Context is all you need

Think about the last real decision your leadership team made in a meeting. Not a status update, an actual decision. Ask yourself why it happened in a room full of people instead of in one person's inbox.

We tell ourselves it is about buy-in. It mostly is not. We put people in the room because each of them carries something no document holds, and somewhere in the conversation one of them leans in and says "wait, that will not work, because," and catches a decision that was about to be made on too little information.

That is what a meeting is for (at its best). It is a context buffer. For as long as we have had organizations, a person in the room has been the thing standing between us and a decision being made on thin evidence.

I want to make the case that the buffer is being removed from a growing number of workforce decisions, that HR needs to be readier for that removal than any other function in your company, and that the reason it lands hardest on HR is itself the warning worth acting on this quarter.

The room knew something the data did not

Here is the shape of the problem, using a case most HR leaders have lived through some version of:

A team gets flagged as high risk. Engagement down, output flat, attrition climbing. To a dashboard, and to any software acting on that dashboard, the move could be out of a playbook: intervene, reassign, get a leader involved. But the HR business partner who had been in the room knew things the data did not show. That team had just absorbed a failed reorg. Their leader was out on medical leave. Most of the frightening attrition number was two loud exits in the same month. The right move was patience, not a blunt instrument.

Same numbers on the table, opposite call. What flipped it was context that no dashboard held and no software had recorded.

Context: everything your organization knows about itself, the quantitative and the qualitative together, in whatever form it happens to live in. Your headcount numbers are context, and so is the reason your pay bands are shaped the way they are, the workaround that keeps onboarding moving, and the judgment your longest-tenured HR business partner applies without being asked. One half of that sits in a database; the other half sits in people, and it has never been written down.

But a person, especially a seasoned HRBP, hesitates in that situation. Something does not smell right, so she checks, or she walks down the hall, or she waits a week. An AI agent, however, does not hesitate. It has no "feeling" about the situation and no hallway to walk down to have a quick chat before deciding. It makes a call and it applies that action quickly and decisively.

You are being sold that capability right now. It is very exciting to talk about how fast agents move and how they can take action at scale, and I am not arguing against buying it. I am arguing that the thing determining whether it helps you or embarrasses you is not the quality of the software (or the marketing team behind the software). It is the quality of the context you can layer into the AI workflow.

Why AI under-delivers in HR, but not in the rest of your company

The models underneath every tool you have been shown were trained on the public record, and on a staggering amount of it. Every published book on compensation design. Every employment statute in every jurisdiction you operate in. Decades of peer-reviewed research on turnover, motivation, and selection. Every management article on performance reviews, and, for good measure, the entire internet's running opinion about what HR is doing wrong. These tools arrive at your company genuinely well read about what has been said outside about your company and about HR.

But they have also never seen your benefits policy. They have not seen your compensation plan (and I would be worried about you if they had). They have not seen the memo explaining why your pay bands are shaped the way they are, or the honest account of why the last reorg stalled in month three, or which of your approval chains are load-bearing and which are ceremonial. None of that is on the internet. For better or worse, most of it is not written down at all, in any system you own, in any form a machine could read.

That gap, between the public record these tools are trained on (a lot of Reddit, by the way) and the private reality of the company you actually run, is the whole of the problem. Every competitor you have can buy the same software you can. None of them can buy your unique organizational context. Of everything in the AI stack you are about to buy, your context is the only piece that is genuinely yours, and in HR it happens to be the piece no one ever captured.

The piece of AI history worth carrying into your next vendor meeting

There is one reference worth having in your pocket, and if you have not run across it, that is not a gap in your reading. It lives mostly inside the engineering world and has only in the past few years started leaking into business conversation.

In 2017, a team of eight researchers, most of them at Google, published a paper with an unusually confident title: "Attention Is All You Need." It introduced the transformer, which is the architecture sitting underneath essentially every AI tool that has been demoed to you since. As a piece of engineering it was the whole ballgame, and if you start listening for the title you will now hear it referenced constantly.

