How do WFM teams use AI? 10 real examples

Chatbots and voice bots are table stakes in support now, so workforce management (WFM) teams are asking a different question about AI in workforce management: where does it fit into the operation itself? The most useful answers come from practitioners who have already built something.
The 10 examples below come from WFM and support operations teams at Flexcar, Checkr, Brex, Whatnot, and FuturHealth. Operations analysts, capacity planners, and WFM leads built most of them, often without help from engineering. Each one solves a recurring problem and shows how much of the work the team chose to hand to AI.
How do AI agents help with contact center scheduling and forecasting?
AI agents help with contact center scheduling and forecasting by acting on WFM data within limits a team sets, such as applying a sick callout to the schedule or moving a meeting to an open slot. Most teams don't start there. They start with AI they drive themselves, like a dashboard built with AI or an analysis in Claude, and hand over more of the work as they learn to trust it.
The examples below fall into four roles, from least to most autonomous:
- Builder: AI helps you make the tool, such as a dashboard.
- Analyst: AI watches the operation and flags what needs attention.
- Assistant: AI drafts changes, and a person approves them.
- Operator: AI makes changes itself, within the guardrails and approvals the team sets.
Only a few of these work as AI agents that act without a person driving each step. When those agents take action in Assembled through its API or the Assembled MCP server, which connects Assembled to AI assistants through Model Context Protocol (MCP), the result is what we call agentic workforce management.
How are WFM teams using AI for forecasting and capacity planning?
1. Long-range capacity planning (Flexcar, builder)
Flexcar's forecast for member growth lives outside its WFM tool, which made it hard to answer the question leadership cares about most: how many people do we need to hire? Danielle Ansley, an operations analyst at Flexcar, used Lovable and the Assembled MCP server to build a capacity planning dashboard. It combines that member forecast with queue volume, contact rates, and current staffing.
The dashboard adds planning assumptions the WFM tool doesn't hold, such as attrition and scheduled slack, plus a heat map that shows where coverage falls short. Flexcar has a one-month hiring lead time and needs three more weeks to get new hires into production, so staying ahead of demand matters. The base version took two to three days to build.
2. Multivariate forecasting (FuturHealth, builder)
FuturHealth is growing fast, so a forecast built on historical volume alone can't keep up. Thomas O'Rear, senior director of operations at FuturHealth, pulls Assembled data through the MCP server and layers in marketing acquisition and revenue targets. He builds forecasts with Claude and the Prophet forecasting library, then pushes the plan back into Assembled at the daily, weekly, and hourly level.
How are WFM teams using AI for intraday management?
3. Real-time queue alerts (Checkr, analyst)
Checkr's first response target for messaging is 90 seconds, which leaves no room to catch a pile-up late. The team's real-time alerting tool watches every queue against thresholds set for each channel. It posts to Slack before a queue breaches, with recommended moves like shifting agents or calling for overtime, and the alerts escalate if nobody acts.
The tool proved itself when a Friday deploy broke the handoff from AI agents to human agents and Monday phone volume hit 10 times the forecast. The alerts got the team from spotting the problem to fixing it in 12 hours, where similar issues used to take days. The tool runs on fixed rules with no language model, a deliberate choice because most of the decisions involved were already predictable.
4. Intraday queue assignments (Whatnot, assistant)
At 7 a.m. each day, Queuetie, Whatnot's Claude-based bot, grades each queue by comparing its backlog with a safe line: the most work the scheduled team can carry. A queue above its line gets the hours it needs to catch up. Every four hours, Queuetie posts who should work which queue based on need, skills, and real-time Zendesk status, and team leads pass the plan on.
WFM and operations leads can also ask Queuetie staffing questions in chat, such as whether six agents can go to training Thursday afternoon. The answer comes from the same queue model.
5. Daily service-level readout (Checkr, analyst)
Each morning, a scheduled Claude task posts a readout of Checkr's service level by queue from the previous day. It also checks Slack for anything that explains a miss, such as an incident or unexpected volume. It only calls out a cause when it's confident the connection is real.
