A recruiter at a small agency stares at a role she has to fill by Friday. She has thirty profiles open in tabs and a headache. Across town, a competitor sends her client a shortlist of candidates scored against every requirement before she finishes her coffee. Guess who keeps the account.
You could be the person behind that shortlist. And maybe you need to be. Maybe your current job feels thinner every quarter, with the tasks you used to own handed to software and the hiring freeze that never quite lifts. Maybe your rent went up again and your savings went the other way. Sitting still has a price too.
Recruiting turns over more than $200 billion a year and still runs on spreadsheets and templated InMails. That will change fast, and the guide's predictions say clients will soon ask every agency which AI tools it uses. The builders who already have a working pipeline and a compliance pack when that question becomes routine get the retainers. The ones still tinkering get the leftovers.
The workflow on this page pulls 200 to 500 relevant candidates in the time a human reaches 30 to 50, scores each one with GPT-4o, and hands back everyone who clears 7 out of 10. At the low end the service sells for $2,000 a month. The tools cost between nothing and $500, and most people see profit within one to three months.
Start tonight with one real job description. Paste it into a chat model and ask it to split the role into must-have skills, nice-to-have skills, minimum experience and location. That structured output is the front door of everything you are about to build.
Here is what the service is, in plain terms. You build and run automated pipelines that handle the top-of-funnel recruiting work that companies either do poorly or pay agencies $15K-$25K per hire to do for them. Your AI agents find candidates on LinkedIn and other platforms, fill in their profiles with contact data and background information, score them against the job requirements using large language models, send personalised outreach, and book interviews. All of that happens before a human recruiter touches anything.
Why is the opening so wide? Because you are up against an industry with 20-30% margins that has barely changed how it works in two decades. A traditional staffing agency assigns a recruiter who searches databases by hand, makes phone calls, and juggles dozens of open roles. Your pipeline does the work of 3-5 junior recruiters at a fraction of the cost, and it keeps running 24/7 while you sleep.
Start with what hiring costs the people you would sell to. The average cost-per-hire in the US is $4,700, and for technical roles it can go past $15,000. Companies with 100+ employees spend an average of $1.1 million a year on recruiting. The average time-to-fill for a position is 44 days, and for every one of those days the company loses productivity, revenue, and often the candidate to a faster competitor.
If you have ever been on the hiring side, you have felt the problems below yourself. AI answers each of them directly:
Your pipeline has four stages: sourcing, screening, outreach, and scheduling. You can automate each one as much or as little as you like.
Sourcing means finding the people who fit the job. A recruiter usually does this by searching LinkedIn, Indeed or an internal database by hand. Your agent takes that over by:
Set up well, your sourcing workflow can find 200-500 relevant candidates per role per week. A recruiter working by hand finds 30-50.
This is where you add the most value. Your agent takes each sourced candidate and scores them against the job requirements using GPT-4o. Here is what that workflow looks like:
Your screening prompt is the heart of what you sell, so give it real time. Teach it nuance. Career trajectory matters more than years of experience. Open-source contributions can make up for a missing degree. Context matters too: a "senior engineer" at a 10-person startup has a very different job from one at Google. If a client asked you why candidate 14 scored a 6, could your prompt's reasoning answer them?
Once your candidates are scored and ranked, your agent writes and sends personal outreach. This usually goes by email (to the verified addresses Clay found), with LinkedIn InMail as a second channel.
Well-personalised AI outreach keeps landing 25-40% response rates. Template recruiting emails get 5-15%.
When a candidate says yes, your agent books the meeting. It connects to the hiring manager's calendar (through Calendly, Cal.com, or the ATS's own scheduler) and offers the candidate open slots. Confirmations, reminders and rescheduling all run on their own.
Your target is to go from "candidate interested" to "interview on the calendar" within 24 hours, with no human stepping in. That speed sells. Most companies take 3-7 days to schedule an interview after a candidate shows interest, and by then the candidate may have taken another offer.
So you have seen what your agent does to a job description once it lands. What it runs on is your next decision, and it is the one that sets whether your setup costs you nothing or the full $500.
You do not need custom software. You can assemble the whole pipeline from tools that already exist:
Total tech stack cost: $400-$800/month, which one client retainer covers.
That is the build settled. Now for the number you put in front of a recruiter every month.
