A Leader's Guide to Advanced Team Structures in an Agentic World (AWS Events)

AWS Events YouTube; Steven Brovich (Amazon since August 1999, ~27 years; second-half career on people/culture; part of Tom's "former C-level executives advising customers through transformation" team). Built on ~100 executive conversations + an HBR / AWS-sponsored AI-readiness study + internal Amazon work across the AI stack. The whole talk is a four-question framework for CIOs deciding how to build, staff, and govern in an agentic world:

Economics — use / compose / build
Talent — expert generalists, not specialists
Structure — IT ops Models A / B / C; the hourglass
Governance — Singapore's January 2026 Agentic AI framework

The one-line setup line: "AI won't take your job. Someone using AI will." (Brovich credits Scott Galloway for the popularisation.) The threat isn't the technology; it's the colleague who learned the tools six months before you did. AI doesn't have a P&L target or ambition; people do.

Empirical grounding — Anthropic March 2026 EI report

Brovich anchors his "what's actually happening in the labor market" section on the March 2026 Anthropic Economic Index report, which he calls "the first serious empirical study of what AI is actually doing to jobs." The findings he leans on:

No systematic increase in unemployment for the most exposed workers since Nov 2022 (ChatGPT launch)
Hiring of younger talent into exposed occupations has slowed ~14% — not a collapse, a slowdown, and juniors feel it first
The most-exposed workers are disproportionately older, more educated, better paid than average — opposite of the old "AI replaces the lowest-skilled first" narrative
The gap between theoretical exposure (what an LLM could do — Alondu/MIT exposure scores) and observed exposure (what AI is actually doing at work) is huge: computer/math at ~33% of theoretical ceiling; office/admin at a fraction of 90%

This is the data anchor under the rest of the talk's structural prescriptions.

Economics

The framework's foundational question. Brovich's claim: small-language-model "build your own moat from proprietary data" optimism doesn't survive the unit-economics test.

Training cost up 2.4×/year, inference cost down 10×/year → the gap (the pricing scissors) opens by 12–24× per year
Frontier-model training cost is now approaching the billions; only a handful of companies can afford it. Inference cost approaches zero.
Three worlds emerging along leverage/differentiation axes:

World	Posture	Leverage	Differentiation	When
1. Use	End-to-end managed (someone else operates the AI; you consume)	Highest	Lowest	Commoditised workflow
2. Compose	Frontier APIs stitched into your context (their intelligence + your workflow)	Medium	Medium	Most enterprise use cases
3. Build	Train / fine-tune your own	Lowest	Highest (and highest cost, lowest speed)	Only at points that truly differentiate

Workflows cut across worlds. Don't try to live in one. Let economics + your actual differentiation drive which world each part lives in. Day 1 — frontier model does everything. Month 6 — economics push some to use, some to compose. Year 2 — high-volume/high-differentiation parts move to build. That's the healthy path. The unhealthy path: "we're a build shop on day 1" → burns the company before they understand their own workflow. See Build vs Buy (Agents) for the existing decomposition this layers onto.

The moats AI erodes (the slide that "scares people"): traditional product differentiation is being eroded; what AI can't replicate gains value — "years of operations, decades of trust, things that cannot be parallelized, things AI can't speed up." The bottom-row moats become more valuable, not less.

Talent

The craft is changing. For 30 years, being valuable in tech meant being able to build the thing (write the code, design the schema, ship the feature). The new craft is being able to orchestrate — point an agent at a problem, evaluate the output, steer the next iteration, know when to overrule.

The expert generalist (Martin Fowler / ThoughtWorks, July 2025)

The new orchestrator archetype Brovich names is Fowler's Expert Generalist — seven characteristics including curiosity, collaborativeness, customer focus, first-principles understanding. "These are exactly what agentic AI amplifies. An agent multiplies a curious person. It doesn't multiply someone who only knows one framework."

Hiring implication: hire for those seven characteristics, not for the framework of the year — "because that framework will change three times before this person's first performance review."

