AI's Impact on Staffing.

Most AI for staffing pitches you will hear in 2026 are inflated. A few are very real. The gap between the two is where staffing firms either compound margin over the next three years or get left behind by the ones that figured it out. This is the operator honest read.

I sold Veritaaq to ManpowerGroup at $250M revenue and 2,500 consultants on assignment. That was one of the largest staffing exits in Canadian history. The thing I would tell my younger self about it now is this: the workflow we ran in 2018 was the workflow Bullhorn shipped in 1999. Five disconnected systems. Every placement built on copy and paste. A back office that grew linearly with revenue. BD that ran on the relationships in a few rep heads.

That is still how most staffing firms run today, whether they are placing IT contractors at a Big Five bank, healthcare professionals into hospital systems, light industrial workers onto warehouse floors, or finance and accounting talent into corporates. Every staffing firm owner I talk to in 2026 has the same question: where does AI actually fit in my business, and what is hype?

I now run Forge AI, a managed service that deploys AI inside staffing firms across recruiting, sales, and back office. We are forward deployed inside a Canadian IT staffing firm with a 30,000 candidate database, and we have shipped the engine from start to finish. The list below is what is actually working at production scale, what is not, and how I would tell a staffing CEO to think about buying any of it.

Before you read on: if you want a fast read on where your firm sits right now, the five minute AI Readiness Scorecard at Forge Research outputs your level and the top three places AI compounds for you. No call required. The essay below is the framework behind it.

The thesis. The cost to serve curve, finally bent.

Services businesses have had the same shape for a hundred years. Cost rises with revenue. Hire more people to serve more clients. Margin caps out at whatever leverage you can squeeze from your back office and your senior bench. Software businesses worked the other way. Cost barely moved as revenue scaled. That is why the public market has always paid software multiples and never paid services multiples.

For the first time in this industry's history, that relationship is bending. AI lets a services business sell more, deliver more, and bill more without adding proportional cost to do it. The recruiter who used to fill 5 reqs a week now fills 15. The BD rep who used to generate 4 meetings a month generates 12. The controller who used to take 5 days to close the books closes them in 8 hours. Same headcount. Multiples of the throughput.

Inside staffing, the implication is direct. The firms that compound margin from here are the ones that bend their own cost to serve curve first. The window is 24 months. After that, the firms that did it have priced the firms that did not out of the market on every contract that comes up for renewal.

The market in 2026: margin pressure is the binding constraint.

Staffing Industry Analysts characterized the recovery from the 2024 downturn as a Lazy W. The market they expected would grow 5% in 2025 actually contracted another 3%. After the 10% drop in 2024, the US staffing industry closed 2025 at roughly $183B, the second straight year of contraction in a market that averaged 5% annual growth for two decades before the pandemic. SIA's November 2025 forecast update called for a modest rebound in 2026, but the flat backside of the W is what operators are actually living right now. Bill rates are flat. Candidate scarcity is real. Recruiter cost per placement keeps climbing. Direct sourcing is rising. Statement of work delivery is eating into traditional contingent staffing. The firms that compound margin from here are the ones that take cost out of the workflow, not the ones that hire more recruiters.

SIA puts it directly in their 2025 Staffing Trends report: AI is now ubiquitous throughout the entire staffing organization, not just talent acquisition. Every department can enhance processes through AI, either by making them faster, less expensive, more accurate, or more informative. Most can do all of those at once. SIA goes further. They signal that 2026 is the year agentic AI starts replacing categories of work that humans used to do, and that staffing firms slow to adapt will lose ground to the ones that move first.

That is the macro. The honest tactical read is that AI compounds inside a staffing firm in three places, and only three.

