TBM 417: Before You Fire All Your Glue People Because of AI

I guess I’m putting together an AI series of sorts. This is #5.

1. Hope, Context, and Control

2. Seeing Everything, Understanding Nothing (The Context Trap)

3. Institutionalized Overload (Now With AI)

4. Legibility and Legitimacy

5. Before You Fire All Your Glue People Because of AI

6. Simple-player to Mutliplayer: Context Through Collaboration

At the day job, I’m deep in the question of how AI can help teams do more effective work, hence all the thinking below. It has highs and lows these days, with enough inspiring stories to keep me curious and hopeful. If you and your team ever want to chat about this, reach out.

---

The hardest part of figuring out how to use AI at work is that our heuristics are right often enough to feel reliable, and wrong often enough to cause real damage.

We develop mental models for where AI fits. Those models work in some contexts. Then we carry them into contexts where they don’t work, and because the logic sounds the same, we don’t notice the difference until it’s too late.

1. People lose their jobs based on these calls.

2. Teams lose capabilities they didn’t know they depended on.

3. Organizations hollow out functions that took years to build, and discover the consequences months later when the problems are structural and the people who could have helped are gone.

The goal of this post is to help you make better calls about where to focus with AI. Where should you lean into the heuristic? Where should you be skeptical of it? What should you look for before committing?

And I want to make a passionate plea: before your organization goes off and fires all of its “glue people” on the assumption that AI can pick up what they were doing, please stop and think about what those people were actually doing. I have watched companies make this move. The ones who got it wrong didn’t realize it for months, and by then the damage was structural. You owe it to yourself, your teams, and the people who held your organization together to understand what you’re really replacing before you replace it.

And I want to make a passionate plea: before your organization goes off and fires all of its “glue people” on the assumption that AI can pick up what they were doing, please stop and think about what those people were actually doing.

The Basic Thesis

There are things we all know we should be doing at work but aren’t. AI can help with many of them. Follow that instinct.

But some of that work was only happening because specific people made it happen, and what those people were doing was far more than it looked like from the outside. The risk right now is that we replace the people, keep the visible artifacts, and lose everything underneath that made those artifacts matter.

This post is about how to tell the difference.

The “Should Do” List

In our work we are surrounded by things we know would help us take our work, and our company’s work, up a couple notches. But things get in the way.

We know we should cross-reference risks in our docs. We know we should keep a well-curated research library with insights linked to decisions. We know we should write good launch notes for our customers. The list goes on. .

And just like in our personal lives, all the stuff we know we should do tends to land in the “aspirational, but the real world got in the way” category.

In some orgs, those things happen sometimes:

Some people bend the curve because they are hired to do that thing, or because they really, really cared about it (and are good at it).

Before AI there were people who curated amazing personal knowledge management systems in Obsidian. They took atomic notes. They Got Things Done (GTD). But this was a rare breed. It was people who really cared about those things, who were particularly dialed in to doing that work.

Every company has those people who valiantly went above and beyond even when it wasn’t their job description: updating docs, closing the loop on outcomes, sending the thank-you note to the customer, dotting their i’s, spending that extra time on great facilitation.

And then there were the superstars, often in short supply, who did things as part of their job description that literally no one else would do because it was “not in their job description.” I know program managers who were magicians at aligning multiple teams towards an outcome without cramping everyone’s style. The product managers loved them because they were doing all the dirty work that didn’t fit in the time anyone else had, in ways that didn’t map to how anyone else identified with doing a good job.

And then there were the superstars, often in short supply, who did things as part of their job description that literally no one else would do because it was “not in their job description.”

In both the case of the informal advocates and the formally appointed role-holders, companies often have a love/hate relationship with the idea that there are people who “glue” everything together. As of late, glue people are getting laid off in many companies because they are the visible externalization of dependencies. Some people imagine AI can simply replace the activities that the informal unicorns and formally appointed connectors were doing.

Which brings us to the topic at hand.

The Heuristic

I’ve been saying recently that people are overthinking how to use AI. I say, very simply, that they should use AI to do the things they know they should do and aren’t doing. Don’t overcomplicate it.

Start with simple examples.

We know we should tailor release notes for different customer segments. That takes time. We don’t have time. So most companies don’t do it. Can AI do this well? Yes. OK. Do it.

