My AI assistant finished a production order for 20 coffee makers in Microsoft Dynamics 365 Business Central.
It took way too long. The screen-based automation caused interruptions, and we briefly lost the connection. In the end, it worked, but not perfectly.
That is the honest starting point for this experiment. I want to understand what AI can actually do inside a business, including the accounting consequences of its actions. A successful demonstration needs to show the problems as well as the completed transaction.
This work took place in our fictional Wheeler Coffee Demo sandbox with synthetic data. No customer records or live manufacturing operations were involved.
Picking up an unfinished order
In the previous experiment, we created released production order 101002 for 20 AutoDrip coffee makers.
That run had already taught us something important about inventory. The order initially had a blank location, while the components were at MAIN. Once we corrected the location with my approval, Business Central automatically consumed 240 screws through forward flushing.
We ended that session with a released order and some component consumption, but no finished output. On November 11, we returned to finish the order.
The first requirement was to understand what had already happened. We couldn’t simply start again and assume everything still needed to be posted. Those 240 screws were already consumed. Completing the order had to preserve that earlier transaction without posting it a second time.
The screws were not duplicated.
That is the kind of detail I care about when considering AI in accounting. Being able to enter a transaction is useful. Being able to continue a partly completed process without duplicating an earlier entry is a much more meaningful test.
There were still decisions for me to make
I made decisions about using November 11 as the posting date and about the wiring time used in the demo. We used the routing-standard time, not measured actual time.
The date deserved particular attention. Our previous session had retained an October 22 demo date. A system can accept a date without that date being the one the business intends to use.
For an accounting workflow, “the software allowed it” is not a sufficient reason to proceed. Someone still needs to decide what period the transaction belongs in and whether the assumptions behind the entry make sense.
AI can help identify and explain those choices. In this experiment, I remained responsible for approving them.
We used the screens, and it was slow
The assistant operated through the Business Central interface. It navigated pages, inspected fields and worked through the application much as a person would.
This let us attempt the workflow without first building a dedicated integration. It also made the process cumbersome.
There were interruptions and a brief connection loss. We had to keep track of the work and establish what state the system was in before continuing. That matters because a connection problem is not evidence that a transaction failed. Repeating an action without checking could create a second problem.
I would not describe this run as a productivity breakthrough. We did not establish a time saving, and the amount of back-and-forth was too high for the routine process I would eventually want.
Still, working through the screens exposed the decisions and checks that a better implementation will need.
Twenty completed units did not settle the costing question
We completed the 20 units and finished the production order. Business Central showed an output ledger cost of $2,840.80, or $142.04 per unit.
But that cost was incomplete. The demo had zero reservoir and capacity costs, so the resulting figures did not represent a complete manufacturing cost.
That limitation belongs in the main story. It cannot be buried beneath a claim that AI successfully finished the order.
Business Central can process the data and setup it has. If those inputs omit costs, completing the workflow does not make the resulting cost complete. An assistant needs to recognize that distinction and bring it to the person reviewing the result.
For me, this is where the accounting department enters the discussion. The work includes reviewing what the transactions mean, whether the underlying setup is adequate and what remains unresolved after posting.
We demonstrated a bounded production workflow. We did not establish that the demo’s cost figures were ready for a business decision.
What I would want in a reusable procedure
I don’t want to explain the entire process from the beginning every time I ask AI to help finish a production order.
I want a procedure that checks the existing order and posted entries, identifies the remaining work, confirms dates and important assumptions, and stops for approval when a decision is needed.
It should also verify the result. In this case, that included checking the completed quantity, ensuring the earlier screw consumption was not duplicated, and calling out incomplete cost inputs.
A direct integration through supported Business Central interfaces, potentially exposed through MCP, is a direction worth testing. It may avoid some of the difficulty of navigating screens.
We have not proved that approach with this experiment. A different connection method would still need permissions, error handling, duplicate prevention and a reliable way to establish what happened after an interruption.
What we actually learned
AI helped us finish a production order in a sandbox, with human decisions along the way. The screen-based approach was slow and interrupted, but we reached the intended operational result without consuming the screws twice.
We also ended with an unresolved accounting limitation: incomplete cost data.
That combination is useful. It gives us a real workflow to improve and specific conditions to test again.
I am interested in building an intelligent company where AI can help carry work through to completion. This experiment makes me more interested in that direction, while also making the controls much more concrete.
Before I would rely on this routinely, I would want a repeatable procedure, better execution and a clear review of the financial result.
Which accounting or ERP workflow would you trust AI to help execute first, and what would you want it to verify before calling the work complete?