6. SLA root cause analysis (FuturHealth, analyst)
When leadership asks why the team missed SLA, Thomas asks Claude. Using the Assembled MCP server, he has it summarize service level for the week, the main areas of variance, and the timeframes to watch. With a lean team, he gets to those answers in about 10 minutes.
How are WFM teams using AI for schedule changes and time off?
7. Sick callouts (Brex, assistant)
Each sick callout at Brex used to mean five manual steps: logging time off in Assembled, updating an attendance tracker, replying in Slack, rechecking queue staffing, and rescheduling breaks, trainings, and one-on-ones. The team's first bot handled the first three steps, which saved about 8 hours a month.
The current version handles the whole flow. It applies the time off, finds coverage gaps, and proposes schedule changes that the WFM team approves in Slack. That saves about 15 minutes per callout, or close to 60 hours a month across roughly 237 callouts, and it took about 7.5 hours to build.
8. Schedule change requests (Whatnot, operator)
At Whatnot, schedule changes arrive as Slack posts: callouts, late arrivals, meeting moves, and time off. Nobody was scheduled to enter them on nights and weekends, so the schedule stayed wrong until Monday. Queuetie now reads those channels at any hour, works out the exact change, and writes it to Assembled through the API.
When a detail is missing, Queuetie asks the lead in the thread, and anything outside policy goes to a person. The model decides what to do, and code checks the plan against hard limits, runs a dry run, makes the change, and reads the schedule back to confirm. Queuetie handles 95% of these requests, with a median of 1.8 minutes from Slack post to schedule, at about 12 cents per decision.
How are WFM teams using AI for reporting and BPO management?
9. BPO invoice reconciliation (FuturHealth, analyst)
Reconciling an invoice from a business process outsourcing (BPO) partner usually means exporting productive hours, matching line items, and writing up variances by hand. FuturHealth built a Claude skill that reads the invoice from Google Drive, pulls hours from Assembled by queue, and produces a line-item variance report the team can send straight to its BPO partner. Teams that want this inside Assembled can use its BPO billing report.
10. Weekly business review (FuturHealth, builder)
Every Monday, a script FuturHealth built with Claude pulls from Assembled (including its voice AI agent), Zendesk, BigQuery, and PostHog, then generates a 98-slide deck in the company's brand template in about 15 minutes. Each functional leader gets a section with their own metrics. Once the script was stable, the team moved it off live Claude sessions so it runs on its own. Read how FuturHealth built all three of its workflows.
What do these examples have in common?
Each team started with a narrow problem it already understood and added autonomy once the first version worked. Christian Shrader, senior manager of operations tools and systems at Checkr, expected to build a fully agentic tool on day one. He was proven wrong about what would pay off first, since getting consistent eyes on queues delivered value sooner.
The teams running these tools in live operations also built guardrails around them. Checkr's tool is read-only. Whatnot's model has no tools, so only code can write to Assembled. Brex learned from a time zone bug that quietly put a week of wrong time off on the schedule, and it now runs a second bot that reviews the first.
The practical lessons were about data and ownership. Clean configuration in your WFM tool comes first, because agent states and event types feed every full-time equivalent (FTE) calculation. Tools also need to live where the whole team can reach them, since a build that runs on one person's laptop is only as reliable as that person.
The best first build is the decision your team already makes the same way every time. Start there, put guardrails around it, and the next step from builder toward operator will be clear.
Watch the teams walk through their builds
Every example here comes straight from the team that built it. In What WFM teams are building with AI: Vol. 1, practitioners from Flexcar, Checkr, Brex, and Whatnot show their tools working, explain how long each one took to build, and share what they learned along the way.
For FuturHealth's workflows, watch Claude for WFM and support leaders. Thomas shares his BPO reconciliation skill and his 98-slide weekly business review live, and explains how he uses Claude for forecasting and root cause analysis.



