How Engineering Teams at Fortune 1000 Companies Code with AI

Big companies don’t just “use Copilot” any more. They mandate tools, build their own, run agents that open thousands of pull requests a week, and report the results on earnings calls. They also run into the same trust, quality and measurement problems as everyone else, just at scale.

This report pulls together what large engineering organizations have said in public: engineering blogs, executive statements, and the big surveys. Every number below links to a source I read, listed at the bottom. Company figures are self-reported and rarely use the same definitions, so treat them as signals, not benchmarks.

Key findings

• Adoption is near universal; trust is not. DORA’s 2025 survey of nearly 5,000 technology professionals found 90% use AI at work, with a median of about two hours a day [2][3]. Stack Overflow found more developers distrust AI accuracy (46%) than trust it (33%) [1].
• The share of AI-written code is now an executive talking point. Satya Nadella put it at 20% to 30% of code in Microsoft’s repositories in April 2025 [6]. A year later, Sundar Pichai said 75% of new code at Google is AI-generated and approved by engineers, up from 50% the previous fall [5].
• The clearest wins are big, boring, well-tested jobs. Airbnb migrated about 3,500 test files in six weeks instead of an estimated 1.5 years [10]. Amazon reports over 4,500 developer-years saved on Java upgrades [11].
• Agents are in production, behind human review. Stripe says its in-house agents produce over 1,000 merged pull requests a week, each reviewed by a person [14]. Goldman Sachs is piloting an autonomous coding agent alongside its 12,000 developers, under human supervision [8].
• Speed and stability move together, not always in the right direction. DORA links higher AI adoption to both higher delivery throughput and higher instability [4].

How adoption actually happens

Public accounts show three recurring patterns. They’re not exclusive; most large organizations run all three at once.

1. Buy and mandate. Coinbase gave every engineer enterprise licences for GitHub Copilot and Cursor, set a one-week deadline to onboard, and CEO Brian Armstrong says some of those who didn’t were let go. He also admits the approach was “heavy-handed” and followed up with monthly sessions where teams share how they use AI [9]. Heavy mandates buy speed of adoption, not skill.

2. Build in-house. JPMorgan built its own coding assistant; CIO Lori Beer said it lifted productivity by 10% to 20% for tens of thousands of engineers [7]. Morgan Stanley built DevGen.AI on OpenAI models because commercial tools handled its legacy languages poorly. It reviewed 9 million lines of legacy code and saved an estimated 280,000 hours by turning old code into plain-English specs that developers then rewrite [12].

3. Platform first, then agents. Stripe’s agents run on the same isolated “devboxes” humans use, with the same tests and rule files [14]. DORA’s data points the same way: 90% of organizations have at least one internal platform, and platform quality correlates with getting value from AI [3].

1. 1

   Weeks 1–4

   Pilot with volunteers

   One or two curious teams, enterprise licences, written data rules. Capture a baseline: cycle time, change failure rate, review load.

2. 2

   Months 2–3

   Team rollout with enablement

   Expand team by team. Pair each rollout with a demo session and a shared prompt or rules file. Coinbase's monthly show-and-tell is a cheap model [9].

3. 3

   Months 3–6

   Target one boring, big job

   Pick a migration or upgrade with clear pass/fail tests. This is where Airbnb, Amazon and Morgan Stanley saw their largest gains [10][11][12].

4. 4

   Months 6–9

   Add review and agents behind gates

   AI review as a second reviewer, agents in sandboxes, human approval on every merge [13][14].

5. 5

   Ongoing

   Measure and tune

   Track delivery stability next to throughput, plus acceptance and dismissal rates for AI comments [4][13].

AI-assisted development in practice

Four uses show up again and again. They differ a lot in how easy they are to verify.

What the successful cases share: a crisp definition of done (tests pass, types check, the build is green), automatic retries with feedback, and a human at the gate. Airbnb’s post is the best public recipe: validation steps, retry loops and large context windows beat clever prompting [10].

Agents in production

“Agent” here means a system that takes a task, changes code across files, runs tests and proposes a pull request without step-by-step prompting. Adoption is real but narrow: in the Stack Overflow survey, about 14% of developers use agents daily at work [1].

The guardrails that recur across the public accounts:

• Isolation. Stripe’s agents run in devboxes cut off from production and the internet [14].
• Scoped context and tools. Rule files applied per subdirectory, and a curated set of internal tools exposed through MCP [14].
• Bounded effort. A hard cap on retries, so a stuck agent hands back instead of burning compute [14].
• Human approval. Every example here keeps a person on the merge button [5][8][14].

Measuring impact

This is where public claims are weakest. Company numbers use different definitions, and the best controlled research is humbling.

Two caveats keep this honest. METR itself says the result may not generalize beyond experienced developers in mature codebases they know well [16]. And in February 2026 METR changed its study design, after finding that many developers no longer wanted to work without AI and avoided submitting tasks where AI helps most, which biases results toward slowdown [17]. The durable lesson isn’t “AI makes you slower”. It’s that self-reported speedups are unreliable, and gut feel is not a metric.

What companies actually report, from weakest to strongest evidence:

What this means for your team

You’re probably not Google. You don’t need a 1,000-PR-a-week agent fleet. But the pattern behind the public successes transfers well to a team of 10 or 200: fix the foundations, aim AI at verifiable work, keep a human at the gate, and measure stability alongside speed. DORA calls AI an amplifier: it magnifies strong and weak practices alike [3].

Here’s where I’d push most teams, in order:

1. Get tests and CI solid before scaling AI. Every big win above relied on an automatic definition of done. Without it, AI just produces more unverified code faster.
2. Pick one migration. A framework upgrade, a deprecated API, a test-library switch. It’s where the ROI is clearest and the risk is contained.
3. Mandate learning, not tools. Coinbase’s lesson is that the sharing sessions matter more than the deadline [9].
4. Treat review load as the real bottleneck. If AI doubles PR volume and reviewers stay the same, you’ve moved the queue, not removed it [4].
5. Protect junior learning. DORA warns that AI can skip the productive struggle that builds expertise, and recommends active mentoring [4].