The Monday Morning Problem
You’re between meetings. Your HRIS team flagged 30 incomplete records in the benefits enrollment system, and they’re due by Friday. Your laptop’s closed. Your phone’s in your hand. You could spend the next hour manually reviewing each record, flagging discrepancies, emailing the gaps back to your team. Or you could take thirty seconds to describe the work you need done and have Claude do it while you walk to your next meeting.
Dispatch makes that possible. You open Claude on your phone, describe what you need—“Flag all benefits enrollment records where Benefits Effective Date is missing or falls outside the 2026 benefit year”—and Claude picks up on your desktop, opens the system, pulls the records, marks them, and leaves a summary in a shared folder. By the time you’re back from your meeting, it’s done.
This is the HR ops promise of 2026: work that used to require someone to be at a desk now happens in the background, initiated from anywhere.
What Dispatch Is (and What It Isn’t)
Dispatch is Anthropic’s feature for assigning long-running background tasks to Claude from your phone. It launched on Windows April 3, 2026, and on Mac March 24, 2026. It’s a research preview available to Claude Pro and Max subscribers.
Here’s the mechanics: You describe a task in Claude on your phone—any task, any prompt. Claude checks whether it can handle the work through existing integrations first. If you’re asking Claude to look at your Google Calendar or post to Slack or check an MCP connector (like Salesforce API or custom company tools), Claude uses those integrations. No desktop needed.
But if the work requires Claude to actually interact with your computer—opening an app, navigating a web interface, finding a file, clicking buttons—Claude switches to computer use. The Claude desktop app, which you keep running on your machine, takes over. Claude remotely navigates your screen, performs the work, and returns results to your phone.
The key insight: Desktop control is the fallback, not the default. Dispatch tries to use APIs and integrations first. Your HRIS system might have a direct API. Your payroll vendor might support OAuth. If Claude can reach the data directly, it will. Computer use is the workaround for systems that don’t expose APIs, legacy software, or custom internal tools.
This distinction matters for HR ops because many of the systems you’ll want to automate—HRIS platforms, payroll tools, benefits systems—increasingly support API access. But many don’t, or the access is locked behind IT approvals. Dispatch handles both.
How Dispatch Works in Practice
The flow is straightforward but worth walking through step by step because what’s happening on your phone differs sharply from what’s happening on your desktop.
You’re in a meeting. Between speakers, you open Claude on your phone and type: “Pull this month’s HRIS export. Find all employees marked as ‘Terminated’ but still assigned to a team. Flag their records in the export file and save to /Shared/HRIS-Data-Quality/Flags.”
On your phone, you hit send. Claude responds with something like: “I’ll take care of this. I’ll open the HRIS system on your desktop, pull the export, scan for the issue, and save the flagged records to your shared folder. Check back in a few minutes.”
Here’s what happens on your desktop in the background: The Claude desktop app, which has been sitting quietly in your system tray, becomes active. Claude opens your HRIS system (or navigates to it in your browser—depending on how it’s deployed). It logs in using your cached credentials. It downloads the monthly export. It processes the data—finding the terminated employees still assigned to teams. It flags those records, either by editing the export file directly or by creating a new summary file. It saves the output to the shared folder path you specified.
You’re still in your meeting. You didn’t touch the keyboard once. Your screen showed Claude clicking, typing, opening windows. If your screen saver came on, Claude waited for it to turn off before continuing. If your laptop went to sleep, Dispatch paused. When your computer woke, the task resumed.
By the time your meeting ends, you get a notification on your phone: “Task complete. Found 7 terminated employees still assigned to teams. Flagged records saved to /Shared/HRIS-Data-Quality/Flags/2026-04-flagged-terminations.csv.”
That’s the promise. And most of the time, it works exactly like that.
Five HR Ops Workflows That Work Today
Not every HR ops task is a good fit for Dispatch. Some are too high-risk, some require human judgment, some break because Claude misinterprets a UI. Here are five that actually work, with honest notes on why they work and what to watch for.
