People First,
AI Where It Pays
Why replacement-first AI programs stall, where human-led teams produce measurable gains, and how to stop weak pilots before they become expensive habits.
The Return Gap
Is Real
Companies are not broadly turning their backs on AI. They are turning away from a weaker idea: that buying the technology, reducing headcount, and waiting for margin expansion counts as a strategy. The evidence now separates adoption from return, and the distance between them is uncomfortable.
A 2026 NBER study surveyed nearly 6,000 senior executives across the United States, United Kingdom, Germany, and Australia. Around 70% of firms were actively using AI, yet more than 80% reported no impact on either employment or productivity over the prior three years. Those same executives still expected a 1.4% productivity gain over the next three years.[1] The ambition survived. The easy ROI story did not.
IBM found the same break in a different sample: only 25% of AI initiatives had delivered expected ROI, and just 16% had scaled enterprise-wide. Half of the CEOs surveyed said rapid investment had produced disconnected technology instead of coherent value.[2] BCG's 2025 maturity work put the share of companies realizing substantial financial gains at roughly 5%. Its recurring operating-model rule is even more useful: about 70% of AI value comes from people and process change, with 20% from technology and 10% from algorithms.[3]
The market is not retreating from AI to people. It is retreating from the assumption that AI creates value by replacing people.
SWITCHCASE STUDIOS · OPERATING THESISCustomer service offers an early view of that correction. Gartner predicts half of organizations planning large AI-driven service workforce reductions will abandon those plans by 2027. In its poll, 95% of customer-service leaders planned to retain human agents, favoring “digital first” rather than digital only.[4] That is not nostalgia. It is an admission that exceptions, trust, escalation, and relationship work were missing from the original spreadsheet.
People over bad AI programs
The useful choice is not “people or AI.” It is whether a specific task can be made faster, better, or less costly with AI while a person retains the context and authority the task requires. If that cannot be demonstrated, the human-led workflow is not a failure to innovate. It is the current best operating model.
Where the Value
Leaks Out
A fast demo proves that a model can produce something. It does not prove that the organization can turn that output into value. Between prompt and profit sit review, rework, integrations, queues, policy, adoption, and the awkward question of what the team will do with the time supposedly saved.
McKinsey's 2025 global survey found that more than 80% of respondents saw no tangible enterprise-level EBIT impact from generative AI. Only 21% had fundamentally redesigned even some workflows, and fewer than one in five tracked well-defined generative-AI KPIs. KPI tracking had the strongest relationship with EBIT impact among the scaling practices studied; workflow redesign showed the largest effect among the organizational attributes.[5] These results are self-reported and correlational, but the pattern is hard to ignore: installing a tool is not the same as changing the work.
The draft arrives faster
The first output feels instantaneous, so the pilot records time saved before anyone measures correction, verification, or exception handling.
Review and rework expand
Fluent output lowers vigilance. Experts inherit more checking, edge cases bounce between teams, and a local gain becomes someone else's queue.
The system stops at the handoff
Weak data, missing permissions, brittle integrations, and unclear ownership keep a polished pilot from surviving production.
Saved capacity has nowhere to go
Even real time savings produce no return when capacity is not redirected to revenue, quality, throughput, risk reduction, or a better customer outcome.
Five Failure Paths That Repeat
RAND conducted 65 interviews with data scientists, engineers, and academic researchers and found five recurring root causes: teams misunderstand the business problem or optimize the wrong metric; the necessary data is unavailable; leaders chase impressive technology instead of a real user problem; infrastructure is inadequate; or the chosen task exceeds the technology's practical capability.[6] Those are not model failures alone. They are selection and operating failures.
Headcount is the first KPI
The business case assumes labor disappears before quality, escalation, demand, and exception work are understood. The savings are booked; the work is not.
Usage stands in for value
Prompts, seats, outputs, or agent runs move upward while cycle time, conversion, correctness, and cost remain unmeasured.
A broken process gets automated
AI accelerates activity inside a fragmented workflow. More work moves sooner, but the constraint at the next handoff stays exactly where it was.
The pilot has no route to production
No business owner, approved data, integration path, risk threshold, or budget ceiling exists. The prototype succeeds on a path the organization cannot operate.
The fifth failure path is subtler: one good result gets generalized to an entire role. AI capability is jagged. A model can be unusually strong on one task and confidently wrong on the next one that looks nearly identical. The unit of AI strategy is therefore the task and its context, not the job title.
AI Pays at the
Task Boundary
The strongest use cases share a shape: the work is frequent, bounded, supported by enough context, easy to verify, and reversible when wrong. A person can accept, correct, or reject the output. AI compresses the repetitive middle while the human owns the objective and the consequence.
