Editorial status — Journal-authored editorial essay produced with generative-AI assistance. It received an automated clarity and source-presence screen; it is not an SCCR paper and was not externally peer reviewed.
The job is no longer the useful unit
Most discussions of artificial intelligence in commercial teams begin with jobs: Will AI replace sellers, marketers, customer-success managers, analysts, or operations specialists? That question is too coarse for operating design. A job is a bundle of tasks, decisions, relationships, permissions, and accountabilities. AI rarely absorbs that bundle all at once. It changes the cost and speed of particular actions inside it.
The practical unit of redesign is therefore not the job but the work packet: a bounded piece of work with an input, an intended outcome, a decision owner, an execution path, and a verification requirement. A prospecting sequence, an opportunity summary, a renewal-risk diagnosis, a forecast exception, and a pricing recommendation are all work packets. Each can be decomposed differently, assigned differently, and governed differently.
This shift matters because organizations can add an AI tool without changing work. Employees may produce drafts faster while approvals, handoffs, data access, and accountability remain untouched. The tool is new; the operating model is not. Sustainable gains require redesigning the packet around what machines can do, what people must judge, and who remains answerable for the result.
Evidence points to tasks, not universal replacement
Recent studies show meaningful gains, but they also show why broad claims about entire jobs are unreliable. In an experiment involving professional writing tasks, access to generative AI reduced completion time and improved average output quality (Noy and Zhang, 2023). A field study of customer-support agents found higher issues resolved per hour, with larger gains among less experienced workers and little benefit for the most experienced group (Brynjolfsson, Li, and Raymond, 2023). These results concern particular tools, tasks, populations, and measurement periods—not every form of commercial work.
The boundary can also reverse within a workflow. A field experiment with management consultants found improved speed and quality on tasks inside the model's capability frontier, but worse performance on a task outside it (Dell’Acqua et al., 2026). The important management lesson is not that AI is good or bad. It is that apparently similar tasks can demand different allocations of human attention.
Labor-market exposure studies make the same distinction at a broader scale. The International Labour Organization's 2025 index treats exposure as the potential for tasks within occupations to change, not as a prediction that whole jobs will disappear. Its headline conclusion emphasizes transformation over replacement and notes that workplace policy and implementation will shape the eventual effect (Gmyrek et al., 2025).
These findings support a cautious operating principle: redesign at task level, test in context, and refuse to turn a local productivity result into a universal staffing formula.
A work packet has four responsibilities
An AI-enabled packet contains at least four separable responsibilities.
- Specification. Someone defines the objective, relevant context, allowed data, constraints, and completion standard. Weak specifications create polished but unusable output.
- Execution. A person, model, workflow, or combination performs the transformation: drafting, classification, calculation, retrieval, comparison, or recommendation.
- Verification. Someone checks factual accuracy, policy compliance, customer context, exceptions, and fitness for the intended decision.
- Accountability. A named owner accepts the consequences of sending, recording, approving, or acting on the result.
Traditional work often places all four responsibilities in one role. AI makes it technically possible to separate them. That separation is useful only when the organization makes it explicit. If a model drafts a commercial proposal but no one owns the specification, the organization has automated ambiguity. If a representative verifies the prose but cannot see the pricing rule or source data, verification is ceremonial. If an agent writes back to the CRM without a clear correction path, execution has outrun governance.
The literature on automation has long distinguished levels and types of human-machine interaction rather than treating automation as a single state (Parasuraman, Sheridan, and Wickens, 2000). More recent organization research similarly argues that augmentation and automation are interdependent rather than clean alternatives (Raisch and Krakowski, 2021). The four-responsibility model turns those ideas into a usable commercial design test.
Revenue work is a chain of consequential packets
Revenue work is especially sensitive because one output often becomes the next team's input. A marketing classification changes sales prioritization. A sales note changes the forecast. A forecast changes hiring or spending. A renewal-risk flag changes customer treatment. A pricing recommendation changes both margin and precedent.
Consider an AI-generated opportunity summary. Execution may be cheap: collect emails, call notes, CRM fields, and product-usage signals, then produce a narrative. The difficult questions sit elsewhere. Which systems may supply data? Which statements are observations, which are inferences, and which are missing? Who checks the customer's commitments? Who may write the summary back to the system of record? How is an error corrected after a forecast meeting consumes it?
Or consider outbound personalization. A model can generate variants rapidly, but the work packet is not complete when text exists. The packet includes audience eligibility, suppression rules, claim accuracy, brand constraints, approval thresholds, delivery, response capture, and learning. Treating “write an email” as the whole task hides most of the commercial system.
This is why the unit of commercial work must include state change, not merely content generation. The output that matters is often a changed record, a customer commitment, an allocated resource, or an authorized action.
Control moves into the workflow
When AI participates in execution, management control becomes partly architectural. Permissions, defaults, queues, confidence thresholds, logging, escalation paths, and interface design shape behavior before a manager sees the result. Research on algorithmic control shows that digital systems can direct, evaluate, reward, and constrain workers through embedded mechanisms (Kellogg, Valentine, and Christin, 2020). Commercial AI can create similar control even when presented as a personal assistant.
That makes three design choices unavoidable.
- Decision rights: Define what the system may draft, recommend, approve, send, or write back.
- Evidence visibility: Preserve the inputs and reasoning traces needed for a person to perform meaningful verification.
- Correction rights: Make it easy to reverse an action, amend a record, and propagate the correction to downstream consumers.
These controls should scale with consequence. A first draft of an internal meeting summary needs less oversight than a discount approval, contract statement, customer-facing commitment, or deletion of account data. The useful distinction is not simply “human in the loop.” It is which human, at which decision point, with what information and authority.
Measure the redesigned system
Adoption metrics are weak evidence of value. Login counts, prompts submitted, drafts produced, or licenses activated show activity. They do not show whether the redesigned work improved customer outcomes or organizational performance.
Measurement should follow the packet. Compare completion time, rework, error severity, escalation rate, customer response, conversion, margin, retention, and downstream correction cost. Segment results by task type and worker experience because average effects can hide who benefits and where performance degrades. Include a control or baseline whenever practical.
The measurement window also matters. Short-run speed can coexist with long-run skill loss, more review work, or data-quality deterioration. Conversely, an initially slow process may improve as teams learn where the capability frontier lies. The experiment is not only a test of the model. It is a test of the work design around the model.
The operating question
AI changes commercial work when it changes who specifies, executes, verifies, and owns a consequential work packet. That is a deeper change than adding a writing assistant and a narrower claim than predicting the end of a profession.
Leaders should start with one packet that is frequent, measurable, and reversible. Map its inputs and downstream effects. Assign the four responsibilities. Set permissions and correction paths. Measure the complete outcome rather than the generated artifact. Expand only when the evidence supports expansion.
The durable advantage will not come from producing the most AI output. It will come from building a commercial system that knows which work to delegate, which work to keep, and how to remain accountable for both.