By TL;DR: Modern IT staffing vetting funnels run four stages: application screen, technical assessment, AI tooling fluency, and behavioral. Quality providers report acceptance rates under 2% per SIA‘s 2025 IT Staffing Report, against 12-30% at legacy agencies. From 10,000 applications, a quality funnel produces roughly 150 placement-ready engineers, a legacy funnel produces 2,500 with weaker fit signals. Gartner‘s 2026 IT Talent Quality Index links acceptance rate inversely with first-year attrition: every 5 points lower acceptance correlates with 18% lower 12-month attrition.
The vetting funnel is the single most important machine inside an IT staffing provider. It defines whether a buyer gets a senior engineer or a relabeled mid-level resume, whether attrition lands under 8% or above 25%, and whether 24-hour matching is a real claim or a marketing line. This article breaks down how a modern vetting funnel actually works, the quantitative benchmarks that separate quality providers from legacy agencies, and how procurement and engineering leaders should read funnel reports during vendor selection. For a complementary view on vendor selection, see our checklist on how to evaluate IT staffing companies.
Benchmarks in this article are sourced from Staffing Industry Analysts (SIA) 2025 IT Staffing Report, Gartner 2026 IT Talent Quality Index, Forrester‘s 2026 Vendor Trust Outlook, Robert Half‘s 2026 Hiring Quality Report, and the US Bureau of Labor Statistics tech occupations data. Funnel numbers represent rolling 12-month averages, normalized to a 10,000-application cohort.
Why the Funnel Shape Matters More Than Headcount
Two staffing providers can both claim a “pool of 27,000 vetted engineers” and deliver wildly different outcomes. The difference lives in funnel shape. Per SIA, the strongest predictor of placement quality is not the size of the top of the funnel but the steepness of the drop at each stage. A funnel that drops 76% at application screen, 70% at technical, 67% at AI tooling, and 25% at behavioral produces a net acceptance rate of roughly 1.5%. A funnel with shallow drops produces high acceptance, large surface-level talent pools, and weaker placement signals.
Forrester’s 2026 Vendor Trust Outlook puts the issue bluntly: a provider claiming “we hand-pick the top 1%” is a 99% rejection rate claim. If the same provider also says yes to 25% of applicants, the math is internally inconsistent. Procurement teams should request stage-by-stage drop rates as part of any vendor evaluation and treat refusal to share them as a quality signal.
Funnel shape also drives first-year attrition. Gartner’s 2026 Talent Quality Index regression links each 5-point reduction in acceptance rate with an 18% reduction in 12-month attrition. The mechanism is intuitive: tighter funnels filter out engineers who would later disengage, hide gaps, or churn for a 10% raise. Buyers experience this as a quieter, more stable engagement.
Stage One: The Application Screen
The first filter is the cheapest and the largest. Quality providers reject 70-80% of applications at this stage; legacy agencies reject 30-50%. The screen does three things in parallel.
Resume parsing for role match. Automated parsing extracts skills, years of experience, last three roles, employment gaps, and stack depth. Quality providers do not rely on parsed tags alone; they cross-check with the application narrative. Robert Half’s 2026 report finds 41% of mid-senior tech resumes contain at least one inflated claim (a tool listed but never shipped, a stack mentioned but never owned). Resume parsing alone catches roughly half of those; a paired narrative review catches the rest.
English fluency baseline. Asynchronous written communication is non-negotiable for distributed engineering. The application screen typically includes one or two short-answer questions scored for clarity, structure, and grammar by humans, not algorithms. Engineers who write in fragments or rely on AI-generated paragraphs get filtered here. SIA notes that 22% of legacy agencies skip this step entirely and rely on a verbal interview later, which is a less reliable predictor of async fit.
Background and right-to-work. EOR-managed providers verify residency, work authorization, and tax status before any technical investment. Mistakes here are expensive in EOR engagements because labor-law compliance audits flag back to the contracting entity.
The application screen is also the stage where ghost candidates are caught. Per Gartner, 6-9% of senior tech applications in 2025-2026 are submitted by recruiters or career coaches on behalf of weaker candidates. Quality providers run identity-verification at this stage to suppress the practice before it consumes downstream interview hours.
