When Answers Become Cheap
You walk into a meeting. Three proposals sit in front of you, each one polished, each one backed by sophisticated analysis. The dashboards look good. The metrics align. Every option sounds reasonable.
But something feels off.
You push on the assumptions. You probe the trade-offs. The answers come quickly, almost too quickly. The reasoning is coherent, the justification thorough. But you can’t locate the struggle. You can’t see what the team agonized over, what they fought for internally, what alternatives they seriously considered before landing here.
Because the hardest part of the reasoning didn’t happen in the room.
They reviewed it. They agreed that the justifications seemed right. But do they truly own it? Are they staking their reputation on this specific path because they believe in it, or because it seemed plausible enough to present?
You can’t tell. And that uncertainty changes everything.
The decision has become yours alone, not because you have the authority, but because you can’t find the conviction in the team that used to translate into shared accountability.
When answers become cheap, conviction becomes the scarce resource.
For the last few years, we’ve been promised a particular future of work. It’s a solitary future where AI becomes the ultimate assistant. Individuals become exponentially more capable, where one person with the right tools does the work of many. In this story, progress looks like independence. The future belongs to the Super IC, the solo genius superhuman with all the right tools.
But inside organizations, something different is happening.
Teams are not becoming more independent. They’re becoming more entangled. The more capable AI becomes, the more people seem to need each other.
Look at what’s happening in practice: nearly every organization experimenting seriously with AI now has channels where people share prompts, compare outputs, troubleshoot approaches. These aren’t casual water-cooler conversations. These interactions carry weight. People are seeking validation on critical decisions: which architectural approach to take, which analysis to trust, which generated strategy to present to leadership. The reliance on peer input has quietly escalated.
At first this feels backwards. Shouldn’t better AI tools reduce reliance on others?
Instead, many leaders describe a strange new pressure. Decisions feel heavier and more difficult even as answers arrive faster.
The New Reality: Plausibility at Scale
AI doesn’t simply retrieve information. It generates plausibility. Every architectural path sounds reasonable, every strategy has a coherent story behind it, every proposal arrives wrapped in confidence. Where organizations once struggled to generate options, they now struggle to choose among many viable ones.
Meetings have changed. Teams move faster through ideation but slower toward decisions. People leave with more clarity about what could work, but less conviction about what they should do.
The problem isn’t too little information. It’s too many plausible paths. When AI can generate sensible justifications for multiple options, the bottleneck shifts. Organizations no longer struggle to produce solutions. They struggle to choose among options that all seem defensible.
This is cognitive abundance: when well-reasoned possibilities multiply faster than the organization’s capacity to evaluate them. The friction isn’t ignorance. It’s excess plausibility.
This observation isn’t theoretical. Companies experimenting with AI have created dedicated channels (Slack workspaces, internal forums) where people share what they’re doing with AI and seek input from peers.
The tension shows up most clearly in frontier AI organizations. Recent research from Anthropic studying their own workforce describes what they call “delegation intuition”: the learned judgment about what to hand off to AI, what to retain, and where human oversight remains essential.1 What’s striking is that Anthropic explicitly worries about where reasoning happens. Their teams have established “social proof” rituals requiring people to share not just outputs, but the reasoning and constraints they provided to the AI, in public channels for peer validation. This isn’t about catching technical errors. It’s an attempt to prevent critical judgment from becoming invisible, quietly outsourced to systems whose conclusions are plausible and therefore easy to accept without collective scrutiny.
What’s notable isn’t just that these spaces exist, but that organizations feel the need to establish governance around them. While these patterns may be more concentrated among early adopters, they’re consistent enough to suggest a structural shift rather than isolated experimentation.
What’s quietly changing isn’t how we generate ideas. It’s what organizations now reward, authorize, and hold people accountable for. Authority is beginning to shift, not toward better answers, but toward those willing to carry responsibility for choosing among them.
What Senior Leaders Quietly Feel
The experience described at the opening becomes recognizable to many executives as they navigate this shift. There has never been more input, more analysis, more possibility. And yet the act of deciding feels lonelier.
Because the real question is no longer “What is the right answer?”
It becomes “Which answer am I willing to stand behind?”
That shift matters. It reveals that organizational judgment is not primarily a cognitive task. It is a responsibility.
