The Hidden Realities of AI Adoption
Kate Niederhoffer (BetterUp Labs) and Thomas H. Davenport (Babson College), hosted by Tom Stackpole (HBR) · article · 2026-07-30 Reader 🖥️
AI SlopCognitive SurrenderEnterprise AI Adoption & ReadinessProfessional IdentityPsychological SafetyTrustAI Workflow Patterns
AI adoption succeeds or fails in an organization's invisible layer — the workflows, trust, and knowledge-validation practices surrounding the technology, not the technology itself — so leaders must treat augmentation as a deliberately governed, segmented, and measured redesign of work (tracking trust as a leading indicator and closing human-verification gaps in automated chains), because leaving that layer ungoverned decouples effort from quality ("workslop"), erodes trust in inter-organizational processes ("knowledge decay"), and hollows out the human judgment the whole system depends on.
practical · mixed — A live Q&A transcript (dialogue, not prose) that develops a theoretical frame — why generative AI is psychologically and organizationally different from prior technology — in service of practical prescriptions: measure trust as a leading indicator, segment AI rollout by risk/function, gate access behind training, and commit credibly to augmentation. The two speakers argue from complementary lenses (interpersonal/social-psychological vs. organizational/process) that converge on the same action agenda.
Structure
| Part | Role |
|---|---|
| Stackpole's framing: AI adoption is usually discussed as a technology or culture challenge, but success depends on a 'hidden layer' of organizational norms, relationships, incentives, and knowledge systems. | thesis-setting |
| The two speakers establish complementary lenses: Niederhoffer's interpersonal/social-psychological view (AI as a version of 'the other') vs. Davenport's organizational/process view (AI integrated with workflows and knowledge management). | framework |
| Convergence on ends: both argue well-being and satisfaction drive productivity, and both favor augmentation over pure automation while accepting automation has a legitimate role. | shared premise |
| Diagnosis of why generative AI is different (Niederhoffer's four factors: mimicry, effort/quality decoupling, heterogeneous invisible usage, natural-language interface), qualified by Davenport's taxonomy distinguishing analytic, generative, and agentic AI as different technologies affecting different people. | theoretical groundwork |
| Automation-vs-augmentation tradeoffs and the individual-access-vs-enterprise-use-case distinction (Johnson & Johnson narrowing from 900 documented use cases to 9). | evidence / practical distinction |
| Concrete diagnosis of the hidden layer breaking down: 'workslop's' negative trifecta (cognitive tax -> emotional toll -> interpersonal trust judgment), 'knowledge decay' in the HR recruiting chain (AI writes JD -> AI resume -> AI cover letter -> AI review -> AI rejection, no human touchpoint), and Davenport's own physician sending an unread AI-generated medical response. | evidence |
| Proposed measurement practices: psychological-safety scales, AI-assisted transcript and social-network analysis, and a preliminary finding that async communication buffers AI's toll on team coordination. | proposed diagnostic method |
| Governance prescriptions and their internal tension: segmentation by risk/function, Colgate-Palmolive's train-before-access model, versus the risk that ungoverned experimentation degenerates into 'token maxing' and redundant effort. | prescription |
| Acknowledged open problem: the junior-talent/entry-level pipeline (Davenport says this concern, raised in 2013, is 'still true today' and unsolved) and Niederhoffer's reverse-bias finding that younger workers trust AI less than senior workers. | complication / acknowledged gap |
| Closing prescriptions for anxiety: a credible commitment to augmentation, explicit no-layoffs messaging, trust-building, and instilling worker agency ('in the pilot seat'). | conclusion / prescription |
Key terms
- Workslop AI Slop
(Niederhoffer) AI-generated content that looks polished and complete but is decoupled from real effort and quality, requiring the recipient to expend cognitive, emotional, and interpersonal work to interpret, fix, or route around it — a proxy symptom rather than the underlying cause, which is the loss of the historical link between visible effort and actual quality. - Knowledge decay
(Davenport) The erosion of trust and reliability that occurs when knowledge passes through a multi-step, AI-mediated, inter-organizational process with no human validation touchpoint at any stage (e.g., the HR recruiting chain where AI writes the job description, tailors the resume, writes the cover letter, reviews, and rejects, with no human ever involved). - Trust as a leading indicator
(Niederhoffer) Team or process trust is not merely a lagging output that results from good or bad AI adoption after the fact, but a signal that changes ahead of downstream problems and should therefore be actively measured and tracked over time as an early-warning system. - Credible commitment to augmentation
(Niederhoffer) A costly, visible organizational signal — not mere messaging — that demonstrates real intent to preserve and develop human roles alongside AI, distinguished from rhetorical claims by requiring investment (time, resources, tolerance for a short-term productivity dip) that a company would not make if augmentation were not the genuine goal. - Human above the loop
(Niederhoffer) Training workers — especially juniors — to develop situational awareness and discernment about when to engage AI tools at all, rather than passively reviewing AI output after the fact ('human in the loop'); a shift from oversight-as-checkpoint to judgment-as-authority.
