Analysis · Product launch

Everyone's shipping an MCP right now. Findem says that's not even the product.

By Matt Charney

When it comes to technology, timing is everything. Don't believe me? Ask Elisha Gray. In 1876, after years of tireless research and relentless experimentation, Gray finally achieved what can only be described as one of the greatest breakthroughs in the history of innovation.

Fully aware of the commercial value of his new invention, Gray rushed to the patent office, ready to finally cash in all that sweat equity he'd accumulated over the years.

When he got there, he received what can only be described as some of the worst news in the history of entrepreneurship: the patent for his invention had already been successfully filed. Thinking this impossible, Gray, thinking the clerk was joking, asked to see the other application.

It didn't take long to find - in fact, it was filed in the same office only a few hours earlier, and certified by the very same patent office worker. The patent holder, named on the form, was a guy by the name of Alexander Graham Bell. The invention was the telephone.

Being first to market has its advantages - categories, it turns out, can get really crowded, really fast.

You know the tells; you start hearing the same acronyms and buzzwords in every vendor briefing; every sponsored blog post or podcast starts pushing content that are essentially variations on the same theme; every panel at every conference has to mention it at least once; and, perhaps most noxious, every VP of Product out there shares that marketing mandated LinkedIn post about how they're approaching building, essentially, a better mousetrap.

Think: "the cloud," "big data," "social recruiting," or, most recently (and, probably, most noxiously), AI, along with derivatives like "agentic," "generative," "conversational" and similar variations on the same theme.

Right now, talent tech is seeing another emerging category becoming commoditized in real time: MCP, or Model Context Protocol, for all you nerds out there. The last six months or so, the MCP has become the recruiting industry's version of "thought leadership," when everyone basically follows the same script and GTM playbook.

Findem's signal to all that noise, delivered this week with the official launch of Findem Studio, refreshingly undermines the core of this cliched category narrative. Their response to their erstwhile competitors who've already rolled out dedicated MCPs for recruiting and hiring?

That's cute, but you didn't understand the assignment.

Context is king, because content abdicated.

In a joint interview with my colleague (and, if I may, top industry analyst) Kyle Lagunas, Findem CEO Hari Kolam and Liv Anderman, the company's head of marketing, made a case that was a bit more contrarian and a whole lot more combative than the usual launch week talking points. Which, as you can guess, I'm totally here for (and bringing the popcorn).

Their messaging framework seems pretty simple. Their assertion is that engineering the actual MCP itself was never the hard part, and that most of what's getting announced right now is more plumbing than product, more infrastructure than architecture.

"MCPs are essentially APIs for AI. They are the piping. They give a model access to data and tools, but access isn't the same thing as intelligence, and no model works across billions of people and company data points at once."

Liv Anderman, Findem

An access card gets you into the building. It doesn't tell you which room to walk into, or which of the ten thousand filing cabinets in that room actually has what you need.

That's roughly the point Hari was making with an analogy he's apparently used enough internally that Anderman name-checked it before he even got to it: think of an MCP as a library card. It gets a researcher (in this case, the AI model) into the building.

The problem is the building has a billion books in it, and "go read all of them" isn't a strategy, it's a way to burn a context window and produce something that sounds smart and means nothing.

"Getting an access card more granular to a particular row of books is also not good enough," he said, "because even hundreds of tons of books is also very, very big. The quality of the MCP is about how precise can you get to point to the exactly the right book the AI model has to read and infer."

That's a real, measurable problem, not just a metaphor. Researchers at Stanford found that large language models handling long stretches of context don't actually use that context evenly. Information buried in the middle of a long document tends to get functionally ignored, even when it's directly relevant to the question being asked, a phenomenon the paper's authors called "lost in the middle."

Dump the whole library into the context window and you're not getting a smarter answer. You're getting a model that can read existing work, not really write new ones (I've clearly never metaphor I didn't like).

The illusion of finished work: why outcomes matter more than inputs.

