AI Is Not a Tool: How Enterprises Can Embrace Multi-Agent Shift

"AI is not a tool." That is the line Allie K Miller, CEO of Open Machine and advisor to Fortune 500 boards, wants every executive to hear. In a live Work Lab episode from Microsoft's Copilot Summit, host Molly Wood pushed Miller on what has changed in enterprise AI over the last year. The answer is uncomfortable: most leaders are still playing by 2024 rules while the game has moved to multi-agent systems that work autonomously for an hour or more.

Treat AI as a tool, Miller warns, and you will budget it like SaaS, roll it out like a standard platform, and measure it purely on productivity. You will never get to new business lines, org restructuring, or process reinvention, because none of that happens when you roll out a productivity tool. Companies that embrace the shift restructure around it. Companies that do not will count small efficiency gains while competitors build new revenue.

The third shift: from chatbots to autonomous agents

Miller maps the last three years into three distinct shifts. First came late 2022 and into 2023, the era of general chatbot adoption. Then the big reasoning models at the end of 2024. The third shift, landing in just the weeks before this recording, is the multi-agent space: AI that takes on many tasks autonomously at decent reliability for an hour or more.

That is why the C-suite is scared: 74% of CEOs worry they will be fired within two years for handling AI wrong, not for ignoring it. The fear has a price tag attached. When agents run around the clock instead of 9 to 5, companies on frontier models can move from spending thousands of dollars per head per year to thousands per head per day, mostly in engineering. Executives she talks to are suddenly deep in conversations about orchestration, efficiency, and token maxing to keep agent costs sane.

We are, she says, in the thick of experimentation, the time to lean into costs, but not company-wide. The mistake is letting 80,000 people burn thousands of dollars a day while you figure out what works.

Build a frontier unit, not a company-wide rollout

Miller's answer is a double path. The whole company, call it the 80,000, keeps learning: department experiments, Teams channels, town halls. Nobody gets left behind. But a small frontier unit of roughly 60 to 120 people gets a completely different mandate.

That unit puts its foot on the gas the moment a new model drops, testing it immediately. It works in teams of two to eight, where Miller says every best idea she has seen started. It carries much higher budgets per head, sometimes thousands of dollars a day for a small team, because its job is to iterate ten versions of a workflow so the one winner can be rolled out to the other 80,000.

The unit has to be cross functional. Miller is blunt: she does not want a frontier unit full of engineers. She wants research, the weirdos in marketing, sales, legal, finance. A prototype that dies the moment it meets a department is worthless, so every one comes with proof that someone inside already checked it. The unit reports through almost no layers of management, and is often run by a non-technical CMO or CFO who simply got hooked. "It is always a passionate weirdo who is using this in their free time," she says, "and probably suffering from sleep loss."

Strategy lives and dies in the middle

The other leadership gap is the middle layer. A Microsoft study of 1,800 employees found that when managers actively model AI use, teams report a 22-point lift in critical thinking about AI work and a 30-point lift in trust of AI. Miller has done more one-on-one C-suite coaching in the last six months than in the previous three and a half years; plenty of executives still whisper that they have never opened one of these tools.

Her warning to boards: do not offload this to a Chief AI Officer. Hiring one person to own the entire strategy was, in her words, one of the worst ideas of 2023: you are handing the reinvention of your company to a brand new hire who does not know your business. The vision has to be set at the top and modeled by managers all the way down. Otherwise you get the polarity she sees everywhere: strategy set in the boardroom, employees told to learn AI on their own, and a terrified middle stuck between the two.

Measure growth, not just productivity

The most common mistake, Miller says, is measuring AI purely on productivity. It is the natural move, and it is exactly wrong. Employees are not motivated by being asked to produce more output at the same cost. Fear-based leadership backfires: one company threatened staff with "figure this out or you are out" and watched productivity plummet for two months.

Instead, pull the AI money out of the standard line items, the salary bucket and the CIO procurement bucket, and measure it on other KPIs: engineering velocity, net new ideas tested, customer trust, employee fulfillment. Growth matters, not just productivity. Miller still cites McKinsey's 1:5 ratio, one dollar on models and tokens for every five on people, but the five means more than upskilling. It also means incentives: cash rewards in a hackathon where people build agent systems and learn by doing.

And be willing to cannibalize. She tells of a real estate company where one analyst, in his free time, built an AI auditing product for client energy usage. It saved clients 1 to 3%, the company took 50% of the savings as its business model, and the product made millions in its first month. Nobody approved it from the top. The company just had to be ready to move people toward net new business creation instead of protecting old workflows.

How AI work actually gets done: gremlin mode

The single-chat mindset is over. Miller tells of a CMO who declared at a board meeting that AI writing "doesn't have heart." His process had been one request to be funny and heartfelt. He had never iterated. Miller calls the fix gremlin mode: take your ten best performing blog posts, feed each to AI to write a brief, open a fresh thread, write the new piece from that brief, then loop it eight times, improving each pass, and hand the perfect version back as the reference point. Six hours of work, but that is what real results look like. It is also why she now hires for three traits: high agency, a strong sense of wonder, and systems thinking. People managers suddenly have an unfair advantage: they have been managing a temporary digital workforce for years, just with humans.

Context is the new moat

Her biggest operational advice: make your whole company queryable. Connect AI to SharePoint, Outlook, your tools. Record and transcribe meetings. Let executives dictate daily context notes, the decisions, the stress, the human nuances that drive most decisions. One executive she knows dictates five to forty minutes every evening, then asks her AI to map three months of notes for patterns in energy and stress. "We are in the era of context engineering," Miller says.

She runs 34 agents herself, with a chief of staff agent named Simon, six direct reports, each managing three to seven subagents, plus temporary agents spun up as needed. She talks to the system, not the screen. A founder she admires walks 18,000 steps a day holding voice conversations with his lead agent. In one Silicon Valley startup, every desk has a $60 microphone and people dictate all day, with agents inside Teams and Slack, weighing in on threads and teaching each other new skills.

If you are the only thing kicking off AI workflows, you are already behind, because your AI works your eight hours while the machines run 24. Her instruction for the next 30 days: build the context layer, make AI proactive so it prompts you instead of the other way around, and stop typing. Dictation is four times faster, and multimodal, image-and-voice communication is where the leverage is. Treat AI as an operating system rather than a tool.

That last part is the whole argument. Coding benchmark scores went from 17 two years ago to 72 a year ago to 94 on the latest frontier models. If you are still making three-year procurement deals based on what shipped last quarter, you are signing a contract with a different world. The moat of 2026, Miller says, is speed of iteration: small teams, fewer layers, lower bureaucracy, and a budget people can operate freely within. That is how enterprises embrace AI rather than merely adopting it.

Source

Video: "The AI shift most companies didn't see coming | Microsoft" (Work Lab at the Microsoft Copilot Summit, 40:39), Host: Molly Wood. Guest: Allie K Miller, CEO of Open Machine.