The one sentence in an AI agent proposal that is the whole contract
The proposal says the AI agent "connects to your accounting system and handles refunds under $500." Read that sentence twice. It is the whole contract.
Until recently, most AI tools only read things: they summarised emails, drafted replies and searched documents. When one of those got something wrong, a person still decided what to do with it. An agent that acts on a wrong reading issues the refund, changes the customer record or posts the entry to your books itself, using permissions you handed it.
Among companies building these systems, 43% now name security and governance as a serious pain, more than any other problem. Most respondents work at large companies. A small business connecting an agent to its books faces the same question with fewer people to answer it.
Before an agent touches money or records, get four answers. What can it do without a person approving? Where is each action logged, in a form your bookkeeper can read? How do you reverse one? And who is accountable when it does something it was allowed to do and should not have?
That last case deserves the most planning: the refund that was within policy and wrong.