekinge, Arnold, Frieder and Roth pooled 81 independent samples in a 2019 meta-analysis. They put the correlation between prehire work experience and later job performance at .06, and at about .07 where the experience was specifically relevant to the job ⁵.

Both figures are corrected correlations, measured at hire, which is exactly the moment a sourcing list and a resume screen read. Experience with the right tasks did no better than experience in general.

An employer's name is cruder than years of experience, because it does not even claim to describe the candidate.

What does predict, on the same corrected scale, is a structured conversation about the work. Sackett, Zhang, Berry and Lievens re-analysed the selection literature in 2022 and placed structured interviews top at .42, ahead of job knowledge tests at .40 and unstructured interviews at .19 ⁶.

The .42 comes with a residual standard deviation of .19 and a lower 80% credibility value of .18, so a structured interview in any one setting can land well below the mean ⁶.

Structure there means the same questions for every candidate, rated on a common scale, and neither part needs a budget.

So spend the outreach message on one decision, asked the same way of everyone on the list. Ask what the candidate stopped a model from doing and why, or what they checked before a number left the building.

Verifying "uses AI daily" takes about ten minutes on a first call.

Exposure is what a target list actually buys. A team that rebuilt a workflow around a model has watched something get rolled back, has argued about what needs a second reader, and holds a working pattern in its head. Sourcing there does raise the odds that a candidate arrives with one.

The list on your screen is still worth having. The change is the column it sits in: notes for a call rather than a shortlist.

Census refreshes the adoption series every two weeks, so the sector figures in it can shift between readings ². None of those readings describes the person you contact.

See a sample report >