What Work Should Never Be Delegated Blindly to SI?

What work should never be delegated blindly to Super Intelligence? Any work in which a wrong answer, hidden assumption or unauthorised action can materially affect safety, rights, money, employment, legal position, confidential information, security, public trust or an irreversible external state should never be handed to machine intelligence without appropriate evidence, boundaries and human accountability.

The important word is blindly. This article is not an argument for keeping Super Intelligence away from difficult work. SI can organise evidence, retrieve sources, compare documents, draft alternatives, detect inconsistencies, prepare decisions and execute bounded actions. The boundary is that capability does not equal authority, fluency does not equal verification, and autonomy does not erase organisational responsibility.

In this eduKateSG workplace series, Super Intelligence is the practical machine-intelligence layer commonly described as AI, generative AI, intelligent assistants, copilots, agents and connected automation. The role-taxonomy article explains Assistant, Copilot, Coworker, Agent and Infrastructure. This page owns the delegation boundary: where a workflow must slow down, expose evidence, keep specialist or managerial authority visible, or stop.

The Seven Conditions for Bounded Delegation

Evidence: important claims can be traced to reliable, current sources.
Scope: the case is inside a defined and tested operating envelope.
Authority: the system is permitted to read, recommend, prepare or execute the requested action.
Verification: an appropriate person, rule, test or system can check the result.
Reversibility: an error can be contained or corrected before disproportionate harm occurs.
Escalation: the system knows when to stop and which owner receives the case.
Accountability: a real person or organisation owns the outcome and the improvement loop.

A missing condition does not necessarily prohibit SI. It usually changes the role. The system may remain useful for retrieval, evidence preparation or drafting while the consequential decision or external action remains under a stronger control.

Authority Should Be Decomposed

Read: retrieve approved information.
Interpret: summarise, classify, compare or extract.
Recommend: propose options or next actions.
Prepare: stage an action without executing it.
Execute: perform an authorised action in an external system.
Escalate: stop and transfer the case to a responsible person.

Many useful workplace workflows should stop at Interpret, Recommend or Prepare. Giving every intelligent system execution authority merely because tools are available creates risk without necessarily creating value.

Current Governance References

Singapore’s IMDA updated its Model AI Governance Framework for Agentic AI on 20 May 2026. The framework emphasises bounding agent powers, meaningful human accountability, technical controls and end-user responsibility; the update also added practical material on automation bias, multi-agent systems and third-party agents. See IMDA’s updated framework.

NIST’s Generative AI Profile, NIST AI 600-1, is a voluntary cross-sectoral companion to the AI Risk Management Framework and frames risk management across the lifecycle of generative-AI systems. See NIST’s Generative AI Profile.

Singapore’s PDPC Advisory Guidelines on the Use of Personal Data in AI Recommendation and Decision Systems provide guidance for organisations and developers using personal data in AI systems. See PDPC’s advisory guidelines.

Prompt injection becomes especially important once agents can browse, retrieve untrusted content and use tools. OpenAI’s March 2026 security guidance describes these attacks as an evolving form of manipulation and argues for constraining the impact even when attacks are not perfectly filtered. See Designing AI agents to resist prompt injection.

1. Physical Safety and Critical Operations

Tasks that can affect physical safety, hazardous environments, critical infrastructure or essential operations should never be delegated blindly. A fluent plan or confident instruction can be acted upon before an error is discovered, so consequence and latency matter as much as average accuracy.

Evidence boundary. Use current authoritative procedures, validated engineering or operational data, sensor state where relevant and qualified domain review. Missing safety-critical facts should remain missing; the system should not infer them from a familiar pattern.

Authority boundary. The ability to describe a safe action does not grant authority to execute it. Emergency, shutdown, isolation and hazardous-work decisions should remain inside the organisation’s established control regime unless a specifically validated automated system is authorised for them.

Why blind delegation fails here. In this class of work, the system can produce an output that looks complete before the underlying state is actually known or before the organisation has made the decision it is entitled to make. The failure is therefore larger than model accuracy. It includes missing context, unearned authority, weak review, and the possibility that a plausible output crosses into the external world before the weakness becomes visible.

