Four panelists, four industries, one shared finding: the companies pulling ahead on generative AI built governance and process discipline to run agents as an operating system, ahead of chasing better model access.
F
ourteen days is the gap between the old way Concentrix onboarded new hires and the agentic pipeline it built to replace it. I heard that case made in person at Ai4 2026, on a panel called "The Generative AI Playbook: Setting Your Enterprise Up for Success," moderated by Rob Pegoraro and seated alongside leaders from General Motors Financial, Franklin Templeton, and UI Health. The company recruits at a scale most enterprises never approach, more than 50,000 new employees in Europe alone this year, spread across 75 locations and 150 languages. Katheryn Harrison, Concentrix's Global Vice President of Strategy and AI Platforms, described a hiring agent that reads resumes, evaluates candidate answers, and then keeps working after the offer letter goes out, setting up payroll and training as the employee moves through onboarding. The Net Promoter Score for that cohort came in well above prior groups, the onboarding itself ran 14 days faster, and Harrison put the resulting output quality at roughly 50 percent higher than what the company saw before.
That kind of number is the reason Harrison's framing for the panel landed the way it did. Enterprises have mostly stopped asking whether generative AI works. The harder question, in her telling, is why so much of it stalls before it ever reaches that kind of scale. The answer she gave came down to operating discipline: treating AI as a system with governance and measurement built in from the start, rather than a collection of pilots waiting to be stitched together.
General Motors Financial Puts Agents Inside the Software Pipeline
Pankaj Jain, CIO for International Operations at General Motors Financial, described the most impactful use case on his side of the business as a generative AI pipeline built for software delivery itself. Project requirements go in, and agents distill them into business requirements and functional specifications. From there, separate agents handle coding and regression testing, moving the work through the development cycle in sequence.
Jain was careful to draw a line most vendors blur. Humans inspect and evaluate at every step, functioning as the quality control layer the agents run through rather than around it. The rollout is still limited to select teams as the company works through what full-scale deployment looks like.
Franklin Templeton's Test for Whether a Job Needs AI
Max Gokhman, who leads AI and digital work at Franklin Templeton, brought a longer view to the panel than most. He spent more than 20 years building models before generative AI had a name for itself, including random forest models used in market timing work ahead of the last financial crisis. That history shaped the caution he offered builders in the room: do not build a solution and then go looking for a problem to attach it to.
He compared early-stage generative AI to a chainsaw. It is fun to use, and that is the danger. Not every job on an enterprise task list needs a chainsaw, and the discipline worth building is the habit of asking whether a job requires AI before reaching for it.
UI Health Treats Every Agent as a New Identity to Secure
Murad Dikeidek, Head of Cybersecurity at UI Health, brought the panel's sharpest security framing. Every agent an organization puts into production carries its own identity and its own access, and a meaningful share of those agents exist for seconds or minutes before they disappear. That short lifespan is what makes agent identity hard to govern. An agent can misbehave and vanish before a security team knows to shut it down.
Ahead of any of that, Dikeidek's team spends real time on unglamorous cleanup work across the hospital's files, removing excessive permissions and oversharing, and enforcing retention limits on data that has sat untouched for years. That is data hygiene work, separate from the agent identity problem above. None of it can be done manually at hospital-system scale, so the team leans on outside tools to handle the bulk of it.
The generational split inside the hospital was its own finding. Younger physicians keep pushing for faster approval and deployment of new AI tools. Physicians further along in their careers are far more hesitant. Dikeidek gets regular requests to greenlight copilot tools for patient data, and his answer stays consistent: he is not convinced that particular category of tool is ready for that particular category of data, and the team evaluates new requests one at a time rather than approving AI broadly for clinical use.
The pattern across all four companies held. Each one scaled AI by building the governance and the human checkpoints that let the model be trusted with real work.
Unknowns and Uncertainties
Two details from the session are worth naming as open rather than filling in with a guess. Jain referenced a specific system name for the software delivery pipeline that I could not confirm with confidence, so I have described the workflow without attaching a proper noun to it. Dikeidek also referenced outside vendors supporting UI Health's data cleanup and permissions work that I could not confirm with confidence. Readers who attended the same session and can fill in either detail are welcome to reach out.
Before you fund the next generative AI use case, ask your team two questions: is there a named human checkpoint at every agent handoff, and does your identity and access process cover agents that live minutes instead of months? A no to either one is the project to fund first.
Bellamkonda, Shashi. Session notes from "The Generative AI Playbook: Setting Your Enterprise Up for Success," panel moderated by Rob Pegoraro. Ai4 2026, The Venetian, Las Vegas, 4 Aug. 2026.
Concentrix. "Concentrix to Speak at Ai4 2026 About Scaling Enterprise AI for Real Business Outcomes." Press Release, 3 Aug. 2026, concentrix.com.
