Northstar Foods: The Atlas Decision
The decks
Elena Park had expected disagreement. That was why she had created five teams.
Hearth & Field, Northstar Foods' largest brand, had been losing relevance for years. Its customers were loyal but aging. Its growth had come from smaller packages, new flavors, and promotions that worked just long enough to become expensive. Park, Northstar's chief growth officer, asked five cross-functional teams to find a new growth platform. She told them they could look beyond the category. She told them not to start with an existing product. She gave them four months.
They had returned with five versions of the same business.
Each proposal combined personalized nutrition, a digital meal-planning service, a subscription offer, and premium shelf-stable food. One team called the concept Daily Table. Another called it Hearth at Home. A third had framed it around parents. A fourth had framed it around adults managing chronic conditions. The research was not identical. The commercial idea was.
Park noticed the pattern after the presentations, not during them. Each team had built a persuasive story. The trouble became obvious only when she asked her chief of staff to strip the branding from the decks and place the five propositions side by side. The consumer need was the same. The product architecture was the same. The route to market was the same. Even the risks were almost the same.
Park called Marcus Bell, the chief information officer, before she called anyone else.
“I have five teams who were supposed to disagree,” she said. “They came back with one answer.”
“Sometimes there is one answer,” Bell replied.
“Then tell me why every team used Atlas to get there.”
A company looking for a different kind of growth
Northstar was not failing. It had $8.4 billion in annual revenue, national distribution, dependable cash flow, and a portfolio of familiar pantry and frozen-food brands. But its advantages were becoming less valuable. Private-label products were better than they had been a decade earlier. Smaller brands had become skilled at serving narrow needs that a large company had traditionally ignored. Northstar could still put a product in nearly every grocery chain in the country. It had become less certain about which product belonged there.
Daniel Ruiz, the chief executive, believed the company had a planning problem before it had a product problem. He had inherited annual planning sessions in which business-unit leaders arrived with settled positions, analysts supplied confirming evidence, and the executive committee chose among proposals that had narrowed months earlier. In Ruiz's first year, Northstar acquired fewer businesses, killed more projects early, and asked more questions before approving capital. That was progress, but it was slow.
Bell proposed Atlas in late 2025. The company built a secure environment around a large language model and connected it to approved internal material: category research, consumer interviews, financial history, retailer reports, and product documents. The tool could summarize, compare, draft, challenge, and search. Bell was careful about the language. Atlas was not an analyst, he said. It was a system that made analysts harder to satisfy.
The strategy group adopted it first. The change was immediate. In acquisition reviews, teams asked Atlas to identify assumptions repeated across the deal model, management presentation, and customer research. In the annual plan, it was used to produce competing market scenarios and to list the conditions under which each business unit's preferred investment would fail. The work was not glamorous. It was useful. Meetings became less performative because executives had to answer objections that appeared before the meeting began.
Ruiz saw enough to expand access. By June 2026, marketing, R&D, sales, finance, and operations all had Atlas licenses. Seventy-three percent of eligible employees were using the system weekly. A proposed snack acquisition was paused after the strategy team used Atlas to link customer churn data, retailer concentration, and a distribution risk that the original diligence work had treated separately. The decision may have been made without the tool, Bell said later. The company simply would have made it with less confidence.
Park was an early supporter. Her teams were drowning in research. Atlas reduced the hours spent locating material and gave junior staff a way to ask a basic question without waiting for a senior manager to be free. She also liked that it could produce a credible counterargument. In marketing, people often fell in love with a consumer story before they had tested it. Atlas made the first draft less precious.
That had been the theory behind the Hearth & Field project. Each team began with different consumers, different category data, and different fieldwork. But they had one thing in common: Atlas was in the room from the first day.
The meeting
Bell brought a prompt log to Park's office. The teams had not used identical instructions. They had used similar phrases: “unmet consumer need,” “scalable growth platform,” “personalized convenience,” “high-margin category adjacency.” They had drawn on overlapping data sets. Bell pointed out that the brief itself was likely to lead people toward health, convenience, and digital services. Those were the market's loudest signals.