Safer operating pattern. Use SI for monitoring, evidence retrieval, incident summarisation, checklists and bounded analysis. Put strong verification immediately before high-impact actions, and make stop conditions conservative when the system encounters conflicting state or an unfamiliar condition.

Illustrative case. An operations assistant can summarise alarms and retrieve the correct runbook, but an isolation action outside the tested envelope should route to the responsible operator rather than being executed because the model’s explanation sounds convincing.

First weak link and repair. Start by identifying the earliest point at which the workflow loses reliable information or legitimate authority. Repair that point before adding a more capable model or more autonomy. If the source is stale, fix source ownership. If the rule is ambiguous, clarify the rule. If review is overloaded, narrow eligible cases or improve verification. If tool permissions are too broad, reduce them.

Measurement. Measure detection quality, false alarms, missed critical events, time to qualified review and the number of cases correctly escalated rather than only the speed of the model response. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

Transfer test. Do not copy the same delegation level into another department merely because the label sounds similar. Compare the consequence of error, source quality, receiver, reversibility, legal or professional obligations and the authority required. Transfer the mechanism only when those conditions are sufficiently similar.

2. Health and Medical Decisions

Health-related work can involve incomplete records, changing patient state and specialised professional responsibility. A model may organise the record accurately overall while missing the one contraindication or recent change that matters.

Evidence boundary. Preserve source dates, current medication state, test results, clinician notes and uncertainty. Do not allow frequency of historical mentions to override a later authoritative update.

Authority boundary. General-purpose SI can support information work, but diagnosis, treatment and other consequential clinical decisions belong inside appropriate clinical processes and qualified professional authority unless a specifically validated regulated system provides a different authorised pathway.

Safer operating pattern. Use SI for record organisation, summarisation, question preparation and evidence retrieval, with clinical review proportionate to the use. Keep the patient’s current state and source-of-record information visible.

Illustrative case. A summary may list a medication found in several older notes even though a later note discontinued it. The workflow needs a currentness rule and clinician review before the summary informs care.

Measurement. Measure clinically material omissions, corrections, source-traceability and whether the workflow reduces preparation burden without increasing the chance that stale information reaches the decision. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

3. Legal Rights, Contracts and Obligations

Contracts, regulatory filings, disputes and decisions affecting rights can create obligations that depend on jurisdiction, facts, timing and professional interpretation. A polished answer can hide a missing authority or an incorrect jurisdictional assumption.

Evidence boundary. Keep source clauses, current rules, dates, governing jurisdiction and material factual assumptions inspectable. Distinguish direct source text from model inference.

Authority boundary. SI can search, extract, compare and draft. It does not acquire professional or signing authority merely because it can produce legal-sounding language.

Safer operating pattern. Use the system to prepare issue lists, clause comparisons and draft language while the authorised legal or business owner interprets consequence and approves commitments.

Illustrative case. A model can flag a changed indemnity clause against the organisation’s standard wording, but it should not declare the contract safe to sign without the responsible professional deciding what the difference means.

Measurement. Track material issues detected, material issues missed, review time saved, source accuracy and downstream rework rather than the number of contracts processed. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

4. Hiring, Promotion, Discipline and Employment Opportunity

Employment decisions affect people directly and can raise fairness, privacy and legal concerns. Models may rely on proxy signals, incomplete records or historical patterns that appear objective while embedding undesirable assumptions.

Evidence boundary. Use job-related criteria, current records, transparent evidence and clearly bounded data. Do not infer personal traits from gaps or from signals that are not relevant to the decision.

Authority boundary. Scheduling, drafting and evidence organisation can be delegated widely; consequential employment judgments require accountable organisational decision rights and suitable human governance.

Safer operating pattern. Let SI reduce administrative load and structure evidence, while keeping criteria visible and ensuring decision-makers can inspect the underlying information rather than a single generated score.

Illustrative case. A candidate with strong non-standard experience may be undervalued by a taxonomy built around conventional titles. The system should surface the mapping uncertainty instead of turning it into an automatic rejection.