Where this gets more interesting, and more uncomfortable if you're a buyer who's already signed off on a few AI pilots this year, is what Findem is calling "finished work." Anderman's framing is that most of what's shipping under the agent label right now produces an illusion of completion rather than the real thing. It reads well. It's structured correctly. It sounds confident. And that, she argued, is exactly the danger.

"You ask any AI for a specific artifact, a list, a succession plan, whatever it might be, you'll get something back that looks like one," she said. "What you can't see is what was the people data that it was working from? Did it hallucinate? Nobody's checked the result before it reached you."

The exact second you realize the flaw, she noted, inevitably arrives at the absolute worst time: right in the middle of a high-stakes presentation, when a stakeholder asks why a candidate appears on a talent map and you realize the profile was completely fabricated.

That isn't some theoretical scenario crafted to score cheap points during product launches. It aligns directly with what researchers keep uncovering about talent acquisition algorithms.

A comprehensive Brookings analysis covering automated candidate sourcing highlighted systemic racial and gender disparities lurking deep within vector retrieval systems—the precise flavor of subtle bias that passes the initial eye test until an audit uncovers it.

Meanwhile, a PeerJ Computer Science paper focused on algorithmic candidate ranking detailed how sequence order and gender attributes skewed model selections. Even worse, findings published by the University of Washington revealed a far darker loop: recruitment professionals actively began replicating these machine-generated biases after repeatedly relying on slick, authoritative summaries.

The software doesn't merely produce a flawed recommendation once. It conditions the human recruiter to adopt those exact same flaws.

Which brings us straight to the fundamental argument Findem is hammering home: true completed outcomes require rigorous standards, rather than functioning as a flashy word for basic output.

Anderman breaks that standard down into three non-negotiables that must occur before any recruiter makes a move: feeding validated information into the engine upfront, executing a proven operational framework instead of hallucinated steps, and delivering complete transparent auditing so talent teams can trace every single data point and inference instead of blindly trusting a black box.

Don't hate the player, hate the playbook.

The core architecture Findem crafted to execute this tailored approach revolves around what they're calling an "expert model," which is a feature I'm particularly excited about, given that it seems perfectly primed to ignite plenty of heated debates over drinks at upcoming trade shows.

Instead of permitting an algorithm to improvise its own workflow on the fly, several pre-packaged agents inside Studio come pre-loaded with specific practitioner playbooks. It essentially licenses proven, real-world frameworks, from mapping out talent landscapes to executing succession strategies, complete with predefined signal hierarchies, giving users a battle-tested blueprint right out of the box.

Kolam described it less as branding and more as provenance. "It is a recipe that has a methodology and strong methodology behind it," he said, and pushed back on the idea that it's static: "The methodology itself is not written in code. It is a living and breathing information that can be tweaked to one's need."

In Hari's view, what these industry experts are providing isn't merely a seal of endorsement—it's an operational blueprint. It defines how the final deliverables get formatted, what benchmarks the scoring model evaluates against, and how talent pools from both inside and outside the company get weighed side by side. From there, buyers can take the baseline agent, clone it, and customize the parameters to fit their needs.

That logic holds up well enough for certain standard use cases, but it leaves open a rather massive question that Findem didn't quite address: finding a principal AI engineer, hiring specialized trade technicians, and running a C-suite search aren't the exact same workflow sporting slightly different attire.

A framework constructed around a single expert's executive recruiting playbook must survive first contact with drastically different sourcing realities to offer genuine utility. Telling buyers "you can tweak and copy it" sounds nice in a slide deck, but that's also the exact boilerplate response every configurable platform offers when pushed on fringe scenarios.

Whether Findem's flavor of customization actually survives real-world deployment across wildly divergent industries is an answer we'll get from field results, not shiny marketing collateral.

The context window problem: what CHROs need to know

If you're sitting in the CHRO seat, the most practical takeaway from our sit-down was watching Kolam demystify all this underlying technical machinery into plain-spoken leverage you can actually wield during your next brutal vendor interrogation.

