The Blockers Are Human: What FinOps Adoption Teaches Us About AI Value

Introduction

The standard explanation for why AI programmes fail to deliver value goes like this: the technology is immature, the models are unreliable, the infrastructure is expensive, the data is messy, the use cases are unclear.

Evidence The evidence points elsewhere. The binding constraint is not the technology. It is behaviour and operating model. Interpretation This is not a new lesson. Cloud FinOps learned it the hard way over a decade. AI value management is learning it again.

The Evidence That the Constraint Is Human

SpeculationMIT NANDA study (2025): 95% of enterprises report no measurable P&L impact from AI investments. The researchers did not attribute this to technical failure. They attributed it to a “learning gap”: the organisational difficulty of integrating AI into workflows, decision-making, and accountability structures.

Important caveat: This finding is widely cited but methodologically contested. The study measured absence of documented P&L impact, not technical failure. Many pilots may have produced value that was not measured, attributed, or captured financially. The 95% figure should be treated as directional evidence of a measurement and integration challenge, not proof that 95% of AI pilots fail technically.

The technology worked. The integration did not.

Interpretation This is the pattern: organisations gave staff access to AI tools, measured activity, and assumed value would follow. The blockers turned out to be human: resistance, misaligned incentives, unclear ownership, fear of displacement, and the absence of forums where cost and value could be discussed together.

Behavioural Outcomes Matter

Evidence Gartner distinguishes three types of AI outcomes: business outcomes (revenue, cost, quality), technology outcomes (performance, reliability, efficiency), and behavioural outcomes (trust, adoption, learning, workflow change).

Evidence Gartner reports that 91% of surveyed CIOs and IT leaders devote little or no time to scanning for behavioural by-products of AI adoption. This is an economic blind spot.

Behavioural outcomes determine:

• whether people use the system appropriately
• whether they trust it too much or too little
• whether review burden increases or decreases
• whether learning transfers or capability erodes
• whether time saved becomes capacity captured

Interpretation Ignoring behavioural outcomes means measuring cost and activity while missing the human factors that determine whether value arrives. For a structured approach to behavioural economics, see The Behavioural P&L of AI.

The Adoption Ladder: From Access to Measured Outcome

Interpretation Adoption is not binary. It is a progression through distinct stages, and most organisations confuse early stages with full adoption.

The adoption ladder has seven rungs:

1. Access: users have credentials and permissions
2. Awareness: users know the tool exists and what it is for
3. Training: users have completed formal instruction
4. Experimentation: users try the tool in low-stakes scenarios
5. Regular use: users incorporate the tool into daily work
6. Workflow integration: the tool is embedded in standard operating procedures
7. Measured outcome: the organisation can demonstrate validated business results

EvidenceIBM's 2026 CEO Study found that only about 25% of the workforce regularly uses AI, despite 86% of CEOs believing employees have the required skills and knowledge. This gap between executive perception and workforce reality sits between training and regular use on the ladder.

Interpretation Most AI value claims assume the organisation is at rung seven when it is actually at rung three or four. That is why time-saved estimates fail to materialise as financial value.

Production-Readiness Controls

Evidence Moving from experimentation to production requires more than technical deployment. It requires operating-model readiness.

Production-readiness controls include:

• Role-based training: not generic awareness, but job-specific instruction on when and how to use AI
• Updated SOPs and runbooks: documented procedures that incorporate AI into workflows
• Usage monitoring: observability of who is using AI, for what, and with what outcomes
• Exception-management design: clear escalation paths when AI output is uncertain or incorrect
• Human authority and escalation: explicit rules for when humans must review, approve, or override AI decisions

Interpretation Organisations that skip these controls often discover adoption problems only after deployment, when usage is low, quality is inconsistent, or errors create downstream problems.

Conviction Distance: The Gap Between Confidence and Reality

Interpretation One of the most useful adoption concepts is conviction distance: the gap between executive confidence, manager readiness, workforce use, workflow integration, and realised outcome.

Measure conviction distance across five levels:

1. Executive confidence: C-suite belief that AI will deliver value
2. Manager readiness: middle management prepared to change workflows and measure outcomes
3. Workforce use: employees regularly using AI in their work
4. Workflow integration: AI embedded in standard operating procedures
5. Realised outcome: validated financial or strategic results

EvidenceBCG's 2026 research found that around 72% of CEOs identify themselves as the main AI decision-maker, and about 82% are more optimistic about AI return. Yet only about 5% of companies in BCG's technology-function study were generating measurable AI value at scale. This is conviction distance in action.

Interpretation Large conviction distance is not necessarily wrong. It can reflect rational strategic conviction in the face of early-stage uncertainty. But it should be visible, measured, and actively managed rather than ignored.

The CAIO-CHRO Partnership

Interpretation Adoption, role design, capacity conversion, and reskilling require shared ownership between the CAIO and CHRO. This partnership is often missing.

EvidenceIBM's 2025 CAIO research noted that some CAIOs view the CHRO as a detractor, despite the CHRO's potential importance to adoption. This is a governance failure, not a personality conflict.

The CAIO-CHRO partnership should address:

• Workforce capacity conversion: how time saved becomes redeployed capacity or cost reduction
• Role redesign: which roles change, which disappear, which are created
• Reskilling and transition: investment in capability building and career pathways
• Adoption measurement: tracking usage, workflow integration, and behaviour change
• Change management: communication, training, and support systems

Interpretation Without this partnership, AI adoption becomes a technology rollout rather than an operating-model transformation. The technology may work, but the organisation does not change.

A Resistance-and-Enabler Map

Here is a simplified map of where resistance shows up, and what the counter is:

Engineering fears surveillance

The resistance: Engineers fear that AI value measurement will become performance surveillance. If every prompt is logged and every output is scored, the fear is that management will use it punitively.