Measurement. Measure administrative time, consistency of evidence capture, reviewer agreement, corrections and any outcome-quality or fairness indicators appropriate to the organisation’s obligations. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

5. Payments, Pricing and Material Financial Commitments

Money-moving actions are consequential and often attractive to automation because they involve repetitive records. A single wrong amount, duplicate action or unauthorised approval can create immediate loss.

Evidence boundary. Amounts should come from authoritative financial systems. Exact arithmetic should be checked deterministically where possible, and the policy or approval threshold used should be recorded.

Authority boundary. Analysis and anomaly detection are different from approval and execution. The model’s ability to explain a transaction does not confer authority to move funds.

Safer operating pattern. Separate interpretation from approval and execution. Use transaction limits, dual control, explicit thresholds, idempotency protection and system-of-record confirmation where appropriate.

Illustrative case. An agent that receives a timeout after initiating a payment should not blindly retry if it cannot tell whether the first transaction succeeded. It should check transaction state or escalate.

Measurement. Track duplicate prevention, exception rate, approval accuracy, reconciliation issues, escaped errors and actual end-to-end processing time. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

6. Privileged Access, Credentials and Secrets

Administrator rights, production access, credentials and secret material change the security boundary of the organisation. Broad machine access creates a large failure radius if the system is manipulated or simply wrong.

Evidence boundary. Verify identity, request purpose, current entitlements, system scope and approval state against the authoritative access-management process.

Authority boundary. Reading a ticket or recommending an access level is not the same as granting it. Privileged writes should remain separately controlled.

Safer operating pattern. Apply least privilege, scoped credentials, temporary access where possible, strong logging, separation of read and write tools and meaningful approval for high-impact permissions.

Illustrative case. An agent can gather evidence for a production-access request, but it should not grant administrator rights simply because a conversational message claims urgency.

Measurement. Track unauthorised attempts blocked, excessive permissions prevented, approval latency, access corrections and the time required to reconstruct who changed what. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

7. Deletion and Destructive Actions

Deletion, destructive commands and irreversible record changes can turn a small reasoning error into permanent loss. Retrying a destructive tool call can also multiply the damage.

Evidence boundary. Confirm target identity, scope, backup or recovery state, current external state and whether the operation is idempotent.

Authority boundary. The system may identify candidates for deletion or prepare a cleanup plan without receiving permission to execute destructive actions.

Safer operating pattern. Prefer soft delete, staged changes, narrow scopes, backups, explicit confirmation and rollback where possible. Stop when action status is uncertain.

Illustrative case. If a file-deletion tool times out, the agent should not assume failure and repeat the call. It should verify state or ask for help.

Measurement. Measure unintended deletions, rollback success, duplicate destructive actions prevented, recovery time and whether the workflow correctly stops under uncertainty. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

8. High-Impact Public Claims and Crisis Communication

Crisis statements, regulatory announcements, allegations, safety notices and investor-sensitive communication can create reputational or legal consequences even when the prose is excellent.

Evidence boundary. Separate confirmed facts, working hypotheses, model inference and unknown information. Preserve source dates and current incident state.

Authority boundary. Drafting is not publication authority. The organisation remains responsible for claims made in its name.

Safer operating pattern. Use SI to assemble timelines, draft wording and compare source statements; require authorised release for material claims and refresh current facts immediately before publication.

Illustrative case. During an outage, SI can prepare a status update describing observed impact, but should not publish a confident root cause while engineering evidence remains incomplete.

Measurement. Track factual corrections after publication, approval latency, retractions, source completeness and whether updates preserve uncertainty appropriately. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

9. Personal Data and Sensitive Records

Personal data introduces privacy, purpose and access constraints. The fact that a system could benefit from more context does not mean it should receive every available record.

Evidence boundary. Know what data is used, why it is necessary, who may access it, how long it remains current and what role it plays in the outcome.

Authority boundary. The ability to process personal information does not create permission to use it for unrelated decisions or disclose it downstream.

Safer operating pattern. Use approved environments, minimum necessary context, role-based access, clear purpose and suitable data governance. Singapore organisations should consider applicable PDPA obligations and PDPC guidance.

Illustrative case. An HR assistant may need approved employment data for a specific workflow, but should not receive unrelated medical or family information merely because it exists in the same system.

