Payment receipts from Outlook and Teams: a controlled collections queue

In 60 seconds: Power Automate can receive an email with attachments or detect a new file in the SharePoint library behind a Teams channel. The flow creates a case, preserves the original, extracts candidate fields, queries orders or receivables, and applies visible rules. Illegible receipts, duplicates, amount mismatches, and missing orders enter a queue with a reason and owner. Copilot can summarize permitted context and draft a clarification, but it cannot establish that funds arrived or decide where to apply them. SharePoint or OneDrive provides staging; the ERP or collections system keeps the official state. Measure time to reconciliation, exceptions, and rework before expanding the scope.

A payment receipt often arrives in a shared Outlook mailbox or a Teams channel. Someone downloads the PDF or image, finds the customer, compares the amount with an order, and copies the result into the ERP. When something fails, the conversation continues in email or chat and the task becomes invisible to the rest of collections.

Microsoft 365 can organize that journey without presenting automated reading as a financial decision. A useful architecture separates intake, extraction, matching, review, and posting. The capabilities and limitations cited in this guide were checked against Microsoft documentation on September 13, 2026; licenses, connectors, and policies still need verification for each tenant.

One map for Outlook and Teams

Both channels can feed the same queue, although they do not store files in the same way.

Intake	Recommended trigger	Evidence added to the case	Caution
Collections mailbox in Outlook	When a new email arrives in a shared mailbox (V2)	Message ID, sender, date, subject, and each attachment	The connection account needs mailbox access; protected email or invalid attachments may be skipped
Collections channel in Teams	File created in the SharePoint library associated with the channel	Site ID, library, file, author, and time	Define one intake channel and folder; do not depend on someone mentioning a bot
Teams message	Trigger for a specific chat or channel only when its text supplies necessary context	Message ID, conversation, and link to the file	The root-message trigger does not receive replies; some variants do not retrigger on edits

The Outlook connector for Power Automate can filter messages with attachments. Microsoft warns that including all attachment content in the trigger itself can cause waits and timeouts when many messages arrive together; its documented alternative is to trigger without the content and retrieve each file with Get Attachment (V2).

In Teams, a file does not really live in a Teams-only folder. Microsoft documents that files shared in a channel are stored in the team’s SharePoint site, while files shared in a private or meeting chat are stored in the sender’s OneDrive. A controlled channel with a SharePoint library trigger is therefore a more durable operational inbox. The Teams connector can collect message context, but its triggers have different scopes: some react only to root messages and others do not react to edits.

The shared journey looks like this:

1. Outlook or SharePoint detects the new item.
2. Power Automate generates a case_id, checks format and size, calculates a file fingerprint, and stores the original in staging.
3. An authorized extraction service returns candidate fields without altering the document.
4. The flow normalizes the fields and queries eligible records in the ERP or accounts receivable system.
5. Deterministic rules compare references, customer, amount, currency, date, state, and duplicates.
6. A valid case becomes ready for the next control defined by Finance; a difference enters quarantine.
7. A person resolves the exception, and Power Automate records the decision and the destination system’s response.

This pattern complements the route for payment receipts received through WhatsApp. The channel, identity, and storage change, but the obligation to explain the match and keep the ERP as the source of truth does not.

Separate orchestration, reading, and decision-making

Putting “AI” in one box hides who did each job. Give every layer a specific responsibility.

Layer	Responsibility	What it must not decide
Power Automate	Trigger, move the file, call services, query candidates, route states, retry, and record results	Whether a financial difference is acceptable by intuition
OCR or document extraction	Propose text and fields with the original value and, when available, a field-level confidence signal	Whether the receipt proves that funds arrived or which balance receives them
Copilot	Summarize a permitted thread, explain context to the reviewer, and draft a clarification request	Approve its own match, change a balance, or authorize a payment
Business rules	Validate formats, states, references, authorized tolerances, and duplicates	Resolve an undocumented exception
Responsible person	Confirm or correct exceptions within their role	Bypass segregation of duties because the flow makes it convenient
ERP or collections system	Maintain orders, receivables, payment application, actor, state, and official history	Treat a folder or summary as a substitute for the entry

Copilot contributes context, not banking certainty. In Outlook it can summarize threads and, in supported clients, PDF, Word, or PowerPoint attachments. In Teams it can summarize conversations, but Microsoft says it cannot summarize images or files shared in the chat thread. An image receipt needs an extraction method selected and tested for that format; the process should not rely on an undeclared product capability.

After extraction, Copilot may prepare text such as “We need a clearer receipt” or summarize why a case has two candidate invoices. Keep the message as a draft until a rule or authorized person decides to send it. The comparison of Power Automate, connectors, and custom development helps choose the mechanism when the ERP lacks a standard connector or the workflow accumulates complex state.

The minimum case schema

The queue needs structured data in addition to the file. Always retain the received and normalized values so a correction can be explained.

Field	Example treatment	Use
case_id	UUID generated once	Correlate the email, file, run, review, and write
source_channel and source_id	outlook + message ID, or teams_sharepoint + file ID	Return to evidence without depending on a display name
received_at and sender	UTC time, address, or authorized internal ID	Age, traceability, and origin controls
file_name, type, size, and hash	Metadata without overwriting the original	Allowed-format and duplicate checks
amount_original, amount, and currency	Read text, normalized decimal, and currency code	Compare without losing the submitted value
operation_date and payment_reference	Payment date separate from receipt date; original and normalized reference	Search and detect reuse
customer_candidate	Internal ID and rule that proposed it	Stop a similar name from becoming an identity
receivable_candidates	IDs, states, balances, and differences	Present explainable alternatives to the reviewer
status, exception_reason, and owner	Closed state, enumerated reason, and accountable person	Operate and measure the queue
decision and system_result	Actor, time, rule or approval, transaction ID, and response	Establish what was attempted and what the ERP accepted

Do not reduce quality to one score. An amount can be read correctly and still belong to the wrong customer. If the extractor exposes confidence, store it per field and use it to prioritize review; do not turn it into permission to write.

