Five stages—task analysis, portfolio choices, controlled design, evidence-based validation and scaled operations—connect business investment decisions with governance across the AI lifecycle. Drawing on published management research, the NIST risk management framework and agent engineering practice, the method brings evaluation datasets, permissions, human review and exit mechanisms into the design stage. Task requirements guide model selection; validation evidence determines deployment scope. Each capability upgrade prompts a fresh review of business outcomes and risk boundaries.
ZHICE / VALUE TO EXECUTIONChoose a stage to explore the work and conditions for progress.
THE QUESTION TO ANSWER
Whose work will AI change, and at which step?
01 / 05
HOW THE WORK PROCEEDS
Separate information retrieval, interpretation, proposal development and execution. Record task frequency, existing quality, time spent and the consequences of errors to identify where assistance is appropriate.
WORK PRODUCTS
Role and task map, current baseline, responsibility and risk register.
CONDITIONS FOR PROGRESS
Tasks can be described, outcomes can be verified and accountable owners are clear.
Across every stageBusiness valueAccountabilityData and evidenceRisk boundaries
Operating evidence informs judgement. The approach evolves with conditions.
THE DECISION CONTEXT
Understand the relationships. Then choose the tools.
Select a factor to see how it connects to the operating decision and shapes our analysis.
A trusted taskZHICE / DECISION LINKS
01 / THE CONNECTION
Defines why to adopt
Define completion and quality first, then compare with the existing way of working.
02 / THE CONNECTION
Tests fitness for the task
Retain representative tasks, source versions and comparison records to test real quality.
03 / THE CONNECTION
Bounds action
Distinguish reading, drafting, submitting and executing; retain approval conditions for consequential actions.
04 / THE CONNECTION
Continually revises judgement
Use exceptions, costs and usage evidence to decide whether to continue, limit, downgrade or exit.
Read the full analysis
Applying AI in an energy business begins by relating technical possibilities to specific tasks. Policy retrieval, maintenance preparation, trading research and production control have different requirements for accuracy, timeliness and authority to act. ZHICE examines the work, its evidence sources and accountability for consequences to identify suitable assistance and judgments that should remain with people. Use-case planning therefore goes beyond value and difficulty: it checks whether results can be evaluated, people will adopt the application and the organization can maintain ownership. A pilot becomes a validation activity capable of supporting a clear decision.
Moving from trial to operation requires knowledge, evaluation and governance capabilities together. ZHICE brings knowledge versions, tool permissions, human confirmation, exception handling and cost records into one design, distinguishing answer quality, task outcomes and the effects of actual use. Each model or knowledge update should check for regressions in critical tasks; each expansion should examine differences in new roles and uses. The aim is a sustainable basis for investment and operating judgments, so the business knows when to proceed, when to restrict use and when to exit.
PRINCIPLES IN PRACTICE
How judgements are formed.
A method makes the relationships between evidence, choices and accountability open to discussion and testing.
01
A model demonstration can reveal an opportunity without establishing an investment case. A correct response may rely on selected documents and questions; shorter drafting time may be offset by checking and rework. ZHICE first defines the task’s starting point, completion state and quality requirements, then establishes a comparison with existing work. Evaluation covers users of different experience and tasks of different difficulty, including human review, system operation and exception handling in costs. The resulting value judgment must withstand scrutiny from both business and finance.
IMPLICATIONS FOR THE ENGAGEMENT
Agree success and stop conditions before the pilot, retaining representative tasks and comparison records. Confirm benefits only from actual evidence within its applicable scope. External research and model benchmarks can inform hypotheses but cannot replace the business’s own validation.
02
Information retrieval, content drafting and business execution need different controls. The more open-ended the task and extensive the tool permissions, the more unpredictable the path to completion may become; checking the final answer alone is insufficient. ZHICE begins with the minimum necessary capability, testing whether fixed workflows and human collaboration meet the need before considering dynamic planning. Important actions retain explicit authorization, with key tool calls and result states recorded. Insufficient knowledge, conflicting rules or abnormal conditions must trigger workable pause and handover arrangements.
IMPLICATIONS FOR THE ENGAGEMENT
The authorization matrix should distinguish reading, drafting, submitting and executing, with the approving party and conditions for each. Wider autonomy requires new validation evidence. A model upgrade or fluent demonstration does not automatically justify broader business permissions.
03
AI operating conditions change with knowledge, models, tasks and people. Cases passed at launch may not cover later situations; fixing one problem may affect tasks that were previously stable. ZHICE keeps evaluation datasets evolving with applications, feeding real failures, human edits and business changes back into validation. Business, knowledge, technology and risk roles have distinct responsibilities, and significant changes require an explainable basis. Governance becomes part of daily operation rather than a one-time launch check or a policy document that is rarely updated.