Measurement. Track unnecessary-data exposure, access violations, correction requests, decision errors linked to stale data and whether the workflow can explain which personal data affected the result. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

10. Sensitive Human Relationships and Tacit Context

Leadership, negotiation, conflict, counselling and important client relationships depend on history, trust and social signals that are rarely fully documented.

Evidence boundary. Distinguish recorded facts from tacit human context. Treat missing relational information as a limitation rather than an invitation for the model to invent motives or likely reactions.

Authority boundary. SI may prepare questions, options and summaries, but accountable people retain relationship-sensitive judgment and the authority to make commitments.

Safer operating pattern. Keep the system in an advisory role, make assumptions visible and let the person closest to the relationship decide tone, timing and action.

Illustrative case. A manager can use SI to structure a difficult performance conversation, but the final approach should reflect the employee’s circumstances and the manager’s real knowledge of the situation.

Measurement. Look at downstream relationship outcomes, clarification needed, promises made correctly and whether the system reduced preparation burden without displacing human responsibility. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

11. Disputed or Contradictory Evidence

When sources disagree, a synthesis model can make the conflict disappear into a smooth narrative. That creates false closure.

Evidence boundary. Keep conflicting sources visible, with dates, definitions and provenance. Identify exactly where the disagreement lies.

Authority boundary. The model can map the dispute but should not silently decide which authoritative source wins unless a clear precedence rule exists.

Safer operating pattern. Return a conflict packet: verified facts, competing claims, missing information and the owner who must resolve the issue.

Illustrative case. Two project systems show different completion dates. The correct output is not a guessed date but a visible discrepancy routed to the project owner.

Measurement. Track conflicts correctly surfaced, conflicts incorrectly flattened, resolution time and whether downstream users can trace the decision to the source that resolved it. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

12. Work With No Reliable Verification Path

A high-consequence output that cannot be checked independently is a poor candidate for autonomous action, even when the model often appears correct.

Evidence boundary. Seek source retrieval, deterministic checks, tests, independent measurements or qualified professional review. If none exists, the uncertainty is structural.

Authority boundary. When verification is weak, keep the system’s role at preparation or recommendation rather than execution.

Safer operating pattern. Narrow scope, collect more evidence, create evaluation cases and delay greater autonomy until a real verification mechanism exists.

Illustrative case. A strategic forecast may be useful for discussion, but if its assumptions cannot be checked and the investment decision is material, it should remain one input rather than an automatic trigger.

Measurement. Track later outcome accuracy, assumption errors, calibration and the proportion of cases that cannot be validated at decision time. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

13. Cases Outside the Tested Envelope

Novel jurisdictions, rare combinations, new product states and unusual customer circumstances can invalidate assumptions learned from ordinary cases.

Evidence boundary. Make the tested categories, required fields and supported conditions visible. Detect when a case violates them.

Authority boundary. A system should not gain authority over an unfamiliar case merely because it can generate a response.

Safer operating pattern. Treat abstention and escalation as success states. Expand the envelope only after representative testing and explicit approval.

Illustrative case. A refund automation designed for ordinary consumer purchases should escalate a novel regulatory dispute instead of forcing it through the routine refund logic.

Measurement. Track out-of-scope detection, false in-scope classification, escalation usefulness and outcomes of newly added case classes. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

14. Work With Slow or Hidden Error Feedback

Some mistakes surface immediately; others appear weeks later through customer churn, audit, legal review or accumulated operational drift. User satisfaction can therefore overstate real reliability.

Evidence boundary. Use delayed audits, sampled review, downstream outcome measures and periodic reconciliation rather than relying only on immediate feedback.

Authority boundary. Keep autonomy conservative until enough real time has passed to observe the error pattern.

Safer operating pattern. Design monitoring at the same timescale as the consequence. Reopen the boundary if delayed evidence reveals systematic errors.

Illustrative case. A classifier may look accurate for months while quietly misrouting a rare but important category that only becomes visible during quarterly review.

Measurement. Compare immediate acceptance with delayed corrections, downstream rework and latent incident discovery. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

15. Systems With Large Failure Radius

Shared infrastructure and broadly privileged agents can spread one failure across many teams, customers or systems. Centralisation increases both leverage and systemic consequence.

