Picture your sales manager reading an agent's summary on Friday afternoon: "Updated the Tan & Co opportunity, sent the follow-up, moved the deal to negotiation." Now picture finding out on Monday that the update was only suggested, the email was sitting in drafts, and the stage never moved.
Nobody did anything wrong in that story. The agent just reported three different kinds of output in the same voice. For SMEs putting agents into their ERP and CRM, that is the gap I would close first.
What the vendors are shipping
Two announcements from this week, and one from October, make this practical rather than theoretical.
Salesforce said on 4 November that Agentforce Coworker plans the steps of a job, coordinates its agents and workflows, and brings back "results and recommended actions for review".[2] Users can now review results and approve actions from their phone or desktop, and Salesforce says approval and audit capabilities are available for supported operations.[2] Scheduled tasks, which run recurring work while you are away, are expected in December.[2]
Infor's distribution update, reported the same day, describes agents that act inside the ERP itself. Its Quoting Agent can convert a quote into an order, its Purchasing Agent generates supplier follow-up messages, and the GenAI Assistant can manage order hold codes.[3] On the finance side, bank statement reconciliation leaves unmatched items for people to resolve.[3]
Microsoft is the most explicit about the line between suggesting and doing. In October, its Dynamics 365 team wrote that its CRM skill definitions "distinguish between analysis, recommendations, proposed changes, and completed actions".[1] The "reconcile my opportunity" skill proposes field-level updates, "draft customer reply" prepares a response for review, and a stage readiness assessment "is not itself a stage change".[1]
Then Microsoft turned that into a test for IT teams: "A proposed update should not appear as a saved change, and a draft should not appear as a sent message."[1] For recurring work, it says teams should also confirm which actions are authorised, when review is required, and how unsuccessful or unconfirmed actions are reported.[1]
That one sentence is a test any SME can borrow, whatever platform it runs on, including one it built itself.
Four labels every agent output should carry
Here is how I would apply it in an SME running sales, service and finance on ERP and CRM.
Analysis. The agent looked at the data and is telling you what it found. "Three invoices from this customer are past 60 days." Nothing changed in any system.
Draft. The agent wrote something a person will send or post. A supplier chaser, a customer reply, a quote. It has not left the building.
Proposed change. The agent wants to change a record and is waiting for someone with authority to say yes. A new delivery date on an order, a credit hold, a deal stage.
Done. The agent changed a record or sent something, and the system confirms it. This label needs evidence: the record ID, the time and who approved it, if anyone did.
The danger sits between the last two. A proposed change reported as done means your team plans around something that never happened. A done action reported as a proposal is worse, because nobody goes back to check it.
Where this bites in ERP and CRM
Take the Infor examples above. A quote converted into an order commits stock and price to a customer. A hold code on an order stops or releases a shipment. A supplier follow-up tells a vendor what you expect from them.[3] These are not research tasks. Each one should land in exactly one of the four labels, and your team should be able to tell which at a glance.
In practice, I would sort the agent's actions into three groups before switching anything on:
- Act and report. Low-risk, easy to reverse. Tagging a lead source, logging a call note. The agent does it and reports "done" with the record ID.
- Propose and wait. Anything that touches money, stock, credit or a customer promise. The agent prepares it as a proposed change and a named person approves.
- Draft only. Anything that leaves the company under your name. The agent drafts, a person sends.
Then check the reporting, not just the permissions. Run a few test jobs and compare what the agent says it did with what the system shows. If its summary ever says "updated" or "sent" when the record shows nothing, fix that before you scale. That is Microsoft's validation test in plain terms, and it costs an afternoon.
Why the wording matters as much as the permissions
Most of the governance talk around agents is about access: what the agent can touch. Access matters. But your people make decisions from the agent's summary, not from the audit log. If the summary blurs "I suggest" and "I did", good permissions will not save you.
Singapore's Parliament touched the same nerve on 4 November. Tin Pei Ling proposed a national registry where anyone could check an agent's identity and what it is authorised to do.[4] Senior Minister of State Jasmin Lau said the government would work with industry on what works in practice and assess whether a national registry is needed.[4] Her closing line is worth keeping: "Agents can act for us, but they cannot answer for us."[4]
That is the point. Someone in your company answers for every action your agent takes. Make sure that person can read, in one line, whether the agent drafted it, proposed it, or already did it.
A small exercise for this week
Pick one agent or automation already running in your sales, service or finance process. Take its last ten outputs and label each one: analysis, draft, proposed change, or done. Then ask two questions:
- Could the person reading it tell which label applied without opening the system?
- For every "done", can you point to the record that proves it?
If the answer to either is no, you have found your first fix. It is usually a change to how the agent writes its summary, not a new tool.
Sources
[1] https://www.microsoft.com/en-us/dynamics-365/blog/it-professional/2026/10/08/bring-microsoft-dynamics-365-into-the-flow-of-work-with-30-new-crm-skills — Bring Microsoft Dynamics 365 into the flow of work with 30 new CRM skills (Microsoft Dynamics 365 Blog, 8 Oct 2026)
[2] https://www.salesforce.com/news/stories/agentforce-coworker-surpasses-185k-users — Salesforce's Agentforce Coworker Usage Soars 21x in Three Months (Salesforce, 4 Nov 2026)
[3] https://erpnews.com/infor-details-ai-agents-for-distribution-erp-warehouse-picking-and-finance — Infor Details AI Agents for Distribution ERP, Warehouse Picking and Finance (ERP News, 4 Nov 2026)
[4] https://www.straitstimes.com/tech/singapore-explores-adapting-web-domain-name-system-to-track-ai-agents-assesses-need-for-national — S'pore exploring adapting web domain name system to track AI agents, need for national registry (The Straits Times, 4 Nov 2026)