Workforce operations
AI Agents
ShiftForge AI
AI agents that handle the busywork of running a hybrid workforce — so the people running it can focus on the decisions that actually need a person.
Timeline
3 months, 2026
Team
Solo
Role
Design, Strategy, Architecture
Operators spend a majority of their time on reactive problem-solving and manual data tracking, with only 7% having real-time end-to-end network visibility.
Challenge
Three things break the same way every time: staff key in line items by hand, invoices stall waiting to match POs and receipts, and approvals sit in someone's inbox.
Root cause: staffing tools track assignments, workforce tools track shifts. No one owns the corrections and matching in between — where the time and money leak.
Design Goals
Auto-build the plan. Let AI match work orders to POs and draft invoices, so a week stands up without hand-keying.
Give every task an owner. Assign clear tasks so invoices don't get lost and exceptions get caught early.
Show overtime before it happens — at the moment of assignment, not on the invoice.
Outcome
Planning setup down ~80% with auto-built plans
Reconciliation moves without chasing people, because the design always shows the next action
Surprise overtime designed out (~100%) by putting the OT forecast on the assignment card
Research
Competitive Analysis
Before designing anything, I wanted to understand this problem and how competitors approached it.
I scored the top 14 vendors in labor operations software on two things: how well they fit a factory floor, and how well they fit a staffing agency's economics. Not one of them scored well on both.
Bullhorn
Staffing OS
5 / 5
1 / 5
2 / 5
TempWorks
Staffing OS
4 / 5
2 / 5
2 / 5
Shiftboard
Mfg WFM
1 / 5
5 / 5
2 / 5
UKG WFM
Mfg WFM
2 / 5
4 / 5
3 / 5
ShiftForge
Orchestration layer
4 / 5
4 / 5
4 / 5
Gap Identification
Staffing tools track assignments: who's placed where, for what rate.
Workforce tools track shifts: who's covering what, for how long.
Neither covers the middle, where corrections and reconciliation happen. A hybrid solution does.
Design Process
Research
Conversations with staffing coordinators and plant supervisors
Mapped weekly workflows and pain points
Identified primary users and jobs to be done
Scope
Mapped user jobs to product features, grouped by role and approval boundary
Ranked by user impact and effort, cut to what could ship as v1
UI Design
Pulled UI references from operations and fintech tools
Studied approval and reconciliation patterns in adjacent products
Prototyping
Iterated through several rounds of HTML mockups, low-fi to high-fi
Landed on a single shared shell across planning, timesheet review, and reconciliation
Handed off to build with Claude Code in Next.js
Personas
Site Managers
Staffing Agency (On-site)
Supervisors
Manufacturing
“
— HR Manager · Staffing Agency
“
— Plant Supervisor · Contract Manufacturing Site
Why an agentic workflow
The operator's week is four repeating jobs across two software systems that do not communicate:
Onboarding a new worker
Assigning them to a shift
Backfilling when someone drops
Reconciling the invoice at week's end
An agent handles the exceptions that break basic automation.
The AI constraints I designed around
Hallucinated citations.
Made-up references. The agent sometimes cited time entries that did not exist, so I designed the app to checks every reference against the real data before showing it.
Version drift.
Inconsistent answers. The same question got different answers from different model versions, so I locked the answer shape to a fixed template the app checks against every time.
Over-narration.
Rambling. The agent wanted to write three paragraphs when one sentence was enough, so I capped the length in both the prompt and the app.
Fake confidence.
Guessing. The agent would confidently invent things it did not know, so I gave it a way to say "I do not know" and made that the only option when it was not sure.
Cost.
Cost per use. Paying a vendor for every question added up fast, so for the pilot the agent runs on the customer's own machine using a free open-source model.
Multi-agent design
The product runs three agents. Each has a bounded scope, a typed output, and a place where a human takes over.
None of the agents can decide anything on their own. They can suggest, they can show their work by pointing at the exact time entries and rate cards they read, but the final approval always belongs to a person.
How I used Claude
On the build. Claude Code handled the technical scaffolding: page routing, design tokens, automated tests, and the local-file agent pattern.
On the design. In the browser, I used Claude to review other products' patterns, test early prototypes against them, and pull in supporting data during research.
On the method. My usual rhythm applied with Claude at each step: discovery, framing, exploration, prototyping, testing, and handoff.
Key Solution
The five AI capabilities that make it work
Assign work without lifting a finger
When a shift needs coverage, a recommendation algorithm ranks available workers across every connected agency — by certification, availability, cost, and distance — and hands a supervisor a ranked shortlist with the reasoning attached, not just a name.
Why it's needed: supervisors told us they often can't even tell which agency a worker on the floor belongs to. Dispatch and cost had to live in the same place.
Human approval when it matters
Overtime, missing punches, geofence issues — all flagged automatically, with a proposed fix attached. But nothing touches pay or a worker's record until a named person clicks approve.
Why it's needed: every dispute we heard about started with someone finding out too late to have a say. This puts the say back in, earlier.
Take action before it's too late
The moment a worker crosses the overtime threshold, it's flagged — not discovered a week later when the invoice lands and it's too late to plan around.
Why it's needed: one customer wrote off $48K in a quarter on overtime they didn't see coming until the bill arrived.
Accountability that highlights exceptions
Before the agent writes a single word of an invoice, it retrieves the real timesheet, rate card, and policy behind that line. It explains what it already looked up — it can't invent a number it can't point to.
Why it's needed: 15–30% of invoices needed a correction before shipping. The fix had to be grounding, not better writing.
Keep track of how it all happened
Every flag carries the evidence behind it. Every approval is logged with who, when, and why. Every action can be reversed by a person with the right permissions.
Why it's needed: today, disputes get settled by memory and email. That's the opposite of a record.
Metrics
I ran multiple simulations with agents and compared the result against the baselines collected from prior user research. The number looks promising, but we need to conduct further testing on staffing agencies in real-world environments.
Next steps
Next step is signing a small group of staffing agencies and manufacturers as design partners. Running the product against their actual weeks is how we measure overtime caught before it billed, hours cut from the invoicing cycle, and follow-ups on exceptions and reconciliations avoided.
