The CEO of an early-stage B2B marketplace builds production integrations via voice notes on his subway commute. A 20-person agriculture tech venture has paused all hiring because commercial large language models (LLMs) now handle more than half of tasks in nearly every business function. At a mature development, security, and operations (DevSecOps) platform, nontechnical staff use AI tools to fix small bugs and rename features, bypassing engineering entirely.
These are not isolated experiments. Over the past several months, we have met with the tech and business leaders at 15 AI-centric companies—spanning continents, industries, and stages of development, from four-person start-ups to established global platforms—to learn what it takes to make AI capabilities truly deliver. We expected to hear 15 different stories. Instead, this diverse group of businesses, independent of one another, seemed to converge on the same fundamental insights about what it takes to successfully place AI at the center of the organization. That convergence is the story.
Earlier this year, we published a strategic playbook on AI-first venture building. This article builds on that playbook—moving beyond frameworks and theoretical strategies to identify the ground-level practices that differentiate winning companies from those that continue to struggle to get real results from their AI efforts. And those struggles are real: McKinsey’s most recent Global Survey on AI finds that while 88 percent of organizations now use AI in at least one business function, only around 1 percent consider themselves fully mature and roughly two-thirds have yet to scale AI beyond isolated pilots.
What do winning enterprises know that their peers do not? What they seem to share is a sense of possibility. Rather than deploying gen AI and agentic AI to simply cut costs and boost productivity, they approach the technology as a multiplier of both ambition and the capabilities required to realize those ambitions. We boiled down what we learned from these leaders into seven essential truths—hard-earned insights that collectively constitute an operating system for getting the most out of AI.
AI is not a tool, it’s a teammate
The real value of AI isn’t doing the same work faster. It’s the ability to amplify the efforts of individuals with agents that function as genuine team members.
When we began our research late last year, leaders tended to describe using AI tools as personal productivity copilots. Months later, they were speaking in terms of having genuine agentic coworkers, with their own names, Slack handles, shared task boards—and the ability to execute tasks autonomously, 24/7. The COO of a Series D fintech, for example, runs a multiagent system to vet new ideas. Any employee can submit one-sentence product ideas via Slack. After initial vetting by a product manager, ten specialized agents work on the idea simultaneously—covering issues such as product definition, back-end feasibility, revenue recognition, and legal compliance—and deliver a comprehensive set of product requirements within hours.
The CEO of a Series A marketplace has an entire staff of personal agents—including an executive assistant agent that replies to emails and manages his calendar, a chief of staff agent that records meetings and autonomously generates and circulates next steps, and an analyst agent that provides real-time data insights from company dashboards. The virtual staff, he says, has boosted his capacity and effectiveness fivefold. We heard similar stories from other leaders we interviewed.
The structural pattern is that jobs are being redefined as collaborations between humans and agents, with work reallocated to amplify what teams are willing to attempt. Engineers now do design and customer research. Nontechnical staff open merge requests and ship internal experiments. A four-person sustainability venture can serve 20 enterprise clients, producing compliance reports in minutes that previously would have taken a law firm weeks. A Series A marketplace went from managing 50 deals per adviser to running 3,000 simultaneously—not by hiring 60 times the staff, but by building agent analysts that handle the volume work while allowing humans to focus on the high-stakes, trust-heavy conversations that truly drive the business. These are not efficiency gains. They are fundamentally different business models, made possible because AI expanded what teams dare to take on.
Potential pitfalls
The shadow side of treating agents as team members is structural dependency. “What externalities are we creating by going into this agentic world that we haven’t foreseen?” asks the agtech venture’s CEO. “By turning more and more over to the agents, what crises are we potentially perpetuating?” The company mitigates that risk with deliberate role design. Humans are channeled into the problems that agents genuinely cannot solve—such as novel scientific judgment and partner relationships—and escalation paths are created for when agents reach the limits of their competence. The human–agent partnership, the CEO says, is not about “replacing humans with AI. It’s about being surgical about where human talent is irreplaceable.”
