For November 2–9, the useful changes are about how AI work becomes part of a recurring accounting process: a saved report, a tested script, a permissioned agent, or a retrievable source document. This issue uses dated vendor release notes, product explanations, and announcements. Those sources document intended capabilities and availability; they do not establish independent implementation results. No hands-on product tests were performed for this issue.
Campfire: save the report, then test the reusable logic
What changed. Campfire’s November 2 changelog introduces Custom Reports by Ember and code-backed Ember skills. The November 4 product explanation describes creating saved transaction reports, income statements, and balance sheets from chat, then refining the same report with follow-up requests. It supports report settings such as entity, period, groups, filters, and formulas. Separately, the changelog says skills can contain scripts, with a builder that tests against a previously closed period and blocks publication until its checks pass.
Evidence status
Vendor-documented shipped changes, with vendor-provided report examples. The concrete reporting change is a saved, editable object inside Custom Reports rather than only an answer or export in chat. The code-backed builder’s checks are described by the vendor; their coverage and independence were not verified here. Passing one historical period does not demonstrate correct behavior on new data or exceptions.
Controller’s lens. Editorial assessment: retain the report definition as well as its totals. A follow-up that changes grouping must not silently change entity, date range, account population, or currency basis. For recurring allocations, deterministic code can make a calculation repeatable while still applying the wrong policy or basis. Use independently calculated expected results and a second period the builder has not seen. Measure reviewer corrections and maintenance work alongside preparation time.
Sage Intacct: available marketplace, future workspace
What changed. Sage set out its vision for AI Workspaces, bringing financial information, agents, tasks, and approvals into a shared experience. Its announcement says the Sage Agent Marketplace is available now, with partner-agent activation and permissions managed inside Intacct. The first AP-focused AI Workspace is planned for limited rollout to selected customers in November. Cash Agent and AR Agent are described as future capabilities.
Evidence status
This is a mix of vendor-documented availability and roadmap. Sage says capability and market availability vary. Its existing Close Agent is described as helping monitor tasks, identify exceptions, and track progress; that description does not establish an autonomous close. The announcement does not provide an independent benchmark for reconciliation accuracy, exception detection, or reviewer effort.
Controller’s lens. Editorial assessment: evaluate one named agent and its enabled actions, not the whole platform vision. Before subscribing, confirm region, prerequisites, permissions, read/write scope, approval behavior, failure handling, and retained evidence. Marketplace activation is an installation milestone; acceptance requires showing that the agent performs the intended workflow and respects the organization’s controls. An Assurance Agent’s risk flag is an input to review, not an audit opinion.
IRIS: document retrieval becomes part of the accounting workflow
What changed. IRIS launched Document Management for accounting practices and professional-service firms in the US and UK, standalone or with IRIS Firm Management. It describes plain-language search, scanned-document search, tagging from a firm’s library, filing rules, version control, and a client portal. IRIS also states that regional AI processing, no model training on client data, retention rules, legal holds, and audit trails are included in the product design.
Evidence status
Vendor-documented launch and control claims, not a verified deployment or legal-compliance assessment. The announcement supplies no measured retrieval accuracy, filing-error rate, or independently tested retention outcome. Its focus is finding, reading, tagging, and managing documents; further AI action on their contents is described as a next step.
Controller’s lens. Editorial assessment: this is relevant to close support and technical-accounting research even though it is aimed at firms. A search result should identify the correct entity, period, version, and source passage. Finding a contract does not establish its accounting treatment. In a trial, plant a superseded agreement and a contradictory amendment, then check whether the reviewer can reconstruct which evidence was used and preserve it under the approved retention policy.
Experiments worth running
Use synthetic or appropriately anonymized data. These are proposed tests, not completed experiments.
Saved-report test: does a follow-up preserve scope?
Test: In an authorized sandbox, build a vendor-spend report from synthetic data for one entity and a fixed period. Ask for a monthly breakdown and a formula row. Reopen the saved report, then rerun it for a second period with independently calculated totals.
Evidence to retain: Retain the original request, successive report definitions, entity/date/currency settings, account filters, formulas, source IDs, exports, expected totals, and reviewer corrections.
Success criteria: All totals reconcile to the independently defined population; follow-ups change only the requested settings; the saved report can be reopened and reproduced. Record any manual repair and confirm actual plan and permission requirements.
Code-backed skill test: a holdout period with difficult allocations
Test: Use synthetic shared costs and an approved allocation policy. Validate against a closed-period fixture, then test an unseen period with fractional cents, a missing department, a zero denominator, and a changed allocation basis. Run only in a sandbox without production posting.
Evidence to retain: Retain the policy version, input snapshots, script version, builder checks, independent expected results, exception output, and approval record. Separate arithmetic tests from checks that the policy and source population are correct.
Success criteria: Allocated amounts sum exactly to the source pool; missing or invalid inputs block completion; changes to the basis require the defined review; repeated runs do not duplicate output. Passing the original historical fixture alone is insufficient.
Document-evidence test: can the reviewer find the governing version?
Test: Create a synthetic set of contracts, amendments, scanned invoices, and a superseded memo for two entities. Include similar filenames and an amendment that changes a key term. Test search and filing under both an authorized and a restricted user.
Evidence to retain: Keep the expected document/version register, access settings, search requests, returned passages and links, filing decisions, version history, and trial evidence for retention and legal-hold behavior.
Success criteria: The correct entity and current document set are retrieved; superseded and contradictory support is identified; restricted documents are not exposed; reviewers can trace each proposed conclusion to preserved support. Document any controls that could not be tested.
The practical opportunity is to turn an AI interaction into a reusable accounting artifact with a clear owner, stable inputs, visible rules, and independent checks. Start with a saved report or bounded allocation rather than assuming an entire close is automated. For a Rebuild Accounting write-up, publish the definition, expected answer, holdout cases, reviewer effort, and remaining gaps. These experiments are proposals, not completed tests. Prepared with AI assistance from sources checked on November 9, 2026; no personal testing, endorsement, or review by Samantha Ye is implied.
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