The hot potato is your approval queue.

Written by Sébastien Simpson

When agents write the reports and file the approval requests, who is budgeting the human’s attention?

A weekly status report used to cost somebody a Friday afternoon, and that cost was the only thing keeping the count down. Nobody ever decided a program needed five recurring reports. They piled up one request at a time, the way gym memberships do, because each ask was small and each answer took an afternoon, which capped how many anyone could make. Orchestrated agents removed the cap. One workflow can now draft the status deck, the risk summary, the executive email and the steering pack, then route each one to a person for approval. Every item in that routing table is a claim on someone’s attention, and the people who built it rarely count them.

That’s the context for a remark I almost missed from L’Oréal’s session at UNLEASH Paris. CEO Nicolas Hieronimus and Chief Human Relations Officer Jean-Claude Le Grand shared a stage on talent, skills and AI, and the most useful instruction was aimed at managers. In SP Research Institute’s account of the session, Hieronimus said executives should review the habit of having subordinates produce large volumes of memos and reports, now that AI makes information easy to get. When the subordinate is an agent, the habit gets worse: the memo costs nothing to request and still costs a reviewer to read.

The review queue is the real ask.

Orchestration moves the work from writing to reading, and a person does the reading. Two weeks ago I wrote about faster output with the same delivery date: production speeds up, the approval path stays as it was, and the finish line holds. What I hadn’t seen measured was what happens inside the reviewer once the agents have names and titles.

A working paper from Boston University and BCG, led by Emma Wiles, an MIT Initiative on the Digital Economy fellow, ran a randomized experiment on 813 HR and finance managers. Each reviewed documents with planted errors, such as job descriptions and budgets, and the documents were identical across groups. Only the story about the drafter changed: an AI tool, a human employee, or an AI “employee” assigned to the manager’s team and listed on its org chart. Across the whole sample the effects were small. Among the 23% of managers whose organizations already put agents on their org charts, the employee story cut the share of errors caught by 17%, made them 44% more likely to ask someone else to review when asking carried a cost, and took nine points of accountability off the manager while adding eight to the AI system. The authors call it a hot potato effect: the reviewer reads less closely and passes the document along. When the same drafts were labeled as a human employee’s, managers in those organizations scrutinized them harder, so this isn’t just what delegation does to people. Under the plain tool framing, managers still caught fewer than half the planted errors in their twenty minutes, so the baseline isn’t strong either.

Naming agents and seating them on org charts is spreading. In a weighted survey of 1,500 senior US managers in the same paper, 14% said their organizations list agents on org or workflow charts, and 17% said agents have been given a name or identity. A reviewer who responds to a colleague-like agent by reading less and asking somebody else lengthens the queue twice. A group that saves five hours of drafting and receives five hours of new approval requests has saved nothing, and its AI usage numbers will still look excellent.

What L’Oréal trained for.

Le Grand’s half of the session was infrastructure. According to the write-up, L’Oréal has organized roughly 600 skills, placed AI-related skills inside that map, and trained about 80,000 people over four years, including people in factories and warehouses. He also described work on the gap between the skills people believe they have and the skills they can demonstrate.

I have a stake in this part. When I was at L’Oréal, I built a Digital MBA for employees and hired Scott Galloway, among others, to help me evangelize a handful of habits, chief among them critical thinking and always-on measurement. Galloway was running L2 at the time, the benchmarking firm he’d founded with Maureen Mullen, whose Digital IQ Index sorted brands into five tiers, from Genius down to Challenged and Feeble. They scored brands from the outside, and by that scale, most weekly status decks I’ve seen land somewhere between Challenged and Feeble. Critical thinking is now the review step. Always-on measurement is how you learn whether the review is real.

A course only tells you who attended. Le Grand’s focus on demonstrated skills is the fix: people practicing the habits on real work, with a manager who leaves them the time.

Critical thinking is the review step.

Critical thinking works at two distances from the work. Up close, it means checking a figure against its source and asking for the reasoning behind a recommendation instead of settling for the summary. One step back, it means deciding which outputs deserve a human at all and where the threshold sits, because a reviewer who reads everything reads nothing closely. The BU and BCG experiment shows what happens without it: the reviewer’s effort follows the story told about the drafter instead of the quality of the draft. The authors’ prescription is a governance one, to pair each agent with an accountable human owner and set minimum review standards for high-stakes decisions. I’d add that none of this comes from a tool tutorial, and all of it can be practiced on real work.