Stage Two: The Technical Assessment
Engineers who pass the application screen enter the technical assessment, typically a 90-120 minute live session covering coding, system design, and language depth. Quality providers reject another 70% of the remaining pool here.
Modern technical assessments have shifted away from leetcode-style algorithmic puzzles toward applied scenarios. A typical 2026 senior backend assessment runs: a 30-minute coding session implementing a feature against a partially-built repo, a 30-minute system design covering scaling, failure modes, and tradeoffs, and a 30-minute language-depth conversation probing the specific frameworks listed on the resume. The format is harder to cheat on because the engineer is reasoning out loud against a live repository, not solving a static problem.
Common failure modes at this stage: surface-level framework knowledge (an engineer claims senior React but cannot describe useEffect cleanup timing), stack inflation (an engineer claims Postgres expertise but defaults to a generic SQL mental model), and ghost-coded portfolios (the engineer cannot reproduce work they claimed to lead). The live-observed format catches all three.
The technical assessment is also where ranking by absolute skill, not relative skill, becomes critical. Legacy agencies often pass candidates whose performance is “good enough relative to other applicants this week.” Quality providers maintain a fixed bar across cohorts. SIA’s 2025 report notes that fixed-bar providers see 32% lower replacement requests in the first 90 days than relative-bar providers, because hiring quality does not drift with weekly application volume.
Stage Three: AI Tooling Fluency
The newest stage in the modern vetting funnel, added by quality providers between 2023 and 2025. Legacy agencies usually do not test it at all. The stage filters for engineers who can ship 30-50% faster using AI-augmented workflows.
What gets tested: live use of Cursor or Claude Code on a small feature build, prompt design for a real engineering task, eval writing for an LLM-backed feature, and basic agent-building competence with LangChain or LangGraph. The format is a 60-90 minute live screen-share session where the engineer narrates their AI workflow while building a small task.
Per Gartner’s 2026 AI Skills Premium Index, engineers who pass an applied AI tooling test ship 35-50% more code per sprint with similar defect rates as engineers who do not. The premium is real and measurable. The challenge for buyers is that AI tooling fluency cannot be self-reported reliably. Roughly 80% of senior engineers in 2026 will list “Cursor” or “Claude Code” on a resume; in live observation, perhaps 25% demonstrate genuine fluency including agent debugging, prompt iteration, and eval discipline.
Common failure modes here: surface-level prompt fluency (the engineer can ask GPT to write a function but cannot iterate on prompts when the first answer is wrong), no eval discipline (the engineer cannot describe how to test an LLM-backed feature), and inability to debug AI-generated code (the engineer accepts plausible-looking code without verification). All three are catchable in a 60-minute live session and impossible to detect from a resume.
Forrester’s 2026 outlook recommends that any provider claiming “AI-native engineers” publish their AI tooling assessment rubric. Without published criteria, the claim is unverifiable. We cover the test methodology in detail in our companion article on AI-native skills assessment.
Stage Four: The Behavioral Assessment
The final stage. Quality providers reject another 20-30% of the surviving pool here. The behavioral assessment tests three dimensions that do not show up in technical screens.
Conflict and ambiguity. Structured behavioral questions probe how the engineer handles disagreement with a tech lead, an ambiguous spec, or a missed deadline they caused. Strong candidates demonstrate ownership, concrete examples, and reflection. Weak candidates default to abstract platitudes or evasive answers.
Async communication. Distributed engineering relies on written async over time-zone boundaries. The behavioral assessment includes one to three written prompts (a code review summary, a status update for a stalled task, a postmortem of a real incident). The signal here is clarity, structure, and completeness, not grammar.
Hidden commitments. The behavioral stage is also where commitment conflicts surface: undisclosed second full-time roles, family obligations colliding with the requested time zone, or visa applications that would interrupt continuity. Quality providers ask these questions directly. Per SIA, 8% of senior engineering candidates carry an undisclosed conflict at this stage; catching it before placement saves the buyer a replacement cycle.
The Funnel Math: From Applications to Placements
The aggregate funnel math is what turns 10,000 applications into a small placement-ready cohort. Per SIA, the typical quality funnel produces 1.5% net acceptance: 10,000 applications yield 150 engineers ready to place across all roles. Legacy agencies running 25% net acceptance yield 2,500 engineers, but with weaker per-role match signals and roughly 2-3x higher first-year attrition per Gartner.
