The Talent-Flow Map: Ranking Where a Company's People Go and Come From

This guide builds a directional talent-flow map for one target company: a ranked list of the companies its people leave for and arrive from, with a defensible inflow-to-outflow ratio per counterpart. It is written for strategy and research teams, talent-intelligence analysts, and operators sizing a market who need to know which competitor is quietly winning talent before headcount or valuation shows it. Follow it start to finish and you will have a leadership-ready map you can defend in a review.

Most talent-research guides read one departure signal, estimate an attrition rate, or tier a list of named rivals. None of those construct the directed company-to-company graph itself, and none compute a per-counterpart net-flow ratio that surfaces a rising magnet before it is obvious. That ratio is the whole point of this document. The rest is the field-level method that makes it hold up.

What a talent-flow map is, and what it proves

A talent-flow map is a directed, weighted graph of company-to-company moves centred on one target: for each counterpart company it shows how many people moved in, how many moved out, and which direction dominates. It proves where a company's people go and come from, ranked, in a way a turnover percentage cannot.

The direction of each edge is defined by transition counts, not by opinion. The edge runs from company A to company B when more employees moved A to B than the reverse. The USPTO's talent-graph definition puts a number on that dominance: if nine people left A for B and one moved the other way, the edge from A to B is weighted 90 percent, meaning 90 percent of movers between the two companies are going one way. That directional weight is the unit of analysis. Everything downstream is aggregation and cleanup on top of it.

The reason this beats every filing-based proxy is that the filings do not contain direction. A WARN notice proves a dated, sized involuntary separation but never says where people went. A DEF 14A turnover rate proves a company-stated magnitude but names no counterpart. Only the person-level graph carries the arrow.

Why the ratio leads and headcount lags

The net-flow ratio is a leading indicator; headcount and valuation are lagging ones. A rising magnet flips the directional weight the moment its inflow from your target exceeds its outflow to your target, and that flip happens while both companies' overall totals still look stable.

That is the single most useful property of the map. Consider a counterpart that historically fed your target more people than it took. As it starts winning, the arrow reverses before either company's headline headcount visibly changes, because a handful of net moves does not move a payroll of thousands. By the time the rival's headcount or valuation reflects the shift, the shift is old news. The ratio caught it first.

This is also why turnover disclosures cannot stand in. Humana disclosed 14.4 percent voluntary turnover, up from 13.4 percent the prior year; Leslie's disclosed 18 percent corporate and 23 percent non-corporate turnover. Both are real magnitudes and both are useless for direction. The SEC's own committee has documented that these disclosures are inconsistent and non-comparable, with companies reporting full-time and part-time, by business unit, domestic-only, or including international. The per-counterpart ratio is genuinely information the filings do not hold.

The fields that date a move, and the dedup rules you must declare

A move is dated off the position entity on a public profile: the company, the title, and the start and end month-year of each role. The direction of the company-to-company edge follows from the ordering of those dates across people. There is no published standard for resolving messy tenures, so you must state your convention in writing before you aggregate.

The absence of a public standard is not a licence to improvise silently. It means the analyst owns the convention and must record it so the map is reproducible. Use these four rules and name them in your scope note:

• One transition per person per ordered company pair. A person who bounces A to B to A to B counts once for each ordered pair, not four times.
• Collapse consecutive same-employer tenures into one tenure. Two back-to-back positions at the same company are one job, not a move.
• Drop overlaps below a set threshold. Where a later role's start precedes the earlier role's end by less than your chosen threshold, treat it as a concurrent role, not a departure.
• Map aliases and subsidiaries to one canonical node. Outflow split across three legal names hides the real magnet.

Each rule exists to kill a specific false positive, and the failure-modes section below names them. The threshold you pick for overlaps is a judgement call; a month or two is reasonable, but the number matters less than writing it down.

The ratio math and the sample gate

Per counterpart company C, inflow is the count of people who moved from C into the target inside the window, and outflow is the count who moved from the target to C. The net-flow ratio is inflow divided by outflow. Above 1 the target is draining C; below 1, C is draining the target.

The trap is small counts. A 3-to-0 flow computes as an infinite magnet and proves nothing. No talent-specific minimum-sample rule is published, so borrow the standard proportion rule of thumb: a proportion's confidence interval is trustworthy only when the number of successes and the number of failures both exceed 5. Applied here, treat a counterpart's ratio as low-confidence until both its inflow count and its outflow count clear roughly 5. Gate first, rank second.

Prefer a rolling 12 to 36 month window. Shorter than 12 months and even large counterparts fall below the gate; longer than 36 and you fold in stale moves that no longer describe the current market. The window is also where geography bites. In Refolk's index the US ML Engineer pool is 7.3x the UK pool, so a UK-headquartered target hits the sub-5 gate on most counterparts far sooner. Same method, wider window abroad.

The procedure, start to finish

Here is the end-to-end method. No canonical procedure is published for this job; vendor pages describe the output but not the field-level steps. This is the constructed version, ordered so that dedup and sample-gating happen before any charting. Realistic budget is one to three working days for a single target with a mature profile source.

Note the ordering disagreement worth flagging: vendor write-ups treat the Sankey visualisation as an early framing step. The statistically defensible order puts dedup and sample-gating before any chart, because a beautiful diagram built on double-counted moves and tiny samples is worse than no diagram. Chart last.

Once you have a clean scope and want the raw transitions without scraping profiles by hand, this is where a plain-language index removes the friction. Refolk returns the dated moves directly, so steps two and three start from structured rows rather than a pile of pages.

Event-stripping: the magnet versus mechanical re-tag test

A genuine magnet shows sustained, individually-timed inflow spread across many months. A reorg, acquisition, or acqui-hire re-tags a large share of people on or near a single date. The test is to cluster each counterpart's moves by month and inspect the shape.

