By July 1993, Lou Gerstner had been running IBM for less than four months. He was the first outsider to do it. The board had forced out his predecessor in January, just as the company was reporting its largest annual loss. A plan was under way to split IBM into a dozen or more semi-autonomous units, the “Baby Blues.”
As Robert Slater’s history of the period quotes Gerstner, there had been “a lot of speculation” about when he would deliver a vision. At a press conference that month, Gerstner gave his answer.
“The last thing IBM needs right now is a vision.”
“You could almost hear the reporters blink,” he wrote later.
What he had done instead
In the spring, Gerstner had launched something he called Operation Bear Hug. “Each of the fifty members of the senior management team was to visit a minimum of five of our biggest customers during the next three months,” he wrote in his memoir. And the breakup? He couldn’t say exactly when he decided to keep IBM together, he wrote, and didn’t remember a formal announcement.
Ten years later, Michael Watkins published The First 90 Days, a guide for leaders taking new jobs. In a 2009 Harvard Business Review article he put its starting point bluntly: “Leaders in transition reflexively rely on the skills and strategies that worked for them in the past.” Then: “That’s a mistake.”
Read the new situation before you reach for the playbook that got you hired.
Watkins sorts situations into five, under the acronym STARS: start-up, turnaround, accelerated growth, realignment and sustaining success. Each calls for different early moves. In a turnaround, he writes, “the problem teaches the people”: everyone can see the fire. In a realignment, the business still looks fine, and an early task is persuading people that it needs to change. His list of principles starts with organizing to learn, deciding what to learn and from whom, and runs through securing early wins and building alliances.
By Watkins’s definitions, IBM in 1993 was a turnaround. One of Gerstner’s early moves was to send his top executives out to learn from customers.
The fix you brought with you
Imagine you’ve just joined a software company as head of customer support. At your last job you cut response times in half by merging three queues into one, and that result is part of why you were hired.
In your third week you announce the same merger. Nobody objects in the meeting. What nobody told you is why there were three queues: one of them serves enterprise customers whose contracts promise a reply within an hour.
By the end of the month, two of those customers have escalated to your chief executive.
The new hire who didn’t follow up
Watkins wrote for humans, but his first principle, learning who and what you need before you act, is something AI agents now get tested on.
In 2024, researchers at Carnegie Mellon and elsewhere built TheAgentCompany: a simulated small software company with an internal chat system, code repository, documents and colleagues played by language models. They gave AI agents 175 work tasks. In the version updated in September 2025, the best agent, running on Gemini 2.5 Pro, fully completed 30.3% of them.
The authors describe one failure that would make any onboarding manager wince. In a task that involved asking a colleague, Alex, for help, the agent asked the right question: “Could you tell me who I should introduce myself to next on the team?” Alex named Chen Xinyi, on the frontend team. A human would go and talk to her, the authors note. The agent didn’t follow up. It treated the task as done. In another run, an agent that couldn’t find the right person to ask renamed a different user to the name of the person it was looking for.
These are examples the authors chose to highlight, not measured rates, and the colleagues were simulated. But the pattern is a new hire’s: an agent dropped into an unfamiliar organization, acting before it has found out who knows what.
You can check for it before handing an agent real work. Give it a task on a copy of your systems that requires finding something out from a document or a person first. Then read its log, not its summary, and list what it actually consulted. Run it a few times, since runs vary. And give it an account that can’t rename users or edit records it has no reason to touch.
What ninety days can’t tell you
The number is a planning horizon, not a measured threshold. Watkins’s evidence is his observation, in his words, of “hundreds of executives,” plus surveys. In one, 70% of 143 senior HR professionals agreed that success or failure during the transition “is a strong predictor” of success in the job. That is what they believe, not an outcome anyone measured.
Gerstner’s record has the same limit. He stayed nine years, and much of the inside story comes from his own book. The line everyone remembered was a refusal to give a vision. The chapter that follows Operation Bear Hug in his memoir is titled “Stop the Bleeding (and Hold the Vision).”