Set your service up in tiers so you can serve clients of different sizes and budgets:
Scale, $5,000/month
- Up to 12 active roles
- Full pipeline with priority processing
- Unlimited outreach volume
- Custom screening rubrics per role
- Dedicated Slack channel with weekly reporting
- Quarterly bias audits
- Best for: Mid-market companies with ongoing hiring needs
At the Growth tier these work out to roughly $500 per active role per month. A contingency recruiter charges $5K-$20K per successful hire. Put that comparison in front of every prospect you speak to, because it does most of your selling for you.
Picture one recruiter on your $1,000 Starter tier. After sourcing and screening costs, that retainer could help pay your rent with the banking app left closed, covered by a pipeline that finds candidates for someone else's open role. Two of them and you reach $2,000, where this page's range begins. What would that first $1,000 cover in your house before anything else?
Now picture the call where you quote that retainer and the recruiter says yes without haggling, because your price is a fraction of one contingency fee. That call changes how you walk around. You stop flinching at the dentist's estimate. You say yes to the weekend away your partner keeps mentioning.
Finding your first clients
You have a price. Whether you reach that first $2,000 month inside one to three months comes down to who you say it to, and how.
Your best clients come from three places:
Fast-Growing Startups (Best Starting Point). These are companies that just raised money and need to hire 5-20 people in the next 6 months. They feel recruiting pain sharply, because they are small enough that the founders or engineering leads are doing the hiring themselves, and they hate it. You find them by watching Crunchbase for recent funding rounds, then writing to the head of talent (if there is one) or straight to the CEO.
Staffing and Recruiting Agencies (Highest Volume). This one surprises people. You sell to recruiters and work alongside them. Plenty of staffing agencies are buried in open requisitions and cannot source fast enough. Your pipeline becomes their sourcing engine. They put their own label on your output and present the candidates to their clients. A single agency client can be worth $5K-$15K/month to you across many roles.
In-House HR Departments (Largest Contracts). These are companies with 200+ employees and their own talent acquisition team, who struggle when a hiring surge hits. They already have budget, process and an ATS. The sales cycle is longer (2-3 months), but the contracts are larger ($5K-$10K/month) and they tend to stay.
Outreach Strategy:
- LinkedIn content marketing. Post 3-5 times a week about what you are learning in AI recruiting, share anonymised pipeline numbers, and comment on hiring posts from the companies you want as clients. This builds your name and brings people to you.
- Cold outreach with proof. Email heads of talent or founders with something specific about their hiring: "I noticed you have 14 open engineering roles on your careers page. Our AI pipeline typically delivers 30+ qualified candidates per role per week at a 35% response rate. Want to see a demo with your actual job descriptions?"
- Recruiting community presence. Join Recruiting Brainfood, SourceCon, and LinkedIn recruiting groups. Be useful, answer questions, and get to know people. Recruiters talk to each other, and one referral can snowball.
- Free pilot offers. Offer a 2-week free pilot for one role so the client can see it work. Deliver results (candidates sourced, response rates, interviews booked) and turn it into a paid retainer. The client takes no risk, and your work does the talking.
Who do you already know who has hired 5 people this year and complained about it? That person is your first pilot.
Staying on the right side of the law
Once someone is paying you, your software is making calls about people's careers. Here is what the law expects of you at that point.
AI in hiring is being actively regulated, and getting this right keeps you legal and wins you clients at the same time.
EEOC Guidelines. The Equal Employment Opportunity Commission has made clear that employers are liable for discriminatory outcomes from AI tools, even when the discrimination is unintentional. Your screening prompts must not use stand-ins for protected characteristics. Filtering by "cultural fit" or "prestigious university", for example, can create disparate impact against protected groups.
NYC Local Law 144. If your clients are in New York City, any automated employment decision tool (AEDT) needs an independent bias audit every year. The audit looks at scoring rates and selection rates across race/ethnicity and sex categories. Even if none of your clients are in NYC, building bias auditing into your service shows you are serious and prepares you for similar laws spreading elsewhere.
Illinois AIPA. Illinois requires employers to tell candidates when AI is used to evaluate their applications. Build candidate notification templates into your service.
GDPR and Data Privacy. If you process candidates from the EU or UK, GDPR applies to you. You need a lawful basis for processing personal data (legitimate interest is the usual one for recruiting), you must answer data subject access requests, and you should not keep candidate data forever. Set a 90-day data retention policy.
