AI Can't Learn What No One Wrote

Most companies think the hard part of AI adoption is the technology. The organizations further along have hit a different wall: when you try to teach an AI system how your organization actually works, you find out nobody ever wrote that down.

This episode breaks down HubSpot's three-stage AI adoption arc and what happens at Stage 3 — where the technology is ready but the organizational foundation isn't. This is where the CHRO has a clear mandate, if they move fast enough to claim it.

What You'll Learn

• HubSpot's three-stage AI adoption framework and why Stage 3 is where most organizations stall
• Why your process documentation describes how you were designed to work — not how you actually run
• The two types of missing organizational knowledge and why one can never just be "found"
• Why decisions about what AI systems handle are organizational design choices, not engineering problems
• Three immediate plays CHROs can run to get ahead of this work before engineers define it for them

Key Quotes

• "When you sit down to teach an AI system how your organization makes decisions, you find out nobody ever wrote that down."
• "The technology is usually ready. The organizational part — that's the part that's not ready."
• "Teaching AI systems how your organization actually operates forces every organization to confront what it was really running on."

Sources for Statistics Cited

• 94% of HubSpot employees use AI weekly — HubSpot Blog: How We Operate as an AI-First Company
• 3,900+ AI tools built by HubSpot employees — HubSpot Blog: How We Operate as an AI-First Company
• Recruiting cut 10 days off time to hire — Source not fully verified
• 80% of scheduling automated — HubSpot: Human-Led, AI-Accelerated Talent Acquisition
  
  
  SEO Summary
  CHROs must lead AI adoption's toughest stage — surfacing undocumented organizational knowledge before engineers define the future of work without them.
  Keywords: CHRO, AI adoption, human capital strategy, organizational knowledge, AI implementation, HR leadership, workforce transformation, CHRO mandate, talent strategy, AI systems

Transcript

Most companies assume the hard part of AI is the technology. Getting the tools, running the pilots, measuring the results. And that's hard. But the organizations that are pushing past tools — the ones trying to build AI systems that can actually run parts of how their business operates — they're hitting an entirely different wall. And when you sit down to teach an AI system how your organization makes decisions, you end up finding out that nobody ever wrote that down. And some of it was never written down on purpose.

Hey there, senior leader, and welcome to the Talent Sherpa Podcast, where senior leaders come to rethink how human capital really works. I'm your host, Jackson Lynch, and today we're going to be talking about a part of the AI adoption process that no one really describes honestly.

Right now, most of what you hear about AI is about tools — which ones to buy, how to get your people using them, how to measure the productivity gains that come with them. And that conversation is real and it does matter. But it's just one step in a much longer journey. The next step, I think, is harder. The organizations that are further along — they're not just helping their people work faster. They are building AI systems that can handle organizational work on their own. Route a client request to the right team, or process an approval without a person touching every step. Run a workflow from intake to resolution based on your organization's own bespoke rules. That's a different kind of AI adoption entirely.

And to get there, you have to do something that sounds pretty straightforward, but it turns out to be anything but — you have to teach the AI system how your organization actually makes decisions. What the rules are. How exceptions get handled. Who has the authority to sign off on what. What happens when something doesn't fit the standard path. And the moment you try to do that, you find out how much of it was never written down in the first place.

So if you are a CHRO or a senior HR leader watching your organization start to move in this direction, this episode is going to tell you what's actually happening and where you fit. And by the end — whether you're an HR person or a senior leader, a C-suite executive, a board member — you will understand why this work is a human capital project, and what to do before the engineers show up asking for documentation that doesn't exist yet.

And hey, before we get into today's episode, I want to say a huge thank you and shout out this week to Rodney from Louisville. Thank you for being a part of this community. And for everyone who is tuning in — whether you're joining from Bangor, Pennsylvania, or Brighton, Massachusetts — I just want you to know I appreciate you being here. I'm really glad you're with us.

And let me also ask you to take 30 seconds and think for a second about a performance problem that you're most focused on right now. And now take a second and ask yourself — is this about a person, or is this about the system that is around that person and how it was designed?

In fact, one of the graduates from the most recent CHRO Ascent Academy said the org design modules were the most immediately relevant, and they said, "We applied them the very same week." The Ascent Academy teaches CHROs to read the difference before they act. CHRO Ascent — and you can learn more at mytalentsherpa.com. Let's find some time to see if we might be able to help you.

All right, let's get into it.

Now, HubSpot's Chief People Officer, a woman named Helen Russell, published something on LinkedIn recently that I keep coming back to. It was really good. She described three stages of what organizational AI adoption looks like. And it's the most honest account of how this actually unfolds that I've seen so far.