1. Recruiter capacity. The minutes per candidate from job order to submit. Most recruiters spend 60% to 80% of their day on admin, sourcing, formatting, and copy paste, and only 20% to 40% on talent. AI flips the ratio. The dollars are unambiguous.
2. BD pipeline. Your placement history is sitting unused as the most valuable prospecting asset in the business. Most BD reps generate two to four qualified meetings per month manually. The right AI stack lifts that to eight to twelve. Same headcount, two to three times the pipeline.
3. Back office. Timesheets, invoicing, payroll reconciliation, MSP remittance matching, month end close. The industry runs three to five FTEs per firm just to keep the books moving. This is the cleanest dollar for dollar replacement target in the business.

Everything else, branding, employer brand, candidate experience apps, chatbots that engage candidates, is mostly noise in 2026. Useful at the margin. Not where the dollars are. Start with the three above.

Recruiting. Where the dollars start.

Sourcing: this is the killer use case.

The single most expensive thing in a staffing firm is recruiter time spent re running the same Boolean search across the same 30,000 candidates for every new requisition. Every staffing firm has this problem. Every staffing firm pays for it in margin. SIA notes that 98% of staffing firms now rely on online job advertising, and that LinkedIn alone accounts for 26% of global online job ad spend. The candidate database used to be a moat. It is not anymore.

What replaces the moat is your ability to retrieve and rank the candidates you already have, faster than anyone else.

First, the problem with how it works today.

Whether your shop runs on Bullhorn, Avionté, JobAdder, Vincere, TrackerRMS, Salesforce, or a custom build, the workflow today is the same. Boolean search in the ATS matches keywords. A recruiter types something like (Java OR J2EE) AND "Spring Boot" AND Toronto and the ATS returns every candidate whose resume contains those exact words.

The trouble is that resumes are not written in a standard vocabulary. A perfect candidate may have written "JVM microservices engineer" on their resume and never typed the word "Java." A senior nurse who spent 8 years in cardiac ICU is a strong fit for a step down unit role, but the words "step down" never appear on their resume. A lead Salesforce architect with 6 years at Salesforce never wrote "CRM administrator" because they outgrew the title in year one.

Boolean search misses these every time. The recruiter ends up manually broadening the search, reading hundreds of partial matches, and quietly knowing that the best candidate for the role is in the database somewhere but the search will not find them. That is the 2 to 3 hour task per req.

First, the vocabulary.

Five terms that show up in every AI conversation. Defined plainly so the rest of this section is readable.

• Large Language Model (LLM). The kind of AI behind ChatGPT, Claude, and Gemini. A model trained on a vast amount of text that learns to predict what words come next. The breakthrough is that this prediction skill turns out to be flexible enough to write, summarize, reason, answer questions, and follow instructions.
• Hallucination. When an LLM generates content that sounds confident and fluent but is factually wrong. The model invents a citation, fabricates a quote, makes up a statistic. This was the central problem with the first generation of LLMs and the reason most early "AI for staffing" pitches failed.
• Embedding. A way of turning a piece of text into a coordinate on a giant map. Resumes about similar work end up at similar coordinates. The map captures meaning, not exact words. That is the part that lets a JVM microservices candidate be found for a Java Spring Boot req.
• Vector database. A database that stores those coordinates and finds the closest neighbors quickly. Pinecone, Weaviate, and Qdrant are common ones. The infrastructure underneath the meaning map.
• Reranker. A second pass model that takes the closest 100 candidates from the embedding map and scores each one in detail against the actual job. Higher precision than embeddings alone, but only run on the small set so cost stays manageable. The fine toothed comb that follows the broad sweep.

Why this was unsolvable until recently.

AI vendors have been pitching this same idea to staffing firms for more than a decade. The natural pushback is to ask why now is any different. For most of those years, the technology was not actually ready. If you tried an AI tool in 2022 or 2023 and it embarrassed you in front of a candidate or a client, that was not user error. The models genuinely did not work yet. The arc below shows what changed, and when.

Figure 01 · Nine years from a research paper to production grade in staffing

How AI fixes it: think of every resume on a meaning map.

The breakthrough that changed everything in the last two years is embeddings. Skip the math. Here is the concept.

