Andrew Siemer, Founder & CEO of Inventive Group, firefighter, veteran. We are a software product team that gets stuff done, brilliantly.getty​The AI productivity numbers everyone is talking about are real. 10x is achievable. Based on what I've seen, some teams are pushing well beyond that.Almost no one gets there by accident. The teams consistently capturing that leverage share something most companies don't yet: the discipline that makes any high-performing engineering team work.After 25 years inside enterprise software and two years running AI coding tools and agentic teams in production, the path is specific.What '10x' Actually Looks Like In ProductionThe wide range in reported AI productivity gains isn't random. A 2025 study of 300 engineers found AI adoption cut PR review cycle time by nearly a third. Our engagements with full guardrails land in that zone, with some pushing 5x to 10x once we layer agent orchestration on well-tested codebases.A METR randomized controlled trial found that experienced developers using AI tools without scaffolding actually slowed down by 19%. Same tools, different operating model. That gap isn't about AI being overhyped. It's about what separates the teams capturing leverage from the teams that aren't.Why The Difference Is Discipline, Not ToolingThe teams getting consistent 3x gains or more share three habits.The first is brutal clarity about what "done" means. AI agents mark work complete the moment the code might compile. Human engineers know "done" means tested, integrated, security-scanned and shipped to staging. Teams winning with AI build that definition into their orchestration explicitly. Every checkpoint that a senior engineer would enforce becomes a gate that the AI has to clear.The second is treating AI agents like new hires, not tools. A CTO friend running 22 iterations of an AI scrum team in a single week found his most valuable insight wasn't a model upgrade. He renamed an agent from "product manager" to "scrum master," and the team's behavior changed completely. The LLM had absorbed each role's definition from the internet, and the label programmed the behavior. Getting leverage means designing AI agent roles with the same care you'd use for an org chart.The third is an experienced engineer in the loop. A 2026 ICSE longitudinal study found that developers using AI assistants produced substantially more code and deleted significantly more. A 2025 study of Cursor adoption found a transient velocity boost alongside persistent increases in static-analysis warnings. AI moves fast. Someone with judgment has to catch what it confidently gets wrong.A Story That Proves The PointLast quarter, I watched an AI agent implement a feature in a container. It couldn't reach the database—a temporary issue. Rather than flag the problem, the agent self-assigned a new task: "I will install PostgreSQL over here and run migrations." I fell out of my seat with a "what-the-hell" moment and reset its thought process. Nobody had asked for another database. Had I not been following the agent's reasoning, I wouldn't have caught it until later.That's exactly the failure mode that experienced engineering judgment catches before it ships. Twenty years of systems thinking—review gates, audit trails, architectural guardrails—is the same discipline that keeps AI efficient, fast and contained. Without it, speed becomes a liability. With it, speed compounds.Where The Real Leverage LivesWe've measured our own engagements over a year of daily use across greenfield and brownfield codebases. The pattern is consistent.With proper foundations (CI/CD, testing, documentation, senior review), AI can deliver a reliable 3x to 5x gain for anyone. Layer in full agent orchestration with defined roles and human oversight at key gates, and the ceiling moves to 5x to 10x. We've built and tested everything from simple OpenClaw setups to a fully custom harness built with Claude Agent SDK in two hours, not because AI is magic, but because two decades of understanding architecture let us execute rather than explore.That's the pattern across every team we've watched succeed. Leverage comes from pairing AI's speed with experienced engineering judgment. Companies that try to capture the speed without the discipline end up in the METR slowdown zone. Companies that invest in foundations capture the upside the headlines actually promise, and then some.What Engineering Leaders Should Do NowIf you're trying to get more out of AI than your team currently is, the path is well-defined. Don't start with a usage mandate—those have already produced cautionary stories. One major cloud provider's internal pressure to adopt AI tooling was implicated in a multi-hour production outage in late 2025, where AI-generated code reached production without proper review. The lesson isn't that AI is unsafe. It's that pushing adoption faster than your review infrastructure can support produces the failures headlines warn about.Start with the foundations. The good news: AI can build most of them for you, if someone with engineering judgment directs it. A senior engineer walking into a brownfield codebase follows a predictable sequence: install CI/CD, add tests where risk is highest, require pull requests and enforce code review standards. None of this is novel. All of it is a prerequisite. AI agents can build this scaffolding faster than humans can, provided someone is specifying what to build and reviewing what gets produced.The real bottleneck isn't your codebase. It's whether you have anyone senior enough to set the standard. Companies that bought AI coding tools, hoping to compensate for missing engineering leadership, ended up in the METR slowdown zone. AI amplifies what exists. If no one's setting the standard, AI happily produces code to no standard at all.Separate the use cases. AI is dramatically effective for boilerplate, scaffolding and isolated tasks. It needs more discipline in integration-heavy work in mature systems. Knowing the difference turns raw speed into predictable gains.Protect your experienced engineers. They're the multiplier on every AI tool you adopt. Their judgment turns AI's speed into compounding value.The teams getting 5x to 10x out of AI today aren't moving the fastest. They're the ones who built the foundation first, paired AI with experienced judgment and let the leverage compound. That path is open to any team willing to walk it.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?