The Cost of Hiring One Employee: 1 Million Yen. That Conventional Wisdom Is Over.

The average hiring cost for small businesses is said to be between 800,000 and 1,200,000 yen per hire. They post job ads, pay agent fees, and spend half a day in interviews, only to lament that “good candidates aren’t coming.”

This structure is now fundamentally changing with AI agents.

In an era where AI tools costing 30,000 to 50,000 yen per month can automate everything from resume screening to initial evaluations, hiring costs can drop from 1 million yen to just 50,000 yen — this is not just about “efficiency”; it’s a game-changing shift.

However, there is one troublesome issue: AI may be “unintentionally” filtering out female candidates.

—

Two-Stage AI Screening Increases Pass Rates by 6 Points

Recent research has yielded noteworthy results. Using AI agents in a two-stage process (initial screening + re-evaluation) significantly improves candidate pass rates compared to a single evaluation.

Let’s look at the specific numbers:

• GPT Model (1 Stage): Pass Rate 33.3%
• GPT Model (2 Stages): Pass Rate 39.3% (+6.0pt)
• Claude Model (1 Stage): Pass Rate 34.0%
• Claude Model (2 Stages): Pass Rate 35.5% (+1.5pt)

In a single evaluation, AI tends to have a “drop if uncertain” tendency. By implementing a two-stage process, candidates on the borderline can be re-evaluated, reducing the chances of overlooking potential talent. It’s common in human interviews to hear, “I initially rejected this candidate, but upon a second look, they were actually good.” The same phenomenon occurs with AI.

What does this mean for small businesses?

The probability of identifying “talented individuals who struggle with resume writing” increases.

Large companies can afford to “drop if uncertain” due to the overwhelming number of applicants. However, small businesses cannot afford to overlook anyone, as they have fewer applicants. A two-stage screening process is particularly effective for small businesses.

—

The Invisible Bias — The Mechanism Behind AI’s “Dropping of Women”

Now we arrive at the main pitfall.

There is a risk that gender bias is structurally embedded in AI resume screening. Research has revealed that the wording of job postings distorts AI evaluations.

Specifically, when job postings frequently use what is known as agentic language (words that evoke leadership, initiative, and dominance), AI tends to assign higher scores to male candidates while lowering scores for female candidates.

Conversely, using communal language (words that evoke teamwork, empathy, and support), such as “We are looking for someone who can collaborate with the team and solve customer challenges together,” mitigates this bias.

Why does this happen? The AI’s training data includes past hiring data and vast amounts of text, which reflect societal stereotypes like “leadership = male” and “cooperativeness = female.” The AI faithfully reproduces these patterns.

In other words, AI is not trying to discriminate; it is merely “learning” and reproducing past discrimination.

This is not a distant issue for small businesses; it is rather serious. Large corporations have specialized departments for promoting diversity and legal teams. However, if small businesses think, “I can relax now that I’ve left it to AI,” they may unknowingly be excluding female candidates — this risk is very real.

—

Three Actions Small Businesses Should Take Starting Today

So, what can be done? Here are three specific actions, not abstract theories.

1. Rewrite Job Posting Language to Be More “Communal”

This can be done at zero cost starting today.

NG Examples:
“We seek someone who can proactively drive business as an immediate asset.”
“Someone who can set high goals for themselves and commit to achieving them.”

OK Examples:
“This job involves working with the team to collaboratively solve customer challenges.”
“There is an environment where you can leverage your experience while growing together with those around you.”

Just changing the wording can alleviate AI evaluation bias and increase the willingness of female applicants to apply. It’s a win-win.

2. Design AI Screening in Two Stages

Do not decide pass or fail based on a single evaluation. Re-evaluate candidates who were marked as “failed” in the initial screening using a different prompt (a slightly altered evaluation criterion).

This alone can improve pass rates by 3 to 6 points. The monthly cost will only double the number of API calls, resulting in an additional few thousand yen. For 100 applicants, the extra cost will only be a few hundred to a few thousand yen.

3. Collect AI Evaluation Results by Gender Monthly

No complex analysis is necessary; a spreadsheet will suffice.

• Pass Rate for Male Candidates: ◯%
• Pass Rate for Female Candidates: ◯%

Just look at these two numbers every month. If the gap exceeds 10 points, there may be an issue with the wording of the job posting or the prompt. Make corrections and check again the following month. By simply cycling through this process, bias will certainly decrease.

—

The Cost Discussion — “Just Because It’s Cheap Doesn’t Mean It’s Sloppy”

The cost of AI screening is dramatically low. While fees paid to recruitment agencies can be 30-35% of the annual salary (120,000 to 140,000 yen for a 4 million yen salary), the monthly cost for AI screening is only 30,000 to 50,000 yen. Even annually, that’s just 360,000 to 600,000 yen.

However, that doesn’t mean you can use it carelessly just because it’s cheap.

Continuing to filter out excellent female candidates with a biased AI screening will accumulate as an invisible cost. The cost of missing out on candidates who could have been hired will far exceed the cost of the tool.

Especially for small businesses in rural areas, the hiring market is a seller’s market. In a situation with few applicants, losing even one candidate due to AI bias is fatal.

—

Don’t Overlook AI’s “Language Cost”

Another practical point to be aware of is that the generation cost of AI agents varies depending on how prompts are written and the language used.

Recent studies have reported that even for the same task, the API call cost can vary by 2 to 3 times depending on prompt design. In tasks like recruitment screening that process a large number of candidates, this difference directly impacts monthly costs.

Specifically, prompts that clearly structure evaluation criteria (e.g., “Evaluate the following 5 items on a scale of 10”) consume fewer tokens and yield more stable results than vague instructions (e.g., “Make a comprehensive judgment”).

Prompt design affects both quality and cost.

—

Conclusion — AI Can Be a Boon or a Bane Depending on How It’s Designed

AI recruitment screening has the potential to reduce hiring costs for small businesses to one-tenth. By reducing oversights with a two-stage screening process and optimizing costs through prompt design, small businesses can establish a hiring process comparable to or better than that of large companies for just 50,000 yen a month.

However, if the design is flawed, it can become a mechanism that unknowingly reproduces gender bias and continues to exclude excellent candidates.

What needs to be done is simple:

1. Review the wording of job postings (zero cost, can be done today)
2. Implement a two-stage screening process (additional cost of a few thousand yen per month)
3. Check pass rates by gender monthly (a spreadsheet is sufficient)

AI is not magic. It’s a tool. Depending on how it’s used, it can become the strongest weapon for small businesses or an invisible bomb.

Don’t just leave it to AI. The difference will be made by how you design AI.