The useful part for you is not the architecture. It is what happened next.

Within a few years, the architecture stopped being what the AI companies fought over. It got settled, published, and shared, and it became something any competent team could build on. What those companies have competed on ever since is context: how much of the customer's actual reality they can get in front of the model at the moment it answers a question. Context windows, retrieval, memory, connectors, and evaluation are all names for the same fight. The industry that produced these tools learned, quickly and expensively, that the model was rarely the thing standing between a demo and something useful. The missing context was.

This is why we have seen stories of AI firms buying and cutting up print books by the millions, the near daily lawsuits around copyright, and the race to feed the world's knowledge into frontier models. The one-line version: the AI industry stopped competing on models several years ago and has been competing on who can find and consume the most relevant context ever since.

Which means the argument I am making about HR is not an HR-specific theory, and it is not a new idea you are late to. It is the conclusion the AI industry just reached, arriving at your function from a different direction.

Attention got the architecture where it needed to go. But for an enterprise organization, context is the part that is "all you need," and it is (unfortunately) the part you have to supply yourself. No one knows your organization like you do.

Your peers in marketing and finance do not have the same problem

Here is where I think HR's situation is genuinely different, and it is the part worth taking to your CEO.

Take marketing. A model trained on the internet arrives already fluent in that domain, because most of marketing's raw material is public by design. Competitor campaigns are public. Pricing pages are public. The entire published literature of advertising, positioning, and brand is public. Customer behavior sits in systems that were purpose-built to record it, and where marketing needs outside data it can go buy it from a dozen vendors. Your CMO has private context too, but the software shows up already knowing the shape of the work.

Now take finance. Finance had a century of regulators forcing it to write everything down. There is a chart of accounts. There is GAAP, and IFRS, and an audit function whose entire job is to verify that the documented version matches the real one. Every public company files a detailed, structured, standardized account of itself every quarter, which means these models have read millions of examples of exactly the artifact your CFO produces. They know what deferred revenue is, and more usefully, they know what your books are supposed to look like, because a regulator required your books to be legible.

Then there is HR. No regulator ever requires you to document why a job is designed the way it is. There is no agreement on a standards body, let alone legal requirement, to define "manager" in a way two of your business units would agree on. There is rarely an audit that checks whether your documented process matches your real one, and if you commissioned that audit tomorrow you already know roughly what it would find. The most consequential parts of this function are tacit and human to human, either by deliberate design or by simple default, and they were never encoded anywhere.

Some of that is not negligence, and I want to be careful rather than pretend otherwise. A great deal of HR context is unwritten for defensible reasons. Employee relations conversations, the real story behind a leader's move, the accommodation quietly granted, the exception approved once and never formalized. Privacy, legal exposure, and basic decency all argue for some of it staying off the record, and I am not proposing you write down everything anyone ever told you in confidence. But there is an enormous middle ground between the genuinely confidential and the merely undocumented, and nearly the whole operating logic of the HR function lives in that middle ground.

So the distance between what these tools know and what they need to know is wider in HR than in any other function you fund. And it is widest exactly where being wrong is most expensive: pay, promotion, exit, leave, accommodation, and investigation. A confidently wrong action in HR does not land on a campaign metric. It lands on a person.

Context is much bigger than the numbers and you have already paid for half of it

I defined context above as everything your organization knows about itself, the quantitative and the qualitative together. What matters here is the split inside that definition.

The quantitative half you likely know well, because you likely funded it over the past decade in the form of analytics and data warehousing. Headcount, requisitions, spans, attrition, survey scores, the warehouse, the dashboards.

The qualitative half is everything else, and most of it is not sitting in a database: the reasoning behind a decision, the knowledge in one person's head, the recorded conversation, the working document, the exception nobody logged.

Both halves are part of the whole of context. Businesses just went after the half that computers could help with first and then called the whole thing "analytics."

