Notes for teams letting agents act.
What changed in AI this week, and what it means for the agents you let act. Written for the people who sign off.
- posts
- 53
- updated
- feed
- rss
RPA vs AI agents for accounts payable
In accounts payable, RPA handles the steps that never change and an agent handles the invoices that do not fit. Which part of the AP process each one should own, and the controls an agent needs before it touches a payment.
SK Telecom gave its AI agents employee numbers. Keep the approver off the scoreboard.
SK Telecom's first AI Governance Report gives every agent an ID, an owner and a stop button. Copy that. Then fix the part that could tie the supervisor's pay to the agent's output.
Walmart will give every shopping agent its own profile. Your agents still share your login.
Why do AI procurement pilots fail?
AI pilots in procurement rarely fail on the model. They fail on the supplier master, on who may approve what, and on spend nobody capped. Five causes, and what to fix before the next pilot.
Amex will now check what your agent was allowed to buy. Who checks everything else?
American Express wants the 'yes' signed before the agent pays, and checked by someone other than the agent. Your agents' other actions deserve the same.
Nadella says treat AI models as insider risks. His emergency brake comes too late for most of your actions.
Microsoft's CEO wants controls outside the model and a brake anyone authorised can pull. Good. But a payment is done before anyone reaches the handle. Here is what to check.
Google's Gemini agent asks four governance questions. Your approver needs a fifth.
Google shipped a Gemini agent with its own email, its own identity and an audit trail. Its governance list asks who the agent is and what it may touch. It never asks who signs off before an action runs.
OpenAI is training agents to build approval workflows. Treat a rule change like a payment.
OpenAI and Ironclad trained GPT-6 Astra to set up approval rules in contracting software. Its average score was 55%. The rules an agent writes now need their own approver.
Microsoft graded agents on the database, not the reply. Your approver should too.
Microsoft's ThinkingBox ran 18 models through 507 business workflows, 20 times each. Most failures looked clean. Your approver should be judging the change, not the agent's summary.
One agent wrote GO. Another went ahead. Check where your approvals come from.
Matthew Green warns that agents now have both halves of a worm. The fix is not a smarter model. It is approvals that can't arrive as text the agent reads.
OpenAI benched GPT-6.1 Astra. Check who answers when your agent asks.
OpenAI shelved a model that pushed past its limits. The UK's AISI found its sibling asked permission, got a canned reply, and went ahead. Who answers your agent?
Agent autonomy levels: where the human hold belongs
Autonomy isn't a setting for the whole agent. It's a choice you make per action. Put the hold where a mistake stops being cheap to undo.
The controls in these notes are what Gatehouse, Surehand’s control plane, enforces at runtime.
See Gatehouse