The leadership move
Stop measuring AI by hours saved. Measure by what the business is now willing to attempt. Give agents names, responsibilities, and escalation paths. Map how jobs are being reordered across your organization and redesign roles to reflect the new human–agent allocation.
Know what to build and what to buy
Build only what makes you truly distinctive. As for everything else, how far you go is a function of your own comfort level.
The first build-versus-buy decision is the easy one. Every AI-first company we spoke with applies the same test to its core proprietary capability: Does this help create a defensible advantage—based on our company’s data, expertise, or intellectual property (IP)—that an off-the-shelf tool cannot replicate? If the answer is yes, build it. The agtech venture builds its own crop-breeding models, IP-landscape scanners, and self-improving R&D agents in-house. The deep-tech materials-discovery company builds its scientific discovery agent on proprietary data. “Our unique selling proposition is the data and the opinion we expose to the world,” the vice president of engineering at a climate intelligence venture told us. “AI can’t replace that.”
The trickier decision is what to do about everything else—the tools and agents that run internal operations. Until very recently, the answer was simple: Buy it all. Specialist vendors delivered polish that internal teams could match and modern tools shipped with native integrations. Many of the companies we interviewed still operate this way and are happy with the results. At the seed-stage AI company we interviewed, for instance, every sales call is recorded, transcripts are automatically posted to a shared workspace, and a weekly digest goes to the team—a process stitched together from off-the-shelf tools, with no custom building. The CEO can ask the connected systems where things stand with any client and get an instant update from across the team. It’s a common ethos at AI-native ventures. “We don’t buy what defines us,” as the tech lead at a digital health scale-up puts it. “We buy what frees us.”
More technically adept organizations, meanwhile, can now draw the line in a radically different place. Thanks to coding agents such as Claude Code and Cursor and similar agent-builder platforms, internal teams can spin up dashboards, automate workflows, and create tailored agents in hours rather than months—sometimes with no engineers in the loop. The agtech chief strategy officer, for example, says that build it yourself is now the default. “I personally think SaaS [software as a service] is dead,” he says. “Trying to integrate a tool takes longer than it takes to build a tool.”
Potential pitfalls
Building custom agents and tools may be inexpensive. Maintaining them is not. It is tempting to let every team and department spin up its own bespoke productivity tool—but the maintenance bill will eventually arrive. Cheap and easy to build today does not equal cheap and easy to own tomorrow.
The leadership move
Build what makes you distinctive. For the operations stack, start with off-the-shelf tools and make AI-native interfaces, integrability, and “swapability” nonnegotiable. Audit the stack quarterly and be prepared to switch out anything that fails those tests. And keep in mind that only the most technical and most disciplined teams should attempt to build everything in-house today—and even they should do so with eyes open about maintenance costs.
Your model isn’t the bottleneck—accessing your tribal knowledge is
Many teams focus on which AI model to run. The ones pulling ahead focus on what their agents can find, and they invest in the knowledge layer that makes the difference.
When you ask an agent a question and it cannot answer, it’s not necessarily the model’s fault. It could be that the answer was never written down or is located somewhere the model cannot reach. In other words, the ceiling on your AI is set by your knowledge hygiene, not your model choice. “It isn’t an AI problem—it’s a knowledge management problem,” says the operations director at an energy tech platform. “AI just makes it visible.” The real gap isn’t technology as much as it is the knowledge infrastructure: unrecorded meetings, unstructured data, and expertise trapped in people’s heads.
For those who get it right, the payoff can be considerable. At the energy tech platform, for example, a knowledge agent that indexes code repositories, pages on the digital workspace Notion, and Slack conversations enables new hires to get up to speed in a matter of days. At the seed-stage AI company, the automated sales pipeline stores all data in a “queryable” knowledge layer. When someone asks, “How far along are we with this specific lead?” they get instant deal context pulled from months of accumulated conversations. “Whenever we have questions, we can just ask our internal knowledge base,” says the venture’s CEO. The feedback loop between product and operations becomes a competitive advantage.
