For most of the last decade, the advantage in software came down to one thing: data. Companies invested heavily in building and leveraging proprietary datasets because they understood that information accumulated over years made their product defensible. Then generative A.I. arrived, and the narrative shifted. Suddenly, the assumption became that whichever company had access to the best models would dominate its industry. Organizations raced to put A.I. into the hands of every team. That assumption hasn’t aged well.
Models improve fast, and they spread just as quickly. Whatever edge a frontier model gives you, your competitor has likely adopted the same model, or something comparable, within weeks. And so we’re back to data. The information your business accumulates over the years is an asset nobody else has. The companies that succeed are the ones able to most effectively leverage their proprietary data for use through A.I. That leaves all companies facing two questions that often pull in opposite directions:
How can we connect A.I. to our data?
How can we protect our data from A.I. misuse?
Connect first, govern later
The first wave of A.I. was about acceleration and answering the first question. Engineering teams embraced coding assistants. Marketing found tools to accelerate content creation. Sales experimented with A.I. for research and outreach. Finance, legal and customer support each began identifying opportunities to automate repetitive work. Individual teams selected the tools that best fit their workflows and connected the internal systems they needed, often before a company-wide A.I. strategy took shape.
The instinct behind all of it was correct. A frontier model with no knowledge of your business is a commodity if largely interchangeable with every other frontier model. Connect that same model to customer histories, internal documentation, proprietary workflows and systems of record, however, and it becomes something competitors can’t replicate.
Thousands of individually rational decisions have added up to an environment that’s increasingly difficult to manage. Different departments rely on different A.I. applications, employees connect those tools to internal systems and new capabilities appear almost weekly. Before long, A.I. is interacting with customer data, code, financial systems and business-critical workflows across dozens of disconnected products.
Most IT leaders can produce an inventory of laptops, cloud infrastructure and licensed software. Far fewer can produce a basic inventory of their agents. Which tools have access to sensitive company information? Which employees have connected external A.I. services to internal systems? Which A.I. applications have become essential to the business, and which are quietly introducing unnecessary risk?
The first wave of A.I. succeeded in driving adoption. What it didn’t create was visibility, without which governance becomes virtually impossible. That’s the state most companies are in as they confront question two: How can we protect our data from A.I.?
What does data protection actually mean?
Protecting enterprise data from A.I. isn’t one problem. It’s two. The first is an external risk. Attackers exploit the connections organizations deliberately create between A.I. systems and sensitive information. The external risk is malicious exploits. An A.I. assistant that reads emails, documents and web pages can be manipulated through prompt injection, where malicious instructions hidden inside otherwise ordinary content cause the model to perform unintended actions. Ask the agent to summarize an inbound email, and the email itself may attempt to convince the agent to expose confidential information or take unauthorized actions. Traditional perimeter security has no answer for this because nothing crossed the perimeter. Your data walked out through a connection you approved.
The second risk is internal—your own organization outrunning its rules. Employees are pasting source code, strategy documents, financial data and customer information into A.I. systems every day. When data moves through a prompt, there’s often no audit trail, no reliable record of where information went and no way to answer customers or regulators who ask whether sensitive data was exposed. Most organizations already have policies governing how data should be handled. A.I. adoption has simply outpaced their ability to enforce them.
One risk comes from outside and one from inside, but they both point to the same underlying issue: organizations cannot protect what they cannot see.
Two questions, one answer
This brings us back to where we started: we need to connect A.I. to our data, and protect our data from A.I. Most companies are treating these as separate programs, and the programs are fighting each other. The A.I. enablement side keeps adding complexity in the name of speed. The security side finds out afterward. That approach doesn’t scale.
Closing the gap means knowing which A.I. systems are connected, what data they can access, what actions they’re authorized to perform and how those permissions are enforced over time. Without that visibility, security becomes reactive, policy becomes difficult to enforce and scaling A.I. across the business becomes a leap of faith.
Creating visibility does more than reduce risk. It allows companies to give A.I. access to the systems that create competitive advantage without handing it everything else. Consider both risks through one lens. The answer to prompt injection isn’t disconnecting the sales assistant from the CRM; it’s scoping the connection so a compromised agent can read this quarter’s pipeline but can’t export the customer database. The answer to employees pasting source code into chatbots isn’t a ban; it’s an approved tool, connected to the same systems, that logs what it touches. In each case, you want the safe path to also be the easier one. That visibility is what keeps enablement and control in balance.
The next enterprise software layer
Every major platform shift has ended the same way. Cloud computing created cloud management platforms. SaaS sprawl turned identity and access management into a foundational first-class enterprise system. Mobile spawned device management as a category. Each time, the layer that lets enterprises see and control the new technology ended up mattering as much as the technology itself. A.I. is reaching that point now, faster than any of its predecessors, because A.I. is the first technology that acts.
The winners of the A.I. age won’t be decided by model access alone. Everyone will have access to capable models. They won’t be decided by data alone either. Plenty of companies are sitting on data they can’t safely use. The companies that pull ahead will be those that connected A.I. to everything worth connecting and could still answer, at any moment, what it’s touching, what it’s allowed to do and where their data went. Everyone is going to use A.I. Not everyone is going to govern it. That’s what will separate the leaders from the laggards.