Another one: we know we should write better updates and capture more context on our initiatives instead of letting stuff simmer in different systems, only to be forgotten and never updated. Can AI make this a more organic activity? Can it remind us to do it, nudge us, make it easier to spot things that are out of date, recontextualize it for different audiences? Yes. OK. Do it.

These seem relatively straightforward because our mental model is: we know we should do it, we know things get in the way, therefore we believe that using AI to assist us would be beneficial. And in a lot of cases that instinct is right.

Where It Gets Interesting

But let’s see how these behaviors chain together.

I have worked with companies that DID value those things, and assembled enough informal advocates and formally appointed job-doers to make them happen. They embedded the underlying behaviors and norms. They saw the benefit, so they kept on it. Without the magical elixir of AI.

And realistically, sometimes it “worked” and sometimes it didn’t. The constraint shifted. “I’m a leader, I don’t have enough time to read this.” “Nothing we are shipping that we are writing the release notes about is all that smart. Why did we prioritize it?” “Even with all of this data and information, it turns out no one is actually using it.”

When Your Instincts Are Right

The heuristic works when the actual barrier to the behavior is friction and cost. Specifically:

When the behavior is already understood and practiced, but painfully.

The activity is mandatory or deeply valued. People do it. But the environment fights them. It takes too long, the tooling is clunky, the workflow has too many steps. The person doing it is essentially absorbing infrastructure cost with their own labor. AI legitimately reduces that cost. Tailoring release notes for customer segments, keeping project status documents current, cross-referencing risk logs. When the only thing standing between “we should” and “we do” is the hours it takes, AI is a direct solution.

When the behavior is valued and agreed-upon, but consistently displaced by competing priorities.

Everyone agrees this matters. It keeps getting bumped. The work isn’t hard or mysterious. It’s just never the most urgent thing. Here AI works because it lowers the effort cost enough that the behavior survives contact with a busy week. It doesn’t need to become anyone’s top priority if it only takes a fraction of the time it used to.

In both of these cases, the instinct is correct: the gap between “should” and “do” really is about cost and time, and AI genuinely closes that gap.

When Your Instincts Break Down

The heuristic breaks when the actual barrier to the behavior isn’t friction and cost, and you don’t realize it.

• When the behavior exists only in pockets, carried by specific people.

• When the behavior is “agreed and valued” but the system doesn’t actually reward it.

• When the behavior was aspirational and nobody actually knows what “good” looks like.

• When systemic forces were actively preventing the behavior.

When the behavior exists only in pockets, carried by specific people.

It looks like the gap is effort: “If only we had more time, everyone would do what Sarah does.” But the reason Sarah does it and others don’t isn’t usually time. Sarah reads the room. She knows when to surface information and when to hold it back. She has the trust and relationships that make her updates credible. She understands the political landscape well enough to frame things in ways that land. She builds shared understanding across teams that don’t speak the same language. She has judgment about what matters that comes from years of pattern recognition.

When you “use AI to do what Sarah did,” you replicate the visible artifact (the update, the summary, the aligned roadmap) while losing the judgment, social navigation, and legitimacy work that made the artifact useful. AI produces the document. Nobody acts on it. Not because the document is bad, but because Sarah wasn’t just writing documents.

When the behavior is “agreed and valued” but the system doesn’t actually reward it.

Everyone says cross-referencing risks matters. But the incentive structure rewards shipping features. The performance review doesn’t mention documentation quality. The people who did maintain risk logs were either intrinsically motivated (and eventually burned out) or occupied a formal role that justified the time (and is now being cut). AI makes it cheaper to produce the risk log, but the risk log enters the same system that never rewarded maintaining it. The constraint wasn’t production cost. It was that the organization doesn’t actually value the output enough to consume it. Making it free to produce doesn’t change that.

When the behavior was aspirational and nobody actually knows what “good” looks like.

“We should be more data-driven.” “We should have a learning culture.” “We should close the loop on outcomes.” These are identity-level desires, how we want to see ourselves, without agreed-upon practice, defined skill, or clear specification of what it means to do it well.

AI can produce something that looks like the aspiration: a metrics dashboard, a retrospective summary, an outcomes report. But without shared understanding of what good looks like, without the capability to interpret and act on the output, and without organizational agreement on what counts, you get artifacts that perform the aspiration without delivering the outcome. Worse: the presence of the artifact can create false confidence that the gap has been addressed.

When systemic forces were actively preventing the behavior.