The agtech CEO, for his part, challenges the orthodoxy that you need a single data lake before AI can work: “A lot of people get wrapped around the axle of ‘you need a single source of truth.’ But the thing that makes data useful is that humans are touching it and updating it all the time.” Some people write in Notion, others use spreadsheets, others live in Slack or on video calls. Rather than forcing uniformity, the company builds lightweight connectors that make all of that data, wherever it lives, queryable by AI. “If somebody’s using a tool, just meet them where they are,” the CEO says.
Potential pitfalls
Your knowledge layer can rot faster than you think—and when it does, agents tend to break. When data gets old and stale, agents will confidently serve up outdated answers, which rapidly erodes user trust. “An agent doesn’t know what is the latest source of truth and what is an outdated document from a year ago,” warns the COO of a Series D fintech venture. It’s a lesson that the tech lead at a digital health company learned the hard way. “If I could change one thing, I’d invest earlier in structuring our content,” he says. “Fragmented data slows down flow and frustrates teams.” The fix is architectural: build connectors to where activity already happens—meetings, Slack threads, working documents—so the knowledge layer stays fresh without anyone maintaining it manually.
The leadership move
Record every meeting. Transcribe automatically. Route outputs to a shared knowledge layer. Make your messaging platform crawlable and connect it to your knowledge backbone so conversations become queryable context. Meet people where they already work and build connectors to capture the knowledge they naturally create. AI is only as smart as what it can find.
Design for the swap, not the stack
The winning architecture is not a monolithic platform. It is a thin governance layer that connects best-in-class components and keeps them interchangeable.
AI-first companies converge on a shared architectural principle: assemble best-in-class tools, wire them into a governed shared layer, and build only the thin connective tissue that makes context secure and queryable. At a global technology platform, for example, engineers query internal wikis, continuous integration/continuous deployment pipelines, and ticketing systems through AI agents connected via model context protocol servers. “I can ask, ‘What services are impacted by this feature?’ and it shows me everything,” the platform’s tech lead told us. That discovery workflow, which reduces hours of manual search to minutes of conversational query, is an architecture outcome.
Model agnosticism is a nonnegotiable design principle. “Everything we build needs to be done in such a way that we can easily rip out a model here and put something else in,” says the Series D fintech COO. The company uses a multimodel gateway, starting with premium models for new workflows, then migrating to more economical options once the pattern is proven. Given how fast the frontier shifts, locking in is a strategic liability.
Potential pitfalls
A connected, composable architecture is an immense source of advantage, but it’s an equally large attack surface. To guard against that risk, the AI biotech company runs a three-tier security model: public data to commercial LLMs, sensitive data to providers with zero-data-retention agreements, and core IP processed only on-premises. “The company would be in mortal danger if this information leaked,” the company’s CEO told us. Security tiering is a design decision, not a bolt-on.
The leadership move
Standardize a governance backbone that includes identity management, permissions, security tiers, and data classification. Connect best-in-class tools around it with lightweight connectors. Build only the thin integration layer that makes context secure and governed. Design for the swap, not the stack.
Trust precedes autonomy
Companies build trust in AI systems through progressive autonomy: AI generates, humans judge, and the system earns more freedom only when it deserves it.
With AI agents, the freedom to operate autonomously is a privilege that must be earned. At the early-stage sustainability venture, for example, the team runs every process manually until repetition demands automation, at which point substeps are automated one at a time until the full process is complete. “Automate slowly,” says the company’s founder. “Do it manually until the pain forces automation. That’s when you know the workflow is ready.” Moving methodically like this surfaces edge cases that an automated pipeline would miss, teaching the team where human judgment is genuinely required versus where it has become habit. “AI is the perfect middle-to-middle tool,” the founder says. “Humans still need to start and finish.”
