AI Succession Crisis Highlights Knowledge Transfer Gaps

IT leaders now view AI knowledge as a critical asset, yet many underestimate how difficult it is to transfer that expertise when staff move on.

Why AI Knowledge Resists Documentation

Traditional software relies on documented code, configurations and workflows that can be read and reproduced. Jamie Dearnley, chief technology officer at Resolver, explained that “you can look at the code or documentation to understand how they work and why they behave the way they do.”

AI systems, by contrast, depend on trained models, data sets, prompts and evolving settings. “A lot of what makes an AI system work comes from testing, trial and error, and the experience of the people building it,” Dearnley added. The result is knowledge that lives in the minds of engineers rather than in static files.

Guy Maliar, CTO of Latent AI, noted that with conventional code a new engineer can reconstruct system behavior because it is deterministic. He said AI “breaks this type of model” because outcomes emerge from a mix of model parameters, prompt engineering and countless small judgment calls. “Since the system is non‑deterministic, you can’t just read it and predict what it’ll do,” the expert continued, highlighting the need for hands‑on familiarity.

Building an Inventory to Capture Hidden AI Assets

Both Dearnley and Maliar recommend starting with a full inventory of every AI tool in use, including shadow solutions that teams have adopted without formal approval. “Once you know what exists, identify who owns each system, which business process it supports, which data it relies on, and what level of risk it presents,” Dearnley said. Without this map, governance and knowledge transfer become impossible.

Creating versioned evaluation sets that define “correct” behavior is another step. Maliar suggested keeping decision traces so successors can see why a model was tuned a certain way. “The evals matter most, since they let a new team confirm they haven’t broken anything, even before they fully understand it,” he explained. This approach treats empirical insights as a deliverable rather than a byproduct.

In the middle of these recommendations, it’s worth noting that organizations often underestimate the breadth of their AI footprint. An inventory that reveals dozens of undocumented models can be a wake‑up call, prompting a shift from ad‑hoc usage to a more disciplined, auditable practice.

Testing Over Documentation: The New Transfer Method

According to Sanjay Kukreja, CTO at eClerx, the inventory must also capture the relationship between models, agents, datasets, prompts and the business processes they support. He stresses that “knowing which models … exist, as well as who owns them and what decisions or actions they support” is essential for continuity.

Evaluation harnesses become the cornerstone of knowledge preservation. If a team cannot answer “how do we know this thing is still working?” through an automated test, no amount of paperwork will compensate. “The eval set is the foundation everything else sits on, since it captures the team’s judgment in a form that actually survives people leaving,” Maliar added.

Mitigating the Black‑Box Risk

The “black box” problem describes the inability to trace how inputs become outputs in complex models, especially deep neural networks with billions of interactions. This opacity can generate technical debt that grows until a model update or data drift triggers a failure with no clear cause.

Maliar cautioned that organizations may either avoid touching a fragile system or rebuild it entirely, each option carrying high cost. “If you can’t explain why the system made a decision, good luck defending it to a regulator, a customer, or even a court,” the CTO warned.

To keep knowledge flowing, Jen Clark, managing director at the Eisner Advisory Group, suggests using generative tools like Claude, Copilot or ChatGPT to turn personal expertise into transferable assets. Recorded walkthroughs, shared repositories and collaborative projects can capture insights as they happen, reducing the need for later reconstruction.

What Lies Ahead for AI Knowledge Transfer

Companies that treat AI expertise as a formal deliverable—maintaining inventories, versioned eval sets and decision logs—stand a better chance of avoiding hidden debt.

They need better tools.

As AI systems become more agentic, the pressure to make them explainable will only increase, pushing firms toward systematic testing and transparent documentation.