“The model did not impose a subscription business,” he said. “It responded to the information it was given.”
Park read the prompts again. Atlas had been asked to list the most promising opportunities, then to develop them, then to improve their commercial logic. It had not been asked what the teams were missing. Nor had anyone required the teams to identify an idea they rejected before asking the system to refine their preferred one.
Priya Nair, head of R&D, joined the discussion later that day. She had worked in food innovation long enough to distrust neat explanations. “The model is not the problem,” she said. “The timing is.”
Nair argued that early-stage innovation depended on awkward work: half-formed observations, contradictory customer behavior, and conversations that did not yet add up to a proposition. Atlas was excellent once a team had a question. It was less helpful when the team had not yet worked out which question mattered. “We gave it a problem before we had earned the problem,” she said.
That comment stayed with Park. Northstar had used Atlas to make strategy broader and more skeptical. In innovation, it may have used the same system to make the first answer more attractive than the second.
What Park found
Park's team reviewed the research. One study in Science Advances was particularly close to what she had seen. People using generative AI produced work that evaluators judged more creative than work produced without it. The group as a whole, however, produced less diverse work. Individual quality went up. Collective novelty went down.
That distinction was not intuitive to every executive. It was easy to see why an employee liked Atlas: the first draft arrived faster, sounded more complete, and exposed fewer obvious gaps. It was harder to see what happened across a portfolio of employees using the same model. A company could become more productive at making good ideas look finished while becoming less capable of finding ideas that did not already resemble one another.
A BCG experiment on product innovation pointed in the same direction. Participants using GPT-4 created ideas that were less diverse than those produced without the tool. A later meta-analysis of 28 studies found a large negative effect on idea diversity in human-AI collaboration, despite no meaningful reduction in average creative quality. Park did not take these findings as proof that Atlas caused Northstar's problem. She saw them as a reason to stop treating the five similar decks as coincidence.
At the same time, she could not ignore what Atlas had done for strategy. Harvard Business Review had described the technology's value in strategic work as an expansion of the number of options a company could develop and examine. That was Northstar's experience. Atlas had made it easier to see alternatives and harder to hide from inconvenient facts. Park did not want to lose that advantage because product teams had used the tool carelessly.
The decision
David Lee, the CFO, asked Park for a recommendation before the October investment committee. Northstar planned to spend $38 million on Atlas in 2027, including data infrastructure, licenses, security, training, and internal development. Lee had two concerns. The first was financial: how could Northstar tell whether Atlas was creating differentiated growth rather than merely accelerating a familiar pipeline? The second was operational: how would the company enforce a different use model without creating a policy that employees ignored?
Bell wanted to improve training and prompt design, but keep access broad. Nair wanted Atlas out of early concept work. Grant, president of Hearth & Field, wanted to move forward with one of the five proposals. Delaying the decision, he argued, would not make the company more original. It would simply leave a declining brand without a growth plan.
Park drafted a third path. Teams would keep access to Atlas. But the sequence of work would change. The first phase of an innovation brief would be human: customer observation, category mapping, problem definition, and an initial concept set. Teams would document the ideas they rejected as well as the ideas they pursued. Atlas would enter after the first concept review, when it could search for evidence, identify blind spots, test assumptions, simulate retailer objections, or improve a prototype.
For strategy, the rule would be different. Atlas would be used early and aggressively to expand scenarios, challenge plans, and surface alternatives. The company would not pretend that strategy and innovation were interchangeable forms of knowledge work.
Ruiz read the proposal on the Sunday before the executive committee meeting. He called Park that evening.
“You are asking us to say that the same tool should be used more in one part of the company and less in another,” he said.
“I am asking us to decide what we want it to do,” Park said.
“And if the five teams came back with the same idea because it really is the right one?”
Park looked again at the decks on her desk. “Then it should survive a process that gives it more than four versions of itself to beat.”
