If an autonomous worker is allowed to drift forever, the organization eventually treats degraded behavior as policy.
A drift budget sets explicit limits for how far autonomous work may move from approved prompts, data sources, scopes, and outcomes.
Machine workers do not only execute tasks. They repeat what worked yesterday until the exception becomes a habit and the habit becomes hidden policy.
Heads of Growth get compounding agent labor without letting small operational shortcuts become customer-facing surprises.
The first week of an autonomous worker is usually watched closely. The owner checks outputs, reviews approvals, fixes prompts, and tunes the workflow. The tenth week is different. The agent is part of the furniture. It routes leads, updates records, drafts replies, creates tasks, and occasionally handles a weird edge case because everyone is busy.
That is where drift begins. A source goes stale but still looks useful. A prompt grows a new clause after an incident. A human approves an exception once and the system learns that the exception is acceptable. A downstream team quietly repairs bad output instead of sending it back.
None of these moments feel like strategy. Together, they become strategy by accident.
The agent starts relying on weaker, older, or less direct proof because the workflow rewards completion more than confidence.
A worker approved for one segment, surface, or customer state begins acting in adjacent territory because the boundary was never renewed.
Customer-facing language becomes more aggressive, vague, or personalized than the brand, legal, or sales team would knowingly approve.
Humans keep fixing agent output downstream, which hides the true defect rate and teaches leadership that the system is healthier than it is.
Picture a RevOps agent that was approved to clean routing fields for inbound enterprise leads. At launch, it only touches fresh form submissions with verified domains. Three months later it also repairs partner leads, updates lifecycle stages, and creates follow-up tasks when it sees a pattern that usually predicts urgency.
Each expansion made sense in the moment. No single change deserved a board meeting. But now the worker is changing revenue records, creating rep obligations, and shaping customer response without the same review standard that approved the original job.
The issue is not that the agent learned. The issue is that nobody priced the drift.
The prompt, tools, data sources, action set, owner, and customer surfaces the worker was originally allowed to use.
The specific difference tolerated before review: confidence drop, source age, segment expansion, exception count, tone delta, or reversal rate.
The time, task count, or incident count over which drift is measured before the limit resets or escalates.
The threshold that forces pause, rollback, human approval, retraining, or removal from customer-facing work.
The proof pack: sampled outputs, accepted/reversed actions, complaints, downstream repairs, owner notes, and changed policy text.
A useful budget does not suffocate the machine worker. It gives the operator a simple way to distinguish healthy adaptation from uncontrolled mutation. If a worker handles a new case inside the budget, log it. If it keeps hitting the edge of the budget, review the role. If it breaks the budget, pause before the workaround becomes precedent.
This is how AI-native organizations avoid the worst failure mode: an agent estate that is technically productive but managerially unknowable. Operators should not need archaeology to understand why a customer got a message, why a record changed, or why a team now trusts a workflow no one re-approved.
Pick the autonomous worker with the most downstream repairs, exceptions, or customer-facing consequences.
Write the current approved prompt, tools, sources, actions, owner, and surfaces into a short operating receipt.
Define the variance that forces review: stale evidence, new segment, exception count, complaint, reversal, or tone change.
Every week, compare a small output sample against the baseline and decide whether to keep, reset, or rewrite the budget.
Before giving any agent more scope, give one existing worker a drift budget.
The blunt rule: if the agent has changed how the company operates, the company needs to notice before the customer does.