Box’s COO on why you should put content at the heart of your AI strategy

We’ve had a rush of reports recently that attempt to quantify the outcomes of enterprise AI investments — and content management provider Box has now entered the game with its own survey, The State of AI in the Enterprise 2025.

This commissioned report surveyed over 1,300 enterprise decision-makers across industries and geographies — providing an interesting snapshot of AI adoption, the differences between winners and losers, and the value of thinking big. As a TLDR; almost 70% of respondents believe that AI will reshape their sector and 90% intend to increase AI budgets over the next year to respond.

But what does it all really mean?

With a background in both cloud (at Google) and management consulting (at McKinsey), Box COO Olivia Nottebohm now runs Box’s customer-facing operations — giving her an interesting vantage point for tracking the operational changes wrought by AI.

I caught up with her to pick through the details of the survey — and to discuss how Box intends to help organizations navigate the challenges it reveals.

Content, content, everywhere, nor any doc to think

I start by asking Nottebohm to summarize how Box will help its customers navigate the transition to an AI-enabled future — and without hesitation she zeroes in on making unstructured data useful for AI:

I think the biggest unlock is the unlock of unstructured data.

As anyone who’s spent more than five minutes trapped in a corporate intranet will know, unstructured data accounts for most of an enterprise’s knowledge — and yet remains the hardest to use. Nottebohm explains:

The majority of enterprise data is unstructured — PDFs, photos, videos, documents, spreadsheets — and this is also the hardest to access and use.

The survey backs up her point. While 94% of companies say that they are already experimenting with AI — and 87% say they have started piloting agents — most remain stuck in early-stage use cases. Box argues that this is because the enterprise content needed to fuel higher-value applications is scattered across messy PDFs, images, spreadsheets and other unstructured formats.

A recent MIT study reached a similar conclusion, finding that 95% of generative AI pilots failed to deliver measurable ROI because they lacked integration with existing business data, including this long tail of messy unstructured data — something Nottebohm notes quickly becomes apparent:

As organizations have started to roll out copilots for example — thinking it would be this big unlock — they’ve discovered the same problems. There is still this messiness. And even beyond structure they quickly realize that maybe governance, permissions, and security also really do matter.

That leads to a second problem, since while 74% of survey respondents cited data privacy and security as a top concern — and 73% listed compliance as a critical criterion — only 24% say they have mature data governance.

Nottebohm views this gap as a prime opportunity for Box:

In an agentic world, content matters. And so this is where we really feel like Box plays a key role — managing the unstructured data needed to drive intelligence and workflows in the enterprise.

This idea of explicitly cleaning up, securing, and governing AI content echoes the concept I recently discussed of building a 'philosophical infrastructure' — the essential task of curating and engineering context in ways that align the behaviour of AI with the organization’s values and goals. In this sense, Nottebohm’s advocacy of data curation and governance could be viewed as a strategic act of self-definition rather than simply a technical clean up job — and it seems clear that Box is keen to position itself as the partner to help underpin that act.

But clear content is not the only thing the survey suggests is driving outcomes for the most successful AI adopters.

All adopters are equal, but some are more equal than others

Beyond the need for well-structured content, Box’s survey also shows a sharp divide between firms achieving high ROI versus those just treading water — with a successful subset already reporting productivity gains of 37% (vs lower and later expectations of 30% gains in three years’ time by the average enterprise).

In particular, the survey suggests that the highest achievers tend to invest more (an average of 25% of IT budgets vs 10%), deploy more targeted solutions (task-specific rather than generic), and be more discerning about model choices (with 83% of leaders saying that selecting the right model for each purpose is critical vs 42% for the rest).

Nottebohm sees the same patterns in practice with successful Box clients:

It’s pretty basic but I’ve observed that successful customers choose a workflow that already exists and where they can quickly imagine how they would just automate specific steps that will remove delays and increase efficiency.

As an example, Nottebohm points to an insurance customer that has automated specific steps in a workflow processing 10,000 claims per week — something with very specific KPIs that made it easy to prove ROI.

It's really just choosing a tangible workflow where people can see the ROI — which then just builds momentum.

Nottebohm contrasts this approach with ‘going wide’ with AI from day one by, for example, rolling out general-purpose solutions:

I think the thousand flowers blooming is a lovely idea because everyone gets to feel it. But then there's huge disappointment when there aren't obvious and measurable leaps in productivity.

This finding is echoed by McKinsey research which found that broad, horizontal deployments — like copilots — delivered only diffuse gains, while vertical, workflow-specific applications were much more likely to generate measurable ROI.

It’s also unsurprising that the survey suggests a strong correlation between specific problem selection and nuanced model choices. When implementing specific solutions, organizations naturally have the context necessary to align model cost and capability with the problem to be solved — and this likely compounds the advantages of practical AI task selection by ‘right-sizing’ the technology used to implement it.

Colleagues, collaborators, co-workers, lend me your ideas

While targeted implementation and solid model choice are key routes to ROI, Nottebohm also explains that framing also plays a critical part in scaling adoption:

I think what people are underestimating is change management. As McKinsey consultants we were always told that execution eats strategy for lunch. So you have to have a story that brings people along if you want to scale.