Imagine every resume in your candidate database gets plotted on a giant map. Resumes about similar work sit close together on the map. Java backend engineers cluster in one neighborhood. Cardiac nurses cluster in another. Light industrial forklift operators in another. Importantly, the map captures meaning, not exact words. So a "JVM microservices engineer" sits right next to a "Java backend developer" even though they share almost no exact keywords.

Then a reranker reads the closest 100 in detail.

The map gets you to roughly the right neighborhood fast. The reranker is the fine toothed comb that goes through the closest 100 candidates and ranks them precisely against the actual job. It is the difference between "this candidate is in the Java cluster" and "this candidate has the exact Spring Boot plus Kafka plus Kubernetes profile we need, with banking domain experience, and they were last active two weeks ago."

Finally, on the top 10 only, Claude writes the recruiter rationale. "Strong match because A and B. Watch for C." This is the part that used to take a recruiter 5 to 10 minutes per candidate. It now runs in 8 seconds.

Why this beats the obvious alternatives.

Two approaches sound simpler. Both fail at staffing scale. The third approach is what Forge runs in production.

Figure 03 · How a shortlist actually gets built in 2026

The unit economics matter. The naive thing, asking Claude to read every resume against every new req, would cost a staffing firm running 50 new reqs a week against a 30,000 candidate corpus 1.5 million LLM calls per week. That collapses any real economic model. The map plus comb pattern delivers LLM-quality matching at non-LLM cost.

If you want the specific AI providers behind each step (the companies that build the embedding models, the reranker models, and the reasoning models the engine runs on), they are listed in the Sources section at the bottom of this memo. For the rest of this memo, the diagram above is the whole story you need to hold in your head.

Recruiter review time goes from 45 to 60 minutes per candidate to two to three minutes. That is the slot where your margin lives.

If you want to see the engine work on a single resume against a single req, the Match Score widget is free, public, and runs the same Voyage and Claude stack we use in production. Paste a resume, paste a JD, see the cosine and the breakdown in 30 seconds.

Multi source sourcing: where the candidate actually lives.

Your ATS is your home base. It is not the whole picture. The best technical candidates are also on GitHub. The senior engineers are on Stack Overflow. The passive professionals show up only via Apollo enrichment. The recently active candidates flag themselves on LinkedIn. Every source is a fragment.

The production grade pattern in 2026 is to pull from all of them, embed them into the same coordinate space, and rank against the req. GitHub commit history reveals language proficiency and project depth no resume captures. Stack Overflow reputation surfaces senior signals that titles miss. Apollo and similar enrichment tools fill in current employer, tenure, and contact data your ATS never had. LinkedIn activity flags candidates currently considering a move. Twitter and Reddit show domain involvement for niche roles.

For an IT staffing firm, this is the difference between sourcing from your 30,000 candidate database and sourcing from the 30 million developers findable across the open web. For healthcare staffing, the equivalent sources are state licensure registries, specialty board directories, and clinical community forums. For light industrial, it is regional certification boards and trade union rolls. For finance and accounting, it is professional association directories and certification registries. Same engine. Much bigger universe per vertical. Same retrieval pattern, embed once and rank in milliseconds.

MSP SLAs and speed to candidate.

If you place into MSP controlled accounts, your competitive advantage is speed. Most VMS submission windows are 4 to 24 hours. The agency that submits first with a quality shortlist wins the placement. The agency that submits at hour 22 with a hand built Boolean search loses, even when the candidates are equivalent.

AI sourcing collapses the time from req in to shortlist out from hours to minutes. That changes the win rate on MSP work materially. For firms doing 60% or more of revenue through VMS programs (Beeline, Fieldglass, IQNavigator, SAP Fieldglass, Workday VNDLY), this is the single highest impact AI deployment in the firm. The same VMS submission deadlines that have been a tax on your operations for two decades become a moat for the firms that are first to respond every time.

Outreach: across multiple channels, with AI personalization.

Most staffing firms still run candidate outreach one by one out of ATS templates or a recruiter inbox. The sequenced outreach tools, Loxo, Gem, Sense, are a step up but most are template engines, not truly personalized.