I include myself in that. I spent 15+ years in this work, first as an HR practitioner and then building People Analytics functions at Meta, Uber, and Nike, and continued that work later at One Model, where I spoke with a few hundred People Analytics teams a year and got an unusually wide view of how it actually gets done within HR teams. So let me say plainly what my field spent the past 20 years doing, because it is the most useful thing a CHRO can understand about the next five years.

Almost none of the value of analytics is created by the visible final step of data science. The final report you got in your hand was important, helpful, and may have driven a decision forward, but the reason that report was valuable happened much earlier.

That value was created upstream, in the unglamorous work of making a number mean something, and that work had a specific shape. Five questions had to be answered before anyone could trust a workforce number: where did this come from, what happened to it on the way to me, how is it organized and against whose definitions, is this one source of truth or 12 spreadsheets that disagree, and have we cut the noise enough that a person can act on it.

Your People Analytics and HR technology teams have been climbing that ladder for two decades, and for those HR teams that invested, it has worked. The best HR teams run on numbers now in a way that simply was not an option in 2010.

But that same climb we went through for quantitative analytics starts again for the qualitative half, using the same disciplines on different material: structured interviews instead of data feeds, transcription instead of extraction, claims tagged with their sources instead of rows tied to a system of record, stored so a person or a machine can query them and see where every answer came from.

Same discipline, new material. That is the single most important sentence for you in this piece, because it tells you what kind of project you are approving. Not a science experiment, not a new department, and not a knowledge management initiative with a taxonomy committee. It is the second half of a build your organization already knows how to do, and the reason it has not happened is that nobody has been told it is the job.

Which brings up the half that has never had an owner at all. Margaret Mead's observation was that what people say, what people do, and what people say they do are three entirely different things, and that maps onto a company almost perfectly. Your systems capture what people do. Your surveys capture what people say they do. What almost no company has built anything for is the third channel, what people actually say, in their own words, about how the work really happens. The operating logic of your function is stored in that third channel and nowhere else.

Trust is the condition, not the compliance step

If your first reaction to "go capture what people actually say" is a concern about surveillance, that is the correct reaction and it is why I do not think this work can be handed to IT.

Three guardrails have to be in place from the first conversation rather than added after the first incident.

Consent and transparency. People should know what is being captured and why. They should be aware when they are being watched and have a choice in what they share. Not buried in a policy update. Told.

Purpose limits. The purpose is better decisions, not monitoring your workforce. The moment this becomes performance surveillance wearing a different coat, you lose the candor that made it worth doing, and you do not get it back. You get one attempt at that trust.

Whose voice gets captured. What you capture becomes what the software believes about your organization. If you only interview the loudest, most senior, most available people, you have encoded their view of the company into every decision made downstream of it. Capture broadly or bake in bias.

Build this the way you would want it built if it were your own words being recorded. Trust is not a footnote on the work. It is the condition that makes the resulting context usable at all, and judging that condition is an HR competency before it is anyone else's.

Why it lands on you

I know how the ownership argument sounds from where you are sitting. Everything lands on the CHRO right now, usually while the budget moves in the other direction, and one more mandate is not a gift.

So let me be precise about what is actually yours here, because most of the work is not.

Your People Analytics and HR technology teams are the right owners. They have climbed this ladder of context once already for the workforce, which nobody else in the company has done. The interviews, the structuring, the storage, and the maintenance all belong to them.

What cannot come from them is the decision that curating organizational context is the job rather than a side project that happens after the dashboards ship. That is a scope decision and a funding decision, and it only gets made at the top of the house. It is also the only part currently missing in most large companies, which is why the work has stalled in so many of them while everyone agrees it matters.

And if HR does not claim the work, the vacuum will get filled by whoever needs the answer first. Finance will need a real model of your workforce the moment it plans capacity or models a reduction, and headcount alone will mislead it. IT will need one the moment it tries to automate a process nobody has honestly described. Both will build a thin version out of system data, because system data is what is sitting there available, and every decision downstream of it will get the human element wrong in the same predictable ways.