Research Response: The Strategic Value of AI Is Not the Same as Its Innovation Value
The Northstar case is built around two propositions that are often discussed separately. The first is that AI can improve strategy-making. The second is that AI can weaken the diversity of thought from which innovation emerges. Taken together, they point to a more demanding management question: not whether a company should use AI, but where in a decision process it should be allowed to shape the work.
AI changes the economics of strategic debate
Strategy work has a familiar weakness. A management team cannot examine every plausible option, and it rarely has the time or political freedom to challenge its own preferred answer thoroughly. AI changes that constraint. It can generate alternative scenarios, trace assumptions across planning documents, compare market positions, and prepare a first-pass critique of an investment case. The result is not better strategy by default. It is a lower cost of being less certain before capital is committed.
That matters because corporate planning is usually narrow for human reasons, not informational ones. Teams settle early. People avoid re-litigating the work of influential colleagues. A plan becomes harder to challenge after it has acquired a financial model, a sponsor, and a calendar date. An AI system cannot remove these dynamics, but it can supply material that makes them harder to ignore. This is why the case shows Atlas working well in Northstar's strategy process. Its value lies in creating more work for management judgment, not in replacing management judgment.
The implication for senior leaders is practical. They should use AI to widen the field before they decide: generate contrary cases, specify what would have to be true for an investment to fail, find assumptions shared by apparently independent plans, and ask which customer or competitor evidence is missing. These are tasks that reward breadth, comparison, and structured skepticism.
Innovation begins before the answer is visible
Innovation is different. At its earliest stage, the work is not primarily evaluation. It is interpretation. A team is trying to notice something that does not yet fit neatly into the category language, customer segments, or financial assumptions it already uses. The first useful thought may be incomplete, awkward, or commercially implausible. That is not a defect. It is often the condition from which a distinct proposition develops.
Generative AI is trained to produce plausible continuations. Plausibility is useful later in the process, when a company needs to test whether an idea can be developed, sold, or scaled. It is less useful when every team turns to the same system to define the initial problem. At that moment, the model's statistical center becomes a quiet organizing force. The team receives a coherent answer quickly. Coherence can feel like insight. It is not always insight.
The research on collective novelty should therefore not be read as an argument for removing AI from creative work. It is an argument for sequence. Human teams should form an initial view of the customer problem before the model is asked to elaborate it. They should use AI to challenge an idea, not to make agreement feel inevitable. In the Northstar case, the problem was not that Atlas helped five teams develop their concepts. The problem was that it entered before the teams had produced sufficiently different concepts to develop.
The operating model that follows
| Work stage | Primary role for AI | Management discipline |
|---|---|---|
| Strategic framing | Generate scenarios; identify assumptions; produce counterarguments | Require an explicit response to rejected alternatives |
| Early innovation | Limited research support; no model-led concept generation | Document human observations, competing problem definitions, and discarded concepts |
| Concept development | Stress-test, refine, research, prototype, and simulate objections | Preserve evidence of the original independent concept set |
| Investment decision | Analyze risk, economics, dependencies, and execution scenarios | Assign human accountability for the recommendation |
This is not a technology architecture. It is a management architecture. It requires different access rules, review practices, and measures for different kinds of work. Strategy teams should be rewarded for the range and quality of alternatives examined. Innovation teams should be measured not only on speed and launch volume, but on whether their concept pipeline remains meaningfully varied. The measure does not need to be perfect at the outset. It needs to make a previously invisible risk discussable.
There is also a regulatory reason to make the distinction explicit. The European Commission's AI Act timeline records that significant applicable provisions, including transparency requirements, began applying on August 2, 2026. Companies will increasingly need to identify where AI is used, who is responsible for its output, and what controls govern that use. A vague enterprise policy will be difficult to defend operationally as well as legally.
The question Northstar leaves open
Northstar's executives do not have to decide whether Atlas is good or bad. That would be the wrong decision. They have to decide whether the company is willing to give the system the same role in every part of the business merely because it is convenient to do so.
The harder choice is to accept that strategic analysis and innovation require different conditions. One benefits from a wider, more disciplined examination of alternatives. The other depends on protecting difference long enough for it to become an alternative worth examining.
