Performance
Time, quality, cost and scope of the product obtained with assistance, discounting review, integration and error correction.
What does it really improve and for what tasks?Case studies · Organizations
Cognitive autonomy also is designed.
An organization does not become dependent because it uses a lot of AI, but when it stops knowing how to decide, verify, learn or continue without it. The dependency can be hidden by good productivity indicators until a plausible error, a crash, a change of supplier or a situation that requires undocumented judgment appears.
It does not stop adoption nor does it propose to keep everything manually. It helps decide where to accelerate, where to introduce control and what capabilities should be maintained because they support security, learning or adaptation.
Indicative overview
Organizational autonomy
Small teams · concentrated knowledge
SME or local business
Few rules, visible leaders and an operational alternative usually contribute more than a complex committee.Illustrative values: show relative priorities, not results of an evaluation.
Dominant prioritySimplify without improvisingMap tasks →
Double control panel
The assisted result alone does not describe the organization.
Responsible adoption measures both the value created and the capacity that remains within the organization. What seems like efficiency today may turn into dependence on a tool, a supplier, or a few people tomorrow. The double scorecard avoids optimizing the product while degrading supervision, collective learning or continuity.
Ability
What people and teams understand, remember, and can reconstruct, plus the tacit knowledge they need to interpret exceptions.
What do we know to do if we withdraw aid?Resilience
Ability to detect errors, absorb a fall, change suppliers and continue operating safely without improvising responsibilities.
What happens when the AI fails or is not there?Distribution
Who receives the benefits, who endures the change, who gets training, and who retains voice, mobility, and learning opportunities.
Does the improvement reduce or widen inequalities?Field evidence · customer service
The average gain can hide very different effects.
In a company with 5,172 agents, assistance increased average productivity by 15% and that of less experienced profiles by 30%. The study also observed changes in quality and retention, and signs of learning in periods without access. It is a relevant result, but it comes from a specific company, tool and task; it does not by itself describe other sectors or the long-term effect.
Brynjolfsson, Li & Raymond, 2025 ↗0%10%20%30%
Average+15%
Less experience+30%
The study also found signs of lasting learning during system outages, although it warns of limits and heterogeneity.Delegation architecture
Don't classify positions. Classify tasks.
The same profession contains operations with very different risks, required learning and verification possibilities. Classifying an entire position as “automatable” erases that diversity. Define a task, identify who receives its effects, and choose a case to see how the control design changes.
Pilot and sampleLow impact · difficult control
Reinforced human decisionHigh impact · difficult control
Reversible delegationLow impact easy control
Assist with verificationHigh impact easy control
difficult to verifyEasy to verifyLow impactHigh impactInternal report
Recommended layout
Delimited assistance
You can automate parts, but you should not erase assertion authorship or traceability.
- Responsible
- Responsible for the document
- Control
- Sources, sampling and approval identified
- Continuity
- Original template and fonts available
!
The matrix guides; does not certify. Legal risk, security, privacy and rights require specific evaluations.
Guidance self-diagnosis
Does adoption create capacity or dependency debt?
Assess observable practices, not intentions or isolated documents. Ask whether there is an up-to-date inventory, authority to disagree, independent verification, unaided practice, rehearsed continuity and participation. The result helps start a conversation and choose priorities; it is not a validated scale or a compliance audit.
Indicative maturity
47/100
Initial government
Controls are already appearing, although they still depend on local initiatives or specific people.
Current priorities
Decision
Define what AI can propose, who decides and what decisions are not delegated.
Ability
Periodically evaluate essential unassisted tasks and schedule deliberate practice.
Editorial heuristics. The six dimensions synthesize principles of human factors, organizational learning, and governance frameworks. They do not allow comparing companies or inferring clinical, legal or financial risk.
Operating system
From principle to verifiable practice.
The cycle references the Govern, Map, Measure, and Manage functions of the NIST AI RMF, and explicitly adds human capacity maintenance. Each layer must leave operational evidence: responsible parties, inventories, tests, thresholds, incidents and exercises that can be reviewed when the model or process changes.
