In the AI Era, HR Should Lead, Not IT: Hire AI Like an Employee
Treating AI as a technology purchase handed to IT is serving the main course as a side dish; what really needs to be broken apart is tasks, not jobs, and HR should lead this, because AI has no feelings while people do.
The argument · tap a timestamp to hear it
Every company in the information business lands in the upper-right quadrant
Mike built a tool with Replit in one night (8pm to 8am), fed it employment research from roughly 15 think tanks and research institutions worldwide, ranked them across four quadrants from least to most disruptable, and checked only one company — One Digital. It landed in the most disruptable quadrant, a blinking red dot, flagging that roughly 25% of roles could be disrupted by AI. His conclusion: any company in the information business will land in that upper-right quadrant.
— Mike SullivanWhat should be broken apart is tasks, not jobs
Mike thinks everyone is looking at the problem through the wrong lens — "jobs" — when what really needs to be broken apart are the tasks that make up each job: everyone has a pre-AI job description and almost certainly will have a post-AI job description. They call this mapping the work genome, are piloting it with four teams, and have brought in three academic institutions to add rigor. His worry: cutting first and thinking about downstream effects later is the easier path, and many companies will take it.
— Mike SullivanHR, not IT, should lead this
Mike puts it bluntly: the group driving this shouldn't be IT, it should be HR. Over the past hundred years HR went from a payroll administrative function to being elevated into the executive suite in the '80s and '90s because talent was the source of intellectual property; now it needs to rise another level — managing not human talent but the organization's human plus AI talent. The reason isn't that AI matters, but that AI has no feelings and doesn't need to support a family, while talent does, so the group assigned to care about people should lead.
— Mike SullivanReducible and irreducible skills
They split human work into reducible and irreducible. Reducible is what can be broken into processes, with clear steps, handed to AI, and AI only gets better at it; irreducible is the "squishy" part — you can't quite explain how it works but it does, coming from experience, life, and signals received from the environment. The key judgment: irreducible skills have no ceiling, and the time saved should be invested on that side. In wealth management, for example, investment modeling, information synthesis, and preparing client meeting materials are things AI does increasingly well; the irreducible part is the human judgment that connects information to a client's real desires.
— Vinay GidwaneyAI employees have résumés, and PIPs too
One Digital "hires" AI as talent: AI colleagues have job titles, job descriptions, even résumés, all managed by human supervisors through a full HR process, and some AI colleagues have even been put on a PIP (performance improvement plan). Vinay says this isn't buying technology, it's "hire the AI into the organization." The direct consequence: when both AI and humans are doing the work, the real question becomes "what is the optimal ratio of human to AI on this task," and that ratio is constantly shifting at both ends — what AI can do is changing, and so is what humans can do. The workforce intelligence score exists to quantify and iterate on this.
— Vinay GidwaneyAnswer engine and collaborative relationship are two different uses
Vinay splits AI use into two categories. One is transactional: treating AI as an answer engine, as an errand runner — "write that email I don't want to write to my kid's teacher," "read this document that's too long." This use simultaneously underestimates AI and yourself. The other is collaborative: you and AI converse, brainstorm, share a goal and push it forward together. One Digital has hundreds of thousands of chat records between employees and AI colleagues, now analyzed by another set of AI colleagues, which can identify when a person shifts from transactional to collaborative — this "amplification score" is the computational driver of the workforce intelligence score.
— Vinay GidwaneyIntelligence lock-in is no less a risk than data lock-in
Vinay opposes companies signing one blanket agreement with Anthropic or OpenAI and having everyone use Claude or ChatGPT. His reasoning: the business these companies want is "renting intelligence to you," so they could raise the rent on intelligence the way a landlord does, or say "unless you give me X, Y, Z, you don't get the smartest intelligence." A few years from now, when humans work alongside AI every day and the organization's shared intelligence lives between people and AI, that company has you by the throat. When people talk about vendor lock-in they mean data locked in a CRM; intelligence lock-in will be a different problem. The solution is owning your own intelligence: being able to hand-pick models per task, not going to a large model for things a small, cheap model can do well, which also optimizes spend.
