The agent has reached the dispatch board
Most agent products begin by automating a task that a person used to perform. WorkJam's July 13, 2026 announcement moves the agent into a more consequential seat: deciding which person should perform which task, on which shift, and in what order. The system is not merely labor inside the workflow. It is beginning to allocate the labor.
WorkJam says its autonomous AI assigns work to the right person on the right shift, confirms completion, and learns from outcomes. It says those decisions operate within shift schedules, compensable time, certifications, labor rules, and each organization's policies and standard procedures. Managers can override decisions, and actions are logged for audit and compliance. The platform connects task management with communications, learning, staffing, capacity planning, and proof of completion.
Those are vendor descriptions, not audited operating results. The announcement provides no deployment count, named production case for the new capabilities, autonomous completion rate, override rate, worker outcome, or independent evaluation. The retailer names in its opening describe WorkJam platform users, not disclosed deployments of this release. Still, the product boundary is commercially important. When software chooses the next task for a person, it has crossed from workflow automation into a narrow form of management. A nicer task list does not make that distinction disappear.
Present tense has an availability footnote
WorkJam's announcement uses expansive present-tense language about continuous prioritization and real-time execution. Its availability section draws a narrower line that a buyer should preserve. WorkJam says the autonomous capabilities are in market today. It separately says the WorkJam AI Task Priority Engine will arrive in the second half of 2026.
The planned engine is the component WorkJam says will read signals across labor, traffic, demand, customer feedback, and an organization's own enterprise AI. It is supposed to produce a dynamic priority score for every task, recalculate that score continuously as conditions change, and attach a plain-language explanation to each recommendation. Those are planned H2 2026 capabilities, not current production evidence.
A procurement document should therefore have two columns: demonstrated now and proposed later. Current diligence can test assignment, rule enforcement, completion evidence, overrides, and logging. Continuous multi-signal scoring, its explanations, and its behavior under rapidly changing conditions belong in an acceptance plan for the Task Priority Engine when it ships. Roadmaps are useful. They are poor substitutes for release notes and replay tests.
Stores, roads, and bank exceptions are converging
Samsara's June 24, 2026 Agent Studio launch shows the same pattern in physical operations. The company says customers can build agents or use more than 15 safety and maintenance templates, toggle capabilities, set permissions, preview behavior, add company policies and documents as a knowledge base, monitor usage, and track outcomes. Its examples include driver assistance, maintenance briefings, and matching moving vehicles to unidentified drivers.
Samsara reports that a driver assistant at a major food distributor saves 30 minutes of communication time per call, while a food-bank maintenance digest saves hours of manual tracking each week. A quoted customer says reporting and data compilation that cost more than six figures annually is now fully automated. These are useful workflow clues, but they appear in a vendor announcement. Samsara does not provide a comparison design, sample size, error rate, or independent audit for those figures on the page.
Smartstream's June 2, 2026 announcement brings the pattern into bank reconciliations, cash breaks, settlement exceptions, and post-trade investigations. Smartstream reports that Tier 1 pilots reduced investigation time per exception from 14 minutes manually to 30 seconds, a 96.4% reduction on the published times; the release rounds it to 97%. It also says Tier 1 clients project 50% to 70% automation in year one. The institutions, number and type of exceptions, observation period, measurement protocol, and denominator behind the automation range are not disclosed on the announcement page; the range is a client projection rather than an observed first-year result. Across the three announcements, the common architecture is company context, bounded authority, exception routing, explanations, and audit trails; only WorkJam explicitly claims autonomous assignment of human work.
Dispatch software inherits management problems
The OECD has a useful name for this category: algorithmic management, meaning software that partially or fully automates tasks traditionally carried out by human managers. Its report currently online replaces the document originally published on February 6, 2025. The employer survey covered 6,047 firm establishments with at least 20 employees, with one response per organization and an ideal respondent described as a mid-level manager or supervisor, in France, Germany, Italy, Japan, Spain, and the United States between June and August 2024.