1. HRIS Export Triage: Flagging Incomplete Records
The task: Your HRIS exports every month. You need to identify records with missing critical fields (hire date, benefits tier, manager assignment) before payroll runs. Usually, this takes an hour of manual scanning.
The prompt: “Download the latest HRIS export from [system]. Filter for records where Hire Date is blank OR Benefits Tier is not assigned OR Direct Manager field is empty. Create a summary CSV with Employee ID, Name, Missing Fields, and Last Updated date. Save to /Shared/Data-Quality/Incomplete-Records-[date].csv.”
Why it works: This task has clear, objective criteria. There’s no ambiguity in what “incomplete” means. Claude scans the data, applies simple filters, outputs a file. The risk is low (you’re not modifying payroll data), and the output is easy to verify.
What Claude usually gets right: Finding the missing values, creating a clean CSV, saving it to the right folder.
What to watch for: If your HRIS system uses unusual field names or color-coding instead of explicit “blank” indicators, Claude might misinterpret the data. Test with a small sample first.
2. Offer Letter Generation and Filing
The task: New hire offer letters need to be customized with role, salary, start date, and manager name, then filed in the correct folder structure.
The prompt: “I’m attaching a new hire requisition [paste details: Name, Role, Salary, Start Date, Manager]. Use the standard offer letter template in /Templates/Offer-Letter-2026.docx. Fill in the bracketed fields with the details provided. Save the completed letter to /Onboarding/Offer-Letters/[Hire-Last-Name]-[Start-Date].docx.”
Why it works: The template is static and known. The data you provide is finite. Claude can substitute text into the template and save it with a predictable naming convention. This removes the manual step of opening the template, filling in fields, saving with the right name.
What Claude usually gets right: Text substitution, file naming, saving to the right directory.
What breaks: If the template uses complex formatting, tables, or embedded fields, Claude might not preserve the structure. Test with a dummy letter first.
Pro tip: For multiple offer letters, batch them: “I’m attaching 5 new hire requisitions. Generate all 5 offer letters and save each to /Onboarding/Offer-Letters/. Reply with a summary of files created.”
3. Benefits Enrollment Data Quality Check
The task: During open enrollment, employees submit benefits selections through a portal. You need a daily check of submissions against your employee roster to catch mismatches (ghost employees, duplicate submissions, missing elections).
The prompt: “Pull today’s benefits enrollment submissions from [portal]. Cross-check Employee IDs against the current HRIS roster in /Data/HRIS-Active-Roster-2026.csv. Flag submissions where (a) the Employee ID doesn’t exist in the roster, (b) the same employee has multiple submissions, or (c) required elections are missing. Create a summary and save to /Benefits-Enrollment/Daily-QA/Mismatches-[date].csv.”
Why it works: This is a data-matching task with clear failure modes. Claude can pull data from two sources, compare them, and identify problems. The output is actionable for your team.
What Claude usually gets right: Downloading data, comparing fields, identifying mismatches.
What to watch for: If the portal requires multi-factor authentication, Claude may not get through. If employee IDs are formatted inconsistently across systems (one includes leading zeros, another doesn’t), matching will fail. Standardize first, or tell Claude: “Match Employee IDs but strip leading zeros and convert to lowercase before comparing.”
4. Onboarding Checklist Orchestration
The task: Every new hire gets a checklist: email provisioned, laptop ordered, benefits enrolled, manager intro call scheduled. You want a daily report of where each checklist stands.
The prompt: “Pull all active onboarding checklists from [system]. For each checklist created in the last 30 days, extract the hire name, start date, and completion status of each item (email provisioned, laptop ordered, benefits enrollment, manager call). Create a summary showing percent complete per hire and flag anyone past day 5 with less than 50% complete. Save to /Onboarding/Daily-Status/Checklist-Summary-[date].csv.”
Why it works: This is lightweight data extraction and aggregation. You’re pulling from a single source, organizing it, and flagging outliers. No complex logic, no modification of sensitive data.
What Claude usually gets right: Extracting data, calculating percentages, identifying records that meet criteria.
