Who Approves AI on Your Factory Floor?
AI arrives in manufacturing through operations budgets, not HR. Employment attorney Burt Garland and Phil Brandt on the internal process gap that turns useful technology into legal exposure.
Picture a camera system installed on the plant floor to cut down on forklift near-misses. It works. It also turns out to produce clean data on who is where, doing what, for how long. A year later, the company is planning a reduction in force, and someone pulls that data to help decide who stays.
Nobody approved that. There was no meeting where a leadership team decided to convert a safety system into an employment decision tool. It happened one reasonable step at a time. That’s the pattern worth watching, because exposure from AI on the factory floor rarely comes from a bad decision. It comes from a decision nobody made.
The problem isn’t the technology. It’s that nobody owns the decision.
Phil Brandt, AAIM’s President & CEO, spent his career in manufacturing before joining AAIM, and he described the dynamic from the inside: when a plant learned it had a new capability, nobody circulated a memo about the boundaries of its use. People used it to the best of their abilities. “Sometimes those things will get used,” he said, “without management or leadership even realizing they’re used that way.”
That was manageable when the capability was a clipboard and a stopwatch. What changed, as employment attorney Burt Garland of Ogletree Deakins put it, is the scale and the sophistication — not anyone’s intent.
The technology also arrives from an unexpected direction. When manufacturers hear “AI in the plant,” they picture robotics. The systems that create employment law exposure are usually something else: dash cams, facility cameras, productivity and monitoring software, scheduling tools, anything that generates a score a supervisor might act on. None of it is sold as HR technology, and all of it gets purchased through operations, safety, engineering, or IT budgets, by people with no particular reason to loop in HR.
Which is the actual gap. Garland’s point throughout the conversation was that the legal analysis doesn’t follow the product category — it follows what the technology collects, what it does with that information, and how the employer uses the output. Somebody has to ask those questions. In most companies, no role owns that step.
“The name on the box doesn’t control. The actual function does.”
Burt Garland
Three questions that belong in your approval process
Garland boiled the pre-deployment analysis down to three questions. What makes them useful isn’t that they’re legally sophisticated — it’s that they aren’t. Any operations leader can answer them without a lawyer in the room, which is exactly why they work as an intake gate rather than a legal review.
- What does it collect?
- What does the system do with that data, and where does it store it?
- How will the output be used in the workplace?
The answers tell you where your compliance obligations begin, and whether you need a deeper review at all. Garland was specific about the line: a safety tool becomes an HR and employment law matter the moment it starts identifying employees, monitoring behavior, measuring performance, tracking time on tasks, or producing information that could be used for discipline or any other employment decision.
Two answers in particular should route the decision to HR or counsel. The first is any sign the system identifies individuals, which can pull you into state biometric privacy law — consent, security, data transfers, retention, destruction. Garland pointed to Illinois’ Biometric Information Privacy Act as one with real teeth, and a live issue for any manufacturer with workers in the state, not just companies headquartered there. The second is any output that will influence hiring, promotion, discipline, scheduling, pay, or termination, which can trigger its own layer of notice, assessment, bias-testing, and human review obligations.
The requirements are also not uniform. Biometric privacy, electronic monitoring notice, and automated employment decision rules all vary by jurisdiction, and as Garland noted, a national employer cannot assume one notice works everywhere. If you need the state-by-state detail, Ogletree Deakins published a primer for manufacturers that maps the specific statutes — the right reference once you know which tools you have.
One caution on consent, because it’s where a lot of companies stop. It matters, but you can’t have consent in a vacuum — depending on the tool and the state, you may also need specific notices, retention rules, security safeguards, and vendor protections. Same with the old standby of posting a sign. If the cameras are a closed system recording to a drive, a camera notice is fine. Most systems aren’t that anymore.
One more thing belongs in the gate: vendor diligence. Most manufacturers are buying these systems, not building them, which raises a fair question about how much risk sits with the vendor. Garland answered it with a comparison every HR leader already understands.
“Think of all the companies that outsource their FMLA. The employer still remains responsible for the actions and decisions of that vendor. They can contract with the vendor to allocate risk, but that does not eliminate potential employer obligations or liability.”
Burt Garland
And it isn’t a one-time gate. Garland’s recommendation was to review the system again whenever new features are activated, because a feature release can change the answer to all three questions without anyone signing anything.
