The Behavioural P&L of AI

Most AI value models count software, infrastructure and estimated time saved. They rarely price the behaviour that determines whether any benefit arrives.

The missing ledger

Enterprise AI programmes usually maintain at least three ledgers:

• technology cost
• delivery milestones
• risk and compliance

They rarely maintain a behavioural ledger.

Yet behaviour determines:

• whether people use the system
• what they use it for
• whether they verify it
• whether they trust it too much or too little
• whether they redesign work
• whether they hide usage
• whether time is redeployed
• whether judgement improves or atrophies

That is an economic blind spot.

Behavioural assets

Calibrated trust

Users understand when the system is reliable and when to verify.

Economic effect:

• lower unnecessary review
• fewer uncorrected errors
• faster adoption
• better decision quality

Appropriate delegation

People give AI work suited to its capability and retain human authority where required.

Economic effect:

• improved throughput
• controlled risk
• reduced rework
• better model utilisation

Learning transfer

AI helps people become more capable, not merely more dependent.

Economic effect:

• capability accumulation
• resilience
• faster onboarding
• reduced key-person risk

Workflow redesign

Teams change roles, queues, handoffs and controls rather than adding AI to the old process.

Economic effect:

• captured capacity
• lower cycle time
• fewer duplicate steps
• stronger scale economics

Honest stopping behaviour

Teams retire low-value uses without treating the decision as failure.

Economic effect:

• capital reallocation
• lower operating burden
• reduced tool sprawl

Behavioural liabilities

Review debt

Outputs are generated faster than people can validate them.

Economic effect:

• hidden queues
• superficial approval
• quality incidents
• delayed delivery

Automation complacency

People stop challenging plausible answers.

Economic effect:

• error propagation
• risk concentration
• weaker judgement

Shadow AI

Users bypass approved tools or processes.

Economic effect:

• data and compliance exposure
• invisible spend
• fragmented learning
• vendor sprawl

Performative adoption

Employees increase visible usage because management rewards prompts, tokens or tool activity.

Economic effect:

• wasted consumption
• distorted metrics
• resentment
• false productivity narratives

Avoidance

People resist tools because they fear surveillance, displacement or loss of expertise.

Economic effect:

• stranded licences
• split processes
• duplicated work
• slower benefit realisation

Capability erosion

Routine cognitive work is delegated before people develop the underlying skill.

Economic effect:

• weaker future workforce
• increased dependency
• poorer exception handling
• reduced resilience

Why the same tool produces different returns

Gartner notes that productivity effects vary by experience and function complexity. OECD reviews similarly show that AI productivity effects are task and context dependent.

Possible patterns include:

• novices benefit from guidance on routine work
• experts use AI to accelerate complex work
• mid-level staff lose time correcting outputs that do not fit domain nuance
• highly standardised roles improve more than ambiguous decision roles
• workers with strong verification skills benefit more than those who cannot judge quality

An average productivity number can hide winners and losers.

Build a behavioural P&L

The term P&L is deliberately provocative. Not every behavioural effect can be booked as accounting profit or loss. The discipline is to identify economic consequences.

Behavioural revenue and throughput

• increased conversion caused by faster or better response
• additional cases completed
• reduced abandonment
• faster product delivery

Behavioural cost

• review time
• correction
• duplicate work
• training
• process friction
• exception handling
• adoption support

Behavioural risk

• error exposure
• compliance breach
• security incidents
• concentration of judgement
• loss of human fallback capability

Behavioural assets

• reusable skills
• trusted workflows
• communities of practice
• evaluation capability
• high-quality adoption patterns

Behavioural liabilities

• tool dependency
• brittle workarounds
• hidden queues
• surveillance distrust
• capability gaps
• unmanaged shadow use

Measures that help

Useful-output rate

Share of AI-assisted outputs that survive review and enter a real deliverable.

Review burden

Human review and correction time per accepted output.

Rework displacement

Whether AI removes rework or moves it downstream.

Trust calibration

Difference between user confidence and actual performance.

Appropriate-use rate

Share of AI use occurring in tasks where evidence shows net benefit.

Workflow capture rate

Share of potential capacity benefit translated into throughput, service, cost or redeployment.

Learning transfer

Whether users improve independent task performance over time.

Shadow-use exposure

Estimated use outside approved systems or evidence standards.

Decision reversal rate

Frequency with which AI-supported decisions are later reversed because of quality or context failure.

No metric should become an individual leaderboard.

Ownership

HR

• capability
• learning
• role design
• employee trust
• workforce transition

COO and process owners

• workflow
• capacity capture
• service level
• operating outcomes

CIO and CAIO

• tooling
• telemetry
• evaluation
• adoption design
• privacy controls

Risk and compliance

• oversight
• material decisions
• human authority
• incident patterns

Managers

• task allocation
• review expectations
• team norms
• local redesign

Employees

• appropriate use
• verification
• escalation
• personal learning

Privacy and surveillance boundary

Behavioural measurement can become harmful.

Principles:

• measure systems and workflows before individuals
• aggregate wherever possible
• collect only data needed for a named decision
• separate learning analytics from performance management
• explain purpose and retention
• involve employee representatives where relevant
• allow challenge and correction
• prohibit raw token or prompt-count leaderboards as performance measures

NIST emphasises explicit human roles, responsibilities and oversight across AI systems. That accountability should protect people as well as the organisation.

Management cadence

Monthly workflow review:

• adoption and avoidance
• useful-output and review burden
• incidents and workarounds
• capacity capture
• training needs

Quarterly portfolio review:

• which behaviours support value
• which liabilities are accumulating
• whether role and process design should change
• whether the initiative should scale, redesign or stop

Conclusion

AI value is not created when a model produces an output.

It is created when people and systems use that output in a way that improves a real outcome.

Behaviour is therefore not beside the economics. It is part of the production function.

Sources

• Gartner, "For AI Value, Focus on Your Use Cases", https://www.gartner.com/en/articles/ai-value
• OECD, "The effects of generative AI on productivity, innovation and entrepreneurship", https://www.oecd.org/en/publications/the-effects-of-generative-ai-on-productivity-innovation-and-entrepreneurship_b21df222-en.html
• OECD, "Using AI in the workplace", https://www.oecd.org/en/publications/using-ai-in-the-workplace_73d417f9-en.html
• NIST AI Risk Management Framework, https://www.nist.gov/itl/ai-risk-management-framework