What breaks: If the checklist tool embeds data in nested fields or uses unusual status labels, Claude might misparse. Clarify the structure: “In the tool, each item shows ‘Completed’, ‘In Progress’, or ‘Not Started’. Look for those exact values.”
5. Offboarding Task Tracking
The task: When someone leaves, you need to track IT teardown (laptop return, access revocation, email forwarding), benefits wind-down (final paycheck, COBRA notification), and record archival. Often these tasks scatter across Slack, email, and spreadsheets. You want a consolidated report.
The prompt: “Scan the offboarding Slack channel for messages about [employee name] from the last 2 weeks. Pull IT teardown tasks (look for mentions of laptop, access, email). Cross-check against our offboarding checklist template in /Templates/Offboarding-Checklist.xlsx. Create a summary report showing: Tasks Completed, Tasks In Progress, Tasks Not Started. Save to /Offboarding/Active-Cases/[Employee-Name]-Status-[date].csv.”
Why it works: This aggregates data from multiple sources (Slack, a local file) and produces a summary. Claude is doing the equivalent of reading your Slack thread and your checklist, then comparing them. That’s tedious for humans but straightforward for Claude.
What to watch for: Slack integration requires the Slack MCP connector to be configured. If it’s not, Claude will try to use computer use to open Slack in your browser, which is slower and more fragile. Ensure your Slack MCP is set up first. Also, Slack conversations can be verbose and ambiguous. Be specific: “Look for messages containing ‘laptop’, ‘revoke’, ‘email forward’, or ‘IT’.”
The Computer-Must-Be-Awake Trap
Here’s the gotcha that will snag you if you don’t plan for it.
The Claude desktop app must remain open and your computer must stay awake for the duration of the Dispatch task. If your laptop goes to sleep, Dispatch pauses. The task will resume when your computer wakes, but if you’re not paying attention, you might think Dispatch froze.
For a task that takes 10 minutes while you’re at your desk, this is fine. For a task that takes 30 minutes and your laptop screen locks after 15 minutes of inactivity, you’ve got a problem. The task will stall. It will resume when you wake the machine (or when you unlock it and the Claude app refocuses), but the total execution time becomes unpredictable.
This is a hard constraint, not a design oversight. Dispatch is built on computer use, which requires the screen to be visible and the OS to be responsive. It’s not a limitation that will go away in the next release.
The implication: Dispatch works for tasks you’re comfortable leaving your machine awake for. A 20-minute HRIS triage while you’re in meetings? Fine; your machine stays awake. A 2-hour payroll automation that needs to run at midnight when you’re not there? Not fine. You’d need to either leave your laptop on overnight (bad practice), or move that work to Scheduled Tasks running on a dedicated always-on machine (which has its own trade-offs, covered below).
For HR ops teams, this means Dispatch is excellent for workday-adjacent automation. Spot a data quality issue at 9 AM, assign it to Claude, get results by 9:30. But recurring, time-critical work (payroll cutoff validation, benefits report generation on a fixed schedule) should not live in Dispatch. That’s when you graduate to Scheduled Tasks, which we’ll cover next.
If you’re tempted to hack around this by leaving your laptop on all night or using remote desktop to keep the machine awake, don’t. The first approach burns energy and isn’t sustainable; the second introduces unpredictable latency and failure modes. Instead, use Scheduled Tasks from the start if the work is recurring and time-sensitive.
Security and Compliance: The Research Preview Reality
Dispatch is a research preview. That word matters. It means Anthropic is actively gathering feedback, breaking changes are possible, and you should not treat this as production-ready automation for sensitive data.
In practice, here’s what that means for HR ops.
Visible PII during execution. When Claude is working on your desktop, it’s opening applications and navigating web interfaces. Your HRIS system shows employee names, Social Security numbers, salary data, health insurance elections. All of that is visible on your screen, and there’s no automatic redaction. If you have a glass-walled office or shared workspace, anyone walking by can see the data. If you’re recording a Dispatch session to debug issues, you’re recording PII.
Mitigation: Run sensitive Dispatch work in private, or dim your screen if you’re in a shared space. Disable screen recording during HR tasks. If you need to share Dispatch output with IT for troubleshooting, sanitize it first (remove SSNs, mask names).
