Agentic AI: What Leaders Wish They Knew Sooner

As AI agents move from the concept stage to deployment in workflows, the reality for leaders is messy. In this short video from the 2026 MIT Sloan CIO Symposium, technology and business leaders share real-world lessons on agentic AI. Their takeaways cover a range of topics, including managing agents like employees and rethinking workflows from the ground up.

— MIT Sloan Management Review YouTube channel

An ~11½-minute multi-speaker compilation from the 2026 MIT Sloan CIO Symposium, hosted by Abbie Lundberg (Editor-in-Chief, MIT Sloan Management Review), asking IT and business leaders “What have you learned this year about how humans and agentic AI work together?” Eleven named leaders give one field-tested takeaway each — making this the wiki’s densest single source on the human-in-the-loop reality of agentic deployment, and a useful counterweight to vendor-keynote optimism.

The eleven takeaways

• Thomas H. Davenport — human oversight is becoming performative. A long-time human-in-the-loop advocate now worried: agents work far faster than humans, so review is “very cursory,” people are “pestered to approve things rapidly,” and “a lot of humans are not going to want to be auditors of what AI is doing.” The wiki’s sharpest statement that rubber-stamp oversight is a failure mode, not a safeguard.
• Melissa Swift — manage agents like employees. The myth is that you hand agentic AI a task and it “magically scatters away and gets it done.” Reality: “just like any human worker… you have to check the output, recheck the output, re-prompt it.” Humans working with agents “is not that different right now from humans working with humans.”
• George Westerman — agents aren’t ready for prime time. The word “agent” is being applied to unsophisticated things — “increasing the hype without necessarily increasing the value.” The forward move: “automate first, then put humans in the right places” and rebuild processes around desired outcomes, not bolt tools onto the steps of the existing workflow.
• Monica Caldas — build human judgment into workflows. Deployed agents in IT operations via an explicit maturity arc: personal-productivity assistance → identify which workflow pieces to reimagine → deploy micro-agents (“not just one agent that does everything”). Lessons: be deliberate, set clear OKRs, define entry/exit criteria, and build a “trust fabric and governance” that keeps humans in the loop “at the right places, not in every place.”
• Michael Schrage — in the loop vs on the loop. Sees a fundamental split: agents for explicit tasks (“Do this!”) vs agents that clarify intent (“Should I do this?”). The split creates the human-in-the-loop vs human-on-the-loop tension. His position: “I am still more comfortable being in the loop… I don’t trust deterministic software agents yet.”
• Max Chan — keep humans in the loop; manage the agent lifecycle. An AI agent “needs to also be treated like a human employee” — a start point, continuous performance monitoring, and lifecycle management (recognising when it “has to come to an end”). And a labour claim: train human employees to leverage AI and “elevate themselves,” while they “replace themselves at the lower level” — training the AI to deliver their old work as they move to “top-line and bottom-line considerations.”
• Keri Pearlson — combine AI and human strengths. From her MIT Sloan AI-security research: don’t use agents to replace people or tasks; design the new job as the combination — agents sort/summarise/find-insights-in data; people do the “subtle, less obvious communication” of one-on-one human interaction. “The marriage… of agents and of people.”
• Meghna Shah — collaborators, not competitors. Beyond content generation: agents do workflow orchestration, scheduling, repetitive work — “you can have 70 processes happen at the same time.” The question shifts from “how is it replacing a human” to “which new job can we now do” given scale, speed, and orchestration.
• Ramesh Razdan — build trust gradually. “We need to earn the trust of agentic AI.” Graduate from small → medium → large; test, experiment, learn, iterate. The metaphor: “giving car keys to a new driver” — learn on local roads, get trained, then run on the highway at full speed.
• Vanessa Escrivá García — humans design, AI implements. A hybrid model where humans design the process and decide what to do; AI is “the way that we are going to implement it.” Humans stay in the loop and “always… have the last word.”
• Kabir Nagrecha — the onus is on the human. The biggest challenge isn’t the AI improving around humans — it’s “the human learning how to interact with AI”: the handoff, the validation, building trust. “Sometimes the onus is more on the human than it is on the AI, as much as we’d like to blame the software.”

Why this matters to the wiki

This source is the practitioner-chorus counterpart to the wiki’s vendor and academic agent sources. Three threads stand out:

1. The oversight-quality problem is now named at CIO altitude. Davenport’s performative oversight and Schrage’s in-loop vs on-loop sharpen the agent-harness human-in-the-loop layer: the question isn’t whether a human is in the loop but whether the human is actually engaging — speed asymmetry can hollow out review into rubber-stamping.
2. “Manage agents like employees” is a live, contested frame. Swift and Chan argue for the employee metaphor (lifecycle, monitoring, re-prompting); the wiki’s BCG source argues against it (anthropomorphising misassigns accountability and invites AI brain-fry). Recorded as a typed contradicts edge — a genuine open debate, not a resolved one.
3. Reimagine-don’t-just-automate, again. Westerman (“automate first, then humans in the right places”; rebuild for outcomes) and Caldas (reimagine the workflow, micro-agents, OKRs) restate the micro-productivity-trap escape at the same altitude as Argenti’s 3×-not-20%.

Dynamic-capabilities (W&W) reading

• digital-transforming/redesigning-internal-structures + improving-digital-maturity — the dominant theme: redesigning workflows and roles around human+agent teams (Caldas’s maturity arc, Westerman’s rebuild-for-outcomes, Escrivá García’s design/implement split).
• digital-seizing/strategic-agility — Razdan’s test-experiment-learn-iterate and Caldas’s entry/exit-criteria experimentation arc.
• strategic-renewal/organizational-culture — the cultural shift to collaborators-not-competitors (Shah), combination/marriage (Pearlson), and humans-learning-to-interact (Nagrecha).
• contextual/internal-barriers — Davenport’s humans won’t want to be auditors and the trust-building friction (Razdan, Nagrecha) are the human barriers to agentic adoption.

Linked entities and concepts

• Promoted to entity this ingest: Michael Schrage (second source — co-authored Kiron & Schrage 2026).
• Dangling (single-source mentions, deferred): Abbie Lundberg (MIT SMR EIC, host), Thomas H. Davenport (Babson/MIT; very strong future-promotion candidate), Keri Pearlson (MIT Sloan CAMS director; strong candidate), Melissa Swift, George Westerman (MIT Sloan), Monica Caldas, Max Chan, Meghna Shah, Ramesh Razdan, Vanessa Escrivá García, Kabir Nagrecha.
• Concepts: ai-agents, automation-vs-augmentation, agent-harness, micro-productivity-trap, enterprise-ai-adoption, ai-employment-effects.

Relationships

• published-by MIT Sloan Management Review.
• contradicts 2026-05-06-kropp-bcg-hbr-dont-treat-ai-agents-like-employees — the manage-like-employees debate.
• supports 2026-06-02-architecting-ai-native-organizations-redesign-work-at-scale-joe-beutler, 2026-05-21-allen-aws-london-exec-forum-agentic-team-structures, 2026-06-12-argenti-hbr-thrive-alongside-ai-mindset-not-skillset, 2026-05-07-kiron-schrage-compound-benefits.