Most recruitment AI doesn’t reduce work, it moves it. Someone still has to configure the tool, manage the exceptions, and explain to a candidate why a form rejected them. The companies that actually reduce triage time have one thing in common: they defined their screening criteria precisely before they touched any automation. The ones that didn’t are still manually triaging inboxes, just with a dashboard in the way.
What AI Recruitment Screening Actually Automates
“AI in hiring” gets used to describe everything from keyword-matching ATS filters to full LLM-powered candidate scoring. These are not the same thing. The useful distinction is: which steps in your current process are purely mechanical, and which require judgment?
The Three Stages Where AI Reduces Manual Work
Intake structuring is the first lever. Most SMBs receive applications as unstructured emails, PDF attachments, or freeform web forms. Before AI can score anything, you need structured inputs. An AI-connected intake form, built with defined fields, required sections, and validation, converts a messy pile into a consistent dataset. This alone cuts 20–30 minutes of per-applicant prep work, when the form is well-designed and applicants fill it out properly. Freeform fallback fields or optional sections reintroduce the mess.
Initial scoring against defined criteria is where the time savings compound. Once applications are structured, an LLM can evaluate each one against a scoring rubric you define: years of relevant experience, specific skills, location eligibility, salary range fit. A well-prompted Claude API call returns a score, a brief rationale, and a pass/fail flag, in under five seconds per candidate. For a role that draws 80 applications, that’s a shortlist of 12 delivered to the hiring manager’s inbox before they’ve had their morning coffee.
Automated routing and communication handles what comes next. Candidates who clear the initial score get an acknowledgment and a scheduling link. Those who don’t pass receive a rejection, sent within 24 hours, not three weeks later. Both happen without a human touching the queue.
What AI Cannot Reliably Do in Hiring
AI cannot assess cultural fit, motivation, or whether someone’s career trajectory makes sense in context. It cannot catch the candidate who undersells on paper but interviews brilliantly. It will not flag that a resume looks templated because the person was coached, or that an answer to a screening question is copy-pasted from a job description.
More practically: any AI screening step that makes a consequential hiring decision, not filtering, but deciding, carries legal exposure. The EU AI Act classifies AI systems used in employment decisions as high-risk. US regulations are inconsistent by state, but New York City’s Local Law 144 requires bias audits for automated employment decision tools. The safe design is AI-as-triage, human-as-decision-maker. Every shortlist AI produces should be treated as a recommendation, not a verdict.
The Two Paths: SaaS Platform vs. Custom AI Integration
The decision isn’t “should we automate?” It’s “how should we automate, given our actual hiring volume and budget?”
When an Off-the-Shelf ATS Makes Sense
If you’re hiring more than 100 people a year, running multiple concurrent roles, and managing a team of recruiters, a platform like Greenhouse or Lever is defensible. You’re paying for the workflow management, the integrations, the compliance tooling, and the support. At that scale, $800–$1,500 a month is a reasonable cost-per-hire reduction.
If you need job board integrations across 10+ platforms, built-in video interviewing, and pre-built compliance frameworks for enterprise clients, buy the platform. Building those from scratch costs more than buying them.
When a Custom AI Screening Workflow Beats the Platform
A 20-person company hiring eight people this year does not need Workday. They need a form that collects structured data, a scoring layer that applies their specific criteria, and a way to deliver a ranked shortlist to the person doing the hiring.
Platforms like Paradox (Olivia chatbot) run $2,000–$5,000 per month at SMB pricing. Greenhouse’s Starter tier starts at $6,000 per year. A custom-built workflow, a structured intake form, an LLM scoring integration, and automated email routing, costs a fraction of that as a one-time build. The business owns the logic. There’s no per-seat pricing. Criteria can be updated without filing a support ticket.
The honest case for custom is straightforward: if your hiring volume is predictable, your criteria are definable, and you’re not managing a dedicated recruiting team, a lightweight custom workflow delivers 80% of the time savings at 15% of the platform cost.
How to Build a Custom AI Recruitment Screening Workflow
This is not plug-and-play. A custom workflow requires clear inputs before any code gets written. The most common failure mode is companies who automate a vague process and end up with automated vagueness.
Step 1, Define Your Scoring Criteria Before Touching Any Tool
The AI can only score what you’ve defined. That means writing explicit criteria: what qualifications are required vs. preferred, which answers to screening questions are disqualifying, what salary range is real (not the posted range, the actual range). This takes 2–4 hours to do properly. Most companies skip it and then blame the tool when the shortlist is wrong.
Good criteria are binary or scored on a narrow scale. “Has 3+ years of direct sales experience” is scorable. “Is a strong communicator” is not, it’s a judgment, not a filter.
Step 2, Build a Structured Intake Form
Replace your freeform application with a form that enforces structure. Required fields. Multiple choice where possible. Short-answer questions with defined character limits. Upload fields for specific documents, not a generic “attach anything” field.