Propositions
- explicitAI functions psychologically as a version of 'the other' — mimicking human presence and behavior — which triggers the same relational cognition humans use for real people, so it can be deployed to either support or substitute for human relationships.conf: high
- explicitEmployee well-being and a sense of meaningful, identity-expanding work causally increase productivity, so how AI adoption is experienced is itself a productivity lever, not only a humane consideration.conf: high
- explicitcorroboratedOrganizations should preserve human employment and steer AI's gains toward augmentation and job fulfillment rather than pure headcount reduction, while still using automation as a legitimate and necessary component of AI strategy.conf: highsupports ←The Acceleration Whiplash · p10 — Both argue organizations should redirect AI's throughput gains toward augmentation and capacity-building rather than pure headcount reduction.
- explicitGenerative AI differs from prior technologies along four axes: deceptive mimicry of human language/thought, the decoupling of effort from quality, heterogeneous and largely invisible usage patterns that erode trust in expertise, and natural language as an unusually low-barrier interface.conf: high
- explicitAI is not a monolith — analytic AI (numeric prediction), generative AI (content), and agentic AI (autonomous task execution) are distinct technologies affecting different people and processes, so blanket claims about 'AI adoption' conflate materially different phenomena.conf: high
- explicitIndividual-level, ungoverned AI access ('AI for everybody') rarely yields measurable productivity, whereas enterprise-scoped, deliberately narrowed process use cases yield measurable value, because value requires whole-business-process redesign rather than mere tool access.conf: highrefinesHow People Are Really Using AI in 2026 · p5 — Offers a mechanism for the 'marginal efficiencies, not transformation' finding: individual/ungoverned AI access doesn't yield measurable productivity, while enterprise-scoped process redesign does — explaining why breadth of individual adoption hasn't translated into business transformation.
- explicitMulti-step, inter-organizational processes with no human touchpoint at any stage drive down trust in the process itself — 'knowledge decay' is fundamentally a trust failure, not merely an accuracy failure.conf: high
- explicit'Workslop' imposes a negative trifecta on recipients — a cognitive tax (confusion, effort to interpret) that cascades into emotional sequelae (frustration, annoyance) that becomes interpersonal judgment (reduced trust in and desire to work with the sender) — so its costs are relational and reputational, not just about wasted time.conf: highextendsManagers Are Struggling to Keep Up with the AI Productivity Boom · p3 — Provides the causal mechanism (cognitive tax -> emotional toll -> interpersonal trust judgment) underlying the workslop prevalence that doc reports statistically (54% of managers receiving AI workslop).