The convergence — the Renaissance Developer (Werner Vogels)

As AI enters the team, two converging movements happen:

Specialists broaden — their specialty isn't enough anymore; they need adjacent domains
Generalists deepen — AI gives them specialist-level depth on demand

They meet in the middle as what Werner Vogels (Amazon) calls the Renaissance Developer. Brovich notes Fowler, Vogels, "Jurgen", and PWC are all converging on the same answer: "the valuable human in the loop is the polymath with steering hands." The opposite of what we've hired for for the last 20 years.

Empirical proof — Anthropic's Feb 2026 hackathon

Anthropic's Build with Claude hackathon (Feb 2026): 13,000 applications → 500 accepted → 277 shipped production code → 21M lines generated. The podium:

3rd place: Dr. Mikall Nettoko, interventional cardiologist (MD, PhD; not a professional dev) — built an AI platform for post-appoint patient care in 7 days, coded between patients and on flights Brussels → SF.
1st place: a lawyer built Crossbeam — a permitting tool for California. Also not a professional developer.

"The top three finishers in Anthropic's own flagship hackathon were not professional developers. They were domain experts, and they beat 13,000 other people."

The lesson: domain expertise + AI beats coding skills alone. Implication for hiring strategy especially over the next two years: "the person who understands your customer, your regulation, your product nuance, who can now build because AI fills the gap — that person is gold."

How to staff a team (the old shape vs new shape)

Old world	New world
Team of specialists, each owns a lane, handoffs between them, coordination overhead	2–3 expert generalists + agents filling gaps where specialist depth is needed
Each owns a lane	Each owns a workflow end-to-end
Coordination + handoffs	Collapses

Brovich calls the result hyperconvergence of the team. "If your team still looks like the side on the left for new AI work, you are building yesterday's team."

Four simultaneous forces leaders must hold in tension

Expert multiplier. Senior people with AI are an order-of-magnitude faster (Project Mantle is the Amazon-internal example to look up).
Bottleneck shifts. Not "can we build it" → "do we have the data and can we decide fast enough."
Verification Tax. AI generates code 10× faster but it's 3× harder to validate. Review bottleneck eats velocity.
Deskilling trap. Juniors using AI ship ~17% more code but understand ~17% less of what they shipped. Faster and less grounded simultaneously.

"All four are true at once. The leader's job is to hold the tension."

Structure

Team shape — the four models

Shape	Description	Brovich's call
Pyramid	Many juniors at base, fewer seniors directing	Most orgs today
Diamond	Cut juniors ("AI replaces them"); bulk middle with AI overseers	The trap — see Designing IT Roles for an AI Era (Talent Strategy POV)
Inverted pyramid (Pod)	3–5 seniors, AI does execution; works for execution but no learning path	Right for new AI work
Hourglass	Pod at top, lean middle, juniors learning the craft on the way up	The learning organisation Brovich wants

The framing: the inverted-pyramid pod is the team shape; the hourglass is the organisation shape that houses the pods. Both can be true at different altitudes. Most companies are heading toward the diamond — "the data on that is brutal. Companies see reducing junior talent as the quickest path to ROI." Inverse: top-dollar paid to seniors with "AI" on their resume. Middle hollows out; top explodes. That isn't healthy.

The protect-the-juniors prescription is the core. "If you stop training your juniors, where do your seniors come from?" Brovich cites Matt Garman (CEO, AWS) in a 2025 interview: "How's that going to work when 10 years in the future you have no one that has learned anything? My view is you absolutely want to keep hiring kids out of college." The CEO temptation: today's juniors aren't this year's problem; they're 2034's, "four CEO cycles from now." That's why this requires leadership intervention.

IT operations — the three models

Brovich's biggest mental-model attack is on traditional IT operations. Non-determinism is a feature, not a bug when agents are involved. "A deterministic agent is a runbook. We already have runbooks. We don't need AI for that." The operating-model shift: tolerant of variance in execution, strict about variance in outcome.

"Just like rivers, agents find their own path. Your job is not to prescribe the route. Your job is to define the riverbank, what the outcome has to be, and let the water find its way."