There is a version of the next five years where HR ends up more central to the company than it has ever been, because context becomes the thing every other function's decisions run on and HR is the function that owns it. That position does not get granted at a table. It goes to whoever does the work.

What to say when your CEO asks

Most CHROs I talk to have already been asked, in front of a board or close to it, what HR's AI plan is. The trap in that moment is to answer with a list of pilots, because pilots invite a follow-up question about results, and the results in HR are going to lag the rest of the company for the reasons discussed above, and that is not HR's fault.

The more defensible answer runs something like this. Our AI results in HR will trail marketing's and finance's for a structural reason: the models arrived already fluent in those domains and they arrived knowing nothing about ours, because the way our work happens was never written down anywhere. That is a fixable input problem rather than a capability problem. We are fixing it in one domain first, in weeks rather than quarters, and we will not automate on top of a process we cannot describe.

That answer does two things a pilot list cannot. It explains the gap before someone else explains it less generously, and it puts HR in the position of naming a company-wide condition rather than defending a function-level shortfall.

The tools coming into your function have read nearly everything humanity has written down. They will never read your organization, because your organization was never written down, and nothing about the way HR has always worked will produce that record on its own. Someone has to decide to go get it.

For a hundred years we got away with a thin picture of ourselves, because there was always a person in the room to supply the rest. Build the fuller picture while those people are still in the room to help you get it right.

Frequently asked questions

Q: What is organizational context, and how is it different from HR data?

Organizational context is everything your company knows about itself, the quantitative and the qualitative together. HR data is the quantitative half: headcount, requisitions, spans, attrition, survey scores. The other half is the reasoning behind a decision, the workaround that keeps a process moving, and the judgment a long-tenured HR business partner applies without being asked, and almost none of it sits anywhere a machine can read.

Q: Why do AI tools underperform in HR compared with marketing and finance?

Because the models arrived already fluent in those domains and knowing almost nothing about yours. Marketing's raw material is public by design, and a century of regulation forced finance to write itself down in standard formats, so the models have read millions of examples of both. No regulator ever required HR to document why a job is designed the way it is, so the operating logic of the function was never written anywhere a model could learn it.

Q: Is this knowledge management with a new name?

No. Knowledge management projects usually open with a taxonomy committee and close with a repository nobody opens. This is the discipline your People Analytics team already applied to workforce numbers, run again on different material: structured interviews instead of data feeds, transcription instead of extraction, and claims tagged with their sources instead of rows tied to a system of record.

Q: How long does a first version take, and what do you get?

Pick one domain, usually workforce planning or talent acquisition, and plan on 10 to 20 structured conversations across two to four weeks. Three artifacts come out of it: a definition set for your core terms that records how each one is known to break, a map of your highest-volume processes as they actually run, and a log of your recurring workforce decisions and who makes them. Then you decide whether to do a second domain with something real in hand.

Q: Should HR or IT own organizational context?

HR. Judging consent, purpose limits, and whose voice gets captured is an HR competency before it is anyone else's, and getting it wrong costs you the candor that made the work worth doing. Your People Analytics and HR technology teams are the right builders; what only a CHRO can supply is the decision that curating context is the job rather than a side project that happens after the dashboards ship.

Related reading

• The End of People Analytics As We Knew It, this argument written for People Analytics leaders.

• Organizational Literacy: The Real HR Transformation Gap, on why data-literate HR functions still cannot read themselves.

• Why HR Needs a Context Layer to Win the Next Wave of AI, the architecture underneath this and what retrieval needs in order to work.

• The Hidden Data Deficit That Will Sink Your HR AI Strategy, Ian O'Keefe on the same gap from the CEO's chair.

• You Cannot Transform HR That You Cannot See, on six HR mandates that all need the same missing input.