Evidence boundary. Map dependencies, affected workflows, permissions, shared data and downstream systems before broad deployment.

Authority boundary. A central platform should not automatically inherit the union of all local permissions. Access must remain scoped by workflow and user role.

Safer operating pattern. Use staged rollout, rate limits, segmentation, monitoring, rollback, dependency testing and incident response proportional to the failure radius.

Illustrative case. A shared agent used by every department should not receive organisation-wide write access because one pilot team needed it for a narrow process.

Measurement. Track blast-radius containment, cross-workflow incidents, rollback success, dependency failures and time to isolate a malfunctioning capability. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

16. Work Exposed to Prompt Injection and Untrusted Content

Agents that read email, websites, documents or third-party tool output can encounter malicious instructions designed to redirect behaviour, expose data or trigger actions.

Evidence boundary. Classify trusted instructions separately from untrusted content and preserve the source of external text. Treat external content as data, not authority.

Authority boundary. Untrusted text should never be able to expand permissions. A website or email cannot legitimately grant the agent access that the user or organisation did not.

Safer operating pattern. Limit tools and data, structure intermediate outputs, sandbox risky work, confirm consequential actions and monitor tool calls. Use separate stages when public web research and private data do not need to coexist.

Illustrative case. A web page may contain hidden text telling the agent to upload private files. The agent should ignore that instruction because the page is evidence for the user’s task, not a source of authority.

Measurement. Track injection attempts detected, policy violations blocked, sensitive actions requiring confirmation, unexpected tool calls and incidents involving untrusted content. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

17. Work That Handles Credentials or Confidential Strategy

Secrets create asymmetric risk. One disclosure of a private key, payroll record or unpublished strategy can be far more damaging than many ordinary generation errors.

Evidence boundary. Know exactly which secret is required, which approved environment may access it and which outputs are allowed to leave the protected context.

Authority boundary. Access to a secret for one task does not grant permission to reuse it elsewhere or reveal it in an answer.

Safer operating pattern. Use minimum necessary access, secret-management systems, scoped credentials, redaction and separation between public and private stages.

Illustrative case. A research agent comparing public suppliers should not receive payroll or authentication data merely because the connector makes those sources conveniently available.

Measurement. Track secret-access events, unnecessary exposures, scope violations, redaction effectiveness and whether tasks can be completed with less sensitive context. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

18. Customer Promises and Commercial Commitments

Delivery dates, refunds, pricing, warranties and service commitments can create obligations and directly affect trust.

Evidence boundary. Use current inventory, shipment status, policy, approval limits and customer-specific records. Do not convert incomplete state into a promise.

Authority boundary. A draft message is not commitment authority. Sending or approving a promise should follow the organisation’s threshold and exception rules.

Safer operating pattern. Automate tightly bounded routine promises with current evidence; escalate disputed, high-value or unusual commitments.

Illustrative case. The system knows an order shipped but has no confirmed carrier estimate. It should state the known status rather than promise arrival tomorrow.

Measurement. Track broken promises, correction messages, escalation rate, approval latency and whether automated commitments remain within policy. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

19. Software Deployment and Production Changes

Generated code can be useful, but production changes affect customers, data and system availability. A good patch is not proof of a safe deployment.

Evidence boundary. Use tests, code review, staging results, dependency state, current configuration and deployment telemetry.

Authority boundary. Editing a branch, merging code and deploying to production are separate permissions.

Safer operating pattern. Grant only the action rights required at each stage, preserve rollback and require stronger approval for critical environments.

Illustrative case. A coding agent can implement a fix and run tests on a branch while the release pipeline or responsible engineer controls production deployment.

Measurement. Track test failures, rollback rate, escaped defects, review burden, deployment incidents and time from issue to safe release. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

20. Procurement and Supplier Commitments

Supplier selection combines pricing, obligations, service risk and governance. A model can make incomparable proposals look neatly ranked.

Evidence boundary. Normalise billing periods, scope, implementation costs, service levels, exclusions and missing terms; preserve source passages.

Authority boundary. The system can prepare the comparison, but award and commitment authority belongs to the appropriate procurement and business roles.