The same skill sits at the center of the recursive self-improvement debate. It came up when the New York City Council spent October 5 on AI risk, as IBTimes reported. I won’t adjudicate the testimony. What reaches a program lead is smaller: enterprises are building a miniature of that loop, with agents drafting, agents reviewing and a person confirming at the end.

Always-on measurement keeps the loop honest.

Measuring the output tells you what the agents produced. Measuring the loop tells you whether anyone is checking it. For every workflow with an approval gate, I’d track five things weekly instead of at the quarterly review: how many items reach each reviewer per day against how many that reviewer can read properly, how long items wait, how often the reviewer changes or rejects one, how many items need rework after approval, and how many known errors the reviewer catches. The fifth borrows the paper’s design. Seed a sample of agent output each week with errors you planted yourself, as a fire drill, and track the catch rate by reviewer and by how the agent is presented to them, since the paper found the presentation alone can move the number. The third is the override rate I proposed last time, tracked per reviewer. When it drifts toward zero, either the agent is excellent or the reviewer is tired, and the seeded errors and the rework count can tell you which.

Review standards loosen, too. The bar that was tight in March slips by October because the drafts have been good lately, and you can’t see that from the outputs. L2 scored brands from the outside, and the brands didn’t get to grade themselves. A measurement the team can adjust isn’t a measurement. In the Sakana experiment I covered, one model played every role in the loop and stayed honest only because two outside models scored the results and kept the scores hidden. A reviewer who agrees with the system’s own grade isn’t an outside judge.

The juniors are still coming.

Both men said L’Oréal will keep hiring young talent. They acknowledged that tasks juniors used to handle, such as data aggregation and document preparation, are changing, and Hieronimus pointed to consumer-facing work as where juniors can spend the time AI hands back. Aggregation and document preparation are also the memos managers keep requesting, and that’s the bind. If managers cut the requests, juniors lose a training ground. If managers keep adding them, juniors spend their year checking machine output for people who’ll skim it.

The way out is to sort the asks by what they feed. A report that feeds a decision stays, and someone owns the decision it feeds. A report that feeds a folder goes. The third kind exists to teach: a junior drafts the analysis, compares it with the agent’s version, and a senior reviews the reasoning with them. Nobody has to read the output for it to earn its place, as long as the review happens.

The reviewer needs a passport too.

In my last post I wrote about ServiceNow’s AI Passport, which gives each agent a signed credential, permissions that narrow as a task moves, and an auditable record of who authorized it. It’s one of the most thoughtful answers I’ve seen to the question of who is accountable when an agent acts. The BU and BCG finding is about something different, which is the story told about an agent. A credential and an audit trail are controls the system enforces, while a name and a seat on the org chart are a story the reviewer absorbs. I’d keep the controls and be careful with the story.

The person at the approval gate deserves the same treatment as the agent: a verified skill level for the kind of decision they approve, a queue sized to what they can read, a named escalation path, and a record of what they overrode and why. I argued then that juniors should get the agent treatment, and the argument extends to everyone who signs off. Agents earn autonomy in steps, and reviewers should earn authority the same way.

Give every leader an ask budget.

Count the recurring requests each leader makes of a team, plus the recurring approval requests that agent workflows send to people, and estimate the hours they consume. Review the total every quarter the way you’d review a financial budget, and track it weekly like any other always-on metric. When a leader adds a standing ask, they retire one. Route by risk: low-risk output passes defined checks and gets logged, and high-risk output goes to a named reviewer who has time to read it. Write down what the orchestrator may never produce or send, the way Hieronimus described L’Oréal’s decision not to use AI-generated fictional characters to sell real products to real people, which, per the write-up, he offered as a case of values doing the work of a policy.

For a program, start with the standing reports. Name the decision each one feeds and who makes it, cut the ones where the answer is nobody, and have the agents generate the survivors from the system of record so a person reviews them instead of writing them. Then spend the recovered hours on the teaching kind of work.

The account I read gives no figure for how many reports L’Oréal’s managers have stopped asking for. I’d like to see that number, and the approvals per reviewer, next to the 80,000 people trained.

Further Reading.

Sébastien Simpson
Next

We're onboarding agents better than interns.