The Mistake: Treating Judgment as a Capability Debate
When a leader makes a decision, they’re not running an evaluation algorithm. They’re stepping into a role that extends beyond the meeting room. The decision carries their name. If it fails, they answer for it to their team, to the board, to the market. The consequences aren’t just analytical. They’re political, ethical, financial, human.
Judgment isn’t choosing the option with the highest expected value. It’s accepting responsibility for the future that choice creates, including the futures that weren’t chosen and the people affected by them.
Even if AI could model every outcome with perfect accuracy, it can’t carry that weight. Responsibility doesn’t get delegated to a system. Legal frameworks may eventually evolve to distribute AI liability differently. But inside organizations today, when a decision fails, someone’s phone rings. Someone walks into the difficult meeting. That hasn’t changed.
AI didn’t create weak decision hygiene. It exposed it.
Someone has to own it.
That’s the problem better AI doesn’t solve. When reasoning arrives pre-packaged, the distance between “this seems right” and “I’m willing to stake my credibility on this” becomes a chasm.
The Un-Delegatable Burden
This reveals the greatest risk in the age of plausibility glut: the Liability Gap. The distance between the machine that invokes the action and the human who must own the consequence. And it’s widening.
Consider autonomous vehicles. When a self-driving taxi causes an accident, the machine made the split-second judgment to swerve. But accountability rests upstream with the leaders who defined the safety parameters, curated the training data, and set the thresholds of acceptable risk. Responsibility hasn’t vanished. It has moved from the driver’s seat to the executive suite.
The same shift is happening inside organizations. As AI takes over more of the strategic “driving,” the leader’s role shifts from practitioner of expertise to curator of liability. You are no longer just deciding what is right. You are deciding what you are willing to be ethically, legally, and professionally responsible for when the driverless strategy hits an unforeseen obstacle.
Responsibility is not a task you can optimize. It is a burden you must carry. No AI can be summoned to testify before a Senate subcommittee or answer to the Board. The anchor, the role that remains accountable when everyone else disperses, provides the one thing the machine never can: someone who has to answer for it.
The Quiet Structural Shift
For decades, organizations built themselves around knowledge scarcity. Information was hard to access, expertise was rare, and the people who had power were those who knew not just what to do, but when and how to apply it. You rose by developing judgment that only came from deep domain mastery. Hierarchy reflected who had earned that expertise.
That logic is breaking down.
When AI can generate sophisticated analysis on demand, when any team member can produce a well-reasoned memo on a complex topic, the visible markers of expertise become harder to distinguish. The person with access to the best prompts can produce output that rivals the expert’s. The junior analyst’s AI-assisted work can look as polished as the senior strategist’s.
This doesn’t mean expertise or judgment disappears. But demonstrable expertise alone no longer determines who gets to decide.
Authority is shifting, quietly and unevenly, toward something else: responsibility. Who is authorized to make the call? Who is accountable when things go wrong? Who carries the consequences?
Without anyone explicitly planning it, organizations are moving from knowledge hierarchies to responsibility hierarchies. Roles start to mean something different. They stop being primarily about what you know and start being about what you’re accountable for. The title matters less because of the expertise it signals and more because of the decisions it anchors.
Why People Reach for Each Other
This creates a real tension.
On one hand, the space of possible solutions is exploding. AI can generate architectural options, strategic frameworks, and analytical approaches faster than any individual can properly evaluate.
On the other hand, responsibility remains stubbornly personal. When the decision goes wrong, it’s not “the organization” or “the process” that answers for it. It’s the person whose name is on it.
The instinctive response isn’t to retreat into isolation and try to evaluate everything yourself. You’d drown in options. Instead, people reach for each other differently. They form communities around the act of evaluation itself.
The shift doesn’t announce itself with a strategic initiative. It starts with friction that’s hard to name. Teams re-litigate the same decisions. A strategy gets agreed to in one meeting, then quietly questioned in the next. Someone brings new AI-generated analysis and suddenly the settled question feels unsettled. The organization isn’t moving forward faster. It’s churning.
At some point, people start adapting without being told to. Someone creates a Slack channel to share AI approaches. A few people start meeting informally to compare how they’re evaluating proposals. Small groups form around specific problem areas, not because anyone mandated it, but because triaging all the options alone became exhausting.
These aren’t normal project teams with assigned deliverables. They’re more like guilds: loose collections of practitioners who share a problem space and establish collective standards without removing individual accountability. Spotify’s scaling model, now over a decade old, made this structure visible.2 Squads were autonomous delivery teams, but guilds cut across them as voluntary communities where people with similar challenges could share knowledge and practices. The guild’s job wasn’t to make decisions for squads. It was to help squads make better decisions.