Safer operating pattern. Separate extraction from evaluation and evaluation from commitment. Escalate ambiguous or non-comparable terms.

Illustrative case. One supplier quotes monthly fees excluding setup while another quotes annual fees including implementation. The comparison must establish a common basis before any recommendation is treated as decision-ready.

Measurement. Track missing terms detected, comparison corrections, procurement review time and downstream surprises attributable to overlooked proposal differences. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

21. High-Stakes Education Decisions

Placement, discipline, progression and high-stakes assessment affect learners and may depend on context beyond the dataset available to the system.

Evidence boundary. Preserve assessment history, curriculum evidence, teacher observation, relevant accommodations and institutional rules.

Authority boundary. SI can organise evidence and identify patterns; consequential educational authority remains with the institution’s accountable process.

Safer operating pattern. Use the system to support review rather than to turn sparse data into an automatic life-changing decision.

Illustrative case. Two recent test scores should not automatically determine placement when longer-term progress and teacher evidence provide important context.

Measurement. Track material decision corrections, evidence completeness, reviewer workload and whether SI improves consistency without suppressing contextual judgment. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

22. Security Response and Automated Containment

Rapid containment can stop attacks, but an incorrect response can disable legitimate systems, destroy forensic evidence or lock out critical users.

Evidence boundary. Combine alerts with asset criticality, identity state, maintenance windows, change records and current incident context.

Authority boundary. Gathering evidence, recommending containment and executing privileged remediation are separate security powers.

Safer operating pattern. Use bounded playbooks, reversible steps, explicit thresholds and escalation for destructive or organisation-wide actions.

Illustrative case. An agent seeing unusual traffic should check a planned maintenance window before isolating a production service.

Measurement. Track true incidents contained, false containment actions, rollback time, escalation quality and the number of high-impact actions that received appropriate approval. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

23. Regulatory and Professional Filings

Filings and attestations can create formal obligations. Correct formatting does not prove factual or legal correctness.

Evidence boundary. Use current filing requirements, authoritative records, jurisdiction, responsible signatories and source evidence for material statements.

Authority boundary. SI can prepare and check completeness; authorised professionals or officers retain filing responsibility where required.

Safer operating pattern. Keep source references and version information visible and require review of material claims before submission.

Illustrative case. A generated regulatory narrative may fit the template while relying on an outdated rule; the source date and responsible reviewer must be part of the workflow.

Measurement. Track filing corrections, rejected submissions, material amendments, review time and source-traceability. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

24. Research and Evidence Synthesis

SI can summarise weak and strong evidence with equal fluency, making volume look like quality.

Evidence boundary. Preserve original sources, study design, population, date, limitations and conflicting findings.

Authority boundary. The system can organise and compare; researchers or qualified decision-makers retain judgment about evidence quality and relevance.

Safer operating pattern. Separate reported findings from model inference and recommendations, and verify material claims against original sources.

Illustrative case. Ten low-quality articles supporting a claim should not automatically outweigh two stronger studies challenging it merely because the model counts mentions.

Measurement. Track citation accuracy, source-quality errors, corrections in synthesis and whether conclusions remain traceable to evidence. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

25. Long-Running Agentic Objectives

Long-running tasks accumulate drift as facts, deadlines, permissions and tool state change. Early context can become stale while the agent continues confidently.

Evidence boundary. Refresh critical state at checkpoints and record which assumptions are still valid.

Authority boundary. An objective granted earlier should not imply indefinite permission to act under changed conditions.

Safer operating pattern. Use time-based checkpoints, revalidation of permissions and explicit stop conditions before consequential stages.

Illustrative case. An agent preparing a multi-day procurement packet should refresh supplier pricing and approval status before final submission rather than relying on day-one data.

Measurement. Track stale-state incidents, checkpoint escalations, resumed tasks after validation and errors caused by assumptions that changed mid-run. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

26. Systems That Learn From Human Corrections

Feedback can improve a workflow, but blindly generalising every correction can encode local mistakes, malicious input or exceptional decisions as new policy.

Evidence boundary. Classify whether a correction is a source update, one-off exception, taxonomy change, reviewer preference or genuine policy change.