If a large share of one counterpart's inflow lands in a single month and coincides with a known acquisition or a WARN filing, exclude or annotate it. Do not let it count as organic pull. Acqui-hires do exactly this at scale: they acquire startups primarily for human capital, and compared with other M&A the deals are smaller and faster, so the re-tag arrives as one clean spike. Microsoft's 2024 deal, in which it hired much of Inflection AI's team including co-founders Mustafa Suleyman and Karen Simonyan, is a documented single-date re-tag that drew FTC and CMA scrutiny. On a flow map that reads as Inflection suddenly draining to Microsoft, which is true only in a bookkeeping sense.

This step is not optional. Between 2020 and 2025 there were 5,700 AI/ML acquisitions with only 21 percent disclosing deal value, so single-date re-tags are common enough that skipping event-stripping will fabricate magnets on any tech-heavy map.

Corroborating the direction with public sources

For the top counterparts, back the profile-derived flow with at least one independent source, or mark it profiles-only. No public source names person-level flows, so corroboration confirms context and events rather than the arrow itself. Use each source for what it actually proves and nothing more.

The table below is the working reference for what to reach for and what it will fail to tell you.

A few specifics that determine how you read each source. Federal WARN triggers at employers with 100 or more employees, requires at least 60 calendar days notice, and covers a mass layoff of 50 or more at a single site; California requires notice for 50 or more regardless of workforce percentage, and New York requires 90 days. That variation matters when you match an outflow spike to a filing, because the same layoff surfaces differently by state. QCEW covers more than 95 percent of US jobs quarterly by detailed industry, and OEWS covers roughly 830 occupations at an annual May reference date. Neither names an employer, so both are strictly context.

Reading skill-filtered flow: the stack-specific magnet

A title-only map hides magnets that pull a specific stack. Filtering the population by skill before you build the graph can surface a counterpart winning, say, TensorFlow talent even while the overall title-level flow looks balanced. Skill tags cluster by employer, not only by person.

In Refolk's index, DoorDash and LinkedIn surface among the top employers of TensorFlow-tagged US ML Engineers but not among PyTorch-tagged ones. That is a real, employer-level clustering signal: a flow map filtered to one framework can reveal a magnet that a title-only map averages away. The two tables below show the population you are filtering.

The practical read: if your target ships on a particular framework, run the flow map on that skill-filtered population as well as the title-level one. The two maps can disagree, and the disagreement is the finding. Meta and Apple top both the US and UK lists, so shared magnets do exist across markets; the framework-level splits are where the stack-specific pull hides.

TARGET: <company>, <population>, window <start> to <end>

NET POSITION: <net-draining | net-gaining>, driven by <top counterpart>

TOP DESTINATIONS (outflow > inflow):
1. <counterpart> - ratio <inflow/outflow> - <magnet | event> - <corroboration>
2. <counterpart> - ratio <inflow/outflow> - <magnet | event> - <corroboration>

TOP ORIGINS (inflow > outflow):
1. <counterpart> - ratio <inflow/outflow> - <corroboration>

RISING MAGNET TO WATCH: <counterpart>, ratio moved <old> to <new>,
inflow falling / outflow rising, first visible <month>

LOW-CONFIDENCE (below sample gate): <list>

How this goes wrong

Most bad talent-flow maps fail in one of seven predictable ways, and each has a specific check. This is the part of the method that separates a defensible map from a persuasive-looking one. Work the list before you present.

• Single-date inflow spike read as a magnet. An acquisition re-tags a whole team in one month and the ratio spikes. Check: cluster moves by month; a magnet spreads across many months, a re-tag does not. Verify against acquisition news or a WARN filing.
• Contract-to-FTE conversion double-counted. One person shows as two moves at the same employer. Check: collapse consecutive same-employer tenures before aggregating.
• Subsidiary aliasing splits a counterpart. Outflow to one company appears under three names, hiding the true magnet. Check: canonical entity mapping before ranking.
• Ratio built on tiny counts. A 3-to-0 flow reads as an infinite magnet. Check: gate on inflow and outflow each above roughly 5 before trusting the ratio.
• DEF 14A turnover taken as directional. A high reported rate is assumed to feed a named rival. Check: never infer a destination from a turnover percentage; the SEC itself notes these are non-comparable.
• Layoff mistaken for voluntary drain. A WARN-driven exit wave looks like competitors winning talent. Check: match the outflow month to any WARN filing and label it involuntary.
• Stale profiles inflate current headcount. People who left still list the target as current. Check: require an end-date or a newer position to confirm a completed transition.

Before you call the map done

Run this checklist before the map leaves your desk. Every item maps to a failure mode above or a sample rule; skipping one is how a wrong arrow reaches a leadership deck.

Keeping the map current

A talent-flow map is a snapshot of a moving graph, so it decays. Re-run it on a cadence that matches the window: a rolling 12 month window refreshed quarterly keeps the leading-indicator property alive without letting stale moves dominate. The point of the ratio is to catch a flip early, and a map you built once and never refresh will miss the flip it exists to catch.

Watch the ratios that are drifting toward 1, not the ones already past it. A counterpart whose ratio has moved from 2.0 to 1.2 over two refreshes is the rising magnet worth naming, even if it has not yet crossed. Pair that with the event calendar: every refresh, re-check the top counterparts against fresh acquisition news and new WARN filings, because a re-tag that was organic-looking last quarter may have a filing behind it this quarter. Geography sets the refresh floor. Where the observable pool is thin, as it is for a UK target against a US one, widen the window before you shorten the cadence, or the map will thrash on sub-5 samples. The discipline is the same every cycle: gate, strip, corroborate, then chart.