Practical Compliance Steps:
- Use structured scoring rubrics that map to job-relevant criteria only
- Track pass-through rates by demographic group (when data is available) to spot disparate impact
- Keep a human in the loop for every final hiring decision: your AI screens, humans decide
- Write down your method so you could explain it to a regulator
- Include compliance documentation in what you hand clients. It sets you apart from competitors
Growing past yourself
This business scales well because adding a new client or role costs you very little. You are mostly reusing the same pipeline with different inputs.
Phase 1: Solo Operator (Months 1-6), $3K-$10K/month
Run 3-5 clients yourself. Stick to one vertical so you build expertise and screening prompts you can reuse. You do everything: sales, pipeline building, client management, and reporting. Your job in this phase is to prove it works and collect case studies with real numbers.
Phase 2: Hire a VA and Specialize (Months 6-12), $10K-$25K/month
Hire a virtual assistant ($800-$1,500/month) to handle data quality, scheduling coordination and client reports. That frees you for sales and the high-value pipeline tuning. Add a second vertical so your income has more than one leg.
Phase 3: Build a Team (Year 2), $25K-$50K/month
Hire a pipeline specialist to run the technical workflows ($60K-$80K/year) and a business development rep to drive outbound sales ($50K-$70K base + commission). By now you are managing 10-20 clients and can add premium services: executive recruiting, retained search with AI support, and recruiting analytics consulting.
Phase 4: Productize (Year 2-3), $50K-$100K+/month
Turn your best workflows into a self-service product. Offer a SaaS platform where smaller companies run AI recruiting pipelines themselves, while your team looks after enterprise clients on the services side. You end up with two income streams: recurring SaaS revenue plus high-margin services revenue.
Be honest with yourself before you plan for Phase 3: do you want to manage people, or do you want the solo $3K-$10K and your evenings back? Both are fine answers, and they lead to very different businesses.
What a typical month looks like
Here is a snapshot of a solo operator 4 months in, with 4 clients on the Growth tier ($2,500/month each):
| Metric | Amount | Notes |
|---|
| Revenue | $10,000/month | |
| Tech stack costs | $650/month | Clay, OpenAI, LinkedIn, email tools |
| Time spent per client | 6-8 hours/month | mostly pipeline optimization and client calls |
| Total working hours | 25-35 hours/month | |
| Net profit | $9,000+/month | |
| Effective hourly rate | $250-$360/hour | |
The margins are high because once you have built and tuned a pipeline for a vertical, adding another client in that vertical takes only 2-3 hours of setup. The AI does the heavy lifting.
Each of your client pipelines typically delivers:
- 200-400 candidates sourced per role per month
- 50-100 passing the AI screening threshold
- 30-50 personalized outreach messages sent per role per week
- 25-40% response rate on outreach
- 8-15 interviews scheduled per role per month
Those numbers are 3-5x what a solo human recruiter produces, and your clients see the difference in their calendars.
Four Growth-tier clients at $2,500 each, the $10,000 month on that table, is a snapshot from month four and well above where most people start. If you got near it, the money after costs could get your eldest's orthodontist appointment booked without a family meeting about it. Build toward it one client at a time.
Mistakes to dodge
Pitfall 1: Over-promising placement rates. You control sourcing, screening and outreach. The client's interview process and offer decisions are theirs. Tell clients up front that you deliver qualified candidates to the interview stage, and that a placement is something you win together.
Pitfall 2: Ignoring data quality. Garbage in, garbage out. If Clay returns bad email addresses, your outreach fails. Spend time on data checks and use waterfall enrichment across several providers.
Pitfall 3: Generic screening prompts. One catch-all GPT prompt for screening gives you mediocre results. Build prompts for each role and industry. Your rubric for a senior backend engineer should look nothing like your rubric for a registered nurse.
Pitfall 4: Neglecting the human element. Candidates are people with lives and worries of their own. Keep your outreach warm and respectful, give people a way to opt out, and never misrepresent the job. Your reputation, and your client's employer brand, rides on how candidates feel about you.
Pitfall 5: Not tracking metrics. Measure everything: candidates sourced, screen pass rate, outreach response rate, interviews scheduled, offers made. Those numbers prove your value to clients and show you where to improve.
If a client asked you tomorrow what your response rate was last month, could you answer from a dashboard? If the answer is no, Pitfall 5 is where you start.
