Agentic AI is a people decision in a technology badge
Published 9 days ago • 8 min read
Sunday 28th June 2026
Agentic AI is a people decision in a technology badge
Most enterprise AI shows no impact on profit, that's not only a technology problem. One reason is workforce design, which is exactly why the CHRO belongs in the room where these calls get made
Machines can take the tasks. The judgment in the hands is the job. Photo by Danny Shives on Unsplash
Hi, welcome to the Trusted Agents Situation Room.
Each week we help business leaders make sense of agentic AI and agentic commerce while the ground is still moving under everyone. We skip the hype and the doom, and stay on the decisions you'll be asked to make, explained in plain terms, in time to make them well.
This week is one of those decisions hiding in plain sight. Most enterprise AI pilots stall, and the easy read is that the technology fell short. Looking deeper, you'll see that work was built around human limits. Redesigning workflows is org design, not IT, which is why this edition is written for the people who own roles, skills and how work is structured. If that's you, you belong in the room where these calls get made.
In 20 seconds
Right now you're being asked to do three things that pull against each other: cut roles where AI can take the load, prove the AI is paying off, and keep the people you can't afford to lose. They pull against each other because the whole conversation is being run as a technology decision, when the question underneath is what your jobs are for.
Agentic AI ["agentic" means software that can act and complete work, not only answer questions] is the first technology that doesn't share the limits your roles were built around. Every wave before it made the tasks inside a job faster. This one makes the role itself the thing to redesign. That is org design. Lead it before it gets decided for you.
Why this is different from every tool before it
What Klarna learned the expensive way
In early 2024, Klarna told the world its AI assistant was handling the work of 700 customer service agents. The figures were real, the savings were large, and for a while it looked like the future arriving early. Then this year the company started hiring human agents back, with its chief executive saying it had cut too far and that quality, and the option of reaching a person, still mattered to customers. Klarna had taken the tasks for the job. The part it designed out was the part that made the service worth trusting. So which part were you ever paying for?
Most companies never get to ask that question as cleanly as Klarna did. They make a less visible version of the same mistake and read the result as a technology problem. MIT's Project NANDA research, The GenAI Divide: State of AI in Business 2025, found that around 95% of enterprise GenAI pilots show no measurable impact on profit or loss. Only about 5% are pulling real value through, yet the cause is not weak models. It's that organisations are bolting capable tools onto workflows that were designed around human work patterns, immediately limiting AI to human constraints.
That is the whole edition in one finding. Old workflow plus new tool gives you a pilot. Redesigned workflow plus a learning agent gives you operating impact.
Why can't you keep the workflow and swap in the tool? Because agentic AI breaks three things you've never had to design around:
It has none of your people's constraints. A person holds a few options in their head, works one domain at a time, tires, and needs to be in the chair for the work to move. Most of your process steps exist to cope with exactly those limits. An agent has none of them which limits the AI.
It acts, it doesn't only advise. Earlier tools told a person something and waited. An agent can complete the whole task. The live question stops being "what can it do" and becomes "what may it decide."
It forces a separation you've never had to make. Once a machine can take the routine layer of a role, you have to say out loud which part was the reason the role existed. Most organisations have never written that down.
Klarna's mistake was avoidable, and so is yours. It comes down to three questions you can answer before you cut a single role. Which part of the job is the task a machine can take, and which part is the reason the role exists? Where must a human keep the final call? And what does the role look like once you redesign it instead of removing it? The next three stories answer one each.
Half the job was never the job
Marek Kowalkiewicz tells the story of a lawyer who heard that AI could review a contract in seconds and said, "That's what I do. Am I obsolete?". The answer is no, and the reason is the point of this edition. Scanning contracts was part of his role, but it was never the reason clients hired him. They paid for someone who could spot a buried risk, read a gray area, and stay steady when a deal went wrong.
Kowalkiewicz calls this the Job Split. Every role has two layers: the tasks, which machines increasingly take, and the outcomes, which stay human. He sets out four moves for any role you look at:
Drop the repetitive tasks an agent can carry.
Defend the judgment, empathy and accountability only a person provides.
Elevate the role by spending the freed time on higher-value work.
Reinvent it as a new human-and-agent hybrid that wasn't possible before.
For a CHRO this is not a thought experiment. It's a job-architecture method. Run it across a role and you get a redesigned role, a re-skilling path, and the start of a new job description, from one exercise.
The "human in the loop" trap
Most leaders reach for one phrase when they want to sound careful: keep a human in the loop (a person who reviews or approves an AI's decision). It sounds like control. It tells you almost nothing. A human standing somewhere near the process doesn't say what that person is for, or which risk they're meant to catch.
