Financial Intelligence · Research Note
What does the inbox know that the accounts don’t?
Can AI connect email evidence with accounting records to establish what happened, whether it was recorded and whether the entry reflects the underlying commercial reality?
Xero · Gmail · ChatGPT · MCP · Financial intelligence · Commercial context · Data coverage · Multi-entity · AI assurance
How this connects
This write-up examines one part of the central thesis:
Are we delivering the business we intended?
It answers one part of the wider question Can AI help finance detect gaps between the accounts and commercial reality—before they distort decisions?
Tested on Xero.
A charge is in the accounts. The invoice is somewhere in email. Or perhaps there is no invoice, and the explanation sits across a receipt, a payment notice and three messages nobody thought finance would need.
The accounts show a financial entry. The inbox may explain what it represents, which entity paid it, when a service ended or why the amount differs.
Finance teams already reconstruct this picture using applications, spreadsheets and manual investigation. This experiment explored what changes when AI can interrogate accounting records and email together—and follow the evidence beyond the first apparent match.
Can that business context help us assess whether financial records are supported, correctly represented and complete?
Xero experiment 04 · 19 November 2026 · ChatGPT + Xero MCP + Gmail connector. No Clearspend tools or matching results used in this run. No financial or email records changed.
The question being tested
Can AI connect email evidence with accounting records to establish what happened, whether it was recorded and whether the entry reflects the underlying commercial reality?
The investigation addressed three connected questions:
| Question | What email could contribute |
|---|---|
| Support | An invoice, receipt, acknowledgement or other evidence behind a recorded transaction |
| Accuracy | Context for questioning a description, amount, timing or entity attribution |
| Completeness | Evidence of activity to check against the accounting records, with explicit limits on any absence claim |
The test searched in two directions:
- Start with the accounts. For recorded transactions with missing support, find the invoice, receipt or correspondence that explains the charge. Where no invoice is found, distinguish missing documentation from a transaction that requires a different kind of evidence.
- Start with email. Look for invoices and payment evidence, then check whether the corresponding activity is represented in the accounting records.
The second direction tests potential gaps beyond the list the accounting system already knows about. Finding evidence of unbilled work would require a further test of contracts, delivery and acceptance; an email invoice alone does not establish it.
Connected components
| Component | What it contributed |
|---|---|
| Xero, accessed through MCP | Available accounting transactions, invoices, payments and journals across two connected organisations |
| Gmail, accessed through its connector | Searchable messages, invoices, receipts, payment notices and commercial correspondence |
| ChatGPT | Directed the investigation, read evidence, compared records and developed explanations across sources |
| Human review | Supplied organisational context, challenged apparent gaps and checked selected items in Xero’s interface |
Clearspend was not used as a source or matching engine in this run. It remained part of the earlier experiment history; this was a direct investigation rather than a controlled, blind comparison of products.
What direct investigation could do
The assistant retrieved supporting documents and explanations for selected recorded items using accounting and email access.
That included formal invoices and receipts, payroll-related evidence, payment acknowledgements and correspondence explaining the purpose or timing of a transaction.
The distinction matters. A supplier invoice may be missing. A bank fee, payroll payment or transfer may need different evidence. A card repayment may combine several purchases. Treating all of these as “find the invoice” sends finance on some impressively unproductive treasure hunts.
The useful capability was the follow-up investigation: read another message, check a payment notice, consider a different supplier descriptor or search the other entity’s ledger.
A matching result could become the start of a question rather than the end of a workflow.
What happened when we searched from email back to the accounts
Some email documents initially appeared to have no corresponding accounting entry.
Human review clarified that the inbox served two entities. Searching the second entity’s ledger found strong corresponding entries for seven previously flagged documents. They were removed from the potential omission list. A separate software invoice also had a probable currency-converted match needing settlement confirmation.
This was a useful demonstration of direct investigation: the assistant could change the search scope and compare the evidence again. It also exposed a mistake in the initial scope. Currency and supplier location had not been enough to assign the correct ledger.
Finding the payment elsewhere does not automatically settle expense ownership or intercompany treatment. It does stop us recommending that the same expense be recorded twice.
Where direct connections still needed more data
Two further candidates were visible in Xero’s reconciliation queue when the user supplied a screenshot. Their corresponding statement lines were not available through the current MCP connection.