- explicitestablishedTrust functions as a leading indicator of AI-adoption health, not merely a lagging output, so organizations should actively measure and track team trust over time as an early-warning signal.conf: highextendsManagers Are Not Overhead: They Are Infrastructure · p4 — Reframes the trust-lift Microsoft's Work Trend Index measures (manager AI-modeling -> +30pt trust in agentic AI) as something organizations should track prospectively as a leading indicator, not only interpret retrospectively as a correlational outcome of manager behavior.supports ←Can an AI-Powered Scribe Curb Physician Burnout? · p9 — Gallani's claim that adoption depends on trust more than technical performance matches the roundtable's claim that trust functions as a leading indicator of AI-adoption health, not merely a lagging output.extends ←Gene Kim on Vibe Coding - Part II | Episode 411 | Troubleshooting Agile · p6 — The HBR roundtable frames trust as a leading indicator organizations should track over time; Kim's unpublished DORA finding supplies a concrete (if undocumented) timescale — ~12 months — for how that trust accrues.extends ←Responsible AI Is Becoming a Growth Strategy · p1 — The roundtable frames trust as a leading indicator of internal AI-adoption health; this article extends the claim outward — trust as a scarce, ecosystem-wide competitive resource affecting regulators, customers, and market position, not just intra-organizational adoption.supports ←To Adopt AI at Scale, Employees Need to Trust Agents · p3 — Both identify trust (not raw capability or accuracy) as the leading determinant of AI-adoption success.
- explicitIncreased asynchronous communication can buffer the negative effect of AI-driven displacement (preferring AI over a colleague) on team coordination, such that teams with high async communication show no measurable coordination toll from AI use.conf: medium
- explicitDeliberate segmentation — by job/persona/function/level for education, and by risk/criticality for process change — is a decisive factor in successful AI adoption, more so than any single tool or blanket policy.conf: high
- explicitGovernance and unstructured experimentation are in tension: without clear communication about why and where AI use is encouraged, ungoverned experimentation degenerates into performative 'token maxing' and redundant duplicated effort across teams.conf: highsupportsAgentics / Tech Things: Tokenmaxxing is dead, long live tokenmaxxing · p3 — Independently corroborates, from enterprise AI governance, that ungoverned incentives around AI usage produce metric-gaming ('token maxing') rather than value — the same phenomenon Meta's token-tied performance reviews illustrate in the tokenmaxxing extraction.
- explicitcontestedThere is a reverse generational bias in AI trust: unlike prior technology waves where younger workers had a native-digital advantage, younger workers currently trust AI less than more senior workers, inverting the usual pattern of generational technology adoption.conf: mediumcontradicts ←Gene Kim on Vibe Coding - Part III | Episode 412 | Troubleshooting Agile · p6 — Kim's observed pattern (oldest and youngest adopt fastest, mid-career senior staff resist) is a different generational shape than the HBR roundtable's 'younger workers trust AI less than senior workers' finding — both invert the naive assumption that youth predicts adoption, but via incompatible mechanisms, worth reconciling rather than treating as the same claim.
- explicitThe most successful managers calibrate roughly equal investment between human-relational behaviors (coaching, relationship-building, alignment) and AI-related behaviors (experimentation, enablement, KPIs) rather than favoring one over the other.conf: medium
- explicitA single, credible message — explicitly committing that no one will lose their job because of AI (accepting some attrition-driven headcount reduction instead) — would materially reduce workforce anxiety about AI more than most other interventions.conf: medium
- implicitDisciplined measurement and experimentation is rare across organizational cultures, which functionally caps how well any organization can distinguish real AI value from noise regardless of its adoption strategy.conf: mediumsupportsHow People Are Really Using AI in 2026 · p6 — Corroborates from the organizational-culture side why quantified AI ROI is rare: most organizations lack a disciplined measurement/experimentation culture in the first place, independent of the social-listening methodology critique.
Arguments
A credible commitment to augmentation requires organizations to accept a short-term productivity dip.
Explicit premises:
- Building the conditions for augmentation requires investment in people — coaching, alignment, and relationship-building — before behavior change occurs.
- This investment has a cost in the near term.
Implicit premises:
- Organizations typically evaluate AI initiatives against short-term productivity metrics.
- If a short-term dip is not tolerated or explained in advance, augmentation-oriented initiatives will be abandoned before they pay off.
Inferences:
- This creates a structural bias toward automation-only strategies — which show faster, more legible ROI — even in organizations that state augmentation as their goal, because automation's gains are immediate and attributable while augmentation's are deferred and harder to measure.