The first stage was about access. You give everyone the tools, create the space to experiment, track who's actually using what. And by the end of that stage, according to HubSpot, 94% of their employees were using AI every week. And here's what I find remarkable: people across the company had built more than 3,900 of their own small AI tools to make their individual jobs easier — and they did that on their own. I would argue that happened because the organization created the conditions for it. That's a culture shift. It's not a technology deployment.

The second stage was where individual productivity started producing real business results. HubSpot mapped their teams — who was ready to go deeper and who needed more time — and invested differently based on where each team actually was. Now I'm going to call this out for HR folks: we tend to want to treat everyone the same. This is another example of where human capital needs to become more capitalistic than socialistic. And that's exactly what HubSpot did. Their recruiting function alone cut 10 days off their time to hire and automated 80% of their scheduling. That's measurable and that's real.

Then the third stage is where they are now, and it's the one that most organizations haven't hit yet. But if everything goes according to plan, they will. And Russell described it as building AI systems that can take your organization's institutional context — their judgment, their processes — and work across the organization based on it.

Let me make this plain. An AI tool helps a person do their job faster. An AI system can do parts of the job on its own — handle process, make routing decisions, execute a workflow — if it knows the rules your organization runs on. The third stage is the work of teaching an AI system what those rules are.

And here's where it gets complicated. Your process documentation — the policy manuals, the org charts, the how-to guides, to whatever degree you have them — they describe how your organization was designed to work. They don't describe how your organization actually works. And in most organizations, these are completely different things. The gap between them is large, and it lives in your people and in the wraparound processes no one has ever written down.

It lives in the person who's been around long enough to know that a certain type of client request needs to go directly to the VP, not through a standard queue. It lives in the arrangement two departments worked out three years ago that no one formally documented because everyone involved just kind of knows. And it lives in the manager who can tell from a single line in an email that a situation needs attention now, and who to loop in before it escalates.

Every organization has things like this. And in most organizations, that knowledge has never been written down. An AI system can't work from knowledge that only exists in people's heads.

Now, a few episodes back — the one Scott and I did, "The ROI Was Never the Tool" — I made the case that the returns from AI investment come from what you build organizationally around the tools, not from the tools themselves. The work of teaching AI systems how your organization actually operates is where that truth becomes impossible to avoid. You cannot build on what no one has yet surfaced.

So here's where I think it usually breaks down.

The first trap is treating this as a documentation exercise. The assumption is that the knowledge is somewhere — in someone's files, an old process manual, a project workflow, maybe a wiki from a few years ago — and the work is just finding it, organizing it, and handing it to the engineering team. That's not how it works. The engineers show up, ask for the documentation, and there isn't any. Or there is, but it describes how the process was designed, not how it actually runs. That gap is where these initiatives start getting stuck.

The second trap is more uncomfortable. Some of what was never written down was not an oversight. Every organization carries informal systems that everybody uses but no one really talks about. The account that gets handled differently because of a relationship that predates the current leadership. The approval step that certain people skip because the person who built that process left two years ago, and the shortcut quietly became the new norm. The informal hierarchy that everyone navigates around because it's faster than following the official structure.

When you try to write down how your organization actually works so an AI system can follow the rules, all of that surfaces. The tendency at that point is to call it a technology problem — a data gap that the engineers can solve. But it's not. The technology is usually ready. The organizational part — that's the part that's not ready. And the people who need to lead this work are not the ones writing the code.

The third trap is the assumption that this is just documentation work, and once it's done, you can hand it over. You can't. You have to make choices. Do you teach the AI system to operate the way things actually get done — shortcuts and all? Or do you fix the underlying process first and then build the system on top of the corrected version? Do you keep a person in the loop for certain categories of decisions because the consequences of getting it wrong are too high? Each of these is an organizational design choice. They're not engineering choices. They don't belong to the engineering team.

Here's the shift as I see it. The work of teaching AI systems how your organization actually operates is hard because it forces every organization to answer a question it's been able to avoid until this moment: what is this organization actually running on? What actually governs the decisions, the exceptions, the handoffs, the day-to-day work? The real rules — the ones that determine how things move and who decides what — rather than the version someone wrote down 24 years ago that's somewhere in a binder.

Surfacing those rules is something like organizational archaeology. You have to dig past the official record to find out how things actually work. And that means going to the people who carry that knowledge, because it's never been anywhere else.