Getting Started Today
- Sign up for Clay (free trial), the OpenAI API, and LinkedIn Sales Navigator
- Pick one vertical you know or can learn quickly
- Build your first sourcing-to-screening pipeline for a real job posting
- Run a test batch of 50 candidates through it and measure the results
- Record a Loom demo showing the full workflow and results
- Reach out to 10 companies in your vertical that are actively hiring
- Offer 2 free pilots and deliver outstanding results
- Turn the pilots into paid retainers and start building your book of business
Recruiting is huge and slow-moving, and it is ready for this. Companies badly want hiring that is better, faster and cheaper, and AI gives them all three. If you build this skill now, you are early, and the people who are early tend to own the market once AI recruiting becomes the normal way to hire within the next 2-3 years.
The tools are there, the demand is proven, and the numbers work. Start building your first pipeline today.
2026 Market Snapshot
An independent Recruiting Businesses market report and Service-as-Software report together paint a clear 2026 picture for you. Hiring is slow, expensive and broken (the average cost of one bad hire is $15,000), and AI is the wedge that lets you, working alone, outperform an in-house team. Service-as-Software founders are running fully-AI recruiting pipelines at a 6:1 revenue-to-cost ratio compared with traditional staffing agencies. Vertical hiring services (Uplers in India, TalentQL in Africa, Remote Talent LATAM, Shepherd in Philippines/LATAM) show that the margin lives in geographic and skill-specific niches.
- $15,000: average cost of one bad hire (CareerBuilder data cited by independent market research)
- 6:1 revenue-to-cost ratio for solo Service-as-Software operators replacing agencies
- Vertical hiring services already operate across India, Africa, LATAM, and the Philippines
- Niche job boards (DeafJobWizard, 4-Day Week, WorkingNotWorking, Women Who Code) are proving that targeting one group of candidates works
- Lead-magnet flywheels (resume reviews, salary calculators, hiring guides) are now the cheapest way to source candidates
A bad hire costing $15,000 is the line your clients already know by heart. How many of them could you name who have paid it in the last year?
Key Players to Watch
The mix of established staffing tech, AI-screening startups and specialist consultants is writing the new playbook. These are worth following:
- Eric Nowoslawski: Clay-based recruiting workflows serving dozens of companies
- Assembly Industries: Full team hiring for businesses, productized model
- JobRack: Remote talent from Eastern Europe
- RemotelyTalents: Sourcing from Ukraine, Romania, Poland
- SecureVision: Hands-on recruiter augmentation for in-house teams
- Deel: Hiring infrastructure across 100+ countries
- Leap Room: AI hiring assistant spanning multiple roles
- Honeit: Call and interview intelligence extraction
- Recruit CRM: Workflow automation for staffing agencies
- talent.ai, The Work In Me, Turing, Dedicare, AI-driven placement engines
- Uplers (India), TalentQL (Africa), Remote Talent LATAM, Shepherd (PH/LATAM), Vertical hiring services
- Ceipal, Arya, Manatal, Globus, XOR: AI screening platforms used inside agencies
- ReachExt, Relocate.me, Wellfound: Lead-magnet leaders (resume reviews, salary calculators)
Predictions for 2026-2027
- Q3 2026: Vertical AI recruiting services (engineering, sales, healthcare) commanding 2-3x premium retainers over generalist agencies
- Late 2026: AI screening becomes table-stakes. Clients explicitly ask "what AI tools do you use?" during procurement
- Mid-2027: Performance-based pricing replaces retainers; recruiters bill per qualified hire rather than per month
- 2027: Geographic arbitrage hiring services (LATAM, Africa, SEA) capture significant share of US small-to-mid market hiring
- 2027: AI interview platforms (Honeit, Leap Room) commoditize first-round screening, compressing time-to-shortlist to 48 hours
If clients start paying per qualified hire by mid-2027, would your cash flow survive a slow quarter? It is worth pricing that scenario now, while you still choose your model.
When screening becomes table stakes, the agencies that adopted it early will look normal and everyone else will look slow. You want to be the person who taught your clients what good looks like before the big platforms arrive with their own version. That reputation is hard to build once the market is crowded, and the crowd is already forming.
Emerging Opportunities
Vertical AI recruiting agency. Pick one role (back-end engineers, B2B SDRs, dental hygienists) and build your whole screening, sourcing and outreach playbook around it. Generalist recruiting is turning into a commodity, and depth is what protects you. The independent market research report's prediction that targeting specific skill sets beats broad approaches is the plain 2026 strategy.
Niche job board + recruiting hybrid. Run a job board for a community nobody serves well (women in DevOps, four-day-week roles, remote-first APAC), then add paid placement on top. You can build supply early by gathering listings from public sources before you charge anyone.