Seventy-four percent of respondents said their firms used at least one tool to instruct, monitor, or evaluate workers. That broad definition can include legacy systems and tools with no artificial intelligence at all, so the number is not an adoption estimate for autonomous agents. Sixty percent of managers using algorithmic management said it improved the quality of their decisions. The survey records managerial perceptions, not an audited productivity study or a survey of worker outcomes.
The concerns are the useful counterweight. Among managers using the tools, 64% reported at least one trustworthiness concern. The most common were unclear accountability at 28%, difficulty following the logic at 27%, and inadequate protection of physical or mental health at 27%. Agentic dispatch does not invent those problems. It can make the queue more dynamic and the rationale less obvious. A manager override is therefore necessary and incomplete. The person receiving a rearranged queue also needs to know why it changed, correct bad source data, and challenge an infeasible or unsafe assignment.
Put terms on the dispatch board
Begin with eligibility. Define which task classes the agent may dispatch, which are excluded, and which workers may receive each class. Encode required certifications, location, shift, paid-time status, site rules, and any individual restrictions the operating policy recognizes. Then name every priority input, its owner, maximum age, and order of precedence. Labor availability, customer demand, safety work, maintenance urgency, service deadlines, and local judgment will eventually disagree. 'The model decides' is not a conflict rule.
Next define capacity and authority. Capacity should include existing assignments, work in progress, expected duration, travel or setup time, and skill level rather than an empty square on a schedule. State whether the agent may recommend, assign, reassign, or cancel work, and which actions require a manager. Give reprioritization a budget: maximum changes per person or shift, minimum notice where practical, protected work-in-progress states, and a named emergency condition that can break the limit. Every change should carry a worker-visible reason based on the actual inputs, not a paragraph generated after the decision.
Finally, define proof, override, and appeal. Specify what counts as completion, who or what validates it, how long the work may be reopened, and what happens when proof is missing. A manager override should stop or change a decision immediately. A worker appeal should provide a route for reporting stale data, an infeasible assignment, an eligibility error, or an unsafe sequence, with a response owner and deadline. Log the original priority, rationale, input versions, override or appeal reason, reviewer, and outcome. An audit log proves what happened. It does not prove that the rule was sensible.
Measure the queue, not the keynote
Start the scorecard with accepted work: first-pass completion, time to verified completion, overdue tasks, reopened tasks, missed service levels, and exceptions by task class. Add dispatch quality: incorrect eligibility attempts, skill mismatches prevented or missed, assignment changes, abandoned work, queue age, manual intervention, and the gap between planned and actual task duration. A large number of assignments is activity. It may also be an efficient way to manufacture interruption.
Track the human effects alongside throughput. Measure priority changes per person and shift, work interrupted after starting, manager overrides, worker appeals, appeal response time, schedule or compensable-time exceptions, and results by site, role, and workflow version. Review the reasons, not only the counts. A falling override rate may mean the system improved. It may also mean people stopped believing the button mattered.
Start in shadow mode: record the agent's proposed dispatches and score them against actual decisions using one completion standard. Give it assignment authority only for a bounded task class; keep reprioritization behind approval until prewritten limits for errors, interruptions, overrides, and appeals hold. Treat the release as a management-policy change with production access, because that is what it is.
The moment an agent decides what a person should do next, it is no longer only automating work. It is exercising a piece of management.
Give that piece a contract before giving it the board.
- WorkJam Launches Autonomous AI That Runs Frontline Work in Real TimeWorkJam, accessed July 14, 2026
- Samsara Launches New Agentic Capabilities to Automate Tedious Operational TasksSamsara, accessed July 14, 2026
- Smartstream launches Smart Agents for back-office operations, proven across Tier 1 pilotsSmartstream, accessed July 14, 2026
- Algorithmic management in the workplace: New evidence from an OECD employer surveyOECD, accessed July 14, 2026