His core point boils down to spatial reality: even the most expansive context window available today caps out around 300 candidate profiles. That's your entire working canvas. Stuff that window with completely irrelevant dossiers—or bury those pristine profiles beneath thousands of junk records—and your expensive AI model will spit back garbage every single time.

"AI is a phenomenal inferencing platform. It's not a search engine."

Direct an algorithm toward flawed inputs and it'll still churn out an ultra-confident response. It just happens to be dead wrong.

Not surprisingly, this exact breakdown is plaguing broader corporate enterprise technology budgets today. A recent study by MIT's NANDA initiative discovered that roughly 95% of corporate generative AI initiatives fail to deliver tangible ROI.

The core issue isn't underperforming algorithms; rather, it's that organizations lack essential surrounding operational pipelines, verified data architectures, structured execution methods, and actionable auditing mechanisms to turn raw processing power into usable results.

Analyst firm Gartner offers an equally grim assessment: the consultancy forecasts that over 40% of enterprise agentic deployments will get scrapped before 2027, driven down by ballooning expenses, dubious commercial utility, and absent governance frameworks.

That's the exact headwind Studio encounters at release, yet it neatly bolsters Findem's thesis that nearly a decade spent meticulously indexing background data matters significantly more than rushing another generic MCP out the door.

As Anderman put it, "we ship our MCP in probably weeks, but the labeled data has taken years, and you can't retrofit that."

Whether that's a durable moat or just a head start depends entirely on how much labeled, structured people data your average competitor actually has sitting around, and most don't have anywhere close to what Findem's built over the better part of a decade.

The questions that didn't get answered

Now, here's where things get way more fascinating than any polished corporate press release. Lagunas came into the sit-down armed to the teeth with tough questions, and several of his sharpest probes never actually received clear answers on the record.

Who actually retains ownership of proprietary information when customers process it through Studio's third-party integrations, and what happens to that repository if Findem gets swallowed in an exit event or a client cuts bait?

Where does Studio's functionality hit a hard ceiling today, and which capabilities got unceremoniously axed from this release because the generated output fell flat?

Those items were on the agenda. They simply ran out of runway before tackling them - and to be fair, it was a relatively short time for what are, ostensibly, fairly complex answers.

Let's be clear that this is not necessarily me taking a jab at Findem (far from it, because from a technical perspective, they're one of the most advanced vendors in our space), because most enterprise technology briefings never encounter those hard questions in the first place, let alone side-step them on live audio.

Yet that's precisely the structural vulnerability any savvy enterprise buyer needs to probe before signing off on next year's contract, not after the check clears.

Anderman did pitch a clever evaluation framework that's worth stealing before your next sales pitch: challenge the vendor to specify what manual operational steps vanish and trace the exact financial differential.

Then force the interface to pull back the curtain on its reasoning by exposing the raw source records, the underlying operational framework, and the human validation trail behind every inference.

"Most tools in the market don't survive that question," she said.

That is precisely the high-stakes wager Findem is placing with Studio. The lingering, unanswered queries around data rights and potential acquisition scenarios are the exact hard questions worth probing next (and precisely what this brief briefing left completely on the table).

Where Findem is willing to take a definitive stand on the record, however, is the sheer magnitude of their strategic gamble.

Kolam's brand of contrarian conviction posits that the talent space is squandering its collective bandwidth obsessing over basic agent assembly. He dismissed this commoditized capability as merely "a prompt away," rather than tackling the grueling, capital-intensive heavy lifting of continuous data validation and structural refinement. And he's right.

"Our sequence of outcomes are the ones that we have high conviction with," Kolam asserted, positioning Studio as a high-stakes gamble that validated, fully executed deliverables—not the shiny UI shell surrounding them—will emerge as the true currency traded in HR tech.

That's a serious claim, placing Findem's reputation squarely on a far higher standard than the endless stream of competitor announcements flooding our feeds this quarter.

Whether buyers actually buy into that vision is something we'll need to evaluate in twelve months, long after the hype cycle around MCPs cools down and we see who constructed an enduring platform versus who was merely riding the latest trend.