Enterprises rushing agents into production without documenting proprietary decision logic risk handing their institutional identity to a handful of large technology vendors, setting up a long-term path to competitive convergence.
Anna Jibgashvili, Corporate VP at New York Life, says the missing layer in most AI programs is digital identity, the risk thresholds, customer-treatment rules, and decision boundaries that remain tribal and undocumented.
She points to years of partial modernization, duplicated sources, and frozen snapshots as the operational reason agents hallucinate, leak information, or contradict business rules at scale.
Enterprise AI programs are increasingly measured by how many agents ship, not by whether the underlying business is ready to supervise them. That framing hides the more consequential risk. When companies push agents into execution before documenting how they actually make decisions, manage risk, and serve customers, they quietly hand over the logic that differentiates them from competitors.
Anna Jibgashvili has spent 15 years working on the enterprise foundations that allow AI to scale safely. Her resume spans New York Life, Guardian Life, Moody's Analytics, S&P Global, and BlackRock, most recently as Corporate Vice President and Data Product Manager for AI and Data. She writes and advises on the disciplines that separate AI-ready enterprises from the rest.
“If you completely give away your digital identity and let third-party tools take over the AI work, you risk losing the differentiating factor that makes your company unique,” says Jibgashvili. Her core warning: institutional identity is the layer most AI programs forget to build.
The unseen identity: Jibgashvili frames digital identity as the proprietary logic that makes one enterprise different from another. "Citibank and JPMorgan have similar data. What differs is how they treat that data, how they calculate risk, how they decide which customer gets an offer. That logic is not digitized in a way an agent can read, and that is the layer being given away." The result is that organizations deploying agents today are often automating execution without encoding the reasoning behind it, a gap that grows more dangerous as autonomy increases.
Three paths forward: Two of them end the same way. Jibgashvili describes three approaches organizations take when scaling AI. The first hands identity entirely to a large vendor whose tools absorb enterprise decision logic. The second keeps a human in the loop as a costly hedge. The third, which she advocates, builds and owns the identity layer internally. Recent global research across 2,000+ senior executives reinforces the stakes. Enterprises that commit to sovereign AI and data platform strategies are generating outsized returns, while those deferring ownership fall further behind. "Two of those paths end the same way," Jibgashvili says. "You lose the differentiating factor, and over time you converge with every other company running the same stack."
The operational reason this matters is the state of enterprise data itself. Years of partial modernization have left most organizations with fragmented ecosystems, duplicated sources, and frozen snapshots that agents now pull from indiscriminately.
Address-change, literally: Jibgashvili offers a concrete example. A customer's current address lives in one system. A marketing database took a snapshot five years ago and never updated. An agent reading both has no way to know which is authoritative. "If the automation is sending sensitive information to the old address, you have a compliance problem you did not anticipate, and you created it with an agent."
The opt-out: The same fragmentation breaks business rules. A customer who told the company never to contact them may be flagged correctly in one table, but invisible in the table the agent actually reads. "The rule exists. It is just not applied across the enterprise. That is not an AI failure. That is a foundations failure."
Data products are necessary. They are not sufficient. Purpose-built data products and structured knowledge bases are the emerging answer. These are curated assets designed to feed agents the most recent, correct, and policy-compliant version of a record. But data products alone do not capture decision logic. "They give the agent clean data," Jibgashvili says. "They do not tell the agent how the company thinks. That second layer is the identity piece, and almost no one is building it."
Governance, redefined: Current governance practice fails agentic AI because it is siloed project by project, she argues. Critical data elements are defined in one initiative and ignored in the next. "We should probably stop calling it governance. The function needs to be redefined. It has to protect decision logic, enforce business rules across every table, and set the boundaries on what an agent is allowed to do. None of that is happening uniformly today." Her own experience trying to stand up governance for a data product was blunt: "It completely failed."
Who owns the work: Roles have not caught up, Jibgashvili notes. Heads of AI are chartered to ship agents, not rebuild the foundation. The subject matter experts who hold institutional knowledge are often the same people being cut. "You cannot digitize what the organization knows if the people who know it are gone," Jibgashvili says. The transformation has to retrain and include them, not replace them. New ownership models and clearly defined accountability structures for AI-driven workflows are the prerequisite, not the afterthought.
The window, in her reading, is short. She estimates enterprises have roughly 12 to 18 months to decide whether they will own their digital identity or rent it from the vendors now absorbing it. The decision will look technical but is really about sovereignty.
"There is still time to do this the right way. Document the identity. Retrain the people. Build the layer the agent reads from, and own it. The companies that skip that work will wake up in a few years and find their differentiation is gone."