Three operations models:

Model	Shape	Verdict
A — Traditional IT ops	Engineering builds → throws over the wall to IT operations	Dead. "If you're operating Model A today, you have a transition plan to execute, not a strategy to debate."
B — Embedded	"You build it, you run it." 3–5 senior engineers run + operate the pod. Multiple deploys/day, sub-hour recovery, 0–15% change failure (Dora Elite territory)	Works at 2–3 pods; breaks at 10+ (duplication kills you)
C — Embedded + platform	Model B pods + a platform providing runtime, identity, observability, guardrails. Platform doesn't constrain — platform enables.	The endgame for medium/large scale

The six failure modes that stack to "95% of AI pilots fail to achieve meaningful impact" in Model A: runbooks are deterministic / agents are not; ticket culture kills context (25% of incident time spent assembling the humans who understand what happened); operators have no authority over the model; they have no authority over the data (75% of failure modes bypass current telemetry); 91% of ML models degrade over time ("you deploy, it gets quietly worse"); ITIL can't keep up.

Governance

The framework Brovich points to is Singapore's Model AI Governance Framework for Agentic AI v1, launched January 2026 at Davos by Minister Josephine Tao (Singapore's IMDA). The first state-backed governance framework specifically designed for autonomous AI agents. Builds on Singapore's 2019 general framework. See Singapore Model AI Governance Framework for Agentic AI for the full breakdown.

Four dimensions:

Risk assessment upfront — structured before deployment
Human accountability chains — every agent action traced to a named human
Technical guardrails throughout lifecycle
End-user transparency — users must know they're interacting with an agent, and the bounds of what it can do on their behalf

Five things that make this framework stand out — first to mandate agent identity management; integrates concrete testing frameworks (AI Verify + Global AI Assurance Pilot); explicitly addresses multi-agent coordination risk; voluntary but directional (Singapore gov + regulated sectors take it as de facto standard); addresses the deskilling trap head-on (must show agents are training the next generation, not replacing them).

"Singapore landed on four governance dimensions. Amazon, building [agent core], launched on the same four. Convergent evolution — different teams, different continents, same answer."

Four questions agents must answer before they act: Who's the agent? What is it allowed to do? Is it performing as expected? Can we audit what it did? Brovich's architectural point: policy enforcement happens outside the LLM loop — at the gateway, before the model sees the request. "You don't ask the agent nicely not to do something."

"Policy is code is the riverbank. The agent can meander, but it can't cross the bank."

This separates policy writers (security) from agent writers (engineering) — both at their strengths. Compare with AWARE Framework (Glean / Cvent) for the enterprise-side technical-controls counterpart.

What you actually do Monday morning (the six-step landing)

Economics. Pick one workflow. Use / compose / build.
Talent. Decide the pod. 3–5 seniors max. If you can't staff a senior-only pod, you're not ready to build — use or compose.
Structure. Are you A, B, or C? Answer honestly. If you're A and calling it DevOps, acknowledge it.
Governance. Build the four-question check at the gateway.
People. Invest in senior domain experts (the things AI can't compress).
Pipeline. Protect the juniors. Do not cut entry-level hiring to fund senior AI talent. Both are true at the same time.

"The companies that win in the next decade aren't the ones with the best AI. They're the ones with the best operating model around the AI."

Connects to your work

This is the most directly applicable executive-tier source in the vault for the user's day job:

The four-question framework (economics / talent / structure / governance) is a clean spine for any Manila IT ELT slide — and complements the existing CIO Agenda 2026 (CXOTalk), Exponential IT, and Designing IT Roles for an AI Era (Talent Strategy POV) frames.
The hourglass vs diamond call is the talent-strategy POV the three-spine model was missing the pipeline view of. Worth being explicit in any IT-LT communication: we are deliberately running the hourglass, not the diamond.
"AI won't take your job. Someone using AI will." is the cleanest FOBO re-framing line the vault has captured — calibrated (it doesn't dismiss the anxiety) and actionable (it shifts the target from the technology to the colleague).
The model A is dead banner is the cleanest reframe for IT-ops conversations the vault has — and reads consistently with the Cvent / Glean / 6000-agents pattern in Governing AI Agents at Scale (Glean + Cvent, CXOTalk).

Sources

Telegram capture #3482 — user-shared YouTube link + transcript
AWS Events YouTube — Steven Brovich keynote