Trust and process-level indicators are more actionable governance targets than content-quality metrics for generative AI. strength: medium
Explicit premises:
- No reliable tool exists to detect whether content was AI-generated, and 'humanizer' tools actively defeat such detection.
- Historically, people have cared more about convenience/quantity than quality (the camera-phone analogy: worse image quality than dedicated cameras, but adopted anyway for convenience).
Implicit premises:
- A metric that cannot be reliably measured is not useful for governance.
- The convenience-over-quality pattern from consumer photography generalizes to professional and high-stakes content.
Inferences:
- If people neither notice nor especially value quality degradation, quality-focused interventions may fail to gain adoption even where technically feasible — creating a mismatch between what leaders should optimize (quality) and what users will actually reward (convenience).
Narratives
- Davenport's primary care physician sent an obviously AI-generated response (produced via a tool called OpenEvidence) to a medical question, apparently without reading it first.
→ Makes concrete, in a real high-stakes professional relationship, the 'knowledge decay' pattern Davenport otherwise describes abstractly for inter-organizational processes — a professional treating AI output as ready-to-send without the human verification step that would normally validate it. - The HR recruiting chain: a company uses AI to write a job description; an applicant uses AI to tailor a resume and write a cover letter; the company uses AI to review, evaluate, and generate a rejection letter — with no human ever involved at any stage.
→ Operationalizes what 'the hidden layer breaking down' means concretely, and grounds the abstract claim that automating an entire inter-organizational process collapses trust in it. - Contrasting governance cases: Meta tying performance reviews to per-person token usage (leading employees to run two agents chatting all day to inflate counts); Johnson & Johnson narrowing from 900 documented individual AI use cases to 9 enterprise-focused ones; Colgate-Palmolive requiring training in effective AI use before granting full access, with post-session value/time-saved surveys.
→ Illustrates the governance spectrum from anarchic (Meta's gamed metric) to well-segmented (J&J, Colgate-Palmolive), giving concrete positive and negative models for the segmentation prescription.
Questions addressed
- ✅ Why is generative AI different from previous workplace technologies?
Niederhoffer's four-factor answer (mimicry, effort/quality decoupling, heterogeneous invisible usage, natural-language interface), though Davenport partially complicates the premise by noting 'AI' spans analytic, generative, and agentic technologies that are not equivalent. - ✅ How would a leader know whether the organization's hidden layer (trust, process, knowledge validation) is breaking down?
Proposed indicators: team trust measurement (as a leading indicator), the workslop cognitive-emotional-interpersonal trifecta, process-trust surveys, and knowledge-decay signals in inter-organizational chains — though the resolution is qualitative/directional rather than a validated diagnostic instrument. - ✅ How should organizations manage the anxiety AI provokes among employees?
A credible commitment to augmentation, explicit trust-building and emotional acknowledgment, instilling worker agency, and (from Davenport) a direct no-layoffs message accepting attrition-driven headcount reduction instead. - ⭕ What happens to the entry-level/junior-talent pipeline as AI absorbs tasks historically used to train juniors into judgment?
Davenport explicitly states this concern was raised in a 2013 article he co-authored and remains unsolved today ('we haven't really figured that out yet... still true today'); no candidate solution is offered beyond noting IBM as an outlier that still hires entry-level. - ✅ Should organizations let individuals freely experiment with AI, or constrain it with rules?
Both speakers converge on a middle path: governed segmentation (clear domains for encouraged experimentation vs. off-limits domains) paired with genuine room for play and job-crafting within those boundaries.
Gaps
- illogicalBoth speakers repeatedly qualify their evidence base as tentative — 'preprints,' 'early days,' 'no published longitudinal data,' 'a preliminary finding' — while in the same breath issuing confident, general prescriptions (measure trust as a leading indicator, adopt async communication as a buffer, pursue a 'credible commitment to augmentation', segment your rollout). The hedges scope the underlying claims as correlational and drawn from single studies or organizations; the prescriptions built on them are delivered with the confidence of established, generalizable causal findings, and neither speaker states what further evidence would need to accumulate before revising the advice. This is sharpest for the async-communication buffering claim: explicitly called 'preliminary,' yet offered live to an audience of leaders as actionable practice.