AI-screening-as-a-service. Sell access to a Clay + Honeit + GPT pipeline that screens, ranks and shortlists candidates for SMBs that have no recruiting team of their own. Charge per role rather than per month.
Lead-magnet driven recruiting. Build free salary calculators, resume reviewers or hiring guides (like Wellfound, ReachExt) that bring you candidates and clients at the same time. In 2026's regulated environment that costs you less per customer than cold outreach.
Owning one narrow vertical, say dental hygienists for every practice in your region, is how your retainers get renewed for years. Repeat work like that, if it leaves you a steady surplus, lets you tell your parents the winter heating bill is handled, paid by one clinic group's retainer. Pick the vertical you understand before you build the next pipeline.
Common Objections & Counterarguments
"AI screening discriminates and creates legal exposure." That risk is real if you ignore it. You answer it with bias-aware tools (Manatal, Ceipal) that keep audit logs, screening criteria written down, and humans in the loop on final decisions. The independent market research report treats employer branding and accountability as competitive advantages in their own right, well beyond a compliance cost.
"Recruiting is relationship-driven. Clients will not trust AI." Clients trust outcomes. Service-as-Software operators win deals because they deliver shortlisted, high-fit candidates faster and cheaper. The AI is the engine under the bonnet, and your client never has to look at it. Lead with case studies (placements made, time-to-hire, retention rates) and keep the tooling talk for later.
"The market is saturated with recruiting agencies." The generalist tier is crowded. The vertical tier is wide open. The independent market research report names Japan Dev (developer-focused, Japan), DeafJobWizard, and WorkingNotWorking as proof that new niches keep appearing. In 2026, specialising beats spreading yourself thin.
"Economic downturns will kill recruiting fees." That holds for retained search at premium rates. You protect yourself with performance-based pricing (pay per hire), recession-proof verticals (healthcare, AI engineers, compliance), and job-board revenue stacked on top of placement fees, so one income source failing does not take you down with it.
Europe now treats recruiting AI as high-risk
The compliance section above covers the American picture. The European one is bigger, and it is the largest single change to this business since this guide was first written.
The EU AI Act sorts AI systems by risk, and employment and worker management sits in Annex III, the high-risk category. That covers systems used to recruit or select people, to filter applications, and to evaluate candidates. Nobody disputes the classification, and it has been in the text since the Regulation was adopted.
What has moved is the date the obligations start. The original schedule applied the Annex III high-risk rules from 2 August 2026. Under the Digital Omnibus package the Commission proposed tying those rules to the availability of supporting tools and harmonised standards, and the widely reported outcome is a deferral of standalone Annex III obligations to 2 December 2027, with AI built into regulated products moving later still. The Commission's own service desk timeline has carried both versions, with the August date marked by a footnote pointing at the Omnibus.
So treat the date as unsettled and the direction as settled. No credible voice is arguing that recruitment AI will stop being high-risk. Check the deadline that actually applies with a qualified adviser before you promise a client anything about it, and build as though the obligations already apply. A service designed around them sells under either date. A service that ignores them becomes unsellable on a day someone else picks.
What high-risk status asks of you
These obligations are heavier than the American bias-audit rules, and that weight is why they matter to your sales as much as to your lawyer. High-risk systems come with requirements for risk management, data governance and quality, technical documentation, record keeping, transparency to deployers, human oversight, and accuracy and robustness. There are registration and conformity steps before a system goes to market.
For a small agency like yours, it means improvised prompt chains can no longer be sold into the EU. When a client's legal team asks for your technical documentation and your human-oversight design, that is an ordinary request under this regime, and it is the rules doing exactly what they were written to do.
Provider or deployer: which one are you?
The Act separates the provider, who develops an AI system and puts it on the market under their own name, from the deployer, who uses one under their own authority. The two carry different obligations, and which one you are depends on how you have set up your service, whatever you choose to call yourself.
If you run an off-the-shelf screening tool on a client's behalf, you lean toward deployer. If you assemble models, prompts and scoring logic into something you brand and sell as your own screening system, you lean toward provider, which is the heavier end. Putting your name on a wrapper over somebody else's model does not obviously make you a deployer, and assuming it does is the most expensive misreading you could make here.
Three things follow for how you write your contracts.
Settle which role each party holds, in writing, before your first engagement. A contract that says nothing leaves the question to be answered later by whoever has the better lawyer, and that answer arrives after something has already gone wrong.
Do not accept unlimited liability for outcomes you do not control. You control the screening logic. What the client does with a shortlist is outside your hands, and a rejected candidate's complaint attaches to the hiring decision as well as to the tool.