Authority boundary. Individual users should not silently rewrite organisation-wide rules through ad hoc feedback unless the system is explicitly designed and governed for that purpose.

Safer operating pattern. Route persistent learning through maintained knowledge ownership, versioned instructions and reviewed policy updates.

Illustrative case. One manager approving an unusual exception should not make that exception the default for every future case.

Measurement. Track recurring corrections eliminated, harmful generalisations prevented, policy-change provenance and disagreement across reviewers. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

27. Systems Optimised Around a Single Metric

An intelligent workflow can optimise the measured number while harming the real outcome. Speed, approval rate or low cost can be gamed by behaviour that shifts work downstream.

Evidence boundary. Use multiple outcome measures that include quality, rework, receiver effort, customer impact, fairness or incidents where relevant.

Authority boundary. A model should not redefine success around the easiest metric to improve.

Safer operating pattern. Measure the end-to-end outcome and inspect unintended consequences before expanding autonomy.

Illustrative case. A support system rewarded only for short resolution time may close unresolved cases, improving the dashboard while degrading service.

Measurement. Track the primary metric alongside correction rate, reopened work, downstream effort and real receiver outcomes. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

28. Work Under Cost Pressure

Cost optimisation can silently weaken safeguards by reducing retrieval, switching models, narrowing context or removing review.

Evidence boundary. Compare cost changes with quality, error, latency and consequence rather than treating spend as an isolated engineering metric.

Authority boundary. Budget pressure should not automatically expand machine authority or eliminate controls that protect consequential actions.

Safer operating pattern. Use lower-cost paths for low-risk work and preserve stronger verification for high-impact cases; reassess when routing changes.

Illustrative case. A company should not remove review from high-value payments simply because the reviewer costs more than model inference.

Measurement. Track cost per accepted outcome, not cost per generated output, and include rework, incidents and downstream correction. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

29. Work After a Model, Tool or Policy Change

A workflow can look identical while the underlying model, source corpus, connector or organisational rule changes. Previous validation may no longer describe current behaviour.

Evidence boundary. Track material versions and maintain regression cases covering ordinary, edge and should-stop scenarios.

Authority boundary. Previous approval should not be assumed to cover materially different behaviour or newly granted tools.

Safer operating pattern. Stage significant changes, rerun representative evaluations and reopen the delegation boundary before restoring full authority.

Illustrative case. A provider update that changes tool-selection behaviour should be tested before an agent regains broad write access.

Measurement. Track regression failures, rollback events, changed exception rates and whether the update altered downstream outcomes. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

30. Work During Organisational Change

People, responsibilities and policies change. A technically stable system can become organisationally stale when the owner leaves or the process is reorganised.

Evidence boundary. Keep current ownership, authoritative sources, approval routes and exception queues attached to the workflow rather than to one individual.

Authority boundary. A role change can invalidate earlier permissions even when the software has not changed.

Safer operating pattern. Reassess delegation when responsibility, policy, department boundaries or operating procedures change materially.

Illustrative case. An automation built around one team’s approval hierarchy should not continue unchanged after authority moves to another department.

Measurement. Track orphaned exceptions, outdated owners, permission mismatches and time required to bring inherited workflows back under clear governance. Also record exceptions, material corrections and escaped failures. A workflow should not be promoted because users like it; it should be promoted because representative cases show that the intended outcome is achieved without unacceptable hidden work or risk.

A Delegation Contract

Purpose: what outcome is the system helping to achieve?
Evidence: which sources or data may it rely on?
Role: read, interpret, recommend, prepare or execute?
Permissions: what may it access and change?
Prohibitions: what must it never do automatically?
Verification: what must be checked and how?
Stop conditions: when must it abstain or escalate?
Owner: who is accountable for the workflow?
Closure: what world-return signal proves completion?
Review trigger: what change requires reassessment?

If a team cannot fill in this contract, the task is not ready for high autonomy. The correct response is to narrow the role rather than hide uncertainty behind a longer prompt.

