AI ROI Starts With Understanding How Work Really Happens

Every leader right now is being asked some version of the same question: Are we ahead or behind on AI? The honest answer, for most organizations, is neither. They are stuck—running AI on top of systems and processes that were never designed to support it, generating outputs that are faster but not materially better and watching the gap between investment and impact refuse to close.

An INSEAD working paper on AI deployment and firm performance found that the companies generating real gains—12 percent more output, 1.9x higher revenue than their peers—weren't just the ones that adopted AI fastest. They were the ones that figured out how to map AI across their production process and redesign the organization around it. That mapping turned out to be a discovery exercise, instead of a technology one. And it is almost entirely human.

AI is only as good as the context it is given. And most enterprises are handing it a picture of the business that is incomplete in a very specific, very fixable way—one that the organizations pulling ahead have already figured out how to address.

When companies first deploy an enterprise context graph—a living map that connects systems, data, logic and relationships across the business—the revelations tend to be immediate and uncomfortable.

In one real-world case, a concept as foundational as "customer" turned out to exist in 27 different places inside a single enterprise, updated by 36 different programs, with only four systems functioning as true sources of record. The context graph surfaced that in minutes. Fixing it manually would have taken months.

But here is the question the graph cannot answer on its own: Why did that happen?

Not how—why. Why did 27 competing definitions survive for years? Who maintained each one, and what were they protecting? What does "customer" mean to the performance marketing team versus the CRM team versus the CFO's office—and why has reconciling those definitions proven so politically difficult?

Those questions matter enormously to the leader in charge. Because the AI agents being built on top of this data will inherit every one of those 27 definitions. Without understanding why they exist, there is no reliable way to resolve them. And without resolving them, personalization at scale, attribution and customer journey optimization all have a structural problem underneath them that no model upgrade will fix.

The context graph finds the problem. Understanding why it exists—and building something AI can actually reason from—requires people who can read an organization the way a strategist reads a market.

The enterprises making meaningful progress on AI right now are not just the ones with better platforms. They are the ones that have invested in understanding how their organizations actually work—as opposed to how they are documented to work.

These are different things. Often very different.

Every organization has two versions of itself: the official one that lives in process documentation, organization charts and system architecture diagrams—and the real one, revealed by observing how decisions actually get made, which data people actually trust, which approvals get bypassed and which teams have built workarounds that have become load-bearing.

Call them desire paths: the informal trails worn into an organization by people finding their own way around the planned walkways. They are invisible in the data. They are obvious to anyone who knows how to observe them.

A context graph built only from documented systems maps the official organization. But the AI agents running on top of it are navigating the real one. That mismatch is where enterprise AI programs stall—and where competitors who close it pull away.

In research conducted with a global pharmaceutical company and a global automotive manufacturer, the work did not start with a data audit. It started with observation.

Common mistake: asking people what they do. Interviews like this produce a description of what employees think they do. Watching what they actually do produces something more useful: a map of which behaviors are serving real functional, social and organizational purposes that the official process doesn't account for.

In the pharmaceutical company, that observation revealed an unspoken rivalry between sales and marketing that was silently distorting how information moved through the organization. No dashboard was tracking it. No KPI flagged it. But it was the reason every new decision-support tool introduced over the previous several years had generated documentation instead of decisions—and why any AI system built without accounting for it would produce exactly the same result.

For a CMO trying to improve marketing effectiveness, that finding is not incidental. It is the finding. It tells you where to invest, what to change and what will determine whether the AI platform you are about to deploy actually shifts performance or becomes another item on the tech stack that nobody fully uses.

At the automotive manufacturer, the same approach showed who people actually listened to, where pushback would come from and what would determine whether the new system caught on or got ignored. Knowing those patterns in advance is the difference between a deployment that changes how the organization operates and one that changes how the organization's organization chart describes itself.

Neither engagement was a data science project. Both produced findings that data science could not.

Most enterprise AI programs treat this as an adoption challenge—get people using the tool, and the value follows. The research suggests it's more specific than that: The real friction is the mapping problem, discovering where and how AI creates value inside your particular production process. That search requires human judgment, not just a rollout plan.

For someone to abandon an existing way of working and genuinely embrace a new one—a new system, a new workflow, a new decision-making process—the appeal of the new approach has to clearly outweigh the comfort of the current one. And that threshold is set by the specific habits, anxieties and informal incentives inside your organization. You cannot read those from a data model. You have to understand them.

The organizations that will win on AI are the ones that figure this out early—that treat human adoption not as a change management footnote but as a strategic input that shapes how the AI system gets built in the first place.

This is what "human in the loop" means in practice—and why it is the operating model that makes the difference.

Publicis Sapient’s enterprise context graph is the shared foundation behind Sapient Bodhi, Slingshot and Sustain. It can ingest, connect and reason across an organization's systems, data and logic at a scale and speed that no team of analysts can match. It surfaces what is documented, maps the relationships between it and gives AI agents the business context they need to do more than complete tasks in isolation.

What it cannot do on its own is encode the third layer—the institutional knowledge that was never written down. The workarounds that have become standard practice. The definitions that everyone uses, and nobody agrees on. The resistance points that will determine whether this deployment is the one that finally shifts performance.

That layer requires people trained in experience research, behavioral analysis and Systems Thinking—people who can observe how an organization actually operates and translate what they find into the structured, queryable context that makes an AI system genuinely useful.

When Publicis Sapient's industry and functional experts work with Bodhi, Slingshot and Sustain, this is the work they bring. The platforms provide the connective infrastructure. The people surface the organizational truth that makes it worth connecting to.

Together, that is what makes the context graph organizationally true, not just technically accurate.

The practical output for enterprise leaders isn't the enterprise context graph itself—it's what the graph makes possible.

When the human layer is in place, the agents built on top of it can do something genuinely different: bring the right analytical capability into the right decision at the right moment, automatically, without requiring a marketer, for example, to know in advance which question to ask or which tool to reach for. That is the difference between AI that sits in your stack and AI that changes how your team operates day to day.

The result isn't marginal efficiency: Campaigns that took weeks can move in days, teams shift from repetitive execution toward strategy and creativity and marketing can demonstrate its impact on pipeline and revenue. And unlike a static tool, the system compounds—every interaction makes the agents more informed, so the longer it runs, the wider the gap grows between organizations that built this foundation and those that didn't.

The map cannot draw itself. In the hands of the right people, it can finally show you where you actually are—and how to get somewhere your competitors aren't yet.