CHAPTERS
Why AI adoption feels urgent—and why ROI is still a black box
Alex Rampell and Russ Fradin frame the central tension: enterprises feel extreme pressure to adopt AI quickly, yet lack credible ways to know whether the spend is paying off. They preview the episode’s core theme—AI ROI measurement is lagging far behind the pace of AI buying.
- •Companies believe they have ~18 months to become AI leaders or fall behind
- •AI is seen as “under-hyped” in capability but poorly diffused inside enterprises
- •Many leaders can report what they bought, not what value it produced
- •The key problem: measuring benefit is as hard as making AI work
From Web 1.0 ad tech to AI: the measurement stack always comes after the spend
Russ recounts his early career in online advertising and the rise of digital measurement infrastructure (e.g., Nielsen-like tooling for the internet). He argues AI is repeating the same pattern: massive budget shifts create a need for governance and measurement tools that accelerate adoption rather than slow it down.
- •Early internet ads faced attribution and “did it work?” questions similar to AI today
- •Digital ad growth required building measurement and planning infrastructure
- •Third-party measurement ultimately helped platforms scale revenue
- •AI’s next wave will similarly require tooling around measurement and governance
The ‘software eating labor’ shift: budgets are moving from people to tools
Alex outlines the emerging macro shift: AI software increases worker output, pushing companies to re-balance labor vs. software spend. That creates a CFO-grade accountability problem—if software spend rises materially, leaders need defensible evidence of efficiency gains.
- •AI enables “hiring software” instead of (or alongside) hiring humans
- •Software budgets may grow dramatically relative to labor budgets
- •Higher OpEx on AI tools demands ROI proof (especially for CFOs)
- •A rising question: what is the productivity baseline to compare against?
What Larridin measures first: tool discovery, usage, and safe enablement
Russ explains Larridin’s starting point: enterprises often don’t even know which AI tools employees are using, licensed or not. The first step is establishing a baseline of ‘what’s in use,’ then helping drive higher, safer adoption across workflows.
- •Many companies discover far more AI tools in use than IT realizes
- •Not all shadow AI usage is bad, but it can be risky without governance
- •Enterprise rollouts commonly fail due to low adoption—AI is no different
- •Driving usage requires employees to feel safe and not fear punishment
The hardest question: measuring AI ROI when “productivity” is fuzzy
They dig into why AI ROI is elusive: surveys are biased, definitions of productivity differ by role, and outputs can be hard to quantify. Russ describes Larridin’s approach: combine traditional productivity research with behavioral usage data to reduce guesswork.
- •Surveys alone are weak: people answer aspirationally and definitions vary
- •You must know whether employees are actually using the tools
- •Larridin combines survey methods with passive usage signals (at aggregate levels)
- •Long-term goal: more passive measurement, but data-sharing constraints remain
The productivity baseline problem: agent vs. principal incentives at work
Alex poses the principal–agent issue: an employee may use AI to finish work faster, but the company only benefits if output increases or costs fall. Russ argues the near-term goal is to build reliable baselines and correlations between usage intensity and work output at group levels, not to micromanage individuals.
- •Individual ‘time saved’ isn’t automatically company value
- •Companies need group-level baselines: heavy users vs. light users vs. non-users
- •FTE-based planning breaks down when AI changes ‘tonnage of work’
- •Managers may eventually adjust staffing or expectations as productivity becomes visible
Goodhart’s Law and why naive metrics backfire (Harvey, Cursor, and leaderboards)
The conversation turns to metric design: once a measure becomes a target, people game it. Using examples like Harvey (legal AI) and developer tools (Cursor/Claude spend leaderboards), they stress measuring real usage plus outcomes—without turning the measurement into a manipulable mandate.
- •Goodhart’s Law: targeting a metric corrupts it
- •Legal AI example: ‘users say it’s great’ is not evidence of value
- •You need passive signals (logins/usage intensity) before interpreting survey responses
- •Developer-heavy orgs may approximate value with spend + managerial judgment, but that doesn’t scale enterprise-wide
Operational metrics that matter: interdepartmental responsiveness as a real-world signal
Russ proposes practical productivity proxies that avoid simplistic output counts like “lines of code.” One promising approach: track responsiveness and service-level behavior between departments (e.g., legal turnaround time, engineering response latency) as a sign AI is reducing coordination friction.
- •Many departments are cost centers; ROI isn’t just headcount reduction
- •Measure whether AI increases throughput of requests and speeds response times
- •Responsiveness can reveal reduced bureaucracy and coordination overhead
- •Keep metrics internal to avoid gaming; focus on behavioral reality over vanity stats
What enterprise leaders say: $700B spend, ‘wasted projects,’ and an 18-month clock
Drawing on interviews with hundreds of IT leaders, Russ summarizes a consistent pattern: spending is surging, confidence in project success is low, and competitive anxiety is high. The lack of measurement itself becomes a strategic risk because leaders can’t distinguish winners from waste.
- •Enterprise AI spend is large and growing rapidly (often cited around $700B)
- •Many leaders believe a majority of AI projects are wasting money
- •Competitive fear drives fast adoption despite unclear payoff
- •Unlike ad tech, AI lacks decades of standardized measurement infrastructure
Employees are anxious, undertrained, and unsure what’s allowed
They highlight the human side of adoption: employees face tool overload, unclear rules, and fear of looking incompetent—or getting fired for misuse. HR and compliance concerns become central blockers to broad AI usage, especially in regulated environments.
- •Workers face many new tools at once, unlike typical annual software rollouts
- •Fear isn’t only job loss—it's making mistakes, violating policy, or looking ‘dumb’
- •Training and clarity lag behind procurement
- •Regulated industries and EU constraints heighten perceived risk
Nexus and ‘safe AI’: wrappers, guardrails, and compliant prompting at scale
Russ describes Larridin’s Nexus product approach: provide a controlled interface around major models so employees can use AI confidently. Guardrails (including policy-aware blocking) aim to prevent prohibited actions—like sharing sensitive HR data or generating disallowed content—so companies can encourage usage without fear.
- •Provide a ‘safe space’ for AI use to increase adoption
- •Wrappers abstract model choice (Claude/Gemini/ChatGPT) into a governed experience
- •Policy/regulatory guardrails block disallowed prompts or data handling
- •Goal: drive usage while building organizational knowledge of what works
Future of work: why AI likely creates more competition, not mass unemployment
They debate job impacts and argue broad job loss is unlikely in competitive markets: if one firm cuts too deeply, rivals will use AI to grow faster and outcompete. AI may push some workers to upskill, enable more solo entrepreneurship, and expand new categories of work (e.g., infrastructure, data centers).
- •Competitive dynamics reward growth; margin gains become competitors’ opportunity
- •Firms generally prefer growth over shrinking headcount (outside some niches)
- •AI may be disruptive for complacent knowledge roles that avoided continuous learning
- •New jobs emerge around new tech waves—entrepreneurship and infrastructure included
AI’s product marketing problem: ‘it does anything’ vs. specific use cases that sell
They close on a go-to-market insight: broad claims (“AI can do everything”) don’t drive adoption; concrete ‘tip calculator’ use cases do. Russ parallels comScore’s early lessons—specific, high-value questions unlock budgets far more reliably than generic platform promises.
- •Horizontal ‘AI does everything’ messaging often fails in enterprises
- •Specific outcomes (e.g., “help you code better”) drive pull and diffusion
- •ComScore analogy: “we know everything” wasn’t effective—specific insights were
- •Successful AI adoption will hinge on crisp use cases and demonstrated value