There are two kinds of missing knowledge in most organizations. The first is knowledge that was always supposed to exist on paper but has gone stale. Process guides written before the latest restructuring. Policies that reflect how a team used to operate. That's a documentation gap. It's manageable.

The second kind is harder. This is the knowledge that was never meant to be written down — not because anyone decided to hide it, but because the people involved never needed to formalize it. It just worked. Everybody knew. There was no reason to document it. That knowledge lives in the habits and memory of your most experienced people. It functions — until somebody leaves, or until you need to teach a system to do what they do. And here we are.

The work of building AI systems that can handle organizational processes forces that knowledge out into the open. And when it comes into the open, someone has to make decisions: what gets handed to the AI system, what gets redesigned before it gets handed over, what stays with the people permanently — because the judgment involved depends on context, relationships, or stakes that are too significant to automate.

These decisions touch how future roles are designed. Who has the authority to decide what. What kind of judgment the organization wants to preserve in their people rather than delegate to systems. That is the domain of human capital. It always has been. AI systems just made it urgent.

So here's what changes when you see it that way. The CHRO who understands this is positioned to lead the work. Human capital holds the human systems where this knowledge lives. The work of surfacing undocumented judgment, examining it rigorously, and deciding what gets preserved versus redesigned — that is organizational development work at its most consequential. Which is, by the way, not what most OD groups do. And the CHRO who claims it early is ultimately going to shape the outcome.

The organizations that move through this fastest have already built the habit of capturing what their people know as part of how they develop and deploy talent. The know-how was already in a form that could be handed to an AI system. The organizations that are stuck — sitting on years of undocumented judgment — the technology is ready. But the organizational foundation is not. Until someone with the right mandate steps in to lead that work, the AI systems are going to keep waiting for inputs they cannot get.

That is the real gap, and it belongs right here in Human Capital's lane.

Now, before your organization's next AI initiative, here's what I would do.

First, map how the work actually runs before anyone looks at your documentation. Most organizations start this work by pulling in their existing documentation and handing it to a project team — usually with technology people. That produces a map of how the organization was designed to operate. What you actually need is a map of how it does operate. So pick the highest-priority AI initiative on your list. Before the engineers touch it, find two or three people who've been around long enough to know how things are actually getting done — the people who know the shortcuts, the exceptions, the relationships that shape how the work moves. Talk to them. Ask what the documentation doesn't capture, or better yet, have them whiteboard it from scratch. The gap between what they describe and what the official record says is your real starting point. And that's what the AI system is going to need to start learning.

Second, get HR into decisions about what AI systems are going to handle before those decisions get made. The timing here is important. Every decision about what an AI system does is an organizational design decision — which calls stay with people, which processes get redesigned before they get handed to a system, which categories of judgment the organization keeps in human hands. Human capital has the tools and the experience to lead those conversations. The question is whether you're in the room before the decision is set, or brought in afterwards to manage what was decided without you. Get ahead of one AI initiative right now. Make sure the organizational design questions get answered before the engineering work makes them harder to revisit.

Third, build a framework for what your organization will and won't hand to an AI system — and bring it to your C-suite and your CEO. Some decisions belong with people. Decisions that depend on knowing the individual on the other side of the call. Decisions where the right answer turns on factors no system can read. Decisions where getting it wrong requires a human to be accountable. Irrevocable bets. Build that framework and bring it to your senior team as a design guide. The impact: it positions the human capital function as the group that knows where to draw the line and why. That's a materially different role than showing up to pump the brakes.

I realize I just handed you three plays on top of a week that already had too many things in it. The knowledge mapping conversation in particular is the one that gets pushed. It feels like extra work layered on top of a project that already has a team and a deadline. It's easier to let the engineers surface the gaps on their own and fill in the holes later. The problem is that by the time the engineers surface the gaps, the framing is already wrong. It becomes a data problem. It should have been an organizational design conversation.

If you take one thing from today, let it be this: teaching AI systems how your organization actually operates forces every organization to confront what it was really running on. And that reckoning is a human capital project. Which means it belongs in our lane.

Thank you for spending time with me today. I really appreciate you being a part of this community of senior leaders who want to rethink how human capital really works. If you're thinking about how to apply this in your own situation, check out getpropulsion.ai — they have AI teammates that can help your leadership team focus on the work that actually drives business outcomes. And if you're a first-time CHRO or preparing to step into the role, I'd love to work with you. All the information is on my website at mytalentsherpa.com. You can also read everything on my Substack at talentsherpa.substack.com.

Thank you for being here. Until next time — keep raising the bar, keep the organizational archaeology moving forward, and keep on climbing.