The World-Return Rule

For consequential external actions, require evidence from the actual system wherever possible. A payment workflow should read transaction state. A deployment workflow should read release status. A CRM workflow should confirm the updated record. A calendar workflow should confirm the created event. This prevents the system’s narrative of intent from being mistaken for completion.

The Review-Capacity Rule

A good human checkpoint can fail through volume. If SI produces sixty consequential cases per hour and a reviewer can meaningfully inspect twenty, the queue grows or review becomes ceremonial. Narrow eligibility, automate deterministic checks, improve evidence presentation or increase appropriate capacity. Do not remove review merely because the system has made review inconvenient.

The Recovery Rule

Autonomy should expand alongside recovery. Ask which actions can be rolled back, which require compensation, which require customer communication and which require incident escalation. If the organisation cannot stop or recover the workflow, keep the action boundary conservative.

What This Article Owns

This page owns the hard boundary between useful delegation and irresponsible delegation. If it disappeared, the workplace series would still explain what SI can do, where it fits, how roles work and how maturity increases—but it would lose the explicit stop rules for high-consequence, hard-to-verify or insufficiently governed work.

That keeps this page distinct from What Should Super Intelligence Do at Work?. Task fit asks whether SI creates leverage. Delegation boundaries ask how much authority the system may receive after the task has been selected.

A 20-Minute Blind-Delegation Audit

Choose one SI workflow that can influence an external outcome.
Write the most harmful plausible error in one sentence.
Identify which evidence should prevent or expose that error.
Name who owns the consequential decision.
Separate what SI reads, recommends, prepares and executes.
Check whether the action is reversible.
Define at least one stop condition and one escalation owner.
Confirm what system-of-record evidence proves completion.
Estimate whether review volume fits real human capacity.
Decide whether the workflow should remain, be narrowed, be demoted or receive stronger controls.

Frequently Asked Questions

What should never be delegated to AI?

Do not delegate blindly when work has high consequence, weak verification, unclear authority, sensitive data, irreversible action or no meaningful escalation. SI may still support preparation under stronger controls.

Is human review always required?

No. Low-risk, stable, reversible and well-validated routines can be automated. Human review should be proportionate to consequence and uncertainty, and deterministic checks may be stronger than manual review for some tasks.

What does blindly mean?

Blindly means the organisation has delegated work without adequate visibility into evidence, boundaries, verification, exceptions or accountability. A highly autonomous system can still be governed if those elements are strong.

What if SI is more accurate than a human?

Accuracy matters, but delegation also depends on authority, consequence, privacy, reversibility and failure detection. Where measured system performance is strong, human involvement can be redesigned rather than retained ceremonially.

Can an agent send emails automatically?

Yes, for bounded message classes with reliable context, approved claims, suitable permissions and escalation. Sensitive commitments, disputes and uncertain cases should route to a person.

Can SI approve payments?

A system may support financial controls or operate within tightly defined automated limits, but material financial authority should follow established approval, security and regulatory controls.

Can SI make hiring decisions?

SI can support recruitment administration and evidence organisation. Decisions affecting employment opportunity deserve strong attention to job-related criteria, fairness, privacy, law and accountable human governance.

How do we know when to stop an agent?

Define observable stop conditions before deployment: missing required data, conflicting sources, tool failures, security anomalies, exceeded thresholds, sensitive categories or uncertainty outside the tested envelope.

Can a model review its own output?

A second pass can catch some errors but is not independent proof. Important claims should be checked against sources, deterministic tests, calculations or qualified judgment appropriate to the task.

What should I read next?

Return to the Workplace Super Intelligence Hub. The next planned article explains how to start using SI without redesigning the entire company.

The Core Boundary

Never delegate authority faster than you can build evidence, verification, recovery and accountability around it. Super Intelligence can read more, compare more, draft faster and increasingly act through tools. Those capabilities become workplace value only when the organisation knows where the system must stop.

Use SI boldly where work is bounded, checkable and reversible. Use it carefully where context or consequence rises. Keep consequential authority visible. Preserve specialist judgment where it is required. Require world-return evidence for external actions. Treat abstention as a sign of a well-designed system, not a failure to be intelligent.

What Work Should Never Be Delegated Blindly to Super Intelligence? | Workplace SI Boundaries and Human Accountability