Agent drift is the gradual divergence between what an AI agent produces and what the Sales Engineer intended. Each individual output might look acceptable in isolation. But over a series of interactions, the agent's understanding of the deal, the customer relationship, and the SE's strategy slowly shifts away from reality — creating commitments, setting expectations, and shaping conversations in ways the SE never approved.

How Agent Drift Happens

Drift doesn't happen because the AI is broken. It happens because the AI lacks access to context that changes constantly:

Context decay. The agent was set up with initial context about the deal — the customer's needs, the product configuration, the timeline. But that context is a snapshot. The customer's priorities shifted after an internal meeting. The SE had a sidebar conversation that changed the approach. The agent doesn't know, and nobody updated it.

Commitment accumulation. The agent sends a follow-up email referencing a deliverable. The customer responds with a clarification. The agent sends another message addressing the clarification but also introducing a new commitment. Over five exchanges, the agent has created a chain of commitments that the SE never explicitly approved — each individually small, collectively significant.

Tone and relationship mismatch. The agent may escalate urgency in its language, make the relationship feel more transactional, or use phrasing the SE would never use with that particular customer. These aren't factual errors — they're relational drift.

Why Drift Is Hard to Catch

The challenge with agent drift is that each output passes a basic quality check. The email is well-written. The commitment is technically fulfillable. The tone is professional. But the SE, if they had seen the full sequence, would say "that's not how I'd handle this deal."

By the time the SE reviews what the agent has been doing, the customer's expectations have already been shaped by five or ten messages that don't reflect the SE's strategy. Unwinding those expectations is harder than if the SE had just handled it themselves.

Detecting and Preventing Drift

Drift detection requires comparing agent outputs against the deal's evolving context — not just checking individual messages for quality. Key approaches include:

Commitment reconciliation: regularly comparing agent-originated commitments against human-originated commitments on the same deal to identify conflicts or divergence.

Intent alignment checks: periodically surfacing agent outputs for SE review — not every message, but at strategic checkpoints where the deal context is likely to have shifted.

Cumulative impact monitoring: tracking the full sequence of agent interactions on a deal, not just individual outputs, to detect patterns of gradual divergence.

The principle is straightforward: AI agents should amplify the SE's judgment, not replace it. Drift detection ensures they stay aligned.