Sometimes the reason people aren’t doing what they “should” do is that the organization, through its power dynamics, political structures, or implicit rules, makes it unsafe or unrewarding to do so. The person who raised concerns in the risk log got punished for slowing things down. The team that invested in documentation was told they weren’t shipping fast enough. In these cases, AI doesn’t remove the barrier. It may actually increase the exposure: now there’s a visible AI-generated artifact pointing at problems the organization has been structurally avoiding.

The Constraint Shift

The deepest version of the break is this: even when AI successfully produces the artifact, the constraint shifts rather than dissolves. Companies that did invest in the “should do” behaviors, with real people, real effort, real norms, discovered this repeatedly:

• “We have beautiful release notes. But nothing we’re shipping is worth writing release notes about.”

• “We have the data. But leadership doesn’t have time to engage with it.”

• “We aligned the teams. But the prioritization didn’t support what we aligned on.”

• “We captured the context. But nobody’s workflow includes consuming it.”

AI makes the first link in the chain nearly free. But the chain has many links, and the downstream links have their own barriers. Barriers that have nothing to do with production cost. They involve capability (can people interpret this?), motivation (do they care enough to act on it?), social dynamics (is it safe and rewarded to act on it?), and system design (does the workflow even include a moment where someone would encounter this?).

The heuristic “do what you should do” ignores all of this (in many situations). It addresses the production side and is silent about the consumption and action side.

Applying Lenses Around Capability, Opportunity, and Motivation

Where we apply COM-B.

Everything above can be sharpened by asking three questions about any behavior you’re considering handing to AI. Is it a capability problem? Is it an opportunity problem? Or is it a motivation problem? The heuristic assumes the answer is almost always opportunity (not enough time, too much friction). When it’s wrong, AI solves the wrong problem.

Capability

Do people know what this behavior actually involves?

For well-practiced behaviors, everyone has a shared understanding of what the output should be. AI can hit that target because the target is defined. But for behaviors that exist in pockets are purely aspirational, there is no shared specification. Different people mean different things by “close the loop on outcomes.” AI will produce something, but without a shared definition of what counts, nobody will trust it.

People also consistently mistake the visible output of a behavior for the behavior itself.

When Sarah sends a well-timed update that keeps three teams aligned, the visible part is the update. The actual behavior is reading the situation, judging what matters to whom, framing it so it lands, choosing the right moment, knowing who needs a private heads-up first. The thinner the organization’s mental model of what a behavior involves, the more likely they are to believe AI can replace it.

And there’s a subtler trap: can people even evaluate whether AI’s output is good?

For well-understood behaviors, yes. You know what a good release note looks like. But for complex, judgment-heavy behaviors, the people evaluating the AI’s output are often the same people who don’t fully understand what the behavior involves. The organization lacks the capability to do the behavior, asks AI to do it, and also lacks the capability to tell whether AI did it well. The gap doesn’t close. It becomes invisible.

Opportunity

Is the environment set up for this behavior to succeed?

Physical opportunity is the heuristic’s home turf. Time scarcity, workflow friction, clunky tooling. When these are the dominant barriers, AI is a direct fit. Take the win.

But the system also needs to consume what AI produces. Even pre-AI, companies that invested in these behaviors found the constraint shifted to feedback loops and visibility. The artifact existed, but nobody’s workflow included a moment to encounter it. AI makes production nearly free. If the workflow doesn’t include consumption, you’ve accelerated the first step of a broken pipeline.

Social opportunity is where the deepest breaks live. Many “should do” behaviors don’t have clear ownership. They lived in gaps between job descriptions. AI can produce the output, but production without ownership means nobody is accountable for whether it gets used. Organizations also frequently endorse behaviors rhetorically that they punish operationally. “We value documentation” but we promote people who ship fast. Making the endorsed behavior cheaper doesn’t change that calculus.

And some behaviors aren’t happening because the organization makes them risky. Surfacing problems, flagging misalignment, questioning priorities. The person who did this before had enough social capital and trust to do it safely. An AI-generated artifact has none of those buffers. A risk report from Sarah, who has ten years of trust with the VP, lands differently than the same report from an AI tool. The information is the same. The social permission is completely different.

Motivation

Why were people doing this in the first place?

The heuristic targets effort cost. But motivation also includes: Do people believe this will produce a good outcome? Do they see it as part of their identity? Is there a competing habit that crowds it out? Each of these can be the binding constraint, and AI-as-effort-reducer doesn’t touch any of them.