No audit trail in the research preview. Once Dispatch completes a task, you see the output, but there’s no built-in log of what Claude clicked, which records it accessed, or which fields it modified. For compliance work (FCRA, GDPR, SOX), this is a problem. You need to be able to prove what an automated system did and when.
Workaround: Create a logging protocol now. When Claude outputs results, have it also log every action to an audit CSV in a shared folder. “File accessed: /HRIS/Employee-Records. Records modified: 3. Timestamp: [date].” It’s manual overhead, but it’s the only audit trail you’ll have until Dispatch adds built-in logging.
UI element misinterpretation. Claude sees your screen as a human does, but it reasons about UI differently. It may click the wrong button, misread a dropdown, or confuse similar elements. For low-risk tasks (data export, reporting), this is annoying. For high-risk tasks (payroll modification, access grants, benefits enrollment), it’s dangerous.
A real example: Claude is looking for an “Export” button and sees an “Archive” button. Both are gray, same size, similar placement. Claude clicks “Archive” instead. Now 500 benefit records are archived instead of exported. Depending on your system, this might be reversible, or it might require IT to restore from backup.
Mitigation: Start with low-risk tasks. Use Dispatch for reporting, data gathering, and flagging. Do not use it for modification, deletion, or submission of data early in your journey with the tool. Once you’ve established reliable patterns and you’ve watched Claude interact with your specific systems, graduate to higher-risk work. And always have a backup or restore procedure in place.
Research preview instability. Breaking changes are possible. A prompt that works flawlessly in March might break in May if Anthropic updates how Dispatch handles file paths or screen parsing. Your scheduled workflows could fail without warning.
Approach: Treat Dispatch as a testing ground, not a permanent automation layer. Use it to validate a workflow idea, gather data on effort saved, and build a business case for formal RPA or a dedicated server-side solution. Once you’ve proven the workflow, move it to a more stable automation platform or Scheduled Tasks with proper monitoring and alerts.
From Dispatch to Scheduled Tasks: The Graduation Path
Once you’ve validated a Dispatch workflow and you want it to run repeatedly on a schedule, you graduate to Scheduled Tasks. This is where you move from “I assigned this task from my phone” to “This task runs every Friday at 9 AM.”
Scheduled Tasks use the /schedule command. You define a task, set a recurrence (hourly, daily, weekly, on a specific date), and Claude executes it automatically on that schedule. The prompt is the same. The logic is the same. The computer-must-be-awake constraint still applies (you’ll need an always-on machine, or at least a machine that’s on during the scheduled window), but you’ve removed the manual dispatch step.
Here’s how it works in practice: You’ve validated your “HRIS Incomplete Records” workflow. Every month, Claude successfully pulls the export, flags missing fields, and saves the summary. Now you want this to run every third Friday at 8 AM, when the export is fresh.
You’d use: /schedule every third Friday at 8 AM. Pull this month's HRIS export. Find all employees marked as Terminated but still assigned to a team. Flag their records in the export file and save to /Shared/HRIS-Data-Quality/Flags.
From that point on, at the scheduled time, Claude automatically kicks off that task on your desktop. No phone, no manual dispatch, no thinking about it. Results land in the folder. If something breaks, you get a notification.
The progression looks like this:
Manual dispatch (phone, ad-hoc) → Dispatch (phone, immediate background work) → Scheduled Tasks (automatic, on a schedule) → (eventually) Server-side execution (always-on, no computer required).
HR ops teams should expect to spend 2-3 months in the Dispatch phase, validating workflows and building confidence. Once you move to Scheduled Tasks, you’re investing in something slightly more permanent, but still fundamentally tied to your personal machine. For truly critical, always-on automation (payroll, compliance reporting), you’d eventually move to a dedicated server or formal RPA platform, but that’s a conversation for next year.
What Actually Works
Here’s what you can reliably count on, based on early Dispatch usage across other domains and what we know about HR systems.
