At a glance
- Most leadership teams now have a budget line for AI agents in finance, supply chain, or service.
- Few can say whether the system those agents will read from runs cleanly every night.
- When batch jobs fail and planners keep side spreadsheets, an agent learns the workaround.
- Leaders who get a return stabilize the live system first, then add agents one function at a time, each with an owner.
The first question about an agent is what it will read. An agent works from the data in your system, and after a difficult launch that data is often the part nobody trusts.
So the order matters. Fix what your people work around, then give them agents to take the follow-up off their desks.
An unstable system teaches agents the workaround
The signs after a hard launch are consistent. Batch jobs fail overnight, releases collide, and custom code conflicts with Microsoft’s updates. Users stop trusting the numbers, and planners drift back to spreadsheets.
The agent can’t see the spreadsheet. A Finance Agent reconciling a ledger with late postings finds variances that are really timing. A Procurement Agent chasing a purchase order the buyer already changed by phone asks the supplier the wrong question.
Distributors report the same barrier. In the first NAW AI Adoption Index, a September 2026 survey of 302 distributors, 68.1 percent named data quality and integration as their top obstacle to AI.
Stabilizing is a short list of unglamorous fixes
Stabilizing means finding what breaks and fixing it at the cause. Our architects look at five things: failed batch jobs, release and branching discipline, custom code that fights Microsoft’s updates, the integrations, and the reports people actually use.
For a specialty life-sciences manufacturer, we started with an audit of the live system and fixed what it found. We then stayed on as senior support across Finance and Supply Chain, Power Platform, and reporting. The client has since scaled that support up about 75 percent.
For a multi-state material-handling dealer, the job was keeping AX 2012 stable while the dealer plans its exit. That meant managed support, database tuning, and field-service fixes, so service kept running while leaders decided what comes next.
Microsoft now changes your system every month
Stability work has no end date. Microsoft stopped publishing Dynamics 365 release plans in 2026; new capabilities now publish continuously to the AI at Work roadmap. The Release Planner retires on November 15, 2026.
That makes stability a run cost. Someone has to test each change against your processes and decide when it reaches your users. Regulated manufacturers feel it most, because each change has to pass their own change control.
Managed services should mean a senior team that keeps the system running and improving, at a predictable cost. For an employee-owned electrical contractor, that includes a four-hour response on priority-one issues.
Then add agents one function at a time, each with an owner
Once the system holds, agents earn their place in specific work. State their status exactly, because it decides what you can plan around.
Agents at work: controllers use the Finance Agent for reconciliation, variance review, and collections, and they still sign off the close. Buyers let the Procurement Agent, in paid public preview, chase supplier confirmations, then decide what to expedite. Service managers can trial the Field Service Scheduling Operations Agent, but it is in preview and not for production use.
Agents also carry a run cost. Every Dynamics 365 agent consumes Copilot Credits, and Premium licenses include 1,000 credits per user per month, pooled across the tenant. 
Copilot Studio disables agents that run above 125 percent of prepaid capacity, so credit use needs a watcher like any batch job.
Govern agents the way you govern employees. Microsoft Entra Agent ID gives each agent an identity, and 
Agent 365 observes and governs what agents do. A team name in the owner column means nobody owns it.
Some work needs no model at all. Microsoft’s own guidance treats agent flows as deterministic, and deterministic steps shouldn’t pay for model inference. Our Helios 2 map sorts each task into Process (stays human), Workflow, or Agent before anyone builds.
When agents can start before the system is stable
The rule has exceptions. An agent that doesn’t touch the unstable transactions can start in parallel. A knowledge agent built in Copilot Studio over manuals and procedures is one example.
If you’re leaving the system soon, don’t build agents on it. Dynamics AX 2012 R3 has been out of extended support since January 2023. On a system like that, stabilize only what keeps customers served and the books closed, and build agents on the system you’re moving to.
Sometimes the instability is in the process. When three plants run the same step three ways, the fix is a decision by leaders: pick one way and hold every plant to it.
What leaders should do before funding the first agent
Ask what the agent will read. If the answer is data your people correct in a spreadsheet, fix that first.
Fix the nightly batch before the dashboard. Failed jobs and colliding releases are the first things users stop trusting.
Budget for the run cost. Continuous updates and Copilot Credits are ongoing; plan for a team that tests and watches both.
Name an owner for every agent. One person decides what it does, reviews its work, and switches it off when it’s wrong.
Keep the call with people. The agent drafts, chases, and reconciles. The controller signs, the buyer expedites, and the service manager sets the schedule.
How we know this
This point of view comes from Ludia’s managed services and Dynamics 365 delivery. Product status is as of Sept 29, 2026; Microsoft now ships changes continuously, so we re-check before every plan.
- Client work described here (life-sciences manufacturer, material-handling dealer, electrical contractor): Ludia client results.
- Distributor AI barriers: Independent Dealer, NAW research finds US distributors slow to scale AI (Sept 3, 2026); NAW AI Adoption Index.
- End of release plans and Release Planner retirement: Microsoft, One always-on roadmap (Aug 25, 2026).
- Finance Agent: Microsoft, Build the future of agentic ERP (Sept 23, 2026).
- Procurement Agent status: Microsoft Learn, Supplier communications features of the Procurement Agent.
- Scheduling Operations Agent status: Microsoft Learn, Dynamics 365 Field Service planned features (updated Sept 3, 2026).
- Copilot Credits in Dynamics 365: Microsoft, Dynamics 365 Licensing Guide (Sept 2026).
- Capacity limits for agents: Microsoft Learn, Copilot Studio billing and capacity management.
- Agent identity and governance: Microsoft Learn, What’s new in Microsoft Entra Agent ID; Microsoft, Microsoft Agent 365 now generally available (May 1, 2026).
- AX 2012 R3 support dates: Microsoft Learn, Dynamics AX 2012 R3 lifecycle.
