Where AI Pays Off and Where It Just Burns Money

In some processes AI cuts costs sharply and speeds up the work. In others, a company decides everything can be automated and ends up losing money. The model is often equally capable in both cases. What differs is how much of the process runs without checks, and who notices when something goes wrong.

Autonomy without guardrails

An agent that runs on its own, with write access and no human oversight, makes mistakes that do not look like mistakes. It overwrites database records because it "understood" them differently. During a legacy migration it converts data using rules that held for most rows and quietly breaks the historical exceptions.

Every step looks plausible. The error surfaces weeks later, when a customer complains or the accounts do not reconcile. At that point the expensive part is not the tokens. It is repairing the data and the lost trust.

Checks that cost less than the error

Guardrails do not have to be elaborate. A few rules go a long way:

• Permissions. The agent reads by default. Writes go through a narrow interface that allows only specific operations.
• Limits. Attempts, time and budget are capped. When they run out, the agent stops and explains what it could not finish.
• Answer keys. Keep a set of cases with known correct results and run it before deployment and after every change. For a migration, that means reconciling record counts and totals between the old and new system.
• Staged verification. A successful run is not a correct result. Output passes a rules check before it moves on.

A human at the end of the process

A human does not need to review everything. They should sign off where an error is expensive or irreversible: publishing, money, customer data, a migration with no way back. The agent prepares, the human approves.

Track how often the human corrects the output. If corrections keep falling, widen the agent's autonomy in small steps. If they rise, the process is not ready to run on its own.

Before automating, ask one question: what happens when the agent is wrong, and who finds out? If the answer is "nobody, until a customer complains", the guardrails are missing.