AI is changing what it means to be a Technical Program Manager, and not in the way most headlines suggest. In the latest LinkedIn Live, David Mantica and Omer Hashmi, CEO of the Technical Program Management Institute, dug into what’s actually different about running AI initiatives versus traditional software programs, and why the TPMs who understand that difference are becoming some of the most valuable people in the room.
Here’s what they covered.
Separating the Value From the Noise
Omer opened with a number worth sitting with: a LinkedIn recruiter recently flagged compensation north of a million dollars as a starting point for a TPM working in AI. That’s the scale of demand right now. But it comes with a catch, because as Omer put it, in AI, the wow factor is easy. Repeatable business value is ten times harder.
He described sitting in on a demo for a large insurance company where a Gen AI agent captured release notes automatically. Everyone in the room was impressed. The real question came after the demo ended: how much time was engineering actually spending on release notes in the first place? Was the juice worth the squeeze?
That’s the muscle Omer says every TPM needs to build in the AI era. Before evaluating a tool, ask what metric it’s actually moving, whether it’s replacing real human effort, and what an acceptable failure rate looks like. Don’t fund the best demo. Fund the workflow economics.
Deterministic vs. Probabilistic: The Core Shift
The biggest structural difference between a traditional software program and a Gen AI initiative, according to Omer, comes down to one distinction: deterministic versus probabilistic behavior.
Traditional software development is deterministic. You define the output, the acceptance criteria, and the exit conditions up front, and Agile’s whole toolkit was built around that certainty. Gen AI doesn’t work that way. There’s rarely a clean acceptance criteria, because the system is built on data science that evolves. Instead of one crisp definition of correct, a TPM now has to balance accuracy, cost, latency, safety, trust, and governance at the same time.
That shift changes how a program gets evaluated too. Golden data sets, adversarial testing, fallback logic, escalation rules, and human-in-the-loop design all need to be defined much earlier than they would on a traditional project.
Omer’s example of what makes the human-in-the-loop piece different: on a traditional platform, a broken workflow usually means a line of code failed to trigger. On a Gen AI system, the code can run perfectly and still fail, because the model misinterpreted the task, used the wrong tool, or generated an answer that sounded plausible but wasn’t grounded in the right context. As Omer summarized it, traditional TPMs manage feature delivery. AI TPMs manage system behavior during uncertainty.
Why “We’re Using Copilot” Isn’t a Strategy
One exchange in the conversation cut right to a common mistake: rolling out a tool like Copilot simply because it’s available, without asking whether it’s solving the right problem. Omer was direct about it. Using a tool like that against the wrong use case is a total waste of money, and the cost of getting it wrong in Gen AI isn’t small. A handful of mistakes can burn tens of thousands of dollars in token consumption.
That’s part of why, as David noted, CFOs are increasingly involved in AI tooling decisions. AI spend isn’t a single centralized purchase anymore, it’s distributed across teams and tools. Omer’s take: this is exactly where a TPM belongs, in the room with the CFO, laying out the pros and cons of each option before the checkbook opens.
Agentic Systems Add a New Layer of Risk
The conversation also covered the shift toward agentic architecture, where systems don’t just generate a response but plan, act, monitor, and iterate on their own. Omer described his own work with an insurance company building an orchestrated agentic solution for claims dispatching, work that used to run through a series of RPAs.
The TPM’s job in that world expands significantly. It’s not enough to understand the end-to-end graph of what each agent does. A TPM also has to understand what tools each agent calls, when human approval is required, what happens when a subtask fails, and what the actual risk is if something goes wrong.
Omer’s example: in a traditional call center chatbot, a bad summary is annoying but low-stakes. You can go edit it. In an agentic system, that same bad summary might trigger the wrong email, update the wrong ticket, or approve the wrong claim. That’s not an annoyance anymore. That’s an operational risk.
Where Deterministic and Probabilistic Approaches Meet
A question from the audience asked where the line gets drawn between deterministic and probabilistic approaches. Omer’s answer: it depends on the organization. In highly matrixed companies, there’s often a data science and AI/ML organization on one side and a traditional software development organization on the other. The TPM sits right at that intersection, translating a business outcome (like solving for millions of dollars in unpaid invoice leakage) into a value stream that spans both worlds.
Someone in the chat pointed out that this sounds a lot like an architect’s role, and Omer agreed it overlaps, along with pieces of business analyst and product owner work. But it goes further. A TPM isn’t just capturing what the business wants. They’re also identifying what the business should be doing that they don’t know about yet, because they understand the technology landscape well enough to see it.
The Two Biggest Mistakes Organizations Make
Asked what trips organizations up most often when they push an AI initiative without strong technical program management, Omer pointed to industry data showing that a large share of AI use cases fail to drive real value. Two patterns show up again and again:
Starting with the tool instead of the problem. Seeing a flashy demo and deciding to implement it before defining what problem it’s actually meant to solve.
Skipping evaluation discipline. Moving straight to execution without a baseline, a real data set, or red-teamed test scenarios in place first.
What TPMs Should Be Building Toward
Closing out the conversation, Omer laid out where he thinks TPMs need to invest going forward. Systems thinking is foundational, since AI is never just a model, it’s prompts, tools, and data working together, and the TPMs who can see across that whole system are the ones who stand out. Just as important is judgment: the real question today isn’t whether something can be built. It’s whether it should be automated, augmented, or left to a human.
Want the full conversation? Watch the recorded LinkedIn Live with David Mantica and Omer Hashmi for the complete discussion, including more on agentic orchestration and where TPM compensation is headed.