The goal is inputs an LLM can process consistently. If two candidates answer the same question in completely different formats, the scoring pass becomes unreliable. Structure reduces that variance before it reaches the AI. It won’t fully disappear, candidates find ways around any form, but it makes the scoring pass substantially more consistent.
Step 3, Connect an LLM Scoring Pass
This is the core of the automation. A prompt is constructed from the candidate’s structured intake response and your defined criteria. The LLM returns a structured output: a score out of 10, a one-paragraph rationale, a pass/fail flag, and any disqualifying factors it detected.
Using the Claude API, this call costs fractions of a cent per candidate. At 100 applicants per month, you’re looking at under $2 in API costs, not per month, total. The prompt needs testing across a sample of real applications before you go live. A well-built system includes a review log so you can audit the AI’s rationale on any candidate.
Custom AI integrations of this type are typically delivered as lightweight web apps or WordPress admin tools, depending on how the hiring team works. See how we scope and build this at designodin.com/ai.
Step 4, Automate Shortlist Delivery and Rejection Handling
The output of the scoring pass feeds two automated paths. Candidates above your threshold get a confirmation email and a calendar link for the next step, phone screen, video interview, or hiring manager review. Candidates below threshold get a rejection email, sent automatically within 24 hours of application.
Both are templated. Neither requires a human to initiate. The hiring manager receives a ranked shortlist, not a pile of applications, with the AI’s rationale attached for each candidate they’re being asked to consider.
Real Costs and Realistic Timelines
What a Custom Build Costs vs. 12 Months of SaaS Licensing
A custom AI screening workflow, intake form, LLM scoring integration, automated routing, and email templates, typically takes 3–6 weeks to scope, build, and test. One-time build cost varies by complexity, but for an SMB use case, it sits well below a year of platform licensing from the enterprise ATS providers.
Greenhouse Starter: $6,000–$8,000 per year. Paradox: $24,000–$60,000 per year. A custom build at that scope: a fraction of the first-year cost, with no recurring licensing. The break-even is typically before month six.
What Ongoing Maintenance Actually Involves
The biggest ongoing task is prompt tuning. As you hire for new roles, your scoring criteria change. Updating the prompt and testing it against a sample of real applications takes a few hours per new role type. This is not software maintenance in the traditional sense, it’s editing a document and running a validation test.
Infrastructure costs are minimal. API usage scales with volume. There’s no server to manage if the workflow runs on a serverless architecture. Annual maintenance is closer to a few hours than a retainer.
Frequently Asked Questions
Is AI candidate screening legal in the US and EU?
AI-assisted screening is legal in most jurisdictions when it’s used as a filtering tool, not an autonomous hiring decision-maker. The EU AI Act classifies employment-related AI as high-risk, requiring transparency, human oversight, and bias testing. In the US, New York City’s Local Law 144 mandates independent bias audits for automated employment decision tools. The safest design keeps a human reviewing and approving every shortlist the AI generates, AI triage, human decision.
How many applicants do you need before automation makes sense?
The threshold is lower than most people assume. If a role consistently draws more than 20 applications and you’re spending more than two hours on initial triage, a structured screening workflow starts paying for itself. The ROI is most obvious at 50–200 applicants per month, high enough to create real friction, low enough that enterprise ATS pricing is hard to justify.
Can I use ChatGPT or Claude to screen resumes without a custom build?
Yes, as a manual process, paste a resume and a scoring rubric into a chat interface and ask for an evaluation. This works and many hiring managers already do it informally. It doesn’t scale past about 10 candidates because it’s still manual. A custom integration automates the data flow: applications feed in, scored outputs feed out, without anyone opening a chat window.
What happens if the AI misses a qualified candidate?
This is the core risk and the reason human review of the shortlist matters. A well-designed system includes a mechanism to review borderline candidates, those who scored just below the threshold, before rejections go out. No screening system, human or AI, has a perfect recall rate. The practical mitigation is to set your threshold conservatively and have the hiring manager scan the near-miss list before the rejection batch runs.
How is a custom AI screening workflow different from Workday or Greenhouse?
Platform ATS tools bundle workflow management, job board integrations, offer management, and onboarding into a single product. A custom AI screening workflow does one thing: it automates the initial triage step. It doesn’t manage the rest of your hiring pipeline unless you build that too. The trade-off is scope for cost and control, you own a narrower tool that does exactly what you need, at a fraction of the price, with no vendor dependency on your scoring logic.
Do I need technical staff to run this after it’s built?
No. A well-built custom workflow is designed for a non-technical hiring manager to operate. Prompt updates happen in an editable document or simple admin interface, not in code. The system surfaces errors and flagged cases through a dashboard or email digest. Ongoing operation requires the same skill level as maintaining a spreadsheet.
If your current hiring process has a human manually reading 60 applications to find 8 worth calling, that’s the problem a custom AI screening workflow solves. If you want to talk through what this looks like for your operation, start a conversation. We’ll be direct about whether a custom build makes sense or whether an off-the-shelf tool is the better fit.