
The video describes a team using a Copilot Studio agent, triggered from an embedded Power Automate workflow, to resolve invoices that were in an error state and raise its touchless invoice rate. This article separates what the presenter reported from what Microsoft documents, and then covers trade-offs and next steps.
This article draws on a video from the MSD365 Community channel.
Context from the video
The video features Don Baldwin, senior director of software engineering at Assurant. According to Baldwin, his finance team was using Invoice Capture but could not get more than a 20% success rate for touchless processing, so the team turned off much of the automation. They then asked a fast track team for help.
Baldwin says they created an agent in Copilot Studio with an embedded Power Automate workflow. The workflow creates a message and sends it to the agent, which uses a set of instructions and resolves the data that was in an error state. He reports the agent was implemented within a week. He also reports that touchless invoicing went from less than 20% for only two vendors to over 55% within the first week for North American vendors, and to over 70% for the entire North America region, with a global launch planned.
These are Baldwin’s reported figures for his own environment. They are not a benchmark and should not be assumed for other organizations.
What Microsoft documents
The supplied Microsoft documentation covers the building blocks and the Invoice capture behavior that creates the problem. It does not document a specific agent-based error-resolution integration for Invoice capture, so treat the end-to-end design in the video as the presenter’s implementation rather than a documented product pattern.
Why error-state invoices block touchless processing
In Invoice capture, an invoice that AI Builder recognizes and maps to invoice fields can be automatically updated and validated according to rules before it is converted to a vendor invoice in Dynamics 365 Finance. If the invoice contains errors or warnings, manual intervention is required to review and correct it. Microsoft Learn
That explains why a high rate of errors and warnings limits touchless processing: each such invoice needs a person to review and correct it.
Copilot Studio agents and workflows
Copilot Studio can create agents and workflows in one studio, and a workflow can call an agent to complete a step. Microsoft Learn This is consistent with the shape Baldwin describes, where a workflow hands a message to an agent.
Calling an agent from a Power Automate cloud flow
The Microsoft Copilot Studio connector lets a Power Automate cloud flow send a message to a standard harness agent. The ‘Execute Agent and wait’ action makes the flow wait for the agent’s response before continuing. In the Message field, the maker enters the instructions for the agent and can add dynamic content from earlier flow steps. Microsoft Learn
The same documentation distinguishes agent harnesses. Agents on the standard harness are called from cloud flows through the connector, while agents on the Microsoft Copilot harness use the agent node in a Copilot Studio workflow and cannot be called from a cloud flow through that connector. Confirm which harness your agent uses before designing the flow.
To build the generic calling step, sign in to Power Automate, add an action in a cloud flow, search for Microsoft Copilot Studio, select Execute Agent and wait, and enter your instructions and any dynamic content in the Message field.
A related Microsoft capability in preview
AI Suggestions, listed as a preview feature in May 2026, is described as an AI-assisted derivation capability that enables touchless invoice processing within Invoice capture. It replaces rigid rule-based derivation with an LLM-powered, Dataverse-grounded, instruction-driven approach. Microsoft Learn
Because it targets the same goal with instructions in place of rigid rules, it is worth evaluating alongside a custom agent.
My Take
The reported result is encouraging, but I would read it as one team’s outcome rather than a proven recipe. The documentation confirms the mechanics of calling an agent from a flow and confirms that errors and warnings require manual review in Invoice capture. It does not confirm how an agent’s output should be returned into Invoice capture or how that interacts with validation, so that design needs to be proven in your own environment.
Trade-offs I would weigh:
- Instruction quality drives outcomes. An agent that follows written instructions is only as reliable as those instructions, so assign an owner and review them like business rules.
- Auditability. Decide how each correction is recorded so finance can see what changed and why, since touchless means no one reviews it at the time.
- Where validation sits. Test whether corrected data still passes through the validation you rely on before a vendor invoice is created in Finance, rather than assuming it does.
- Reported results are context-specific. The week-long build and the move past 70% touchless are Baldwin’s reported results; your vendor mix and error types will differ.
- Preview risk. If you consider AI Suggestions, it is documented as a preview feature, so check its status and support terms before production use.
For background on the platform, see Microsoft Dynamics 365 explained, and for the wider AI picture see Dynamics 365 AI and Copilot.
Practical Implications
If your finance team has reduced automation because too many invoices need correction, consider a structured pilot:
- Measure the baseline. Record your current touchless rate by vendor and list the most frequent error and warning types.
- Define a small scope. Start with a limited set of vendors and error types so results are easy to review.
- Write the instructions. Capture how your finance staff currently correct each error as explicit agent instructions.
- Build and test the flow. Use the Execute Agent and wait action to pass invoice data and wait for the agent’s response, and test how the result re-enters your Invoice capture process.
- Keep a manual fallback. Route anything the agent cannot resolve to your existing review queue.
- Review before expanding. Sample corrected invoices for accuracy before adding vendors or regions.
If you want help scoping an Invoice capture automation or comparing a custom agent with built-in capabilities, our Dynamics 365 consulting services team can help design the pilot and its controls.
Frequently Asked Questions
What problem does the Copilot Studio agent address in Invoice capture?
It targets invoices with errors or warnings, which Invoice capture requires someone to review and correct manually. In the video, an agent following written instructions resolves data in an error state so more invoices can be processed without manual handling.
How does a Power Automate flow call a Copilot Studio agent?
Use the Microsoft Copilot Studio connector's Execute Agent and wait action in a cloud flow. The flow waits for the agent's response, and you write the agent's instructions in the Message field, optionally adding dynamic content from earlier steps. This applies to agents on the standard harness.
Will I get the same touchless rate that was reported in the video?
Not necessarily. The over 70% touchless figure for North America is the reported outcome in one environment. Vendor mix, invoice quality, and error types will affect your results, so pilot and measure.
Is there a built-in alternative to a custom agent?
Microsoft lists AI Suggestions as a preview feature in Invoice capture, described as an LLM-powered, instruction-driven approach to derivation that supports touchless processing. Because it is a preview, confirm its status and support terms before production use.
Free tool
Calculate Your Dynamics 365 Migration ROI
See how much you could save with a free, instant ROI estimate. No signup required.



