For the week of October 19–26, three October 22 releases offer concrete changes to evaluate across close, integrations, reporting, and technical workpapers. The common thread is a move from answering questions toward preparing or coordinating work. The evidence below is official vendor documentation and a vendor-reported customer testimonial, not independent benchmarking. No hands-on product tests were performed for this issue.
Tipalti: AP execution and MCP data access are different capabilities
What changed. Tipalti announced AP Agent to coordinate invoice capture, coding, approvals, payments, and reconciliation, with configurable human oversight. It describes field-level explanations, action logs, and learning from corrections. Separately, its new MCP connection makes AP and payment data available to external enterprise AI tools. The product FAQ sets an important boundary: MCP is read-only at launch. It cannot approve invoices, send payments, or change Tipalti records.
Evidence status
This is vendor-documented functionality and an announced product direction, not a reproduced end-to-end implementation. MCP initially supports Claude Enterprise and requires an admin-approved connection, an appropriately permissioned Tipalti account, and approved data scopes. Tipalti says each request is checked against user permissions. The announcement does not establish that every AP Agent capability is enabled in every customer account; confirm entitlement and rollout before planning a pilot.
Controller’s lens. For month-end close, start by comparing AP records with ERP balances through separately authorized connections. Define the as-of time, entity, currency, posting status, and population before matching. Read-only access reduces execution risk but does not prove completeness or a successful reconciliation. Test AP Agent's payment and approval boundaries separately from the MCP reporting connection; one product's advertised autonomy does not expand the other's permissions.
CoCounsel: executable calculations and finished workpapers
What changed. Thomson Reuters launched the next generation of CoCounsel Tax, Audit & Accounting. Announced capabilities include code execution, multi-tab Excel workbooks with live formulas and an assumptions tab, document drafting and updates using firm templates, MCP support, and expanded integrations. The release says the platform is available to US customers, with additional capabilities to follow.
Evidence status
The launch documents vendor-described capabilities, not independently verified correctness. A named Sikich executive reports improved efficiency and consistency, but the release supplies no quantified benchmark or controlled comparison. Claims that the system chooses the correct authority or gets professional judgment right remain claims to test, even when calculations are reproducible.
Controller’s lens. A controller can evaluate whether an AI-assisted workpaper preserves source facts, policy choices, formulas, assumptions, and supporting authority. Reproducing a calculation does not validate its accounting treatment. For a prepaid-expense test, have the reviewer establish the capitalization policy and allocation convention first. For technical memos, verify that cited guidance exists, applies to the facts and period, and supports the conclusion. Retain the original inputs and reviewer changes alongside the final file.
Spotlight: calculate first, then write the narrative
What changed. Spotlight released Enhanced AI Executive Summaries. Its documentation describes calculating figures before generating commentary, one-click topic creation, custom prompts, highlights and recommendations, and the ability to review, edit, or discard suggestions. The enhanced summaries can feed an Action Plan; action plans themselves were already listed in its April release notes.
Evidence status
Dated release notes and a product explanation support the availability announcement and intended design. The vendor's stronger claim of no hallucinations is not established by the material reviewed. Separating calculations from language is a useful design choice; it does not demonstrate that generated explanations or recommendations are factually supported.
Controller’s lens. For reporting and flux analysis, verify two layers independently: the numbers and the explanation. An accurate expense increase does not establish whether headcount, timing, price, or reclassification caused it. Require the narrative to distinguish observed movement from a supported driver and an unresolved hypothesis. Existing commentary also needs review when the underlying period or data changes.
Experiments worth running
Use synthetic or appropriately anonymized data. These are proposed tests, not completed experiments.
AP close test: can a connected assistant prove its population?
Test: Use synthetic AP and ERP datasets with 30 invoices and five planted differences: duplicate, missing posting, wrong entity, cutoff difference, and partial payment. With authorized product access, repeat under a restricted user and attempt a non-destructive write request.
Evidence to retain: Retain population counts and totals, as-of timestamps, source IDs, permissions, prompts, tool calls, exception output, and reviewer corrections. Keep product tests in a sandbox; exported fixtures alone do not test connector permissions.
Success criteria: All five differences identified without false clearing; both populations reconcile to control totals; restricted records stay inaccessible; the read-only connector refuses writes. Measure preparation and review time separately.
Workpaper test: formulas that survive review
Test: Prepare a fictional prepaid schedule with ten contracts, including mid-month starts, a cancellation, and missing dates. Provide a reviewer-approved policy and calculation convention, then request an Excel workpaper and supporting memo.
Evidence to retain: Keep source contracts, independently calculated expected results, generated formulas, assumptions, exact model/product version, cited guidance where used, and the review record.
Success criteria: Calculations match the approved expected results; source-to-row traceability and editable formulas remain intact; missing facts are escalated; no invented authority or unsupported accounting conclusion appears.
Flux narrative test: correct numbers without invented causes
Test: Give the reporting assistant two synthetic periods containing a reclassification, a timing shift, and an expense increase whose cause is intentionally withheld. Ask for an executive summary and action plan.
Evidence to retain: Retain the source reports, independent variance calculations, withheld-fact register, prompts, generated commentary, and a sentence-by-sentence evidence assessment.
Success criteria: Every amount ties to the reports; the unexplained increase remains unresolved; hypotheses are labeled; no causal assertion is presented as fact without support. Recommendations do not imply actions were approved or performed.
Evaluate autonomy, access, calculation, and explanation as separate acceptance decisions. Start with a bounded workflow and an independently prepared expected answer. A connected assistant, an editable workbook, and a fluent narrative each remove friction—but none alone establishes a reliable close. The workbench proposals above have not been run. This issue was prepared with AI assistance from sources checked on October 26, 2026; it does not imply personal testing, endorsement, or review by Samantha Ye.
Prepared with AI assistance for Rebuild Accounting. Sources were checked for this issue; product availability and terms can change. Editorial assessments and proposed tests are identified separately from vendor documentation. This issue does not claim hands-on testing or personal endorsement.
← Back to all issues