Claude can: Extract data from web-based HRIS, pull reports, filter and organize data into new files, identify mismatches and anomalies, create and save documents based on templates, navigate simple web interfaces, handle basic multi-step workflows (open app → log in → pull data → save file).
Claude struggles with: OCR on images or scanned documents, multi-factor authentication beyond SMS codes, complex nested UI (deep dropdown menus, modal dialogs that layer), real-time Slack or Teams conversations where context is implicit, making judgment calls that require domain knowledge you haven’t explicitly stated.
For HR ops specifically: Dispatch shines with data cleanup, reporting, and file orchestration. It’s weak with anything that requires reading between the lines, understanding company culture, or making exceptions based on context you haven’t encoded in the prompt.
Example of “shines”: “Flag all employees with hire dates in Q1 2026 who don’t have a signed offer letter in their file.”
Example of “struggles”: “Identify employees who were likely hired as contractors but should be FTEs based on their work pattern.”
The second example requires context and judgment. The first is a structured query. Dispatch handles the first effortlessly. For the second, you’d set up the framework (pull contractor vs. FTE field, flag mismatches based on a specific rule you define), and Claude would execute it. But Claude won’t generate the rule for you from observation.
You’ll also find that the more recent and modern your HR systems, the better Dispatch works. A cloud-based HRIS with a clean web interface is much easier for Claude to navigate than a legacy desktop app with a cluttered UI from 2008. If your tech stack is enterprise but current, Dispatch works well. If you’re running legacy systems, you’ll hit more friction.
The early wins are modest but real. An hour a week of manual data triage becomes automatic. A monthly report that took 90 minutes now takes 5 minutes to trigger plus 10 minutes to review. These aren’t revolutionary, but for ops teams juggling competing fires, they compound.
Security by Workflow, Not by Technology
Here’s the honest bit: Dispatch is not automatically secure or compliant. It’s a tool that amplifies what you already have.
If your HRIS systems are already locked behind VPN and multi-factor auth, Dispatch gains that security. If your local network has file permissions in place, Dispatch respects those. If your company policy says “never export PII to shared drives,” you can’t use Dispatch to bypass it (nor should you).
What Dispatch adds is speed and scale. You can now do in 30 seconds what took an hour. The security model is the same; the scope is wider and faster. So your compliance and security thinking needs to tighten, not loosen.
For HR ops teams subject to regulations (FCRA, GDPR, SOX), this means:
Set a clear data-risk hierarchy. Low risk: internal reports, aggregated data, non-identifying metrics. Medium risk: individual records without PII, transactional data. High risk: anything with SSN, salary, health info, or disciplinary records. Use Dispatch liberally for low-risk work. Be cautious with medium. Avoid high-risk data in Dispatch until the tool matures.
Establish output validation. Before Claude’s work touches production (before a report goes to an executive, before flagged records go into a payroll system), a human reviews it. Dispatch should be fast, not faster than safe.
Document the provenance. When a Dispatch task creates an output file, include a header in that file: “Generated by Claude Dispatch on [date] at [time]. Reviewed by [human] on [date].” This creates an audit trail even if Dispatch itself doesn’t.
If you’re in a heavily regulated industry (financial services, healthcare, public sector), Dispatch is a research preview for learning purposes, not production automation. Run it in a sandbox, not against live systems. Validate the workflow, then implement it in a formal automation platform once you’re confident.
The Gartner Context: Why HR Ops Is Accelerating Now
There’s context for why HR ops teams are looking at tools like Dispatch now.
Gartner reports that 82% of HR leaders plan to deploy agentic AI within 12 months. By 2030, Gartner estimates, 50% of HR activities will be automated or performed by AI agents. That’s not hype; it’s structural demand. HR ops teams are bottlenecks. They’re doing work that’s repetitive, rule-based, and doesn’t require judgment. It’s a perfect target for automation.
In 2024, 26% of HR teams used AI for at least one task. In 2026, that’s 43%. It’s a steep climb because the ROI is real. One ops person plus one automation tool can do the work of two ops people if the tool is handling 30% of the volume.
