Ninety percent of developers now use AI and individual throughput is up — yet org-level delivery stays flat almost everywhere. This module explains why, maps the mandate landscape you are expected to translate, states the honest base rates nobody puts on slides, and defines the two-half job the rest of this track trains.
Rahul is a director at a Bengaluru GCC — sixty engineers, four EMs, a US retail parent that wants agent adoption scaled and headcount flat. Eight months ago he did everything the playbooks said: bought Copilot seats for everyone, ran a kickoff, tracked adoption weekly. Adoption hit 85% in two months. PR counts rose about 30%. His slide practically wrote itself.
Then the parent org's VP asked one question on the quarterly call: "So what shipped faster?" Rahul pulled the delivery data and found the answer he could not say out loud: nothing, measurably. Cycle time was flat. Change-failure rate had crept up two points. One team's code churn — lines rewritten within weeks of merging — was climbing quarter over quarter. He had proven his engineers use AI. He had no evidence the organization was any better at delivering software, and some evidence it was getting quietly worse at keeping quality.
Rahul's situation is not an anomaly. It is the modal outcome of AI adoption in 2025–26, and understanding why is the founding insight of this entire track.
DORA's 2025 research — roughly 5,000 respondents — found something that should reframe how every engineering leader thinks about AI. Adoption is near-universal: about 90% of developers use AI at work. Individual task throughput is genuinely up, on the order of 21%. And yet organization-level delivery performance is, for most orgs, flat. The individual gains are real. They are just not compounding into org outcomes.
DORA's explanation is the mirror effect: AI is an amplifier of the system it lands in, not an independent force for improvement. Fast code generation flowing into a slow review process produces a bigger review queue, not faster delivery. More PRs hitting a flaky CI pipeline produce more reruns. Speed injected into a system with weak batch discipline and unclear ownership produces motion, not progress.
Individual output rises. The review queue grows, batch sizes balloon, rework climbs, and incidents follow the untested surface area. Leaders see activity metrics rise while delivery metrics sit still — exactly Rahul's dashboard.
The same tools land on small batches, fast feedback loops, strong version control and quality platforms — and individual speed compounds into org throughput. The difference is not the model or the vendor. It is the surrounding system.
Here is the leadership punchline: DORA's AI Capabilities Model names seven organizational capabilities that determine whether AI amplifies performance or dysfunction — and every one of them is owned by leadership, not by individual contributors.
No engineer can buy their team a clear AI stance or a quality platform. Those are budget, policy, and org-design decisions. Which means the flat org-level results are not an AI problem. They are a leadership backlog.
AI amplifies the system it lands in, and the seven capabilities that decide which way it amplifies are all owned by leadership. If individual gains aren't becoming org gains, the missing work is yours.
While the evidence stayed nuanced, the mandates did not. A handful of highly public company policies now set the tone for boards and CEOs everywhere — and in every one of them, the manager is both the enforcement layer and the translation layer.
| Company | Mandate | What lands on the manager's desk |
|---|---|---|
| Shopify (Apr 2025) | "Reflexive AI usage is now a baseline expectation"; AI use factored into performance and peer reviews; teams must prove AI can't do a job before requesting headcount | Translating "reflexive usage" into fair, observable expectations; running prove-AI-can't cases without blocking legitimate asks |
| Microsoft (Jun 2025) | "Using AI is no longer optional — it's core to every role and every level"; managers weigh AI use in reviews | Deciding what "weighing AI use" means in a calibration room without turning telemetry into a score |
| Coinbase (Aug 2025) | One week to onboard AI coding tools; non-compliers dismissed; ~40% of daily code AI-generated with mandatory human review; "no vibe coding in financial systems" | Enforcing the pairing that matters: mandated use plus mandated verification, with explicit no-AI zones |
| Duolingo (2025) | AI use evaluated in hiring and in performance reviews | Redesigning interview loops and review rubrics so the evaluation is defensible |
Read the Coinbase row carefully, because it contains the shape of the whole era: the same mandate that requires AI generation also requires human review of everything generated, and carves out zones where AI-first development is banned outright. Mandate plus verification plus boundaries — that is a governance design, and someone has to build it. That someone is you.
