AI agents are starting to handle money, so who watches the market owner's bills?
IBM's new identity controls for AI agents point to a practical question for commercial landlords: who signs off when AI touches contracts, bills and collections?

AI agents are moving from writing copy to handling contracts, bills and collections. IBM's new identity and control capabilities for agents make the question concrete for small commercial landlords and market operators: who is accountable when the agent gets it wrong, and who signs when it touches money.
Why AI agents suddenly need an "identity"
According to PR Newswire Asia, IBM has introduced identity and control capabilities for AI agents. The core logic is to give each agent a governed identity, extend key management to AI agents, enforce least-privilege access, and revoke permissions immediately when an agent attempts an unauthorised operation.
The line in that announcement cuts close to the bone: companies are deploying AI agents faster than they can verify them.
Put in operating terms: when you let AI chase a rent payment, in what capacity is it sending that notice? Can it change the bill amount along the way? If something goes wrong, can the log show who did it? These questions did not matter when AI was helping write copy. Once it touches contracts, bills and receipts, they become mandatory.
What this means for market operators
Running a commercial complex or a farm-produce wholesale market is, at bottom, the repeated confirmation of a pile of small tasks: which unit is up for renewal, which tenant is behind on payment, whose utility bill is due this month, which leasing lead has gone three days without follow-up. These tasks used to depend on people watching them, and when people get busy, things slip.
This is exactly the stretch that AI can take over. But how it takes over matters. SHEYU AIMARKET (舍予AI商业租赁系统) works this way: an AI steward handles leasing follow-up, bill collection and meter reading and billing, and every write operation passes through a two-gate "preview then confirm" flow with a full audit trail. In other words, the AI first shows you what it intends to do; it executes only after you confirm, and every step leaves a record.
What this mechanism solves is not whether AI can do the work, but whether you dare sign off while it does.
The threshold for AI in an operating system was never whether it can do the job, but whether you dare accept the result once it has.
How do markets without a tech team keep up?
This is the real situation for most small and mid-sized market operators: they know AI is useful, but they have no IT department to build a permission system or an audit log.
One pragmatic choice is to use a system that has already built these mechanisms into the product, rather than assembling one from scratch. SHEYU AIMARKET runs contracts, bills, commissions and meter reading on a database transaction engine, so a lease takes effect and the bill is issued, and a lease termination settles the account. In plain terms, the numbers reconcile and cannot be quietly altered, so reconciliation no longer means checking three sets of ledgers against each other.
A boundary needs stating: this system is a leasing operations platform. It manages asset and lease control, leasing CRM, contracts and bills, and property and IoT matters. It does not decide whether a given tenant should get a rent reduction, and it does not replace sitting down with a tenant. AI provides preview and execution; decisions remain with people.
What to prepare now
There is no need to rush into a system. Do one thing first: write down clearly which tasks can go to AI and which must carry a human signature.
For example, collection notices can be drafted by AI and sent after your confirmation; changes to bill amounts must be made by a person; leasing leads can be flagged by AI, with a person making the follow-up call. Once that line is clear, check whether a tool matches it. If it does not, change tools; if it does, then discuss going live.
According to Anthropic's official guidance, most coding and knowledge work starts at Opus 5.5 medium effort, and only the hardest long-horizon agent work moves up a tier. Even vendors talk about using models in tiers; operating scenarios should use AI in tiers too: routine tasks to the machine, signatures to people.
There is no standard answer here, but one thing is certain: fixing permissions after an AI has chased the wrong bill costs far more than thinking it through now.