Dispatch doesn’t change the overall trend. But it changes who gets to participate. You don’t need an IT team, an RPA vendor, and a 12-week implementation project. You need a Claude Pro subscription and 30 minutes to test a workflow. That’s why Dispatch matters: it’s bringing automation within reach of the people who actually do the work.
For HR ops teams, this is the year to start experimenting. The tooling is here. The capability exists. The only question is whether you’ll learn it on your own terms or wait until someone else has built the workflows and you’re training to use them.
The Trade-Off
Here’s the core trade-off, stated plainly: You gain speed-to-deploy. You lose auditability and reliability.
Dispatch lets you automate a workflow in an hour that would take IT six weeks to build in RPA. But RPA comes with audit logs, error handling, retry logic, and the kind of governance that compliance teams sleep better knowing about. Dispatch comes with whatever you can cobble together.
For HR ops teams, the calculation is this: Are you optimizing for agility or governance? Right now, most HR ops teams optimize for agility. You’re running lean. You’re doing work that would ideally be split between two people. You don’t have six weeks to wait for IT. You have Friday’s data quality deadline.
Dispatch is for you. Use it.
But use it as a testing ground, not a destination. Dispatch proves that automation is possible. Once you’ve proven it, you graduate. You move repeatable, mission-critical work to Scheduled Tasks (which you can at least set up with retry logic and notifications). Eventually, you move truly critical work to a dedicated machine or an external platform that doesn’t depend on your laptop staying awake.
The research preview status is actually helpful here. It gives you permission to treat Dispatch as experimental. You’re not setting up a permanent system. You’re validating ideas. That’s healthy for this stage of the technology.
Getting Started: The First Dispatch Task
If you’re ready to try Dispatch, here’s how to start.
Pick a task that fits these criteria: It’s something you do weekly or more frequently. It takes 20 minutes or longer. It’s low-risk (data reporting, not data modification). The systems involved are cloud-based or at least web-accessible. No unusual authentication beyond your normal login.
Examples: Weekly benefits enrollment audit, monthly HRIS data quality check, daily onboarding status report.
Set up the Claude desktop app if you haven’t already. Keep it open. It’s fine to minimize it. On your phone, open Claude and describe the task in detail. Include the specific system names, file paths, what you’re looking for, and where you want the output.
Then send it and let go. Don’t watch your screen. That’s the whole point. Go do your actual job. Check back in 10-15 minutes.
The first time, review the output carefully. Did Claude find what you asked for? Did it save the file in the right place? Did the format match what you expected? If the answer to all three is yes, you’ve got a workflow. Run it a few more times and observe patterns. When you’re confident, move it to Scheduled Tasks.
This is the SCALES method in action: Spot a repeatable task (weekly triage), Launch a Dispatch workflow, Evaluate what Claude produced. If it works, you keep it. If it breaks, you iterate. That’s how you build reliable automation.
Key Takeaways
- Dispatch lets you assign background tasks from your phone while Claude works on your desktop, freeing you to focus on other work. It’s available to Pro and Max subscribers on Mac and Windows.
- Claude tries existing integrations first (Slack, Google Calendar, MCP connectors) before resorting to computer use. Desktop control is the fallback.
- Five proven HR ops workflows: HRIS export triage, offer letter generation, benefits enrollment QA, onboarding checklist tracking, and offboarding task orchestration.
- The computer-must-be-awake constraint means Dispatch works for day-of-work automation but not for time-critical overnight jobs. Move recurring work to Scheduled Tasks instead.
- Dispatch has no built-in audit trail in this research preview. Log outputs and review them manually until the tool matures.
- UI element misinterpretation is possible. Start with low-risk work (reporting, not modification) and validate before expanding scope.
- Once a Dispatch workflow is proven, graduate it to Scheduled Tasks for automatic recurrence, then eventually to a dedicated server or formal RPA platform for mission-critical work.
- 82% of HR leaders plan agentic AI deployment within 12 months. Dispatch is how ops teams start that journey without IT overhead.
- The trade-off is speed-to-deploy versus auditability and reliability. For agile HR ops teams, Dispatch wins. Treat it as a testing ground, not a permanent system.