Layered matching, not general similarity

Matching should return evidence and differences rather than a free-form answer. A practical sequence is:

1. Reject unsupported formats and separate documents without minimum fields.
2. Stop a processed file or reference by checking the hash, reference, and ERP history.
3. Resolve the customer through agreed identifiers; email, name, and conversation supply context only.
4. Query orders, invoices, or open items in an eligible state.
5. Compare currency before amount; never offset currencies without an explicit rule.
6. Compare reference, amount, date, customer, and state, retaining each rule’s result.
7. Route zero candidates, multiple candidates, or one candidate with out-of-policy differences to review.
8. Recheck the balance, state, and duplicates immediately before preparing or executing any authorized write.

The proposal should show the candidate record, matching fields, differences, query time, and allowed action. Our guide to control levels for AI agents maps cleanly to these steps: extraction is reading, suggesting a document is a recommendation, and preparing an application is a draft. Posting a payment affects cash and balances, so it cannot bypass financial controls or exception review.

Four exceptions the pilot must handle

Quarantine is a worklist, not a folder named “Review.” Every entry needs an enumerated reason, evidence, owner, and deadline.

Reason	What the review card shows	Permitted resolution
Illegible	Original, missing or uncertain fields, and source channel	Request another file or transcribe under the required control
Possible duplicate	Hash, reference, amount, date, and earlier case or payment	Confirm the duplicate or justify why it is another transaction
Amount mismatch	Balance, received amount, currency, and difference	Keep pending, associate under policy, or request clarification
No order or receivable	Proposed identity, search criteria, and rejected candidates	Correct identity, match manually, or return to sender

Add multiple_candidates, currency_mismatch, unsupported_file, and destination_failure if they appear in the sample. Do not automatically turn one person’s correction into a new rule. Analyze its frequency, document the policy, and test the change first.

Teams can present the exception with a case link and the minimum information required to decide. Avoid copying the full receipt into every notification: that broadens access and creates more copies subject to retention. If the decision happens in a card or message, its result must return to the case and system of record; it cannot remain as an isolated “approved” in chat.

SharePoint is staging; the ERP preserves truth

A controlled SharePoint library can retain the original file, case_id, processing state, and a retention date. OneDrive can support individual work, but depending on the personal space of the person who shared a file weakens continuity, permissions, and employee-offboarding behavior.

None of those folders should define that a payment has been applied. The sources of truth divide as follows:

• Exchange or Teams preserves the original communication under company policy.
• SharePoint temporarily preserves the file and processing metadata.
• The queue preserves the owner, reason, state, and exception resolution.
• The ERP or accounts receivable system preserves the balance, match, application, reversal, and official identifier.
• The bank or payment provider preserves evidence that the movement occurred when an authorized integration exists.

A receipt sent by a customer is evidence to review, not bank confirmation. A successful Power Automate write is not enough either: retain the ERP response and run a later reconciliation across received cases, decisions, and created transactions. This reduces the cost of re-entering data between systems without sacrificing traceability.

Permissions, retention, and models

Use a managed connection account with access only to the required mailbox, channel, library, and operations. Separate who edits the flow, reviews exceptions, and can post or reverse payments. Avoid the process owner’s personal account, and log changes to connections, rules, and versions.

Before the pilot, agree on:

• who can view the original and which fields may leave the queue;
• how long email, chat, staging files, extracted data, prompts, and logs remain;
• which sensitivity or retention label applies to each location;
• which connectors may exchange data inside the environment;
• whether names, account details, bank references, or attachments may be sent to a specific model;
• how applicable access, correction, deletion, and legal-hold requests are handled.

Microsoft states that prompts, responses, and Graph data used by Microsoft Copilot are not used to train foundation models, and that the experience respects tenant identity, permissions, and policies. This does not replace a lawful basis, classification, or internal decision about purpose and minimization. It also does not authorize pasting a receipt into a consumer account, enabling web search, or sending it to an external connector.

Power Platform can apply DLP policies to connectors, and Microsoft Purview offers retention policies and labels for supported locations. Verify the actual license and configuration: a documented feature is not proof that it is enabled in your tenant.

A small, measurable pilot

Start with one shared mailbox or channel, one entity, one currency, and one receipt type. Run the first stage in shadow mode: extract, search, and propose while the team keeps its current procedure. Compare results before the flow may prepare a write.

Metric	Definition	What it helps decide
Time to reconciliation	From receipt to the correct accounts receivable match	Whether the flow reduces total waiting, not only transcription
Exception rate	Quarantined cases ÷ received cases	How much of the chosen scope still requires judgment
Rework	Corrected, reopened, or re-entered cases	Whether automation removes or shifts work
Queue age	Open time by reason and owner	Whether human review has real capacity and an achievable SLA
False positives	Proposed or accepted matches that validation finds incorrect	Financial risk and rule quality

Segment results by Outlook or Teams, PDF or image, customer, extractor, and reason. Do not publish an OCR, savings, or autonomy rate from a demo. Use synthetic documents for development and an authorized, minimized sample of your own documents for validation; never copy PII or real receipts into the article or uncontrolled environments.

The pilot can advance when the team can explain every match, resolve the queue within its deadline, detect duplicates, and reconcile results against the ERP. Kiia designs this route with Power Automate as the orchestrator, Copilot as bounded assistance, and Finance as the owner of rules and exceptions.