Leading companies also identify their quality ceiling—and then hold the line, while treating any benchmark as a current read, not a fixed rule. One health tech founder, for example, reports being offered an 85 to 90 percent success rate on an agentic solution, which he rejected outright. “We need to operate at 99.999 precision; in healthcare, you can’t go ‘80 percent is good enough,’” he says. In software development and back-office workflows, by contrast, most of the leaders we talked to put the ceiling at about 70 to 80 percent—noting that in most cases, AI can get you there reliably; beyond that, human judgment is where real value concentrates. “It’s not just using AI,” said the head of engineering at a company that develops AI for contact centers. “It’s knowing when not to.”
Potential pitfalls
Autonomy without guardrails backfires fast, and failure often arrives in the places you least expect. When negotiating a contract with a customer, for example, execs at the AI-for-doctors company ran email chains and proposed changes through an AI agent built to read and answer emails, which immediately crafted this response: “All fine re: your changes. Let’s go.” In fact, the proposed terms required additional negotiation. Fortunately, a human worker flagged the error before the response was sent. “We turned that off immediately,” the CEO says. The lesson: AI should suggest, not act, until trust is earned in that specific context.
The leadership move
Define where human approval is mandatory and encode it into workflows today. Measure full cycle time (generation plus review), not just generation speed; gains at the generation step, after all, can vanish in the review step. Build feedback loops that let the system earn more autonomy over time. The goal is not permanent human in the loop; it is building the trust that makes autonomy safe.
Centralize the platform; decentralize the tasks
No centralized AI department can drive transformation. What works is when platform teams govern the infrastructure and business teams solve their own problems on top of it.
The AI operations director at the large energy tech platform made a deliberate choice early on. “I decided that I am not the expert,” she says. The idea that one person can understand and optimize AI use across multiple business functions, she learned, tends to collapse the moment that workflows change faster than prescriptions can keep up. The tech lead at a digital health scale-up reached the same conclusion: “It makes zero sense to have one person oversee ten business units they don’t understand.” Instead, each business unit at the venture decides how AI will support its goals. A small AI guild connects leaders for pattern sharing, but ownership stays with the teams.
At many of the organizations we talked to, the technology team owns governed access to models, composable architecture, security guardrails, and connective infrastructure. The business teams own the problems to be solved and the strategies needed to solve them. That separation lets the platform stay current while business teams move at their own pace.
Leading companies also enable anyone to become a builder, with the idea that the people driving the most impact are those who apply AI to their own workflow friction. The health tech CEO gives everyone one to two hours daily for free experimentation. The Series A marketplace CEO assures staffers, “You can figure it out. You don’t have to ask engineering to build you something.” McKinsey research has found that workers were three times more likely than leaders expected to report that AI helps them perform 30 percent or more of their daily tasks. The barrier to scaling, in other words, is not employee reluctance; it is leaders not enabling fast enough.
Potential pitfalls
Decentralization without a platform is chaos, but centralization without specificity can lead to a different kind of failure. The head of engineering at a Series C scale-up describes the former: “Sometimes people use unapproved automation tools, and then we have to close those accounts until a proper approval is done.” A staff engineer at a mature technology platform describes the latter kind of failure: “I believe we moved too broadly too early, trying to build an ‘agent that could do everything’ rather than focusing on specific use cases first. We should have shipped smaller, faster.” These two distinct failure modes have the same answer: a real platform, governed centrally, with clearly bounded scope, that lets business teams build freely within it.
The leadership move
Appoint a platform owner with explicit authority over governance, architecture, and security guardrails, and document which decisions belong to business teams before the platform goes live. Build lightweight sharing rituals (a dedicated showcase channel, weekly demo slots, a shared prompt library) so local wins compound into reusable patterns without bureaucratic overhead. Measure platform adoption at two levels: activity signals (token usage, tool adoption rates) are useful leading indicators; the real metric is what business teams built and deployed on top of the platform.
Adoption is a flywheel, not a rollout
Successful adoption isn’t a rollout with a deadline. It’s a flywheel with four reinforcing layers: role modeling, sharebacks, measurement, and hiring.
Most organizations have deployed AI tools. Far fewer have built the culture and muscle to use them at scale. The companies that close the gap build a flywheel.