Read next: Read the primary BetterUp/Microsoft Work Trend Index studies underlying the workslop and manager-effectiveness findings — already critiqued for the same causation-vs-correlation gap in extract-lenny-noam-segal-tech-worker-ai-sentiment-2026#g1 and extract-managers-are-not-overhead-they-are-infrastructure#g1 — to see whether the underlying data supports prescription-level confidence or only correlational-level confidence. - incompleteTrust is proposed as a 'leading indicator' of AI-adoption health that organizations should track — but neither speaker specifies a threshold, trend, or baseline that would count as a warning sign, nor distinguishes a genuine leading indicator (trust changes BEFORE downstream harm) from trust simply lagging behind whatever caused the harm. Without a demonstrated lead-lag ordering, 'measure trust' is a plausible slogan without an operational definition.
Read next: Look for a longitudinal study design establishing lead-lag ordering between team-trust measures and later productivity, quality, or attrition outcomes. - uninformed'Credible commitment to augmentation' is treated as an achievable, meaningful organizational signal, but neither speaker defines what makes a commitment 'credible' versus merely rhetorical — no criteria are given, and no example is offered of a company that made and kept such a commitment under real headcount or cost pressure. Davenport's 'no layoffs' message and Niederhoffer's three-part framework remain proposed prescriptions, not documented cases.
Read next: Find case studies of companies that publicly committed to an augmentation-over-automation stance and track whether they held to it through a downturn or cost-cutting cycle. - incompleteThe entry-level/junior-talent pipeline problem is explicitly named as unresolved by Davenport himself, referencing a concern he first raised in a 2013 article that remains unanswered today. Neither speaker proposes even a candidate organizational design beyond citing IBM as an outlier that still hires entry-level workers — 'human above the loop' is offered as an aspiration for junior discernment but not operationalized (what tasks, what training, what timeline).
Read next: Look for organizational-design proposals or empirical tracking of entry-level hiring rates alongside AI capability growth, and any company documenting how it builds judgment absent traditional apprenticeship tasks.closed byPodcast: How Will We Train Developers If AI Does the Routine Work: A Conversation with Scott Hanselman — That gap notes neither speaker proposed a candidate organizational design for the junior-talent pipeline problem beyond 'human above the loop.' This document proposes a concrete one: a formally named, nursing-style preceptor role, evaluated on trainee growth rather than shipped code, with company-size implementation guidance. - incompleteThe reverse-bias finding — younger workers trust AI less than senior workers, inverting the usual generational-adoption pattern — is asserted without a causal mechanism. Is it because younger workers face more direct job-security exposure from AI, more skepticism from spotting flaws as heavy users, or a cohort/timing effect specific to entering the labor market during this technology's rollout? The finding is stated as fact but left unexplained.
Read next: Read primary survey research on AI trust broken out by age/seniority to test candidate mechanisms.
Closes gaps elsewhere
- How People Are Really Using AI in 2026 · g2 — Davenport names the marginal-vs-transformative bottleneck as structural, not temporal: individual, ungoverned 'AI for everybody' access does not yield measurable productivity, while enterprise-scoped, deliberately narrowed use cases (e.g., Johnson & Johnson's shift from 900 documented individual use cases to 9 enterprise-focused ones) do — locating the gap in organizational selection/redesign of use cases rather than in model or adoption maturity.
Related documents
- Agentics / Tech Things: Tokenmaxxing is dead, long live tokenmaxxing
- Managers Are Struggling to Keep Up with the AI Productivity Boom
- How People Are Really Using AI in 2026
- 5 Things High-Performing Teams Do Differently
- Podcast: Will Agentic AI Bring Fantasia's Sorcerer's Apprentice to Life?: A Conversation with Tracy Bannon
- AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy
- The best response to AI slop, infinite advice, and online noise is from Robin Williams
- #5: AI Workflow Patterns: The Real Unit of AI Adoption in 2026
- Amy Gibson's Post
- Anders Aukland's Post