Your CEO will forward one of these memos. Your job is not to re-forward it. Your job is to translate it into policy, rubrics, review standards and measurement that are fair, enforceable, and honest — which is Modules 4 through 6 of this track.
Before you promise anything upward, anchor on what large surveys actually show. These numbers are your negotiating position with every executive who asks why you aren't 10x yet.
These figures are survey self-reports, each with its own sample and framing — treat them as base rates for setting expectations, not precise forecasts for your org. Their strategic value is that they pre-empt hype: a leader armed with them can commit to realistic targets and hit them.
The gap between the mandate landscape and the base rates is the defining tension of your job right now. Executives are mandating universal adoption while the median org has no measured ROI. Leaders who resolve that tension with hype get found out in two quarters. Leaders who resolve it with honest systems — policy, measurement, enablement — become the people boards trust. This track exists to make you the second kind.
Everything in this track sorts into two distinct bodies of work, and it pays to keep them separate in your head.
Research suggests around 60% of a typical EM's tasks are already AI-assistable: meeting synthesis, 1:1 prep, feedback drafting, status reporting, planning scenarios, decision memos. Modules 2 and 3 build this personal pipeline. The payoff is hours per week returned to the work only you can do — and firsthand fluency, because you cannot govern tools you've never seriously used.
Policy, review standards, agent governance, honest measurement, ROI cases, rollout, upskilling, org design. No vendor teaches this neutrally, because every vendor's version concludes with buying their product. Modules 4 through 8 build it vendor-neutrally. This half is the differentiator — the work the seven DORA capabilities demand and the mandate era assumes you know how to do.
Half A makes you credible and gives you time. Half B is why this track exists. Weight your attention accordingly.
Every playbook in this track — every prompt, every workflow, every lab — operates under two non-negotiable rules. They are introduced here once and assumed everywhere.
The privacy boundary and the human-send gate are the two rules this entire track assumes. When any later module seems to skip them, it hasn't — it is relying on them.
For leaders in Indian services companies and GCCs, everything above lands with extra force, in three specific ways.
First, the delivery-manager double bind. Your company mandates AI adoption for margin; your client's contract prohibits client code in public AI tools. Both are enforceable, both are yours to reconcile, and the reconciliation — per-client approved-tool matrices, boundary rules, MSA/SOW AI addenda, audit evidence packs — is a concrete skill this track teaches in Module 5. A "client says no" variant exists for every playbook.
Second, GCC leadership demand is the growth story. India hosts roughly 2,100+ GCCs employing over two million professionals, and the centers are shifting from delivery engines to AI nerve centers, with VP-engineering and product ownership sitting in India. Around 80% of newly launched GCCs put 15–25% of headcount into AI/ML within eighteen months. The leaders those centers need are exactly the Half-B profile: adoption at scale, measurement a parent org trusts, talent-war-aware org design.
Third, pyramid economics. The services business model is a junior-heavy pyramid, and AI compresses exactly the billable junior work the pyramid is built on. Junior-role redesign and upskilling at scale — Modules 7 and 8 — are not optional enrichment for Indian delivery leaders. They are existential business questions, and Gartner's finding that 80% of the engineering workforce must upskill through 2027 lands hardest on orgs with six-figure engineer counts.
By the end, you will have authored an AI-usage policy and red-teamed it, set review standards that caught a seeded defective PR, defined a five-rung agent autonomy ladder, designed measurement that survives a skeptical CFO, phased a rollout with kill criteria, ruled on a tiny-team restructure, and written a board narrative from deliberately mixed metrics. Not read about these things — produced them, against scenarios with planted traps.
Rahul, from the opening, is where you might be now: adoption proven, impact unproven, quality drifting, hard questions incoming. The rest of this track is the difference between his current answer and the one he needs.
The AI-era leadership job has two halves — using AI in your craft, and building the org system that converts individual gains into delivery outcomes. The first buys you credibility and hours; the second is the work only leadership can do, and it is the half this track weights.