The Slack channels and hallway huddles emerging today serve a similar function. They compare approaches, pressure-test assumptions, challenge reasoning. They build collective intelligence without diffusing responsibility. The decision-maker is still on the hook, but can draw on the community when the space of plausible options exceeds what any individual can reasonably navigate alone.
This becomes standing infrastructure. Ongoing communities where people develop shared judgment, build common language around what good evaluation looks like, and calibrate their thinking against each other over time.
But when a high-stakes decision emerges, the pattern intensifies. People don’t just consult the community. They mobilize around the problem itself.
Role-Anchored Swarms
The role stays fixed. Someone owns the decision. There’s a clear name on it, a person who will answer for it if it goes wrong.
But the group that engages with the problem becomes fluid. When a consequential question surfaces, people swarm it. Not because they’re assigned to it, but because they have perspective, because they see risk the role-holder might miss, because they care about the outcome.
An engineering leader at a Series C SaaS company with around 150 technologists described this pattern emerging organically. When asked how they knew AI adoption was working, they pointed to their Slack channels. Not because people were sharing productivity tips, but because of what the conversations revealed.
The team was discussing a recent scaling issue, and the thread quickly moved beyond the technical solution into something more revealing. People were sharing the context they’d provided to Claude, comparing the different options it had generated, debating the criteria for evaluation. Senior staff developers weren’t just reviewing the proposed solutions. They were challenging assumptions embedded in the AI-generated proposals that weren’t actually part of the constraints.
“We realized,” the leader explained, “that the real conversation wasn’t about which solution to pick. It was about recognizing what hidden assumptions we’d adopted by letting the AI frame the problem in the first place.”
The swarm doesn’t follow the org chart. It might include someone two levels down who has deep context, a peer from another function who faced something similar, someone newer who asks the naive question that turns out to matter.
They explore the problem space together. They pressure-test assumptions. They surface options that weren’t in the initial framing. They challenge the reasoning, not to undermine the decision-maker, but to expand their field of vision.
Then the swarm disperses. The decision-maker decides.
The goal was never consensus. It was error correction. It was making sure that the person who carries responsibility for the decision had access to the collective intelligence of people who actually engaged with the problem, not just the polished analysis that survived the filtering process.
On paper, the decision looks individual. But it rests on two layers of invisible infrastructure: the standing guilds that build shared judgment over time, and the dynamic swarms that mobilize around specific decisions. Neither is documented or officially sanctioned. Yet both are becoming essential to how consequential decisions actually get made.
When Swarms Fail: The Institutional Antibodies
But swarms don’t emerge smoothly everywhere. A top-50 global bank learned this the hard way.
The engineering leadership felt proud about their forward-looking AI adoption. They saw the grassroots Slack channels forming, the informal evaluation groups coalescing around complex problems. They recognized the pattern and decided to encourage it, creating space for guild-like structures to emerge organically.
Within weeks, they hit institutional antibodies.
The Architecture Review Board flagged that decisions were being made outside the formal review process. Compliance escalations piled up because the informal evaluation groups weren’t documented in the governance framework. Long-standing SOPs didn’t account for fluid participation in decision-making. The swarms kept forming, but they kept running into roadblocks designed for a different model of work. They’re still navigating this collision between informal evaluation and formal accountability.
The tension revealed something important. Simply expecting grassroots behaviors to support a guild model isn’t realistic outside the earliest startups. The swarms don’t just need permission. They need guardrails that make sense within existing institutional structures. Leadership must be deliberate about which formal processes to adapt, which to preserve, and how to make collective evaluation legible to compliance, legal, and governance functions that weren’t built for it.
This also exposes the power dynamics that swarms introduce. When an informal evaluation group surfaces concerns about a senior leader’s proposal, what happens? When swarms form around some decisions but not others, who gets swarmed and who doesn’t? When participation is voluntary and organic, who gets excluded, by design or by accident, from the conversations where institutional knowledge is being built?
The risk isn’t just exclusion. It’s that swarms can become shadow governance. Healthy swarms are inclusive, psychologically safe, and documented. Shadow governance is exclusive, political, and opaque. The difference matters, but it’s often invisible until the organization hits a crisis.