Some behaviors were carried by people who identified with doing them. When AI takes over the output, the organization loses the motivational engine that made the behavior happen with care and judgment. And in organizations where “should do” behaviors have been repeatedly championed and then abandoned, people develop a rational skepticism. They’ve learned that the organization doesn’t follow through. Cheaper production won’t convince them. Demonstrating that the systemic conditions have changed might.

How the picture shifts

The critical insight across all three lenses: as you move from well-practiced-but-painful behaviors toward aspirational-or-contested ones, the dominant barrier shifts from physical opportunity to a combination of capability, social opportunity, and motivation. The heuristic “use AI to do what you should” only addresses the physical opportunity layer. It is silent about the rest.

For behaviors where the capability exists, the social environment supports it, and the motivation is present except for effort cost, AI is a near-perfect intervention. For behaviors where the capability is thin, the social environment is ambivalent or hostile, and the motivation was carried by specific individuals rather than organizational systems, AI produces artifacts that enter a system unprepared to use them. The artifacts look like progress. The underlying conditions haven’t changed.

The Glue-Person Risk

This brings us to the most consequential version of the problem. The people who carried these behaviors, the informal advocates, the formally appointed connectors, the “glue,” were rarely doing what it looked like they were doing.

From the outside, it looked like they were writing updates, maintaining docs, sending summaries, and running alignment meetings. Activities. Tasks. Things that AI can plausibly do.

From the inside, they were:

• Reading weak signals about when an intervention would land and when it would be resisted. Timing and context that requires years of accumulated pattern recognition.

• Navigating legitimacy. Knowing whose endorsement was needed to make an artifact credible, when to surface something publicly versus privately, how to frame a message so it wasn’t threatening.

• Building shared language across roles and teams that organize reality differently. Translating between engineering’s mental model and sales’s mental model and leadership’s mental model.

• Running informal feedback loops. Noticing when something was drifting off course, deciding whether it was worth intervening, calibrating the right amount of correction.

• Absorbing role ambiguity. Stepping into gaps between job descriptions that nobody had formalized, taking ownership where ownership was unclear.

• Making the invisible visible. Surfacing friction, cost, and coordination failures that the organization had normalized and stopped noticing.

When organizations say “AI can do what the glue people did,” they are looking at the output layer and missing everything underneath. The risk isn’t just that the AI-produced artifacts will be lower quality. It’s that removing the people and assuming AI covers it eliminates the judgment, social infrastructure, and adaptive capacity that made those artifacts matter, while creating the appearance that everything is still being handled.

The cruelest version of this: the glue person’s work was often invisible because they were good at it. Things ran smoothly. Information flowed. Teams stayed aligned. Nobody noticed the work because the work’s purpose was to prevent problems from becoming visible. Remove the person, replace the visible artifacts with AI, and for a while nothing seems different. The problems surface months later, disconnected from the cause, and nobody connects the breakdown to the removal of the person whose job was to prevent exactly that kind of breakdown.

The cruelest version of this: the glue person’s work was often invisible because they were good at it. Things ran smoothly. Information flowed. Teams stayed aligned. Nobody noticed the work because the work’s purpose was to prevent problems from becoming visible.

So What Do You Do?

The heuristic is still good. Use AI to do the things you know you should do. Start there. But hold two things alongside it:

Diagnose before you automate.

Ask: why isn’t this happening? If the answer is genuinely “it takes too long and we don’t have time,” go. That’s the sweet spot. If the answer involves “nobody’s really sure what good looks like,” or “the org doesn’t actually reward this,” or “it works when Sarah does it but not when anyone else tries,” the barrier isn’t cost, and AI-as-effort-reducer won’t solve it. You need to understand the actual gap before you can know whether AI addresses it.

Respect the iceberg.

When someone was carrying a behavior successfully, formally or informally, assume you’re seeing 20% of what they were doing. The artifacts were the visible output. The judgment, relationships, timing, framing, and political navigation were the invisible 80% that made the output useful. Before you replace the person with AI-generated artifacts, understand what the other 80% was and what happens to the system when it disappears.

Watch for constraint shifts.

When AI makes something easy to produce, ask: what was the next bottleneck, even before AI? If companies that were already doing this work still hit limits (and they did), then making the first step free doesn’t clear the path. It just moves you faster to the next wall. Know what that wall is before you celebrate solving the first problem.

The advice is sound. The instinct is often right. But “do what you should do” is only as good as your understanding of why you weren’t doing it in the first place.