The counter: Make measurement about systems, not individuals. Aggregate data. Focus on workflow-level outcomes, not person-level activity. Build trust before building dashboards.

Finance distrusts soft benefit claims

The resistance:Finance has heard “productivity improved” too many times without seeing it in the P&L. They want hard numbers, not anecdotes.

The counter: Agree on attribution methods in advance. Use the distinction ladder to separate activity from outcome. Report attribution coverage as the scoreboard, not claimed savings.

Centre of excellence fears being a bottleneck

The resistance: The CoE wants to enable, not gatekeep. But without standards, they fear chaos. With too many standards, they fear becoming the thing that slows everything down.

The counter: Hub-and-spoke model. The CoE sets standards, provides shared infrastructure, and maintains the attribution framework. Business units own their outcomes and report back.

Business units disown the spend

The resistance:Business units say “we did not ask for this” when AI spend shows up in their allocation. They resist accountability for something they did not control.

The counter: Showback before chargeback. Make cost visible first. Build shared understanding of unit economics. Only move to chargeback when there is genuine ownership.

The workforce fears displacement

The resistance: This is not irrational fear. Some roles will change. Some will disappear. Pretending otherwise destroys trust.

The counter:Be honest. Name what is changing. Invest in reskilling. Make the augmentation case credible by showing real examples, not generic claims. Do not wave away displacement concerns with “AI will create new jobs” without showing what those jobs are.

The FinOps Lessons That Transfer

Cloud FinOps spent a decade learning how to govern variable, decentralised spend without strangling experimentation. Evidence The lessons transfer directly to AI value management:

Showback before chargeback

Make cost visible before making it punitive. Chargeback without context breeds gaming. Showback builds shared understanding first, then accountability follows naturally.

Community over mandate

Top-down mandates without practitioner buy-in create compliance theatre. Real change comes from communities of practice that share learnings and build shared standards.

Executive sponsorship is non-negotiable

Evidence FinOps Foundation research consistently shows that successful programmes have visible executive sponsorship. Without it, FinOps becomes a reporting exercise that no one acts on.

Crawl, walk, run

Do not skip stages. Organisations that try to implement advanced optimisation before establishing basic visibility fail predictably. Maturity is earned, not declared.

Incentive Design That Does Not Backfire

Interpretation The hardest part of AI value management is incentive design. Get it wrong, and you create gaming. Get it right, and you create alignment.

Never reward activity

Do not reward prompt count, model usage, or tool adoption. These are activity metrics. They can be gamed. They do not prove value.

Evidence The tokenmaxxing lesson: A backlash emerged in 2026 against treating raw token consumption as a performance metric. DORA and Nature Machine Intelligence both warned that rewarding token volume creates perverse incentives—teams game the metric by running unnecessary prompts, inflating context windows, and optimising for activity instead of outcomes.

Reward attribution coverage improved

Reward teams that improve their attribution coverage: the proportion of AI spend for which they can demonstrate a business outcome. This is hard to game because it requires proving value, not just claiming it.

Reward laggards retired

Reward teams that kill initiatives that are not working. This creates a culture where stopping is not failure, it is discipline.

Reward cost per successful outcome falling

Reward teams that improve unit economics: the cost per successful outcome. This aligns cost optimisation with value delivery.

Name the J-Curve Dip in Advance

Evidence AI value lags investment by design. This is the productivity J-curve: returns dip before they climb. The dip is normal and predictable.

Interpretation The most expensive failure is cancelling during the dip. Organisations that do not name the J-curve in advance treat the dip as failure and cancel before the curve turns.

The defence: tell sponsors in advance

Before the programme starts, tell sponsors:

• Returns will lag investment by 6-18 months
• Productivity may dip before it climbs
• This is normal, not failure
• The dip is where most programmes get cancelled
• We will measure attribution coverage as the scoreboard, not immediate ROI

Interpretation Naming the J-curve in advance turns it from a surprise into a milestone. It gives sponsors permission to hold steady through the dip.

As AI Goes Agentic, Adoption Includes Agents

Speculation This section is speculative. As AI becomes agentic, adoption stops being only about people. Agents need objectives, guardrails, and risk tiers. This is a management problem, not only a technical one.

Agents need objectives

What is the agent trying to achieve? What counts as success? What are the constraints? These are management questions, not technical ones.

Agents need guardrails

What can the agent do? What can it not do? What requires human approval? These are governance questions.

Agents need risk tiers

Not all agents should have the same level of autonomy. High-risk decisions need tighter controls. Low-risk decisions can be fully automated. This is a risk management question.

SpeculationThe organisations that figure out agent governance early will have a durable advantage. The ones that treat agents as “just another tool” will struggle.

The Argument: Augmentation Optimism Versus Honest Displacement

There are two competing narratives about AI and work:

The augmentation case

AI removes drudgery, raises the ceiling on what individuals can achieve, and creates new roles that did not exist before. The workforce becomes more productive, not smaller.

Evidence for: GitHub Copilot studies show developers completing tasks faster. Customer service agents handling more complex cases. Analysts producing more insights.

The honest displacement case

Some roles will change. Some will disappear. Pretending otherwise destroys trust. The workforce that believes AI is being done to them, not with them, will not adopt.

Evidence for: Agentic AI can do work that previously required people. Per-seat pricing is breaking because one person with agents can do the work of five. That is not augmentation, that is substitution.

The synthesis: both halves are needed

Interpretation The credible story is both halves: AI will augment some roles and displace others. The organisations that succeed are the ones that are honest about both, invest in reskilling, and make the augmentation case with real examples, not generic claims.

Morale is a mechanism. A workforce that believes AI is being done to them will not adopt. A workforce that believes they are being equipped to do better work will.