The first layer of that flywheel is role modeling: leaders go first, visibly. The Series A marketplace CEO builds production integrations himself and makes AI fluency part of performance reviews. “If anyone here is not tinkering,” he says, “they’re probably cooked.” The Series D fintech COO blocks out Friday afternoons for company-wide hacking sessions; even the CEO has been forced to build his own agent. “If you are not spending significant time thinking about how you scale yourself,” the COO says, “then you’re not up for the job.” The takeaway: When leaders use AI and share the results, they give everyone else permission to experiment.
The second layer is conviction through sharebacks. None of the companies we looked at relies on mandates alone; instead, the mechanism for genuine adoption is social proof. The energy tech platform runs AI guild talks, town halls featuring success stories, and monthly AI challenges for non-engineering teams. “You can’t sit on people’s shoulders and tell them to use it,” the platform’s operations director says. “If people share success stories, that’s what works.” At the seed-stage sustainability company, effective prompts and workflows are wrapped into reusable custom GPTs and distributed to the team. “Whenever someone builds a great prompt or workflow, we share it,” the company’s founder told us.
The third layer is reinforcement through measurement. “We even have a leaderboard of which department uses the most AI,” the DevSecOps staff engineer notes. The lesson: What gets measured gets repeated.
The fourth layer is hiring for the right DNA. The Series D fintech COO probes every candidate on AI experimentation. If a candidate describes using AI only for summaries, he says, “I literally think, ‘OK, then our interview is over.’” The health tech CEO describes the bar as “more a willingness check than a skills check.” The deep-tech CEO applies the same standard across functions, even in nontechnical roles: “If someone said, ‘I’ve never used AI before and I do everything manually,’ that’s pretty much a no-go.”
This four-layer flywheel constitutes a self-reinforcing system that compounds over time. When leaders role model AI use, they give teams permission to experiment. When experiments succeed and are shared, they create social proof that accelerates adoption across the organization. Measurement makes that adoption visible and creates accountability. And hiring for AI fluency ensures that every new joiner raises the baseline, making all three prior layers more effective. The flywheel stalls when any single layer is missing.
Potential pitfalls
Two failure modes bracket the flywheel. The first is exhaustion. “People can actually get burned out because there’s so much opportunity to do so many things now,” the deep-tech CEO says. The second is forgetting that change, especially rapid and radical change, can be disorienting. “For most people, it’s quite scary,” says the operations director of the energy platform. “Be kind about it. Shift to using AI to solve problems. Ask people: ‘What’s your biggest pain point? What do you really hate doing?’” Solve the pain first; adoption follows.
The leadership move
Measure AI adoption by department, publish leaderboards, and tie AI fluency to performance reviews and hiring criteria. Audit which of the four flywheel layers (role modeling, sharebacks, measurement, hiring) is weakest in your organization. That is where to invest first. Without all four, AI tools get adopted in pockets and never compound into organizational capability.
The operating model gap is widening
These seven truths are more than a list of best practices. They are an agentic system—and one that meshes with McKinsey’s Rewired playbook for AI transformation: Treating agents as teammates (Truth 1) immediately raises the question of what to build versus buy (Truth 2). Building requires getting the knowledge layer right (Truth 3), which depends on a composable, governed architecture (Truth 4). Operating safely requires trust built incrementally (Truth 5). Scaling requires the right organizational design (Truth 6). Sustaining it requires adoption as a cultural flywheel, not an IT rollout (Truth 7).
The gap between companies running on this system and companies standing up their first chatbot is not a technology gap. McKinsey calls this “the largest organizational paradigm shift since the industrial and digital revolutions.” Significantly, the 15 companies in our research arrived at the same truths independently, under different competitive pressures, in different industries, at different scales. That convergence suggests that these seven principles are not theoretical constructs but the operating system that emerges when AI-first teams figure out, under real conditions, what actually works. The companies that delay are not only standing still; they are ceding ground to competitors who are compounding their advantages with every cycle.
Related Articles