Where AI compounds: a 30 minute SLA agent that fires immediate cross channel blitzes (email plus SMS plus InMail plus voice for top tier candidates), with per channel messages generated from the actual JD plus the actual candidate profile, and response detection that pauses the sequence and alerts the recruiter the moment someone replies.

Where it falls flat: AI generated cold outreach that does not reference anything specific. Candidates can smell template LLM slop instantly. The floor for outbound has gotten higher because of it. The bar is no longer "personalized." It is "obviously written by someone who read my resume."

Phone screening: voice AI is finally good enough for the gates, not the qualification interview.

The honest read on voice AI in 2026: it is good enough to run the knockout gates (availability, rate, work authorization, location, clearance, interest) in a five to seven minute call. That is a real time saver because the gates are where 60% of candidates fail and you do not want a recruiter eating that hour.

Where I would not let it run yet: the deep qualification interview. The ten to fifteen minute conversation where you are probing technical fit, motivation, communication, and reading the human signals. Voice AI is closer than it was even six months ago, but the miss rate is still meaningful, and the cost of a false positive (a candidate who passes the bot and bombs the client interview) is high. Today, gate with AI. Let humans qualify.

Submission: the unsexy win that pays for everything else.

Every staffing firm has its own submission template. Branded Word doc, custom resume reformatting, ten different VMS portals each with their own quirks. Beeline, Fieldglass, IQNavigator, SAP Fieldglass, Workday, plus client direct submissions. This is the most boring slot in the workflow and one of the most valuable to automate. Reformatting plus tailoring plus composing the submission email plus creating the submission record in the ATS is a 20 to 40 minute task that AI does in under one minute, deterministically. Do not skip this one.

Sales and BD. Your placement history is a prospecting weapon.

Most BD inside a staffing firm runs on the relationships in a few rep heads. When those reps leave, the pipeline leaves with them. The biggest unlock in BD AI is not outbound copy. It is account mapping.

Every staffing firm has years of structured data sitting in their ATS. Every placement, every submission, every department, every contact at every client. Most firms never look at it. Run a model across it and you get a structured view of every account: tier, departments penetrated, white space, buying patterns, expansion signals. That is the input the BD desk has needed for 20 years and never had.

From there, lead enrichment (Seamless, Apollo, Explorium, Clay) finds the VPs, Directors, and Hiring Managers in the white space departments. Across channel sequenced outreach personalized off the account map context lifts qualified meetings per rep from two to four per month to eight to twelve.

Beyond mapping: intent signals.

Account mapping tells you where to expand inside your existing clients. Intent signals tell you when. The data is mostly public. You just need agents listening for it on a 24 hour cycle and triaging it for the BD desk every morning.

The signals that move the needle for staffing BD:

• Layoff announcements (counterintuitively, they precede new hiring waves in adjacent departments inside 60 days)
• Funding rounds and IPO filings (signal hiring expansion across the board)
• Job posting cadence on the company careers page (volume up means BD opportunity right now)
• Executive movements (a new CTO usually means a new tech stack decision and a new hiring pattern within two quarters)
• SEC filings and earnings call transcripts (mentions of growth, new product, geographic expansion, hiring plans)
• M&A activity (acquisitions create integration hiring waves, divestitures create displaced talent your competitors are not yet tracking)
• New office openings, relocation announcements, real estate filings (geographic hiring plays)

Agents that monitor these signals and alert your BD desk turn cold outbound into warm outbound. The rep is not asking "can I sell you something." They are saying "I saw you announced X, here is what we learned placing twelve senior engineers into similar situations." The win rate on warm signal BD is three to five times cold.

Caveat: BD outreach is the place where AI sloppy is the most damaging. A staffing firm CEO sending generic AI cold emails to a director at a bank torches the relationship before it starts. The right pattern is AI drafts, then human approval, then send. Not AI sends. Do not skip the approval gate. SIA notes that AI use cases in sales and business development are now standard across the industry, but the firms winning are the ones treating AI as draft accelerant, not autopilot.