They were reclassified as present in the queue with strong corresponding email evidence, rather than potentially absent from Xero.
Xero’s public API does not expose raw unreconciled bank statement data through the ordinary public interface, as explained in its published API policy. A reconciliation-queue export or separately supplied bank/card data would extend this investigation’s coverage.
Two earlier candidates remain unresolved, with possible combined-card-payment explanations requiring detailed statements. Other documentation questions in the original transaction review also remain open. This was not evidence that every transaction had been supported.
What the email context changed
The evidence helped distinguish different situations that can otherwise collapse into the same exception label: a missing document, a payment recorded in another entity, a charge awaiting reconciliation or a possible combined settlement.
Those situations require different responses. Searching for another invoice will not resolve an entity-scope error. Adding an expense will not help if the payment is already represented elsewhere. An arithmetic match will not establish how a card repayment was allocated.
The experiment therefore went beyond document retrieval in selected cases. It used context to revise the interpretation of the records—and to withdraw apparent gaps when the evidence contradicted them.
The starting point and method still mattered. ChatGPT accessed Xero and Gmail directly, without using Clearspend’s matching results. That describes how this investigation was conducted; it is not a conclusion about which applications finance teams should use.
The financial leadership implication
The value extends beyond locating a PDF. Finance needs to establish whether a charge is supported, what it represents, whether it belongs to the expected entity and what needs to happen next.
Direct investigation offers a possible shorter route from uncertainty to an explanation and an appropriate action. It could also help identify recurring causes of missing documentation rather than repeatedly clearing the same exceptions.
This run demonstrated selected evidence retrieval and explanation. It did not prove completeness, autonomous resolution, a faster close or an economic return.
The next test should measure time to a reviewer-accepted explanation, false alarms, unresolved items and human effort, then track the outcome of any resulting action.
Current conclusion
Email contains business evidence that can change how financial entries are understood, supported and questioned. Connecting that evidence with accounting records allowed selected transactions to be explained and apparent gaps to be reassessed.
The test did not establish that every entry was correct, identify every missing invoice or prove that all business activity was recorded. Organisational context, source coverage and human challenge remained essential.
The wider opportunity is to help leadership judge whether the financial picture reflects what is actually happening in the business—and reach a sound explanation while there is still time to act.
Supporting evidence
A read-only direct investigation found selected invoice and payment support. Seven earlier potential omissions had corresponding records in another entity. User-supplied reconciliation-queue evidence explained two more. One separate FX match remains probable; two earlier candidate documents remain unresolved. Other transaction-support questions remain open.
Sources
- Private Experiment 04 evidence — 19 November 2026
Read-only accounting and email investigation, cross-entity corrections and user-supplied reconciliation-queue evidence. Entity identities, amounts, document numbers, mailbox links and screenshots are withheld.
- Xero Developer Centre — Reconcile via the API
Xero response dated 6 May 2026 explains its public-API restriction on raw unreconciled bank statement data.
Counterarguments
No controlled product comparison, complete audit or proof of accounting completeness. Familiar cases and human corrections influenced the result. The current connector cannot retrieve the raw reconciliation queue. Economic benefits remain unmeasured.
Current conclusion
Email evidence helped explain selected transactions and revise apparent accounting gaps. The experiment supports investigating support, accuracy and completeness across sources, without proving every entry correct or every omission detectable.
How my thinking has changed
The central question is what business evidence adds to financial understanding. Direct accounting and email access is the method; the finding concerns the context needed to interpret whether records reflect commercial reality.
The experiment behind this article
Research question
Other experiments on Xero
Experiment 03 · 18 Nov 2026
Can connecting email with accounting records reveal commercial context that helps explain invoicing issues and identify the right intervention?
Read the article: The invoice tells you what was billed. What does the business know that finance doesn’t?Experiment 02 · 17 Nov 2026
Which unsupported transactions can document matching resolve, and which require a broader investigation?
Read the article: Another missing invoice—or a transaction that needs a different explanation?Experiment 01 · 16 Nov 2026
Can a collections workflow help resolve payment delays—and what additional evidence is needed when reminders do not address the blocker?
Read the article: The invoice is overdue. What is actually holding up payment?Planned
Xero JAX