The second use is where good companies get caught
The riskiest moment usually isn’t deployment. It’s the second use, and Brandt named the scenario directly: a system installed for process improvement, with no employment purpose in mind, that later becomes a data source for a downsizing decision. Who do we keep. Who do we not. The consent, the notice, and the internal justification were all scoped to the original purpose — and nobody re-scoped them, because nobody experienced the second use as a new decision.
“Hold true to the purpose and the intent of why you’re using that system, versus letting it become a second-hand or third-hand tool for adjacent decision-making.”
Phil Brandt
Garland’s practical fix: if you change how you’re using a tool that’s already deployed and that employees already know about, make sure they know about the change too. That’s a communication step, not a legal one, and it’s the kind of thing a process catches and good intentions don’t.
Underneath it all is the point Garland kept returning to: the employer owns the employment decision regardless of what produced it. Responsibility doesn’t transfer to the algorithm, which is why bias testing, impact assessments, notices, and documentation are becoming standard parts of AI governance rather than optional extras.
The process has to reach the frontline supervisor
Brandt made the observation that closes the loop: technology decisions get made in a room, at a board or leadership level, and then trickle down. Whatever governance you build has to survive that trip — and it fails at the moment a supervisor looks at a score and treats it as settled.
“An AI score is not automatically a fact.”
Burt Garland
Garland was careful about the framing. Skepticism isn’t quite the word, he said — caution is. As more people work with these tools, more are running into AI that’s flat-out wrong, or hallucinating outright. Supervisors need to understand the limits of the tool they’re handed, how its output is meant to be used, when human judgment is required, and when to route something to HR or legal.
That’s also what separates real human oversight from the appearance of it. “We always have a human make the final call” isn’t meaningful review if the human is confirming a number they don’t understand. Garland’s standard: the manager has to understand what the output represents, question it when appropriate, weigh other relevant information, and have genuine authority to override the system.
Frequently asked questions
Who should own AI approval — HR, IT, or operations?
Operations will usually initiate the purchase, and that’s fine. What matters is that a defined step exists between evaluation and deployment where someone answers the three questions and decides whether HR or legal review is needed. The failure mode isn’t the wrong owner. It’s no owner.
What should HR do before a new AI-enabled system goes live?
Garland’s sequence: inventory the system and identify which employees it affects and in which states, determine what data is collected and where it’s stored, review notices and consent, assess security and vendor terms, decide whether outputs will affect employment decisions or stay limited to process improvement, train supervisors, and document the compliance work. His shorthand for the last two: document, document, document, and train, train, train.
Our tool was installed for safety, not employment decisions. Do these questions still apply?
Yes, and arguably more so — that’s the profile most likely to drift. The original purpose doesn’t govern; the current function does. If the output has started informing decisions about people, the obligations follow the new use, not the original justification.
Governance isn’t the brake
It would be easy to read all of this as a reason to slow down. Brandt raised that tension on the show and then answered it: good governance is what lets you deploy with confidence instead of hoping nothing surfaces later. Garland agreed — the goal isn’t to stop AI through compliance, it’s to deploy it intelligently through compliance. And none of it requires a new department. It requires one defined step in a process that currently doesn’t have one.
There’s usually real enthusiasm around what a new AI system will do for the business, and it’s often justified. It just shouldn’t be the only voice in the room. If you’re evaluating a system now, AAIM members can reach the Solutions Team to think it through before anyone flips the switch. Burt Garland can be reached directly at burton.garland@ogletree.com.
Want to hear the full conversation?
Phil and Burt opened the episode with three developments worth knowing: a New Jersey paycheck-splitting scheme that cost two related companies $457,500, a new bipartisan paid leave bill in the Senate, and an NLRB ruling that an employee’s public criticism of his company’s DEI program was protected activity.
Related from This Week at Work: Jay Samit on AI: 5 Ideas Every Leader Should Hear · EEOC National Enforcement Plan and Employers · Comp Time: What’s Legal, What’s Not · All This Week at Work episodes
AAIM in the Community
July 31: Central Illinois HR Conference
Aug 4 – 7: MOSHRM State Conference
Aug 24 – 26: SHRM Indiana Conference
Aug 30 – Sept 2: HR Florida Conference & Expo
Oct 9: Greater St. Louis SHRM Conference
Training
Contact Us
Headquarters
12851 Manchester Rd
Suite 150
St. Louis, MO 63131
Toll-Free: 800.948.5700