Nottebohm argues that the best way to frame AI is as a ‘capability expander’ — a tool that doesn’t just make existing ways of working more efficient but rather enables people do things that were not previously possible. Beyond being a philosophical preference for growth over efficiency, Nottebohm also argues that this way of thinking encourages employees to support AI adoption:

If you announce plans to replace 20% of the workforce, it’s not a recipe for success. People won’t embrace agents if they believe they’re a replacement.

Instead Nottebohm emphasises the need to upskill employees to use AI while also celebrating their achievements when they do. As an example, she cites one Box customer who created an internal ’AI fair’ where employees demo the agents they’ve built — and are rewarded for their ideas:

People present the agents they've created at a walk-through fair — and it's not focused on efficiency or cutting headcount but on sharing innovation. So people share ideas knowing they won’t be replaced — and innovation happens everywhere, not just in IT.

Nottebohm explains that the same pattern is visible inside Box, where product teams have built a Box-trained agent to handle 20,000 customer success queries a year on their behalf — leaving them free to focus on more interesting work.

This need to upskill and encourage employees is clearly visible in the survey, with 80% addressing AI skills gaps by upskilling, and 87% already experimenting with agents — creating demand for employees to learn to build and use them.

But the message from Nottebohm is simple — large scale AI adoption won’t happen without proper change management. And the foundation is to frame AI as a tool which leads to better work, not redundancy.

A journey of a thousand workflows begins with a single step

While the survey results reflect enterprise ambitions, Nottebohm is also keen to bring us back down to earth. She uses the example of Box’s five-stage maturity model — spanning simple human assistance to fully autonomous agents — to clarify where organizations are today:

When I’m in a room with CIOs and ask where they are, the majority raise their hand at stage one. They’re still trying to figure it out.

This reality contrasts sharply with the survey results, where ambition appears to run far ahead of practice. 41% of organizations say they are piloting fully autonomous operations in select domains, 60% expect AI transformation within two years, and around a quarter of all business processes are projected to be AI-augmented within three years. So, while self-reported expectations are racing forward, most enterprises still appear to have the hard work of scaling beyond pilots and generic use cases ahead of them. As Nottebohm explains:

This reality check is a good reminder. We have all these conversations. We listen to the news. We read, listen to podcasts. But at the end of the day, CIOs are still trying to figure this out and so we're trying to meet our customers where they are.

Nottebohm goes on to explain that this does not mean that progress is impossible, however, citing one Box customer who has moved from manual loan applications to a fully automated workflow, and another that used an agent to help onboard 50,000 employees following a merger.

You have to get people together, agree on how it’s going to work, and then put some professional services hours in. But when it’s done, it feels automagical.

Despite this optimism, Box is clear that the higher rungs of its adoption curve — encompassing agents from the “high autonomy” quadrants I’ve described in an agentic taxonomy — remain largely aspirational, at least for now. And while Nottebohm expects progress to happen in these higher levels sooner than we might think, she also argues that giving enterprises the tools to quickly move through the most immediately practical parts of the curve is what really matters right now.

For Box, therefore, its maturity curve is not just a vision of the future but rather an attempt to ground customers in reality — while also providing a plausible map of what could come next. It’s also a reminder that this progression depends on a solid content foundation — precisely the layer Box is betting its future on.

My take

Combining the survey data with Nottebohm’s reflections provides a useful snapshot of where enterprises believe they are with AI right now. And while adoption is claimed to be near-universal — i.e. 94% of organizations claim to be experimenting already — the differences that separate winners from laggards are becoming clearer. And these differences matter when almost 70% of those surveyed believe — rightly or wrongly — that AI is going to transform the basis of competition in their sector.

Setting aside greater spend — which in my experience is no guarantee of greater success — the report suggests that winners are more intentional. They focus on targeted workflows, a plurality of model choices, and explicit change management.

And as Nottebohm rightly points out — it’s not really that complicated when you think about it like that.

But when 41% of organizations say they are piloting autonomy — and most CIOs still admit they are at stage one — it shows how quickly ambition can run ahead of capability. For enterprise leaders this is a real risk, potentially leading them to invest in advanced use cases without the governance, content discipline, or change management frameworks that AI leaders have put in place to make them work.

Within this bigger picture — and true to its heritage — Box wants to position itself as the partner who can help you put this scaffolding in place. By pitching itself as a central, secure repository for curated content, it hopes to become a single hub around which agents, automation, and collaboration can be layered. In one way, it’s a stance that speaks directly to the overwhelming number of companies who expressed concerns about data privacy, security, and compliance in the survey — but it’s also a bet that the most valuable AI use cases will be built around centralized content repositories rather than federated systems that leverage data wherever it happens to lie. Agree or not, it is at least a coherent philosophy in an AI marketplace that’s often more smoke than fire.

And the resulting pragmatism about its customers’ likely progress is refreshing. Rather than promising instant leaps to stage five autonomy, Box is emphasizing content strategy, data discipline, and incremental expansion — positioning itself to meet its customers where they are in order to help them move deliberately from retrieval, to discrete workflows, to orchestrated processes, and — eventually — beyond.

For buyers, therefore, the real question is not how fast they can reach full agentic autonomy — but whether they believe the content scaffolding offered by Box will help them sustainably make the climb. And that answer will ultimately test Box’s bold bet that the future of enterprise AI will be built on a content-first foundation.