What's at stake, from trivial and easily corrected to irreversible?
What are you optimising for, from speed and scale to accuracy, compliance or innovation?
Put those two on a grid and the right kind of human involvement falls out of the box you're in. Low stakes, optimising for speed: let it run, sample for anomalies. High-impact failure in a compliance setting: an expert reviews and can override. Irreversible consequences: a person holds the decision and the agent supports. This is the work HR co-owns with risk and the business, because it defines who is accountable for what an agent does, role by role.
Do you want Situation Room updates delivered to your inbox?
Both tests stay abstract until you run them on a real role. Here is the redesign Klarna skipped, done on a single job. Take a loan approvals officer. The visible job is forms keyed in, documents checked and re-checked, a queue worked case by case, capped by one person's hours. Strip that away and the real role appears. She is a steward of capital and trust. She represents the bank to the customer and the customer to the bank, and she stands behind the decision with regulators.
Run the two frameworks on her:
Job Split. Drop the data entry and repeat checks. Defend the accountability of explaining a decision to a person. Elevate the role with real-time risk and scenario analysis. Reinvent it as a "trust architect" who oversees a set of AI loan agents.
Oversight. High-impact failures, optimised for accuracy. Experts review the AI's decisions, watch for patterns, and adjust the parameters.
Here's the part a machine can't reach. An agent can read income flows and a credit score. It has never had to choose between a mortgage payment and feeding a family. The officer, freed from the admin, has time to listen, to weigh compassion against the data, and to back the borrower who looks risky on paper and turns out to be the most reliable customer on the book.
The same pattern holds across very different roles:
It doesn't stop at the role
Redesigning roles is the start. The larger shift is the one MIT Media Lab points to in its work on social orchestration for agentic systems. The paper names a "Human Connectivity Barrier": a person can hold only so many relationships and hand-offs in their head, and almost every coordination rule in your organisation is shaped by that ceiling. Agents don't have it. They can coordinate across thousands of interactions through shared rules [protocols, meaning agreed ways for agents to interact and hand off work].
So agentic AI doesn't only change who does the work in a role. It changes how work is coordinated between roles. The reorg you'll eventually face isn't a smarter version of today's org chart. It's a different coordination design, and the human limits that drew the current one were invisible until now.
The pilots stall on workflows built for human work patterns. Redesigning those is org design, not IT, and it puts people strategy at the centre of the AI conversation.
What this means for HR's operating model
This is where the easy slogans get dangerous. Two cautionary signals are worth holding side by side.
IKEA took the savings from a customer-service bot and put them back into people, retraining roughly 8,500 call-centre staff as remote interior-design advisers rather than cutting them. The substance of that decision holds up: a cost saving turned into a growth capability. The tidy "they never made a single cut" version that travels with it does not survive a look at 2026, so reach for the reskilling logic, not the halo.
Duolingo went the other way in public, announced an AI-first push, and met enough employee and customer backlash that it softened the message. The reskilling case was sound. The slogan was the mistake. The difference between those two is exactly the thing an employer brand and people function should be watching.
The practical work for HR sits in four places. Job architecture: which roles get the Job Split first, and what the redesigned version looks like. Accountability: who owns the decision when an agent acts, mapped by the stakes test, not a blanket checkpoint. Reskilling: real pathways into the elevated work, decided before a downturn forces it. And what you select and promote on, because once the tasks are handled the differentiator is the judgment, the steadiness, the care. Call it the human umami if you like. It's the flavour you now hire for on purpose.
Pick one role and run the Job Split on it: Drop, Defend, Elevate, Reinvent. One role, one hour.
For that role, write down the outcome it exists to deliver, in one sentence, without naming a single task.
Map its decisions on the stakes test and decide where a human must hold the call.
Find one workflow where a current pilot is stuck, and ask whether the workflow, not the model, is the blocker.
Name the reskilling path into the elevated version of the role before anyone asks "are we cutting heads."
Two questions
Two questions to put to your leadership team:
Which of our roles are mostly scaffolding around a human outcome we've never written down?
When an agent makes a decision in our name, who is accountable, and have we said so out loud?
Where Trusted Agents comes in
This is the work we do with leadership teams: not building the agents, but helping you see which roles to redesign, where accountability has to sit, and how to sequence it without breaking trust with your people or your customers. If you're starting to map this for your own function, a short advisory conversation is the fastest way in. We're also building a more interactive way to work through this with us, and I'll share it here first when it's ready.
If you want to push on agentic AI without losing control of what matters, start here and book a 30 minute conversation with us.
The situation room will be taking a break over the summer, enjoy the holidays