IMPLICATIONS FOR THE ENGAGEMENT
Every application needs owners who maintain use, knowledge, evaluation and exception records, with the ability to initiate a pause or downgrade. Periodic reviews must determine specific improvements, scope adjustments and resource commitments so that governance changes actual operations rather than merely producing reports.
FROM INQUIRY TO EXECUTION
Five stages, with clear progression.
Each stage produces an outcome and establishes what the next step requires. Expand a stage for activities, evidence requirements and review criteria.
01
PHASE / Whose work will AI change, and at which step?
Analyze tasks
Task analysis turns broad AI ambitions into units of work that can be discussed. ZHICE separately observes information gathering, interpretation, proposal development and business actions, identifying the evidence people need, their responsibilities and whether current problems arise from capability, process or knowledge conditions. A sufficiently clear task boundary gives value and risk assessment a common subject.
How the work proceeds
Separate information retrieval, interpretation, proposal development and execution. Record task frequency, existing quality, time spent and the consequences of errors to identify where assistance is appropriate.
01
Document complete tasks with the people who perform them. Specify inputs, steps, completion states and approval responsibilities, distinguishing judgments and actions that must remain with qualified professionals.
02
Select representative samples by employee experience, task difficulty and exceptional conditions. Check current quality, effort and review processes rather than choosing only easily demonstrated successes.
03
Identify information gaps, conflicting rules, cross-system handovers and unclear accountability. Determine which issues first need management or knowledge improvements and which are suitable for AI validation.
Stage review
Tasks can be described, outcomes can be verified and accountable owners are clear.
01Tasks can be clearly described and evaluated. Introducing AI does not obscure outcome accountability, and key roles have confirmed the scope.
02Samples are reasonably representative, initial causes have been distinguished, and non-technical prerequisites have arrangements in place.
02
PHASE / Which use cases merit investment, and which should wait?
Choose use cases
Use-case selection turns task analysis into an investment portfolio. ZHICE considers value, knowledge and data readiness, evaluation capability, willingness to adopt and operating consequences, comparing rules, analytical models, retrieval and generative AI. Priorities must explain why to invest now and when to change the decision, rather than substituting a composite score for management judgment.
How the work proceeds
Assess business value, knowledge readiness, testability, willingness to use and operating risk. Compare process improvement, rules-based systems, analytical models and generative AI.
01
Compare business effects, implementation conditions and failure consequences for each use case. Mark evidence gaps in uncertain assumptions, distinguishing work ready for validation from work needing foundations or further observation.
02
Design technical and management alternatives. Check whether process improvement, rules-based automation or existing analytical tools can solve the problem, and explain AI’s additional value and costs.
03
Agree scope, accountability, budget and exit conditions for the first validation. Identify shared knowledge, access and evaluation capabilities to avoid duplicated foundations across departments.
Stage review
Benefit assumptions have a basis and critical prerequisites are confirmed.
01Every selected use case has a business owner and validation objective. Critical prerequisites are confirmed or have a completion plan.
02Investment judgments consider costs and risks together, with workable scope reduction, adjustment or exit if the pilot falls short.
03
PHASE / Where should capabilities and permissions sit?
Design controls
Controlled design determines how capabilities enter real work. ZHICE brings knowledge sources, task paths, tool permissions and human confirmation into one architecture, stating when the system informs, drafts or requests approval. For field operations or important business actions, consequences should determine control boundaries rather than the default settings of an office assistant.
How the work proceeds
Design knowledge sources, retrieval and tool interfaces, human–AI responsibilities and confirmation points. Distinguish advice, work-order drafting and execution, using only the permissions the task requires.
01
Check knowledge sources, versions, applicable equipment and access rights. Design retrieval, citation and conflict handling so users can return to the original text and understand applicability.
02
Choose retrieval, fixed workflows or dynamic agents according to the task. Specify each tool’s inputs, outputs, exceptions and permissions, retaining only the business actions that are necessary.
03
Design human review, approval, pause and takeover points. Distinguish ordinary assistance from high-impact uses, covering insufficient knowledge, access anomalies and tool failures.
Stage review
High-impact actions have explicit approval mechanisms, and knowledge and tool access boundaries can be checked.
01Knowledge applicability and tool authorization can be checked. Important actions retain necessary approval, with clear handover conditions.
02The design covers conflicting information, insufficient permissions and tool failures without relying entirely on the model’s judgment during execution.
04
PHASE / What evidence supports launch or a wider rollout?