— Vinay GidwaneyWriting software will become like using a spreadsheet
Vinay thinks engineering is no longer "the longest pole in the tent," no longer a special skill held by a small group of prima donnas who need to be protected and paid top dollar — AI can already write software. If AI can write software, the question becomes: how do you build a governance framework so technology development spreads through the organization and captures the innovation happening among employees. He predicts writing software will become as common as using a spreadsheet: some people are great at Excel, some aren't, but everyone uses spreadsheets. He also flags the need for governance — they lived through "spreadsheet hell," where determining whether a family trust was appropriate ended up spawning six different Excel files and took years to get back to a single tool. The good news is AI can help with governance too, like reading code and finding two people doing the same thing, then merging them.
— Vinay GidwaneyIf the CEO isn't lit up, the whole company won't change
Vinay says one pattern in their data is very hard: if a manager is activated and has a high AI amplification score, their team immediately follows. So his blunt advice to listeners is — if your CEO or boss hasn't had that "oh moment," your organization won't change. It's not enough to say "I believe in AI, we deployed AI": a CEO can't write about AI in the shareholder letter while doing their own job the way they did for the past thirty years. They have to actually change their own work, because that flows downward, gives employees permission, removes fear, and makes using AI no longer a stigma.
— Vinay GidwaneyIn their own words · checked verbatim
we don't see headcount when it comes to our people. we see faces
Mike Sullivan13:10
we don't care about the AI. The AI has no feelings. The AI doesn't need to put food on the table or live a happy life. Your humans do.
Mike Sullivan34:29
So we actually hire the AI into the organization. They have job titles, they have job descriptions, they even have resumes.
Vinay Gidwaney44:39
intelligence lock in is going to be a whole other problem
Vinay Gidwaney52:41
AI can write the software now. So, if AI can write the software, how do you build a governance framework within your organization that allows for a proliferation of technology?
Vinay Gidwaney1:05:48
if you have a CEO or your boss who has not had their oh moment with AI yet, your organization's not going to change.
Vinay Gidwaney1:24:08
Our biology hasn't changed. The technology to train our biology has changed.
Vinay Gidwaney1:28:10
Figures
| Share of roles software estimates could be disrupted by AI | 25% | 12:09 |
| Number of fully AI colleagues deployed | the 14th | 31:29 |
| Volume of chat records between One Digital employees and AI colleagues | hundreds of thousands | 47:41 |
| Number of financial planning strategies/entries One Digital has written | well over a hundred | 54:42 |
| Example confidence level for meeting notes matching a strategy | 96% | 55:43 |
| Human mile run (fastest time in the early days of the technology) | about 12 minutes | 1:28:10 |
| Human mile run (fastest time now) | 3 and a half minutes | 1:28:10 |
| Vinay's estimate of the share of the AI future still unknown | 99.9% | 1:26:10 |
Glossary
- work genome
- A method for breaking each job into tasks and mapping out pre-AI and post-AI job descriptions.
- reducible / irreducible
- The part that can be broken into processes and handed to AI, versus the part you can't quite explain but that works, coming from experience and life signals.
- workforce intelligence score
- A metric that folds the amplification score together with token spend to quantify the output and cost of human-AI collaboration.
- amplification score
- The computational driver that identifies when an employee shifts from transactional AI use to collaborative use.
- intelligence lock-in
- The risk of being held by a model vendor when the organization's shared intelligence lives between people and AI.
- PIP
- Performance Improvement Plan; One Digital has put some AI colleagues on one too.
How to listen
HR leaders, COOs and business-line managers driving AI adoption, especially companies that treat AI as a technology purchase handed to IT.
The segment from 1:02:46 to 1:05:48 on what small companies should do — mostly experiential.