Back office and finance. The cleanest dollar replacement in the business.

This is the slot most staffing firms underestimate. Three to five FTEs per firm are doing structured, repetitive work that AI does in minutes. Unlike the recruiter replacement debate, nobody gets emotional about replacing the timesheet chasing process.

Timesheet validation

Direct client timesheets get pulled from the ATS billing module. VMS timesheets get reconciled against the ATS placement records. Validate hours against placement terms, flag exceptions to ops, auto approve clean submissions. Most firms run this manually with chaser emails every Monday. AI compresses it to a clean queue your AP person reviews in 20 minutes.

Invoicing

Validated hours flow into structured invoice payloads. Line items, bill rates, the right HST/GST by province (Ontario 13%, Quebec 14.975%, BC 12%, Alberta 5%), the right legal entity, the right MSP billing format. The invoice lands in QuickBooks (or whichever accounting system the firm runs) automatically. The same workflow that took a controller two days a month finishes in minutes.

Payroll reconciliation

Every hour billed should equal every hour paid. The reality is most staffing firms have a 1% to 3% leakage between payroll and their ATS. Missing payments. Rate mismatches. Withholding errors. Cross checking payroll against ATS placement records weekly catches it. For a $40M firm, that is $400K to $1.2M of margin you did not know you were leaving on the floor.

MSP remittance matching

VMS clients pay bulk remittances through the MSP. Matching line items to individual invoices in the ATS, identifying full pays vs partial pays vs shortpays, and flagging shortpays for collections is the kind of work that takes a person two days every two weeks. AI does it in minutes. The shortpay recovery rate goes up because nothing falls through the cracks.

Month end close

The three to five day close becomes a sub eight hour close. Revenue accruals, intercompany eliminations, work in process reconciliation, journal entry posting, trial balance variance analysis. All of it is structured. All of it is repetitive. All of it is the cleanest possible AI replacement. Gartner predicts 50% of organizations will use AI to replace bottom up forecasting by 2028. Inside staffing, this should be even more widespread because the workflow is already structured.

The agentic shift. What virtual employees mean for staffing demand.

SIA flagged this clearly: 2026 is the year agentic AI starts working for organizations at scale, not just answering questions. Adecco announced its Salesforce Agentforce deployment in late 2024, with agents engaging candidates 24/7, reviewing resumes, and producing shortlists. Adecco receives 300 million resumes a year. They are deploying agents because they have to.

For staffing firm owners, the agentic shift cuts two ways.

On the supply side, you can run more reqs per recruiter, more accounts per BD rep, more month end closes per controller, with the same headcount. Operating leverage that was capped at 1.2 to 1.5x before is now 3 to 5x. That is the win.

On the demand side, the same agentic capability that helps you also replaces work your contractors used to do. Customer service. Tier one IT support. Junior finance roles. Some bench engineering. The contingent labor categories most exposed to agentic substitution will see softness. The categories most insulated, the senior technical roles, regulated healthcare specialties, roles that require physical presence or judgment under uncertainty, will hold value or grow.

The strategic implication for staffing operators is that the firms that compound from here are the ones that move up market, lean into statement of work delivery, and operate their own businesses with agentic leverage. SIA's data backs this up: 44% of staffing firms now cite IT solutions and SOW as their top acquisition preference, far more than any other workforce solution segment. The firms that stay in pure contingent staff augmentation while the demand mix erodes are the ones who get squeezed.

The data layer is everything. And most firms are not ready.

The thing every AI vendor will skip telling you: their model is only as good as your ATS data. If the ATS is full of duplicates, stale records (180 days untouched or more), candidates with no resume on file, and free text fields where structured fields should be, the embedding space gets noisy and retrieval quality collapses.

Most staffing firm ATS systems have 15% to 35% data hygiene issues. Before any AI sourcing engine ships value, that has to get cleaned up. Vendors who do not talk about this are setting you up to blame "the AI" when the real problem is the corpus.

