Least Agency: How Much Power Should Your AI Agents Get?

Most finance teams think about AI agents in terms of what they can do. The better question is what they are allowed to do. New global guidance published last week gives a clear answer: give each agent the least power it needs, and add more only once it has earned it.

The guidance is written for banks and other financial firms, but the logic fits a 30-person trading company just as well. Here is what it says and how to apply it.

What the FSB published

On 15 October 2026, the Financial Stability Board (FSB) released its final report on sound practices for responsible AI adoption. It sets out 12 sound practices, covering everything from board oversight and data governance to human oversight and third-party risk. The FSB presents them as a toolkit that firms can tailor to their own context and risk profile, rather than a rulebook.

According to Reuters, the FSB strengthened its recommendations on agentic AI after feedback on its June draft. It also dropped an earlier idea of treating AI as synthetic employees. The final report instead says firms should recognise that "AI agents do not reason like humans and may interpret pre-defined goals in unexpected ways."

That one line explains most of what follows. An agent told to "collect overdue cash" may find shortcuts that a sensible credit controller would never take.

Least agency, in plain English

The report takes a familiar IT security idea, least privilege, and applies it to agents as "least agency". Each agent gets only the permissions it needs to do its job. Three points stand out.

Autonomy is a spectrum. The FSB notes that agents range from those doing simple admin tasks to those running whole multi-step workflows. Not every agent needs the same freedom, so firms should agree a risk tolerance for each use at the outset.

Risk grows faster than access. The report warns that the risk from an agent's access to tools and systems can multiply, not just add up, as its permissions expand. An agent that can read invoices is low risk. One that can read invoices, email customers and edit customer records is far riskier than those three powers taken one at a time.

Access can grow in stages. Firms may widen what an agent can reach as they gain confidence that it works safely. Trust is earned, not assumed.

Alongside this, the report lists practical controls. These include banning certain actions outright, and setting clear, risk-based triggers for when an agent must get human approval, such as transactions above a set value or actions that cannot be undone. Others are breaking a goal into steps with a check at each one, and asking agents to summarise what they did and why, then cross-checking those summaries against other records.

What this means for your business

Most SMEs in Qatar, the wider GCC and India are not the direct audience for FSB guidance. But the thinking behind it is sound for any business that lets software act on its money, customers or tax filings.

For example, take a hypothetical Doha distributor rolling out an AI agent for receivables. In month one, it only reads the ledger and drafts reminder emails for a person to send. Once the drafts prove reliable, it sends routine reminders itself, while anything to a customer with a disputed invoice still goes to a person. Changing customer records or offering settlement discounts stays off the table until there is a clear track record.

An Indian SME using an agent to prepare GST reconciliations could follow the same path. First it reads and matches. Then it flags differences for review. It never files anything without a named person signing off.

The aim is not to slow things down. It is to make each new power a deliberate decision, recorded and reviewed, rather than a setting someone ticked during installation.

A checklist for setting agent permissions

• Write a one-line job for each agent. State the outcome it owns and what it must never do.
• Give it an autonomy level. For example: read-only, drafts for approval, acts within limits, or acts freely. Start most finance agents at one of the first two.
• List every system and tool it can reach. Remove anything it does not need for today's job.
• Set approval triggers in numbers. Any payment, credit note or write-off above a value you choose needs a person's sign-off. So does anything you cannot reverse.
• Look at combined access before widening it. When an agent gets a new tool or data source, review everything it can now do together, not just the new piece.
• Keep a log of its work. Record what triggered each action, what the agent looked at and what it did. Spot-check the log against your ledger every month.
• Promote on evidence. Move an agent up a level only after a set period of clean results, and write down who approved the change.

If you use NexCFO for AR collections or AP, give those agents an autonomy level and approval triggers in the same way as any other tool.

Give each AI agent a small job, clear limits and a way to earn more, and it becomes far easier to trust.