In a field study of 5,172 customer-support agents, access to an AI assistant increased issues resolved per hour by about 15%. Less-experienced and lower-skilled workers gained the most, escalation requests fell, and the evidence suggested that workers learned from the system.[7] This is augmentation at its most efficient: the tool codifies patterns from strong performers, surfaces them inside the workflow, and leaves the human responsible for the conversation.
The “jagged frontier” experiment with 758 consultants explains why task selection matters. On work inside the model's capability frontier, AI users completed 12.2% more tasks, worked 25.1% faster, and improved human-graded response quality by roughly 30% or more, depending on the treatment. On a task outside that frontier, AI-assisted consultants were 19 percentage points less likely to reach the correct answer.[8] Speed and polish are useful only after correctness survives.
Choose the Work Before You Choose the Tool
| Work pattern | Best operating model | Proof signal |
|---|---|---|
| High-volume, bounded, repeatable Classification, summarization, extraction, first drafts |
AI-assisted Automate the repeatable core; sample and audit output. |
Net cycle time falls while correctness and exception rates stay inside tolerance. |
| Pattern-rich, expert-verifiable Support guidance, code scaffolds, research synthesis |
Human in the loop AI proposes; a qualified person accepts, edits, or rejects. |
Review time plus creation time beats the human-only baseline. |
| Ambiguous, relationship-heavy Negotiation, leadership, sensitive service recovery |
Human-led Use AI for preparation or recall, not final judgment. |
Trust and outcome quality improve without flattening context or accountability. |
| High-consequence or irreversible Safety, legal commitments, material financial decisions |
Human-controlled Strict validation, approvals, logging, and a clear fallback. |
No critical incidents; false-positive and false-negative rates remain within an explicit risk floor. |
A preregistered field experiment with 776 Procter & Gamble professionals adds another useful pattern. Individuals working with AI used 16.4% less time and produced work comparable to conventional two-person teams; teams with AI used 12.7% less time. AI also helped participants cross functional boundaries, allowing commercial professionals to produce more technically balanced ideas and technical professionals to produce more commercially balanced ones.[9] The opportunity was not “remove the team.” It was “give more people access to the team's combined range.”
The practical fit test
Favor AI when context is available, output is verifiable, mistakes are reversible, frequency is high, and someone can act on the result. Favor people when the goal is ambiguous, context is tacit, trust is the product, or an error creates an irreversible consequence.
Spot the Failed Path
Before It Scales
AI programs are unusually good at looking alive. Output streams in, dashboards fill up, and users report feeling faster. None of that proves a net improvement. A useful pilot is designed to disprove itself early.
METR's early-2025 randomized trial is the cleanest warning. Sixteen experienced open-source maintainers completed 246 real tasks in repositories they knew well. With AI, they took 19% longer even though they expected to be 24% faster and later believed they had been 20% faster.[10] METR now cautions that this is a historical result for early-2025 tools and that later tools are likely more capable, though its newer data remain too selection-biased for a precise estimate.[11] The durable lesson is measurement, not a permanent verdict on coding assistants.
DORA's 2024 analysis found that greater reported AI adoption was associated with lower delivery throughput and stability, possibly because generated work increased batch size and review load. DORA explicitly treats these as associations, not proof that AI caused the decline.[12] Again, the operating signal is downstream: if creation accelerates but review, deployment, or recovery gets worse, the system did not get faster.
A Simple Green / Amber / Red Gate
| Status | What the evidence says | Decision |
|---|---|---|
| Green | Net business outcome improves; total cycle time falls; quality and risk are non-inferior; users adopt the workflow without hidden expert rescue. | Scale deliberately. Expand one task class or cohort at a time and keep the control group long enough to detect drift. |
| Amber | Local speed improves but review or queues expand; gains appear only for one cohort; adoption is brittle; cost rises faster than value. | Re-scope. Narrow the task, improve context, retrain users, redesign the handoff, or return authority to a person. |
| Red | No net gain after stable use; total cycle time worsens; quality breaches tolerance; critical security, privacy, or compliance risk appears. | Stop or roll back. Preserve the learning, remove the workflow, and redirect budget to a better-defined problem. |
Perceived speed, no net speed
Creation time falls while checking, repair, waiting, or downstream support rises. Measure start-to-accepted-output, not time-to-first-draft.
One cohort carries the gain
Novices improve while experts slow down, or one task performs while adjacent work fails. Segment results before averaging them into a scale decision.
The queue moved downstream
One team ships more output into an unchanged approval, QA, legal, or deployment constraint. The bottleneck was relocated, not removed.
Economics worsen with scale
Inference, integration, observability, review, and exception costs grow faster than the value captured. Fully loaded cost—not token price—is the denominator.