These aren’t edge cases. They’re the reality any organization will face as these patterns spread. The swarms persist because they solve a problem the formal structure cannot: when AI makes it easy to generate convincing proposals, you need a way to stress-test them that goes beyond the official chain of command. But persistence alone doesn’t make them efficient, politically comfortable, or easy to coordinate with existing governance. The challenge isn’t whether swarms will form. It’s how to make them work within institutional reality.
The Failure Mode: Deciding Alone
What happens when leaders don’t have access to these swarms? When they make consequential decisions in isolation, relying solely on AI-generated analysis?
The risk isn’t just bad decisions. It’s a specific kind of badness that compounds across the organization.
There’s the local optimization trap. A leader optimizing within their own context, their department, their quarter, their immediate constraints, might find what looks like the best available option. AI can help them find it faster. But that locally optimal decision might be globally suboptimal. When every leader is making reasonable choices in isolation, the organization drifts toward mediocrity, with decisions that make sense individually but conflict or undermine each other systemically.
And there’s the averaging effect. When decisions are made primarily through AI-assisted analysis, they tend toward the statistically probable, the middle of the distribution. AI generates what seems reasonable based on patterns in its training. But “reasonable” often means “average.” The fringe proposals, the big, audacious ideas that might energize people to think differently, get smoothed out. They don’t survive AI-mediated evaluation because they can’t be easily justified within existing patterns.
This is particularly dangerous for innovation. The proposals that change trajectories are often the ones that sound unreasonable at first. They require human conviction to champion them, the kind of conviction that emerges from genuine wrestling with a problem, not from reviewing AI-generated options. When decision-making becomes primarily about choosing among plausible AI outputs, organizations lose their capacity to pursue game-changing bets.
Without swarms, without the collective pressure-testing, the naive questions, the cross-functional perspectives, leaders make cleaner, faster, more defensible decisions that are quietly less ambitious and less integrated than what the organization actually needs. Safe choices compound. Decline doesn’t arrive through dramatic failure. It arrives through a portfolio of reasonable decisions that collectively starve the organization’s capacity for necessary risk.
This is why people swarm problems and stress-test each other’s reasoning. Not because collaboration became trendy, but because exercising responsible judgment alone becomes impossible when you’re drowning in plausible options. You can’t properly evaluate everything AI can generate, but you’re still the one who has to own the decision.
Two Worlds, One Organization
It doesn’t happen evenly across an organization. Most end up operating in two modes at once, often without explicitly acknowledging it.
Some parts stay execution-focused. Clear objectives, established processes, well-defined roles. Decisions get made through normal hierarchical channels. AI gets used as a productivity tool, faster drafting, better formatting, quicker analysis, but it doesn’t fundamentally change how authority works. This tends to show up in operations, finance, and HR, functions where work is more standardized and decision criteria are clearer.
Other parts start operating evaluation-first. Decisions take longer even as information arrives faster. People expect to weigh in even when it’s not their formal responsibility. The informal networks matter as much as the formal structure. AI isn’t just a productivity enhancer. It’s actively changing what it means to make a well-considered decision. Engineering, product, and strategy teams often shift here first, not because they’re more sophisticated, but because their work involves higher uncertainty, more novel problems, and decisions where the “right answer” isn’t obvious from established patterns.
The difficulty is managing the boundary between these worlds.
Someone from the execution-focused part brings a question to the evaluation-first part and gets frustrated by what feels like endless discussion and lack of clear ownership. Someone from the evaluation-first part brings a carefully evaluated decision to the execution-focused part and gets frustrated when it’s challenged by people who weren’t part of the evaluation process.
Neither side is wrong. One is optimized for speed and clarity of authority, the other for thoroughness and shared evaluation. But they create friction when they intersect, and that friction is intensifying as AI use deepens unevenly across the organization.
This split isn’t new. Organizations have long struggled with tension between innovation teams and operations teams. But AI is changing what that boundary means. It’s no longer about a dedicated innovation group versus everyone else. Any team using AI heavily for strategic decisions starts operating evaluation-first, while teams using AI purely for efficiency stay execution-focused. The boundary cuts across functions in unpredictable ways, creating coordination challenges that org charts can’t capture.
The Tacit Knowledge Paradox
As these evaluation communities deepen, knowledge itself changes shape.
Mastery stops being something you acquire through documentation or training programs. It spreads through participation: being in the conversations where people are actually stress-testing AI outputs, seeing what questions experienced practitioners ask, understanding which concerns turn out to matter and which don’t. Learning accelerates for people inside these networks. They develop judgment faster because they’re constantly seeing their reasoning tested against others.
