Your finance lead opens a chat window after lunch and asks an AI agent, "Where are we stuck on the close?" The agent answers in neat bullets. Then she still has to leave that chat, open the ERP, find the period-close checklist, hunt the three exceptions, and rebuild the same picture by hand.
That is the expensive pattern in SMEs that already say they use AI in finance. The agent lives in one place. The work lives in another. So the team gets clever answers and the same queue of invoices, reconciliations and approvals they had before.
On 5 November, Sage described a different shape for this. At Sage Future for Partners, it set out AI Workspaces for Sage Intacct: a place where finance teams work with agents "alongside relevant financial information, tasks and workflows".[1] Jon Fasoli, SVP of Sage Intacct, put the job cleanly: agents do the routine work, and finance people "review and approve work according to the custom controls their business built into the Sage AI platform".[1]
That is the design I would copy, whatever ERP or CRM you run. Put the agent next to the invoice. Not in a side chat that forgets your controls.
What Sage is actually shipping
Parts of this are already shipping. It is not only a slide.
Sage says its always-on finance agents already cover AP, period close, assurance, time entry and nonprofit benchmarking, with a Finance Intelligence Agent for natural-language questions on Intacct data. Cash and AR agents are planned later. The first AI Workspace, focused on accounts payable, is planned for limited rollout with selected customers in November. Partner-built agents can already be found and activated through the Sage Agent Marketplace, with permissions assigned in an agent management console.[1]
Look at what the AP Agent is meant to do: process invoices, match them to purchase orders, identify exceptions, and route work for approval so people focus on the items that need attention. Close Agent monitors close tasks, spots exceptions and tracks progress. Assurance Agent flags potential risks and control issues.[1]
That is not "ask me anything about finance". That is an agent sitting inside a known process, with a known exception path and a known human step.
Why this matters in Singapore right now
CPA Australia's Business Technology Report 2026, drawn from 1,679 accounting and finance professionals across five markets, found Singapore leading on appetite for agentic AI: 56 per cent of organisations expect to increase their use of it over the next 12 months, the highest result in the survey.[2] AI use itself is near universal here at 93 per cent.[2]
The same survey shows the gap that bothers me. Only 12 per cent said AI is deeply embedded throughout their organisation.[2] And the barriers respondents named first were overreliance on AI, reduced human oversight, data privacy concerns and the pace of change.[2]
Demand is high. Embedding is thin. A chat bolted onto Excel will not fix that. An agent that lives inside AP matching, close tasks and CRM follow-ups might.
How other stacks are drawing the same line
You do not need to be on Sage Intacct to use this pattern.
Microsoft's Account Reconciliation Agent in Dynamics 365 Finance works from an Account reconciliation workspace. It evaluates exceptions and recommends an action, such as creating a journal entry. A person can accept that recommendation, or choose reverse, link transactions, or accept without change. Every exception is logged so you can see what users, automation or agents did.[3]
Dost's AskDost, announced on 2 November, connects to ERP data and prepares invoice, approval, payment and reconciliation workflows, but executes them only after a person approves, with an audit trail of what changed and who approved it. It lists native connections to SAP, Dynamics 365 Business Central, Sage and Oracle.[4]
Different vendors. Same operator lesson: the agent prepares or recommends inside the finance system of record; a named person still owns the irreversible step.
How I would design it in an SME
Here is the habit I would set before buying another "finance AI" seat.
1. Pick one process, not one model. Start with invoice matching, bank reconciliation, collections follow-ups, or CRM quote-to-order. Write the steps on one page: what the agent may prepare, what it may propose, and what only a person may release. If you cannot name those three lists, you are not ready for the agent.
2. Give the agent a desk in the process. In AP, that desk is the unmatched-invoice queue and the PO match screen. In close, it is the task list and the exception board. In sales ops, it is the stalled opportunities and the draft follow-ups waiting in CRM. The agent should open to the same records your team already trusts, not a private transcript.
3. Route exceptions, not essays. Borrow Sage's AP pattern: match what you can, surface what you cannot, send those to a person with authority.[1] Your reviewer should see the invoice, the PO, the variance and the recommended action in one place. They should not have to reconstruct the story from a chat.
4. Keep the approval where money and promises move. Payment runs, credit notes, credit holds, stage changes that commit stock or price, customer emails under your letterhead. The agent prepares. A named role approves. Sage frames it that way for Intacct.[1] Microsoft and Dost draw the same line in their own products.[3][4]
5. Check the trail on Friday. Pull ten agent-touched items. For each one, can you see who or what acted, what was proposed, who approved, and what the ERP or CRM record shows now? If the trail breaks, fix the logging before you scale the agent.
What this is not
This is not an argument against chat. Chat is fine for "explain this variance" or "draft a supplier chaser". It becomes a problem when the chat is the only home the agent has, because then your controls, queues and audit history stay elsewhere.
It is also not a call to let the agent close the books alone. Sage's own framing keeps professionals reviewing and approving against the controls the business set.[1] CPA Australia's respondents already worry about reduced human oversight.[2] Keep the person. Move the agent to where that person already works.
A small exercise for this week
Take your busiest finance or sales-ops process. Draw three columns on a whiteboard:
- Agent prepares (match, draft, recommend)
- Agent proposes (waits for a named approver)
- Person only (pays, releases, sends, commits)
Then open the tool you use today and ask: does the agent sit next to those records, or in a separate window? If it is a separate window, your first improvement is not a smarter model. It is a better desk.
Put the agent next to the invoice. Keep the approval with the person who answers for the money.
Sources
[1] https://www.sage.com/en-us/news/press-releases/2026/11/sage-intacct-sets-out-its-vision-for-an-agentic-finance-platform-with-ai-workspaces — Sage Intacct sets out its vision for an agentic finance platform with AI Workspaces (Sage, 5 Nov 2026)
[2] https://www.cpaaustralia.com.au/about-cpa-australia/media/media-releases/singapore-leads-region-in-race-to-next-generation-ai — Singapore leads region in race to next-generation AI (CPA Australia Business Technology Report 2026, media release)
[3] https://learn.microsoft.com/en-us/dynamics365/finance/general-ledger/acct-rec-agent — Account Reconciliation Agent (production ready preview) - Dynamics 365 Finance (Microsoft Learn)
[4] https://www.finextra.com/pressarticle/111529/dost-launches-askdost-to-give-finance-teams-their-own-ai-agents — Dost launches AskDost to give finance teams their own AI agents (Finextra, 2 Nov 2026)