The risk grows with autonomy. Gartner predicts more than 40% of agentic-AI projects will be canceled by the end of 2027 because of escalating cost, unclear business value, or inadequate risk controls.[13] That is a forecast, not an observed failure rate, but the decision rule is sound: autonomy should be earned with evidence at each wider boundary.
Do not ask whether the AI worked. Ask whether the whole system of people, model, review, handoffs, and consequences worked better.
THE MEASUREMENT RULEA 30-Day
Course-Correction Loop
Long transformation roadmaps hide weak assumptions. A short proof loop forces them into the open. Every pilot should begin with a named business owner, one exact task class, a human-only baseline, a target business metric, a quality and risk floor, a fully loaded cost ceiling, a stop date, and a person authorized to roll it back.
Define the job at task resolution
Document the trigger, input, context, decision, output, recipient, exception path, and consequence of error. Record the human-only baseline before anyone touches the model.
Run a matched pilot
Compare human-only and AI-assisted work on representative cases. Keep model, prompt, workflow, cohort, and review policy stable enough to learn something.
Measure the downstream system
Count review, rework, escalations, waiting, incidents, total cost, and customer impact. Trace the output until it is accepted—not merely generated.
Scale, re-scope, or stop
Apply the pre-agreed green, amber, and red gates. No “strategic potential” exception should overrule a missed quality floor or absent business outcome.
Turn learning into the next move
Document the task boundary, cohort, evidence, failure mode, and decision. Reuse the data and controls; do not preserve a weak use case just to preserve momentum.
The Minimum Viable Scorecard
| Measure | What to include | Failure signal |
|---|---|---|
| Business outcome | Conversion, resolution, throughput, revenue, loss avoided, or another outcome with an owner | Activity rises; the business KPI does not move |
| Total cycle time | Creation + waiting + review + rework + exception handling | First output is faster; accepted output is not |
| Quality & correctness | Task-specific rubric, defect rate, false positives/negatives, expert acceptance | Fluency improves while correctness crosses the risk floor |
| Fully loaded cost | Models, infrastructure, integration, review, monitoring, support, incidents | Cost per accepted outcome increases with usage |
| Human system | Adoption, override rate, confidence calibration, training, escalation quality | Experts silently rescue the system or users route around it |
Post-deployment monitoring is not cleanup. NIST's guidance treats functionality, operations, human factors, security, compliance, and broader impacts as continuing responsibilities after an AI system goes live.[14] The same go/no-go discipline used in a pilot should remain available in production: manage drift, tighten scope, lower autonomy, route work back to a person, or decommission the system.
The goal is not to save every AI pilot
It is to save the company from scaling the wrong one. A stopped pilot with a clear result is cheaper—and more innovative—than an ambiguous prototype kept alive by sunk cost.
The companies that earn a return will not be the ones that choose AI most aggressively or people most sentimentally. They will choose the work precisely. They will put AI where it compresses real friction, keep people where context and consequence demand them, and measure the entire system closely enough to change direction before hope becomes infrastructure.
People first. AI where it pays. Evidence at every boundary.
Bibliography
- Yotzov, I., Barrero, J. M., Bloom, N., et al. Firm Data on AI. NBER Working Paper No. 34836, February 2026. Read the working paper. Executive-reported estimates; not yet peer reviewed.
- IBM Institute for Business Value. 5 Mindshifts to Supercharge Business Growth. May 2025. Read the study summary.
- Boston Consulting Group. AI Transformation Is a Workforce Transformation. February 2026. Read the report. Maturity and value findings are vendor-defined and correlational.
- Gartner. Gartner Predicts 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI. June 2025. Read the forecast.
- McKinsey & Company. The State of AI: How Organizations Are Rewiring to Capture Value. March 2025. Read the survey. Self-reported, correlational results.
- Ryseff, J., DeBruhl, B., & Newberry, S. The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND, 2024. Read the report.
- Brynjolfsson, E., Li, D., & Raymond, L. R. Generative AI at Work. The Quarterly Journal of Economics, 2025. Read the study.
- Dell'Acqua, F., McFowland, E., Mollick, E., et al. Navigating the Jagged Technological Frontier. Organization Science, 2026. Read the paper.
- Dell'Acqua, F., Ayoubi, C., Lifshitz-Assaf, H., et al. The Cybernetic Teammate: How AI Is Reshaping Collaboration and Expertise in the Workplace. Harvard Business School, 2025. Read the field experiment.
- METR. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. July 2025. Read the study.
- METR. Update on Measuring AI Productivity Uplift for Experienced Developers. February 2026. Read the update.
- DORA. Impact of Generative AI in Software Development. 2024 Accelerate State of DevOps Report. Read the report. Associations are observational, not causal.
- Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. June 2025. Read the forecast.
- National Institute of Standards and Technology. New Report Tackles Monitoring Deployed AI Systems. March 2026. Read the guidance summary.