Let the People Who Do the Work Build the Agents
5 min left
I’m part of building an approach at my company where the people who know the work best use AI to build solutions to their own problems in their own area.
They’re the subject matter experts. They know their part of the work better than anyone. So they’re the ones who build.
We prototyped the idea in a hackathon we ran. The question underneath it is simple to ask. Who should build the agents that run the work?
This week two large organizations put their answers in public.
Philstar interviewed Patrick Meneses, chief digital and technology officer at Security Bank in the Philippines. He said the bank has trained more than 3,000 employees on agentic AI. The article reports that staff have developed 3,100 AI agents for business and operational work. SiliconANGLE reported that AT&T has nearly 500 agentic workflows in production.
3,100AI agents staff have developed for business and operational work at Security Bank
A bank and a telecom company. The same question underneath.
Who gets to build the agents and who answers for them?
Here’s why I think the builder should be the person who does the work.
Every time someone builds a tool for somebody else there’s a translation step. The expert explains the work. Someone turns that into requirements. Someone else builds to the requirements. Then the expert tries it and runs into the thing they never mentioned because it was so obvious to them they didn’t think to say it.
I think a lot of good tools lose the detail that made them worth building right there in that handoff.
When the expert builds it the requirements and the builder are the same person. They don’t have to explain the exception that shows up once a month. They already know it. They know which output looks right and is wrong anyway. They can adjust it as soon as they spot the problem.
That’s the knowledge workers idea as I see it. To me AI has made the distance between knowing a problem and building something for it a lot shorter. The people with the deepest knowledge of a problem should get to use that.
Security Bank seems to be leaning the same way. According to Philstar the bank’s training reaches past programmers and data scientists to employees in business, operations, finance and risk. The article’s own example is an operations employee who knows a process well and might develop an agent to improve it.
I want to be clear about what this approach is not.
It’s not anyone building whatever they want with nobody watching. There are real guardrails. And when the work is security intensive or data sensitive, specialists in security and data are brought in.
I think that pairing matters. You want the person who knows the work and the person who knows the risk working together on anything sensitive.
AT&T draws a similar line in public. SiliconANGLE reports that individual employees can create personal productivity agents on their own. Workflows that affect customers have to go through the company’s Business Agent Foundry and pass security, privacy and compliance reviews. Every production workflow starts with human oversight. Agents only operate on their own after accuracy reaches the mid to high 90% range within a risk profile the company accepts.
- Business Agent FoundryWorkflows that affect customers have to go through the company’s Business Agent Foundry
- security, privacy and compliance reviewsand pass security, privacy and compliance reviews.
- human oversightEvery production workflow starts with human oversight.
- Agents only operate on their ownafter accuracy reaches the mid to high 90% range within a risk profile the company accepts.
What I like about that is the line follows the consequence. If a personal agent that drafts your notes gets something wrong, you’re usually the one who finds it. A workflow that reaches a customer puts the mistake in front of someone else first. So the review gets heavier as the stakes go up.
Letting more people build shouldn’t spread the accountability thin. A named person should still answer for every decision that reaches a customer.
Meneses put the accountability piece plainly. “AI can recommend, analyze and assist, but accountability cannot be delegated to an algorithm.”
I believe the building should keep spreading out into the business, closer to the people who do the work.
Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, said something similar in the SiliconANGLE piece. “I feel privileged to have the team that I do, but in the long run, my team shouldn’t exist,” he said. “What we do should be completely distributed.”
That’s a striking thing for a leader to say about their own team. I don’t know exactly how AT&T will get there and I won’t guess.
I’d add one thing from where I sit. Specialists in security and data still have a place as more people build. In our approach they’re brought in for the security-intensive and data-sensitive work, and I think that role matters more as the number of builders grows.
We’re still refining our version of this.
If you’re thinking about this for your own organization, these are the questions I’d start with.
Who is allowed to build an agent today and who decided that?
Do the people building know when to bring in a specialist in security or data?
Where does a personal agent end and a customer-facing workflow begin?
What does an agent have to prove before a human steps back?
Whose name is on it when something goes wrong?
None of these need a perfect answer on the first day.
Security Bank and AT&T gave their answers in public this week. Yours will depend on the work your people do.
Sources
Keisha Ta-Asan, “Security Bank trains more staff, develops AI agents,” The Philippine Star, November 15, 2026. https://www.philstar.com/business/2026/11/15/2570124/security-bank-trains-more-staff-develops-ai-agents
Paul Gillin, “AT&T uses agentic AI to reinvent business processes from the ground up,” SiliconANGLE, November 14, 2026. https://siliconangle.com/2026/11/14/att-uses-agentic-ai-to-reinvent-business-processes-from-the-ground-up/