Two practical questions to ask before you buy anything.

• Will the vendor audit my ATS first? The honest ones do this in week one. If a vendor is ready to demo their AI on your data without auditing it, the demo is going to look better than production ever will.
• Where do the embeddings live? Embeddings are the flywheel. You want them on infrastructure you control or contractually own, not locked inside a vendor database where you cannot migrate them. Get the answer in writing.

If you want a quick sense of where your data sits, the readiness scorecard includes data quality and ATS hygiene questions in its 14 question diagnostic. The output flags whether your data is the bottleneck or whether you are ready to layer AI on top.

What I would not let AI do yet.

• Auto send outbound BD emails to clients. Drafts, yes. Sends, no. The damage from one off tone AI email to a Big Five bank director is bigger than the time saved on every email AI ever drafted for you.
• Run the deep qualification interview. Voice gates, yes. Full qualification, no. Not in 2026. Maybe in 2027.
• Make pricing decisions on placements. AI is great at modeling rate scenarios and surfacing margin tradeoffs. Accepting or rejecting a final offer is a human call.
• Touch payroll authorization. AI reconciles. AI flags. AI does not approve a payment run.
• Make hiring or rejection decisions on candidates. The EU AI Act classifies recruitment AI as high risk specifically because of this. Even outside the EU, treat candidate selection as a human decision supported by AI, not made by AI.

Regulation. Do not get caught flat footed.

The regulatory environment matters more than most staffing operators realize. The EU AI Act came into force in August 2024 with phased rollout through 2026. Recruitment related AI systems (matching, facial recognition, performance tracking) are explicitly classified as high risk, with strict compliance obligations: risk assessment, high quality datasets, activity logs, documentation, human oversight, accuracy and security standards.

Penalties are large. Up to €35M or 7% of global revenue, whichever is higher. California already imposes $5,000 per violation per day under its state AI Act. Other US states are following. Canada continues debating its comprehensive privacy and AI law.

The two practical things this means for a staffing operator.

• Pick AI vendors who can show you their compliance posture. Risk assessment documentation, dataset quality controls, activity logging, the works. If they cannot, you inherit the liability when something goes wrong with their model running on your data.
• Audit the AI you deploy regularly for bias and error. Independent algorithm auditors exist. Use them on anything making selection or ranking decisions. The audit cost is small relative to the litigation exposure.

How to actually buy any of this.

The AI for staffing vendor space in 2026 is loud. Most pitches will promise the moon, demo on sandbox data, and quote you a SaaS license. Five questions that separate signal from noise.

1. Have they actually placed a candidate? Most AI vendors selling to staffing firms have never sat in the recruiter chair. The ones that have built operating muscle in staffing will give you specific war stories. The ones that have not will give you framework slides.
2. Will they OAuth your real ATS, or only a sandbox? Demos on sandbox data are the oldest trick. The vendors with real product will run their engine on your live ATS (read only) on the call.
3. What are the unit economics? Ask the embed cost per candidate, the LLM cost per req, and the projected monthly bill at your scale. Vendors who do not know are pricing on guesses.
4. What is their ATS data hygiene story? If they do not have an answer, you are going to be paying them to deploy AI on top of bad data. That does not work.
5. Are you buying software or are you buying outcomes? The honest 2026 answer is: if you are buying software for your team to operate, you are buying friction. The right model is managed service, where the vendor runs the workflow and you pay for the outcome (placements, qualified candidates, qualified meetings, hours saved). That is the model Forge runs.

Closing. The next 24 months.

The next 24 months are the window. The staffing firms that compress recruiter admin and back office cost first will price aggressively against the ones that do not, and the gap will compound. This is not a speculative call. It is already happening at the firms that started 12 months ago. SIA's framing in their 2025 trends report is the same: AI is now pervasive throughout every part of a staffing firm, and the question is no longer if, it is how fast.