But this creates real risks.
If you’re not in those Slack channels, not part of those hallway huddles, you miss the calibration that everyone else is getting. The knowledge becomes sticky in ways that are hard to transfer. New people joining the organization, or people in other functions, can’t just read the documentation to catch up. The documentation doesn’t capture what actually matters.
The organization becomes fragile. Too much depends on who’s in the room, who happens to be in the right Slack thread, whose perspective gets included in the swarm.
Some organizations are starting to experiment with treating decision artifacts differently. Not just logging what was decided, but preserving the reasoning: what was considered, what was traded off, why other options were rejected.
This might look like a decision log that captures both the options that were actively considered and, critically, the decisions that were deliberately deferred. Not “we chose A over B and C” but “we chose A over B and C, and we explicitly decided not to decide on X and Y yet because we need more information about Z.” The deferred decisions often matter as much as the made ones. They signal what the organization is still uncertain about, what conversations are still open, what might need to be revisited as conditions change.
These shared narratives capture the evaluation process, not just its output. They make visible what would otherwise remain tacit: the assumptions that were tested, the concerns that were raised and addressed, the naive questions that turned out to matter, the expertise that was brought to bear.
The goal isn’t perfect documentation. It’s to leave enough of a trail that someone who wasn’t in the room can understand not just what was decided, but how the decision-makers thought about the problem.
These practices are still early. Most organizations are experimenting, not executing at scale. But the attempts signal an important recognition: the old ways of preserving institutional knowledge don’t work when that knowledge is being generated in fluid, AI-mediated evaluation communities.
The Tension Remains
Organizations are in the middle of a structural shift. Not in what they decide, but in how decisions get made and where accountability actually sits.
AI isn’t eliminating the need for human judgment. It’s making judgment inescapable. When every option comes with sophisticated reasoning, when analysis is abundant and cheap, the bottleneck moves entirely to deciding what to actually do. And that’s a fundamentally human responsibility.
Organizations aren’t moving toward flat structures or role-less collaboration. If anything, roles are becoming more important, but for a different reason. Not because they represent expertise or knowledge, but because they anchor accountability. Someone has to own the decision, and that ownership is what gives the role its weight.
But individual decision-making isn’t enough anymore. The space of plausible options is too large. Collective evaluation becomes the infrastructure that makes individual accountability possible.
This is the actual shift. Organizations are moving toward responsibility-centered operating models. Roles define who decides. Shared evaluation expands what’s possible to decide well. The combination allows people to exercise judgment over problems that would overwhelm any individual, while maintaining clear lines of accountability.
The tensions this creates won’t resolve cleanly. The distance between AI-generated analysis and the people who own outcomes will widen as AI capabilities advance. The boundary between execution-focused and evaluation-first work will keep shifting. The power dynamics of who gets swarmed and who doesn’t will need ongoing negotiation.
These tensions aren’t problems to solve. They’re conditions to navigate.
Navigation requires holding multiple truths simultaneously. AI makes us more capable while making decisions feel lonelier. Collective evaluation is essential while individual accountability remains non-negotiable. The reasonable path is often the wrong one, but the unreasonable path requires conviction we can’t generate algorithmically.
The defining challenge of the next era isn’t knowing more than others or having better AI tools. It’s learning how to operate in an environment of cognitive abundance without being trapped by the plausibility glut it creates.
The swarms, the guilds, the decision artifacts, the new relationship between roles and responsibility: these are early experiments in navigating that challenge. The questions they raise, how to maintain coherence across operating modes, how to make collective intelligence legible to governance, how to preserve conviction in an environment of overwhelming plausibility, will be worked out differently across contexts, by practitioners willing to experiment and name what they discover.
Leaders navigating this shift might start by asking: Where are the informal evaluation networks already forming in my organization? What decisions are being made in isolation that shouldn’t be? And which of our formal processes are blocking the collective intelligence we need?
That work is already underway. What it looks like in practice is the subject for another conversation.
Footnotes
-
Huang et al. (2025). “How AI Is Transforming Work at Anthropic.” Anthropic. https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic ↩
-
Kniberg, H. & Ivarsson, A. (2012). “Scaling Agile @ Spotify with Tribes, Squads, Chapters & Guilds.” https://blog.crisp.se/wp-content/uploads/2012/11/SpotifyScaling.pdf ↩