Validate with evidence
Evidence-based validation examines answers, actions and the effects of use together. ZHICE builds samples from real, appropriately authorized tasks, covering routine, boundary and failure conditions, and works with business experts to judge applicability. Validation also examines human review, task costs and environmental changes, distinguishing established conclusions from evidence that remains insufficient for launch or expansion.
How the work proceeds
Build evaluations from real tasks processed under appropriate authorization, covering normal, boundary and exceptional cases. With business experts, examine results, action traces, user experience and full costs.
01
Connect tasks with evaluation criteria. Combine automatically verifiable checks, professional judgment and user feedback, with appropriate human calibration of model-based grading.
02
Repeat critical-task validation in a suitable controlled environment. Examine final states, cited evidence, tool actions and exceptional behavior, recording failure conditions and unresolved issues.
03
Compare the complete effort of existing and AI-assisted work across roles and tasks. Bring quality, operating, cost and risk evidence together in a launch recommendation.
Stage review
Previously agreed task standards are met, with explicit decisions on unresolved issues.
01Previously agreed task and operating conditions are met. Important failures have been addressed, and residual risk decisions are recorded.
02Value assessment includes review and rework. Conclusions remain within the sample and use-case scope, with explicit conditions for expansion.
05
PHASE / How will performance and accountability hold as use expands?
Operate at scale
Scaled operations manages continuing change after expansion. New roles, knowledge versions, model capabilities and external services may alter performance and risk. ZHICE assigns lasting responsibility for knowledge maintenance, regression evaluation, exception handling and value reviews, adjusting scope as evidence accumulates. Scale is demonstrated by the organization’s ability to manage these changes consistently, not simply by more accounts or deployments.
How the work proceeds
Establish knowledge updates, model change controls, regression evaluation, user feedback and incident reviews. Track effects by role and use case, adjusting rollout scope in stages.
01
Record actual failures, human edits and takeovers. Attribute issues to knowledge, retrieval, models, tools or processes, assigning owners to resolve and verify them.
02
Run relevant regression checks for model, knowledge, prompt and tool changes. Examine task quality, permission boundaries and full costs, downgrading, restricting or pausing use when needed.
03
Have business and operations teams regularly review value and differences in adoption. Update training, task scope and investment plans, treating each new use case as a business decision requiring renewed validation.
Stage review
Ownership remains active, exceptions can be contained, and benefits and costs can be reviewed.
01Knowledge, evaluation and operations ownership persists. Exceptions can be detected and handled, with executable pause and takeover arrangements.
02Upgrades and expansion have supporting evidence. Benefits and costs can be reviewed, and new uses do not bypass required governance decisions.
ANALYTICAL FRAMEWORK
Exhibit 1 | Weigh evidence readiness against the consequences of action
Priority depends on whether a task can be evaluated and failure controlled. High potential value does not automatically justify greater autonomy in a high-impact use case.
Swipe to explore the complete figure
01
Build evidence; retain human judgmentBusiness impact is high but validation conditions are insufficient. Narrow the task, strengthen knowledge and evaluation, and define failure consequences. Demonstration success does not replace a launch assessment.
02
Proceed with staged authorizationValidation conditions are relatively strong, but action consequences remain material. Combine professional review, required approvals and a limited scope, checking accountability and recovery arrangements throughout.
03
Explore with a clear learning goalAction consequences are lower, but evidence is weak. Use small experiments to clarify task and knowledge needs, record failures and avoid premature commitments to scale.
04
Validate actual value firstTask boundaries and evaluation conditions are clearer. Conduct comparisons and controlled trials while checking review effort, adoption and full costs to establish a basis for expansion.
An original ZHICE qualitative framework. Evidence readiness increases from left to right; impact increases from bottom to top. Positions express decision approaches, not measured scores or expected investment returns.
ANALYTICAL FRAMEWORK
Exhibit 2 | Turn actual failures into reusable evaluation and operating capability
A continuous loop from user feedback to regression review avoids repairs aimed only at individual examples and detects regressions hidden by model updates.
Swipe to explore the complete figure
01
Capture actual issuesRecord tasks, relevant knowledge, operating conditions, human edits and key actions. Retain enough evidence for review while handling information within authorization requirements.
02
Assign the cause and ownerDistinguish outdated knowledge, retrieval omissions, model errors, tool failures and process accountability issues. Identify the responsible role and dependencies requiring joint work.
03
Revise capabilities and rulesAdjust knowledge, applications, permissions, training or working practices. State the reason and scope of each change, rather than attributing every issue to model capability.
04
Retest and reviewAdd failures to evaluations, check related tasks and exceptions, then decide whether to restore, expand, restrict or pause. Feed the decision back into actual use.