You do not need to figure all of this out yourself. The architecture and the patterns are stable enough now that the work is mostly deployment, not invention. But you do need to start, and you need to start with the workflows where the dollars are unambiguous (sourcing, back office, BD) before the cosmetic stuff (chatbots, employer brand AI).

Three concrete next steps if you are reading this and want to act.

1. Take the five minute readiness scorecard. You will get a level and the top three places AI compounds for your firm. No call required.
2. Run the Match Score widget on a real req. Paste a job description and a candidate resume. See the cosine, the breakdown, and the rationale that the production sourcing agent would generate. Free. No signup.
3. Start a free 30 day Bullhorn trial. We OAuth your Bullhorn read only, embed your candidates, and run the live sourcing agent against your real reqs for 30 days. Bring a real open req on the setup call and we will rank candidates from your own database in front of you.

And if you want to talk about running it as a managed service inside your firm, that is what Forge AI does. Recruiting, sales, finance, forward deployed inside your operations, outcome based pricing, no software your team has to learn. The architecture is in place. The engine is in production. The next move is yours.

Jordan

Founder, Forge AI. Previously founder and operator of Veritaaq (sold to ManpowerGroup) and Notch (sold 2026). Building the AI operating system staffing firms wished they had.

Sources and references.

Specific date and stat claims in this memo, with the underlying sources. We will keep this list current as new data lands.

• US staffing market data. Staffing Industry Analysts (SIA), US Staffing Industry Forecast, March 2025 and November 2025 updates. SIA reported the US staffing market fell 10% in 2024 and contracted another 3% in 2025, closing 2025 at roughly $183B. SIA US Staffing Industry Forecast: March 2025.
• Online job advertising and LinkedIn share. SIA North America Staffing Survey 2024 and SIA Online Job Advertising Market 2024.
• Adecco Group + Salesforce Agentforce. Salesforce press release dated December 17, 2024. The Adecco Group Scales Recruitment with Salesforce.
• ChatGPT launch. OpenAI launched ChatGPT as a research preview on November 30, 2022. One million users in five days. Reached 100 million users in two months, the fastest consumer product growth in tech history at the time.
• Transformer paper. Vaswani et al., "Attention Is All You Need," Google Research, June 2017. The architectural foundation of every modern LLM.
• GPT-3 release. OpenAI, June 2020. The first model where emergent capabilities at scale became unmistakable.
• GPT-4 and Claude 1 release. Both released on March 14, 2023. First meaningful quality jump beyond ChatGPT for many practical workflows.
• Claude 3 Opus. Anthropic, March 4, 2024. The reasoning breakthrough that made grounded, multi-document AI recommendations reliable enough for production staffing workflows.
• OpenAI o1. OpenAI, September 12, 2024. First publicly available model trained specifically on reasoning rather than next token prediction.
• Cohere Rerank 3 and Rerank 3.5. Rerank 3 released April 11, 2024. Rerank 3.5 released December 2, 2024.
• Voyage voyage-3-large. Voyage AI, January 7, 2025. Standard embedding model in most production staffing AI deployments. Voyage AI announcement.
• EU AI Act. Entered into force August 1, 2024. Phased rollout: prohibited practices effective February 2, 2025; codes of practice May 2, 2025; transparency obligations August 2, 2025; high risk system obligations (which include recruitment AI) effective August 2, 2026. Penalties up to €35M or 7% of global turnover. European Commission announcement.
• Bullhorn founding. Bullhorn was founded in 1999. Three of the systems referenced in this memo (Beeline, Fieldglass, IQNavigator) shipped between 1999 and 2003.
• Gartner forecasting prediction. Gartner predicts 50% of organizations will use AI to replace bottom up forecasting by 2028, cited in SIA Staffing Trends 2025.
• Forge production deployment. Sourcing Agent live in production at a Canadian IT staffing firm with a 30,000 candidate database. Embedding cost data, latency numbers, and unit economics described above are drawn from the live deployment as of May 2026.

Last updated May 2026. Forge Research will revise sources and stats in place as new data is published. If you spot a number that needs updating, write us at hello@forge-ai.ca.