Some friends were talking about doing AMA (Ask Me Anything) sessions. I thought I’d try to figure out how to do that on my own. So I asked Gemni some questions about my views. The answers were pretty good.
It’s more than a little weird to be able to go to a public utility and ask it what I think. It’s even weirder when the answers are close, accurate, and succinct.
Let me know in the comments if you’d find a real time version of this… an AMA in a webinar-ish format.
In the rest of the article are these questions and their answers. The title is a link to the full Gemni answer. The basic text is key bits from the Gemni Answer. Then, my commentary and improvements are marked with JS:
Why should HR be thinking about AI governance?
What is the best way to get started with AI Governance in HR?
What is Data Quality?
Why does the testing of AI instances matter for AI in HR?
How do you test an AI instance?
How does Sumser see failure and failure rates in HR AI instances?
What is a talent community?
Why should HR be thinking about AI governance?
”AI governance isn't a bureaucratic checklist or a philosophical discussion about ethics. It is an operational necessity. It is the process of embedding clear authority, liability, and human safety guardrails directly into every software system that touches an employee or candidate.”JS: AI governance will eventually feel like financial governance…who has authority to decide, responsibility for execution, accountability for results. It will be well institutionalized when it disappears into the work. Until then, governance is likely to feel stiff and over-bearing. Building new habits can feel like that.
What is the best way to get started with AI Governance in HR?
”Before you can govern anything, you must know exactly what tools exist and what they are doing. Sumser states that organizations must catalog every single instance where an AI tool or algorithm is utilized. For every tool listed in this registry, you must explicitly document:Who owns and operates it.
What does it do? What doesn’t it do?
Who created, maintains, and improves it.
Who troubleshoots it when things inevitably go sideways
Who has the authority to over rule and/or shut down the AI instance.”
JS: Building the first registry is detective work. Inevitably, a well executed initial registry will create a process that requires entry in the registry before an AI project can get started. Now that we are a couple of years into the wild west approach to AI implementation, the need to understand, prioritize, budget, and monitor the various entrepreneurial projects in our organizations.
Revising and updating the registry is the start of governance. The real trick here is to get the organization’s members to let the system know about new projects. This, in turn, suggests that finance may be involved because they can control spending on projects .
What is Data Quality?
”For John Sumser, superior data quality isn't achieved by making data neat; it's achieved by making data accurate to the chaos of human reality. High-quality HR data must retain its origin, allow for conflicting viewpoints, and treat missing information with the same analytical weight as completed data.”“When you eliminate outliers and squish complex human professional lives into generic data fields, you lose the unique differences that actually make individuals valuable to an organization.”
JS: Importantly, some of the most important data can be mined from empt fields in the database. The problem with current ideas about data quality is that they smoosh together things in pursuit of completeness. When an employee doesn’t answer a question or leaves a field blank, they are telling you something. Forcing data into those blank spots using machine generated ideas/characteristics damages the integrity of the data. (I’m looking at you skills inference).
Why does the testing of AI instances matter for AI in HR?
”Sumser urges HR leaders to stop wasting time on broad, philosophical debates about hypothetical AI futures and focus heavily on risk containment.”JS: What matters most is understanding all of the ways an AI instance can fail. One tests AI instances in order to understand where and how they fail. That allows the organization to mandate firm limits, understand what access the AI instance should have, and to plan what to do in the event that the instance fails. Each AI instance could fail in unique ways.
How do you test an AI instance?
“Test “Work-as-Done” (The Complex Reality) rather than “Work-as-Imagined“
”Feed the AI edge cases, internal political contradictions, incomplete datasets, and ambiguous text. Instead of assuming the technology can automatically replace a worker, Sumser states the burden of proof is on the machine to demonstrate it can handle the “ugly bits” of the job that nobody ever writes down in the official workflow diagram.“Map the “Competency Envelope”
You must systematically push the AI instance until it fails to map its “competency envelope”—the exact operational boundaries where its reliable logic stops and its hallucinated guesswork begins. Testers must find the precise niche topics or data constraints that trigger system breakdowns”Cumulative Error Stress-Testing (Chained Agent Validation)
Errors accumulate exponentially across an AI chain. Even if individual agents are 99% accurate, the overall system reliability plummets as the chain gets longer.How to test it: Test the entire data lifecycle from end to end. Measure how a tiny data leak or mathematical error at Step 1 compounds by the time it reaches Step 5. Testing must explicitly reveal how components corrupt one another over time.
Intentional “System Failure” Simulations
Run controlled exercises where you intentionally trigger AI errors—such as feeding it a deeply biased historical dataset or letting it make an egregiously incorrect automated ranking decision. Use these simulations to build human response protocols so your human operators are trained on exactly how to spot, override, and correct the machine when it behaves unpredictablyJS: The thing that’s missing from this answer is that you have to test each instance (and the system it lives in) many thousands of times. That will give a distribution of failures. That’s where the analysis of the instance’s safety and reliability should begin.
How does Sumser see failure and failure rates in HR AI instances?
“Sumser rejects the idea that AI failures are entirely unpredictable, random events. Instead, he views them as an engineering reality following a distinct statistical curve, where the most severe corporate danger does not live where most people expect it to.”
”Every aspect of AI has a quantifiable failure rate. Good AI governance means shifting attention away from rare, catastrophic edge cases and building continuous reliability testing explicitly designed to catch the heavy distribution of quiet, subtle errors that erode an enterprise's data integrity over time.”
JS: There are three classes of AI ‘failure’.
Obvious Catastrophic Failures: These are rare easy for the user to see. They involve things like completely hallucinating a law or blatantly making a clearly biased recommendation. Most failure rates are talking about this kind of error.
The Drift (the real danger): These ‘errors are much harder to spot because they sound almost right. Slight shifts in nuance will cause the system to drift and deliver skewed input to the next system in the chain.
Cascading failure: The AI instance always sits inside of another system. Small, indistinguishable errors flow easily.
What is a talent community?
Here are a few pearls from Gemni’s answer:“Most corporate “talent communities” are not communities at all; they are glorified, passive resume databases masked by marketing language”
“A real community requires reciprocity, mutual benefit, and ongoing communication. A community is an ecosystem that gives value back to the job seeker.”
“Sumser strongly cautions companies about relying on automated AI ranking systems to sort through their talent pools. If an organization uses an unvetted algorithm to segment and rank members of its talent community, it risks baking structural bias into its hiring pipeline, inadvertently screening out highly qualified non-traditional candidates.”
JS: In the end, a vibrant talent community that delivers routine value to its members is a new class of stakeholder/employee → Prospective employees. This makes the Talent Community a sort of club for the organization’s farm teams.
Image by Zyanya Citlalli