Last year, we wrote about Scout, Traba’s AI interviewer.

To recap: the light industrial staffing industry—the lifeblood of the global supply chain—is inherently difficult because it is extremely nontrivial to find the volume of consistent, reliable workers who are capable of working the specific job requirements, and are able to pass all of the gates (interviews, drug screens, background checks, certifications, etc.) necessary for massive industrial organizations across the country.

To wit: the conversion rate from interested applicants to filled positions is often 1-2% — twice as competitive as Harvard.

To be clear here — sufficient workers that match these qualifications absolutely exist across the country — as such, to fulfill our mission of making the global supply chain operate at peak efficiency, we must meet them where they are. To do so at the scale and speed necessary for enterprise industrial operations, automated AI at scale is the only viable solution.

Scout manages the vetting aspect of this funnel — it communicates with workers over voice and text, asks logistics questions and role-specific attribute questions, and decides whether they’re qualified for a shift. Nearly half a million worker interviews in, it’s been a resounding success.

But vetting is just one part of the human funnel. Much more goes into ensuring the success between great workers and great businesses.

But first: what do we mean by “Physical AI”?

Physical AI represents the domain of AI systems built to act directly in the real world rather than live entirely in the digital space. In a present universe in which foundational labs seem capable of absorbing every pure-digital task conceivable, “the moat of the real world” becomes the subject of immense scrutiny. The real world comes with a mountain of overhead — hardware, latency, messy human behavior, and human consequences when AI misbehaves — fundamentally different missions that frontier labs are unlikely to take on. NVIDIA’s GTC 2026 keynote declares “the ChatGPT moment for robotics,” while Bezos’s secretive Project Prometheus raised another $10B betting on a similar thesis.

But physical AI is about more than machines. Presently, our global supply chain’s efficiency is determined by the productivity and agency of the humans who show up day in and day out. In 2026, physical AI is about perfecting the autonomous processes that vet, qualify, train, retain, resource, pay, and support human operators.

At the crux of all of our decision making is a simple tenet — construct your agentic systems with the premise that their decisions directly affect the lives of real people. A suboptimal vetting call costs a strong worker a shift, or sends an incompatible worker to a business. A misfired worker SMS or call at 6am loses trust (and sleep). A bad assessment of a worker’s pay rate preferences inevitably escalates into manual operator support down the line. There is no SWE-bench to optimize against here, as correctness is subjective.

Fortunately, because we control the entire productivity loop (worker engagement → onboarding / employment → vetting / qualification → requirements → selection → performance / execution) we still can construct objective functions by evaluating empirical results.

As a first order measurement, a worker who stays on the job for multiple weeks is likely a successful placement, and likely to remain on the job for a prolonged period of time (months).

Ultimately, the goal is to optimize a funnel such that every worker initially engaged becomes a consistent worker (one who remains on the job without voluntary or involuntary dismissal for a period of multiple weeks).

In reality, this will certainly not be the case, but if we can construct agentic solutions to optimize every step of the funnel, we can approach our best case scenario.

The Industrial Staffing Funnel

So what is the funnel that defines a worker lifecycle? Roughly it can be approximated to the following series of steps:

Business Posting — A business posts the exact requirements for an industrial job opportunity

Engagement — A worker is engaged for said job opportunity, through ads, referrals, in person engagement, or otherwise.

Onboarding — A worker signs up via the Traba app and provides key information to determine if they are eligible for any number of job opportunities

Application — A worker formally applies for the opportunity via the Traba app

Vetting — We evaluate the worker for the qualifications relative to the specific job opportunity

Requirements — Workers need to fulfill requirements (drug screens, background checks, orientations and tours, live interviews, etc.)

Matching — We decide on the final set of workers to bring on shift

Pre Shift — Confirming with the worker all details to be successful on the job

On Shift — Guiding and tracking the worker en route to shift, handling clock in / clock out, and performance management

Post Shift — Performance evaluation, post shift feedback, payments and invoicing

To optimize for our end goal, every step of the funnel will need to be driven and executed by agents.

In this series of blogposts, we’ll be expanding on key agent infrastructure and development that has contributed most tangibly to our success in building the fastest growing industrial staffing operation on the planet.

We’ll be looking at agent development from a problem statement perspective. e.g.:

• How do we construct a pliable eval that lets us iterate on vetting interviews safely?

• How do we conversationally manage an entire worker lifecycle such that we can push opportunities, resolve worker issues, dig into worker performance, etc?

• How do we optimize when and how workers receive communications so they receive valuable information without being overwhelmed?

• How do we build a self ingesting knowledge base that can keep up with shifting customer demands without messy human games of telephone?

• How do we convert unstructured facts about workers into finite qualifications to search from?

For each interesting problem that we’ve had to unwind, we’ll expand on a chapter in this blog series. We’re starting with the one that sits underneath all the others — the eval problem.

How do you trust an AI to make decisions about a real person’s livelihood, and to continue to keep this trust as the roles, the questions, and the models all change beneath you?

See you in Beyond Scout Chapter 1: SEER.