Per-role math is what matters at the engagement level. A buyer searching for a Senior LLM Engineer in offshore Asia draws from the relevant slice of the placement pool. From the 150-engineer quality cohort, perhaps 12-15 are LLM specialists; from the 2,500-engineer legacy cohort, perhaps 100 carry the LLM tag but only 8-12 would pass a quality-funnel AI tooling test. The headline pool size is not the right comparison; the slice that matches the role is.
Time-to-shortlist math is downstream of funnel shape. Quality providers running an always-on funnel can return 3-5 shortlist candidates within 24-48 hours because the pool is already pre-vetted at depth. Legacy providers pulling from a wider, shallower pool typically need 5-10 business days to verify which “vetted” candidates would actually pass a buyer’s bar.
How to Read a Provider Funnel Report
Procurement should request a funnel report as part of any IT staffing RFP. A useful report includes five numbers per stage: applications entering the stage, drop rate, time-in-stage, top failure modes, and acceptance rate to next stage. Five red flags to watch for:
Acceptance rate above 10%. Almost always a signal that one or more stages are skipped or rubber-stamped. Ask for the missing stage data.
Application screen rejection under 50%. Suggests the top of the funnel is over-filtered before applications arrive (which is fine if disclosed) or that resume parsing is being relied on without narrative review.
Missing AI tooling stage. A 2026 funnel without a dedicated AI tooling assessment is not a 2026 funnel. Either the provider has not adapted or the stage is rolled into the technical assessment without distinct scoring.
No behavioral stage. Functional skill without behavioral fit is the single biggest driver of 90-day attrition per Forrester. Skipping it is a cost-cutting decision the buyer will pay for downstream.
No 12-month attrition follow-through. A funnel that does not track placed engineer retention has no feedback loop. The best quality providers track 90-day, 6-month, and 12-month attrition and feed it back into the funnel bar.
For deeper questions to ask in vendor evaluation, see our companion checklist on the 15 questions to ask IT staffing providers.
Failure Modes That Survive the Funnel
No funnel is perfect. Three failure modes survive even tight quality funnels and need engagement-layer monitoring.
Skill atrophy. An engineer vetted on a stack two years ago may have drifted. Quality providers re-vet engineers on a 12-month cycle. Buyers should ask about re-vetting cadence.
Quiet quitting at the 6-month mark. Behavioral assessment catches obvious red flags but cannot predict mid-engagement engagement decay. Quality providers run 6-week pulse checks and quarterly engagement reviews to catch this early.
Role mismatch. An engineer technically qualified for the role may be a poor fit for the specific team culture or product stage. Funnels do not see this; engagement managers do. Quality providers re-match without a replacement fee if mismatch surfaces in the first 30-60 days.
Why Quality Funnels Cost More to Run
SIA estimates that running a quality vetting funnel costs $1,200-$2,400 per placed engineer in vetting overhead. Legacy funnel costs run $300-$600. The 3-4x cost gap is real and is recovered in two ways at quality providers: longer engagement length (24-36 months average vs 8-14 months at legacy), and lower replacement reserve (engineers placed through a tight funnel get replaced less often, freeing margin).
Per Forrester, the structural economics favor quality funnels when engagement length exceeds 12 months. For shorter engagements, legacy funnel pricing is competitive but quality risk is higher. The break-even crosses around 12-15 months of engagement.
For buyers, the implication is straightforward: if you expect a 12+ month engagement, the higher vetting cost embedded in a quality provider’s rate card pays back via lower replacement cycles and higher engineer continuity. Both translate to lower total cost of engineering capacity over the engagement lifetime.
Start an Engagement With a Tight-Funnel Provider
Second Talent operates a four-stage vetting funnel with a net acceptance rate of roughly 1.5% across senior engineering roles in Asia. We test on Cursor and Claude Code at the AI tooling stage, run live-observed system design at the technical stage, and re-vet active engineers on a 12-month cycle. Our funnel report is published on request to any RFP participant, with stage-by-stage drop rates, time-in-stage, and top failure modes documented.
Common starting points:
- How to evaluate IT staffing companies
- 15 questions to ask IT staffing providers
- Developer rate card by country
- AI Staffing & Augmentation Services overview
Matching in 24 hours. $0 upfront. Pay only when you make a hire.
