ZharfAI Team
An AI agent is software that pursues an objective through model inference, memory, tools, and repeated actions. Calling it a teammate can help people understand interaction, but it must not obscure accountability. An agent does not accept legal responsibility, manage its own incentives, or understand organizational context the way a person does. A named human and an accountable organization remain responsible for deciding where it runs, what it may do, and whether its output is used.
Adding agents is therefore not equivalent to adding headcount. It changes how work must be decomposed, queued, observed, reviewed, and recovered. If a process has unclear ownership, automation can make the ambiguity travel faster.
Begin with a work unit: an invoice evidence packet, a support case classification, a release-note draft, a tested code patch, or a list of contract deviations. A work unit has an input, expected artifact, acceptance criteria, deadline, authority boundary, and owner.
Avoid broad roles such as “general operations agent.” They mix incompatible data, permissions, and evaluation criteria. A narrow role is easier to test:
The unit of management is not how human the conversation feels. It is whether the system produces a defined artifact under controlled authority with inspectable evidence.
This is especially important when one agent delegates to another. Delegation should not silently expand permissions. The parent process must pass the minimum task context, define the expected return, preserve provenance, and remain responsible for the combined result.
Jobs are bundles of tasks, relationships, judgment, and accountability. The International Labour Organization’s 2025 global exposure index estimated that one in four workers is in an occupation with some generative-AI exposure, while only a smaller share falls into its highest exposure category. The authors emphasize that transformation is more likely than complete replacement because many occupations still require human input. Exposure is a capability estimate, not a forecast of layoffs.
For each workflow, classify tasks by:
Good early candidates are frequent, bounded, reversible, and checkable. High-variance work can still use an agent for evidence gathering or drafting while a person retains judgment.
Do not automate the easiest steps if that leaves people with a queue composed entirely of stressful exceptions. Work design should consider cognitive load, skill development, job quality, and fair distribution of benefits—not only minutes removed.
A role charter is the operating contract for one agent configuration. It should include:
Use a responsibility matrix. The business owner is accountable for the outcome and residual risk. The system owner maintains the agent and integrations. The operator starts or supervises work. The reviewer has enough domain expertise and time to challenge the result. Security, privacy, legal, HR, or worker representatives participate where the use case affects their responsibilities or people’s rights.
NIST’s AI Risk Management Framework Core calls for documented roles, clear lines of communication, trained personnel, executive accountability, and differentiated responsibilities for human-AI configurations. It is voluntary guidance, but its central point is practical: “human oversight” is not a control until the person, decision, information, authority, and response time are defined.
People should understand an agent’s operational state without reading every generated token. A shared queue can use explicit states:
Show the objective, owner, start time, last verified step, tool currently in use, external side effects, budget consumed, blocking condition, and next action. A spinner labeled “working” is not observability.
Every work item needs an idempotency strategy. If a process retries after a timeout, it must know whether the first payment, email, ticket, or database update already happened. Use stable operation IDs, deduplication, transaction boundaries, and reconciliation rather than relying on the model to remember.
The agent-observability guide covers traces, events, evidence, and alert design in more depth.
A useful handoff contains:
“I could not finish” creates another investigation. “The supplier record lacks a tax identifier; no payment action was attempted; source rows 18–21 are attached; procurement must confirm the legal entity” reduces work.
Human approval should be placed before the consequential action, not after it. The reviewer needs the original evidence and material alternatives, not merely the agent’s summary. Approval interfaces should make it easy to reject, edit, request more evidence, or stop the workflow.
For irreversible or high-impact work, use dual control, scoped authorization, and a rollback or compensating process. The guide to human approval design explains how to avoid turning review into a ceremonial click.
Autonomy is not a single switch. Use levels tied to consequence:
Higher autonomy requires stronger evaluation, narrower permissions, more reliable state, and faster incident response. It should be earned through production evidence, not granted because a demo completed one workflow.
Separate capability from permission. A model may be able to compose SQL but receive read-only database access. It may be able to draft a refund but only a person can authorize money movement. See secure tool permissions for agents for concrete least-privilege patterns.
Agent success rate alone can improve while the business process worsens. Use a balanced scorecard:
Track denominators and uncertainty. Ten flawless pilot cases do not establish a 100-percent success rate. Separate model failure, tool failure, bad input, permission error, and unclear process; each needs a different remedy.
