Which Decisions to Automate, and Which to Never Delegate to an Algorithm
As AI tools proliferate, there's a growing temptation to automate not just tasks but decisions. Some of this is valuable. Much of it is outsourcing accountability under the guise of efficiency.
The automation fallacy in decision-making
As AI tools proliferate, there's a growing temptation to automate not just tasks but decisions. Some of this is genuinely valuable. Much of it is a way to avoid the discomfort of uncertainty, outsourcing accountability under the guise of efficiency.
Decisions that should be automated
Certain categories are genuinely suited for automation: rule-based operational decisions with clear thresholds (fraud flagging, inventory reorder, email routing), decisions where speed matters more than nuance, and decisions where human judgment adds noise rather than signal.
The test: if a decision can be fully described by a set of conditions,"if X > threshold, do Y", and the cost of occasional errors is recoverable, automation is appropriate.
Decisions that should never be automated
Strategic decisions involving irreversible commitments, hiring, positioning, pricing strategy, fundraising, belong to a different category. Not because AI can't generate a recommendation, but because accountability cannot be automated. When a key hire fails 8 months in, someone needs to own that decision, learn from it, and carry that knowledge forward.
Automated decisions are orphaned decisions. No one owns them. No one learns from the failures. Over time, this degrades organizational judgment rather than improving it.
The hybrid approach
The most powerful model isn't automation OR human judgment, it's AI-assisted human judgment. Use AI to gather signals, surface patterns, flag anomalies, and stress-test assumptions. Use human judgment to weigh incommensurable values, absorb accountability, and make the final call.
This isn't a compromise, it's genuinely more accurate for complex decisions than either pure automation or unassisted human judgment.
A useful heuristic
Ask two questions about any decision you're considering automating. First: is the decision recoverable? If yes, automation carries low risk. Second: does the decision require carrying forward a lesson? If yes, keep a human in the loop, not to override the AI, but to own the learning.