Govern
Purpose, limits and responsibility
Connect each use to a legitimate purpose, an acceptable level of risk, and a person with real authority.
- ✓Brief and understandable policy
- ✓Prohibited or reserved uses
- ✓Life cycle manager
Minimum test
A person can say who is responsible for each assisted decision.
Literacy and supervision are not synonymous with a generic course. Article 4 of the European AI Regulation calls for measures adapted to knowledge, experience, training and context; article 14 requires effective human oversight for high-risk systems. This case study does not determine legal classification or prove compliance.
Consolidated text ↗Development and organizational justice
Transition does not affect everyone in the same way.
Exposure depends on the tasks and also on access, training, language, disability, job security, experience and the real possibility of participating in the redesign. Measuring only the average can obscure who loses practice, endures more review, is excluded from learning tasks, or lacks a safe alternative.
ILO · global index, 2025
1 in 4Workers are in an occupation with some potential exposure to generative AI.
Exposure means possibility of task transformation; it is not equivalent to job replacement.
Highest exposure3,3%
of global employment
Global employment in the most exposed categoryBy sex does not explain causality
Women4,7%
Men2,4%
Consult methodology and scope ↗Train before demanding
Literacy should be tailored to tasks, risks, and experience level, with protected time for practice and feedback on real cases. A generic course does not demonstrate competence.
Participate before imposing
Those doing and receiving the work detect tacit knowledge, exceptions, review burdens, and invisible costs that rarely appear in a demonstration.
Support differentially
Accessibility, language, digital competence, age, caring responsibility and employment status may require different supports to achieve an equivalent opportunity.
Share the benefit
Productivity must also translate into learning, quality of work, reasonable time and opportunities for progression, not just more expected volume.
Team simulation
Rehearse before you need it.
A short practice allows you to discover dependencies that do not appear in a policy: concentrated knowledge, missing permissions, symbolic verification or an alternative that no one has tried. Choose an incident, bring together the functions involved and respond without looking for blame; the goal is to redesign the system.
Simulation · 35 minutes
AI stops responding for 45 minutes
The volume of work continues to come in and several tasks no longer have visible alternative procedures.
Closing questionWhat part of the work could continue today without consulting any assistant?
Identify blocked processes, affected people and decisions that cannot wait.
Activate the approved alternative and reduce the range before improvising with another tool.
Assign priorities, authority and return criteria to normal service.
Record what knowledge was missing and schedule a new continuity test.
First cycle
From zero to a governed adoption in 90 days.
Inventory real uses, select three priority tasks and appoint those responsible.
Compare assisted and unassisted results, incorporate controls and listen to the teams.
Approve criteria, train by function, rehearse incidents and review indicators.
Evidence and frameworks
What does this case study support?
Empirical results, classical frameworks, exposure indices, and normative references are separated because they answer different questions. An experiment can estimate performance on a task; a framework organizes controls; A standard establishes obligations. No single study demonstrates a general cognitive loss caused by AI or proves that an organization is well governed.
Actual deployment · a companyAverage productivity increased by 15%; in the least experienced profiles, 30%.Brynjolfsson, Li & Raymond · QJE, 2025 ↗Knowledge work · experimentImprovements depend on whether the task falls within an irregular technological frontier.Dell’Acqua et al. · Organization Science, 2025 ↗Collaboration · meta-analysisPersona and AI together do not automatically beat the best of both separately.Vaccaro et al. · Nature Human Behaviour, 2024 ↗Assisted decision experimentThought-provoking features reduced overdependence more than a simple explanation.Buçinca, Malaya & Gajos · PACM HCI, 2021 ↗Automation classic frameworkLeaving the person alone to make exceptions may require more training, not less.Bainbridge · Automatica, 1983 ↗World employment task indexThe exposure is uneven and describes potential for transformation, not automatic loss of employment.OIT · Working Paper 140, 2025 ↗Voluntary framework governanceGovern, Map, Measure and Manage offer a shared language for managing risks.NIST AI RMF 1.0 and Generative AI Profile ↗
!
Organizational evidence still has limits. Many studies cover tasks, tools, and short periods. That is why this case study proposes to measure locally, publish limits and review decisions when the context changes.