Keep the documentation as you go. Rebuilding a risk-management file and a data-governance record after a client asks for them costs many times more than keeping them up. It is the same discipline as bookkeeping, and it fails for the same reason: nothing forces you to do it until something does.
If an EU client's lawyer emailed you tomorrow asking whether you are a provider or a deployer, what would you write back? Draft that answer before you sign anyone in Europe.
Why compliance is your moat
Your first instinct will be to treat all of this as cost. For a small operator like you it works closer to the other way round, and this is the strategic point of the whole section.
Recruiting is a crowded service business where people usually compete on price and speed. Both get competed away, and neither survives a client's procurement process. Regulatory competence stays yours, because most small agencies will never do the work, and the clients who need it most are the ones with the biggest budgets.
The practical version is plain and cheap to start. Structured scoring rubrics tied to job-relevant criteria. Pass-through rates tracked by group where the data exists. A documented human-in-the-loop step where a person makes the decision and can say why. Retention limits you actually enforce. A written method someone non-technical could follow.
That package takes you a few days to build, and it turns a conversation about hourly rates into a conversation about risk. When an enterprise client compares two agencies and only one can hand over documentation, price stops being the deciding factor.
It also decides which clients you can serve at all. Regulated employers, public sector bodies, anyone hiring across the EU, and any employer big enough to have an employment lawyer will ask. Those are the clients paying the retainers described earlier in this guide. The ones who never ask are the ones paying the least.
Here is the cost of waiting. Every month you hold off, another builder writes the bias audit, the retention policy and the plain English method, and walks into the enterprise meeting you wanted. Rules like these reward the person who prepared before anyone asked. Draft your one page method tonight, while it is still a head start.
Who should skip this
If you are unwilling to engage with the regulation, choose a different AI service. Recruitment carries real weight as an automation target. It sits in the high-risk tier in Europe, it falls under bias-audit rules in New York City and notification duties in Illinois, and it produces decisions that affect people's livelihoods, decisions they have standing to challenge. A guide that sells this as a simple prompt-engineering business is describing a different business from the one you would actually be running.
If you hope to sell on speed alone, the tools themselves will compete with you. Applicant tracking systems and sourcing platforms are absorbing the same automation, and a service whose only claim is faster screening ends up competing against a feature.
If you have no access to hiring managers, start there and leave the technology for later. Trust with the people who make hiring decisions is the real constraint here, and no amount of tooling replaces it. The operators doing well mostly came from recruiting and added AI, far more often than the reverse.
If you cannot explain a rejection, do not automate rejections. When your system cannot give a job-relevant reason a specific candidate was filtered out, you have built something you cannot defend to a client, a candidate or a regulator. Working most of the time falls short of the standard you will be held to.
The number that protects you
"Track pass-through rates by demographic group" appears in every compliance checklist, including the one above, and almost nobody tells you what threshold turns a rate into a problem. There is a specific and long-standing one.
Under the US Uniform Guidelines on Employee Selection Procedures, a selection rate for any group that is less than four-fifths, or 80%, of the rate for the group with the highest rate is generally treated by the enforcement agencies as evidence of adverse impact. It is a rule of thumb rather than a legal definition, and it is the number that starts conversations.
Here it is worked through. If your screen passes 50% of one group and 35% of another, the ratio is 0.70. That is below four-fifths. At that point you need a job-relatedness justification for the criteria producing it. Quietly adjusting the prompt and moving on is the wrong response.
Three things make this usable in your day-to-day work.
Measure at every stage, right through to the final one. Adverse impact builds up. A screen that looks fine at sourcing, fine at keyword filtering and fine at scoring can still produce a combined ratio well under the threshold, and looking only at the end hides which stage caused it.
Small samples produce meaningless ratios. Twelve candidates tell you nothing about selection rates, and a ratio worked out on them will swing wildly from month to month. Pool across engagements and time before you draw conclusions, and say so when you report.
Know what you may lawfully collect, and where. Demographic data for adverse-impact analysis is handled differently from data used in the decision itself, and the rules change between jurisdictions. In the EU it is special category data with its own limits. Collecting it carelessly to run a compliance check gives you a second problem on top of the first.
Why go to this trouble? Because it turns a defence you would have to make up on the spot into one you can hand over on request. If you can show your ratios by stage, over a meaningful sample, against job-related criteria you wrote down in advance, you are in a completely different position from an agency that simply insists the model is objective.