An illustration of ZHICE’s AI operating approach. A feedback loop does not guarantee continuous improvement; each change still needs testing against applicable tasks and a decision by an accountable owner.
EVIDENCE & ACCEPTANCE
Make deliverables useful to decisions.
Go beyond document names to define purpose, evidence and verification. Agree the work scope against actual conditions at project initiation.
DECISION TOOL
Trusted AI value: turn capability assessments into stage decisions
Swipe to explore the complete table
Stage and scope
Core outputs
Evidence for validation
Leadership decision
Analyze tasks: Define input, judgment, action and responsibility boundaries around specific roles and work outcomes.
Produce the task map, baseline and problem classification, distinguishing suitable assistance from conditions that first need improvement.
Provide representative tasks, operating records and role confirmation, specifying completion states, quality requirements and known exceptions.
Confirm tasks worth further study, responsibilities that cannot be transferred to the system and foundational work still needed.
Choose use cases: Compare value, implementation readiness, evaluation capability and consequences to allocate adjustable validation funding.
Produce the use-case portfolio, alternatives and pilot agreement, specifying budgets, owners and conditions to continue or exit.
Provide value hypotheses, dependencies and evidence gaps, explaining priorities and the choice of technology.
Approve initial validation scope and resources, commission additional evidence for unready cases and defer unsuitable directions.
Design controls: Connect knowledge, tasks and authorization within explicit human–AI collaboration and approval boundaries.
Produce application architecture, knowledge specifications and a permissions matrix, with human review, exceptions and takeover practices.
Provide knowledge provenance, tool behaviors and scenario cases demonstrating that important information and actions can be checked.
Determine whether experimental design and accountability are sufficient, and proceed to controlled validation or add necessary controls.
Validate with evidence: Examine content, actions, adoption and costs together, defining evidence sufficient for a limited launch.
Produce evaluation findings, failure cases and launch recommendations, distinguishing conditions met, restricted uses and required remediation.
Provide task results, key actions and expert judgments, comparing human review, rework and operating costs.
Use pre-agreed standards to launch, limit use or continue remediation, recording residual risks and conditions for later expansion.
Operate at scale: Manage changes in knowledge, models, tasks and roles, feeding actual experience into renewed validation.
Establish operating ownership, change regression and value reviews, defining improvement, expansion, downgrade and exit arrangements.
Provide incident handling, repair verification, adoption differences and cost records, demonstrating evidence behind scope changes.
Decide whether to continue investment or adjust scope, restrict or pause where needed, and retain takeover and recovery arrangements.
This table describes consulting work and decision evidence. It does not promise returns or replace professional, safety or investment approvals. Launch and expansion require judgments specific to the use, site conditions and applicable requirements.
OUR PRACTICE
Explore the critical questions.
Each practice examines diagnosis, solution design, implementation and delivery requirements in detail.
We translate management research and technological advances into working methods for energy businesses. The frameworks below are ZHICE's synthesis of the research and industry questions.
Application in the ZHICE approachApply trustworthiness and risk management across AI design, development, use and evaluation, specifying intended use, accountability and continuing monitoring.
Scope and limitationsA voluntary risk management framework, not proof of legal compliance for projects in China. Material described as a concept note on the 2026 critical infrastructure Profile page is not treated as a formally released standard.
Application in the ZHICE approachIdentify risks specific to generative AI and incorporate trustworthy knowledge, information security, human review and operational feedback into application design.
Scope and limitationsA cross-industry risk guide. Controls must be selected for the task, data, people and system environment rather than applied mechanically.
Application in the ZHICE approachDistinguish predefined workflows from agents in which models dynamically choose steps, keeping architecture complexity proportionate to task needs.
Scope and limitationsAn engineering practice article, not an industry standard. The official site notes that some tool information has changed; only its architecture selection principles are used here.
Anthropic EngineeringDemystifying evals for AI agents · 2026-01-09
Application in the ZHICE approachEvaluate task outcomes, action records and performance across repeated runs, combining automated checks, model-based assessment and human judgment.
Scope and limitationsGeneral agent engineering experience cannot replace on-site validation, safety assessment or professional judgment in the energy industry.
Erik Brynjolfsson / Danielle Li / Lindsey Raymond · The Quarterly Journal of EconomicsGenerative AI at Work · 2025-02-04
Application in the ZHICE approachDraw on perspectives on differences in employee experience and learning effects to test AI outcomes by role, proficiency and task difficulty.
Scope and limitationsThis reference uses the final paper published in QJE in 2025. The study concerns customer support workers at one company; its average productivity change is not extrapolated into benefits for energy businesses.