Measure counterfactual value against the current process. If people could complete the task faster without supervising the agent, the automation may be negative despite producing correct work.
Evidence should make managers cautious about universal claims. In a peer-reviewed study of more than 5,000 customer-support agents, Generative AI at Work reported a 15-percent average increase in issues resolved per hour after access to an AI assistant, with larger gains for less experienced and lower-skilled workers. The setting had repeated conversations, measurable outcomes, and historical examples. It does not prove the same gain for every occupation or autonomous agent.
The preregistered jagged technological frontier experiment involved 758 consultants. On 18 tasks selected within the tested model’s capability frontier, participants with AI completed 12.2 percent more tasks, 25.1 percent faster, with higher assessed quality. On a separate complex task designed outside that frontier, AI users were 19 percent less likely to reach the correct solution. The paper’s March 2026 revision remains specific to those tasks, participants, and model conditions.
An RCT by METR with 16 experienced open-source developers and 246 repository tasks found early-2025 AI tools increased completion time by 19 percent in that sample even though participants expected acceleration. In a February 2026 update, METR said newer data suggested improvement but was too affected by participant and task selection to support a reliable current speedup estimate.
The lesson is not that AI always helps or hurts. Value depends on task, worker experience, tool quality, workflow redesign, and measurement. Run controlled local comparisons and revisit them as tools and adoption change.
Workers should know when and how agents affect their work, performance evaluation, customers, or employment. Give them a channel to contest outputs and report hidden labor, unsafe shortcuts, monitoring concerns, and inaccessible tooling.
Do not use acceptance of AI suggestions as a proxy for employee performance. A careful worker may reject more suggestions because they detect errors. Metrics built from tool usage can create pressure to accept bad work or move effort off the record.
Plan training by role: operators need limits and escalation; reviewers need failure patterns and evidence literacy; system owners need evaluation and incident skills; leaders need to understand residual risk. In the EU, Article 4 of the AI Act establishes an AI-literacy obligation for providers and deployers, while high-risk systems can carry additional human-oversight duties. Applicability and sufficiency depend on the facts and require current legal advice.
Involve affected staff in task mapping and pilot design. They know where data is unreliable, exceptions hide, and informal coordination keeps work safe. If the agent learns from employee work, address attribution, privacy, retention, access, and incentives explicitly.
Daily: review failed, stuck, high-impact, and over-budget work; reconcile side effects.
Weekly: sample accepted work, review false escalations and missed escalations, update known failure patterns, and compare quality, correction, latency, and cost by task class.
Monthly: review permissions, data access, model and prompt changes, incidents, worker feedback, and whether the use case has expanded beyond approval.
Quarterly or risk-based: retest protected evaluations, confirm owners and training, exercise rollback and fallback, reassess vendor dependencies, and retire automations that no longer create value.
After an incident, preserve evidence, contain authority, reconcile external effects, identify both technical and organizational causes, and update the charter, tests, and process. Avoid blaming the final reviewer if workload or interface made meaningful review impossible.
There is no safe universal ratio. It depends on exception frequency, consequence, review depth, task duration, interface quality, and whether agents act concurrently. Measure queue delay, missed issues, interruption load, and review quality; reduce concurrency before oversight becomes nominal.
Names can help navigation, but role, version, owner, and authority must remain more prominent than persona. Friendly language must not imply human judgment, confidentiality, or responsibility that the system does not possess.
Retire or redesign it when the process no longer exists, owners disappear, model or vendor support ends, risk exceeds tolerance, correction cost erases value, or a simpler deterministic system performs better.
Choose a frequent, bounded, reversible task with accessible evidence and a motivated human owner. Begin in observe or draft mode, establish a baseline, and expand authority only after measured performance supports it.
Productivity estimates are not directly comparable across these settings. They use different tools, workers, tasks, outcomes, and research designs. Local evaluation and worker input remain necessary.
Beyond chatbots and copilots, autonomous AI agents are emerging as the new digital workforce. Discover how they differ from traditional AI and how they are transforming industries.
Read More
When an agent can buy, the payment system must bind identity, intent, item, payee, budget, receipt, and dispute rights into one controlled transaction.
Read More
Production voice agents succeed through timing, turn-taking, confirmation, recovery, and a clean path to a human—not voice synthesis alone.
Read MoreSee the daily briefing and the operational guides. This page is an archive note, not an invitation to start a project.