Which controls should come before AI-assisted finance work?
Start with source checks and permitted actions. Extracting the right amount from a document does not authorise a payment.
The experiences presented in Barcelona distinguish preparation by an assistant from approval by a person. That distinction gives a finance team a way to scope a first task with a result it can review.
What can AI prepare for a finance team?
AISB26 sessions provide examples of document preparation, analysis and follow-up. These eight uses distinguish what an assistant prepares from what a person decides.
How can AI prepare invoices and payment follow-up?
An assistant can prepare an invoice draft from an approved quote, contract or export. References, amounts and recipients must be matched to the source before issue. The accounts team needs a clear way to handle missing information.
One European fintech presentation described agents that prepare and recommend, with human approval for payments. A workshop can practise the same boundary using invented documents: prepare the item, trace each field and flag what is missing.
How can the team work on supplier invoices and expenses?
A first task is to extract the supplier, date, amount and entity, then compare them with a purchase order or expense policy. A discrepancy should remain visible and have a named person responsible for resolving it.
An “approved” statement printed on an invoice cannot authorise its own payment. Approval and payment details must be checked in the system designated by the organisation, independently of the document being processed.
How can people ask questions about company figures?
An analytical assistant needs approved metrics and sources. A response should let the controller trace the period, scope and definition behind a figure. If the information is absent, the assistant should identify the gap.
A planning case at AISB26 examined written business context: strategy, objectives, rules and definitions accessible to the assistant. This illustrates what a team needs to document before expecting useful answers from a model connected to its data.
How can AI help prepare forecasts and scenarios?
An assistant can help formulate an assumption and explain a variance using a governed planning model. Calculations stay in the system designed to perform them and must be reproducible.
The controller checks assumptions, units and scope. Well-written commentary does not establish that a forecast is correct. The explanation must point back to the data and modelling choices, with uncertainty kept visible.
How can AI prepare a procurement or supplier decision?
An assistant can assemble approved information about volumes, contracts and deliveries into a comparison brief. The procurement owner retains the decision and commitments made to the supplier.
Before connecting actions, the team needs to examine data freshness, permissions and stopping conditions. An exercise using invented orders can identify these requirements without granting write access to a production system.
How can AI support risk and compliance work?
An assistant can assemble documents, summarise relevant facts and identify missing material. The analyst must be able to trace the evidence and decide what happens next under the applicable procedures.
A central-bank presentation distinguished different uses and treated sensitive information separately. That experience does not provide general permission to send banking or personal information to an external service. The approved environment must be decided for the task.
How can AI prepare reporting and management commentary?
A controller can request a first draft based on a validated export and explanations provided by the team. The assistant must distinguish an observed change from a possible explanation.
A reusable format can specify the expected sources, variances to examine and questions to leave open. Human review then covers the figures, the explanations and the information omitted, before the note is shared.
How should finance measure the cost of an AI-assisted task?
The cost of the work includes licences, consumption, preparation, checking and rework. Compare it with accepted results, using criteria set before the trial. A generated response is not automatically a completed task.
An AISB26 keynote examined this question at organisational level. A finance team can start with a record of attempts, usable outputs and review time, without extrapolating a small experiment to the whole company.
Which safeguards need to be verifiable?
Safeguards need to be visible in tools and responsibilities. The table is a scoping proposal; it does not assert that every organisation mentioned used an identical system.
| Safeguard | Practical check |
|---|---|
| Human approval | Define who can authorise a payment, issue or amendment. |
| System of record | Check approval and account details independently of the processed document. |
| Approved metrics | Record metric definitions, reporting periods and scope. |
| Business context | Assign an owner for rules and updates. |
| Traces | Keep the information needed to investigate and recover from an error. |
| Sensitive information | Choose an environment consistent with classification and contract terms. |
| Ownership | Name a process owner and provide a way to stop actions. |
What should the team learn before using these methods?
The team needs to describe a task, prepare sources, write business rules and check the result. An accounting owner or controller supplies the criteria for accepting or rejecting an output.
A proposed WeGoTo finance workshop can start with an approved export and a task chosen with the CFO. The work is to document a use case, begin a business-context file and define approvals. Format, trainer and outputs are specified in the quote.
Which limits should the team keep in view?
A demonstration does not establish reliability. Test missing information, contradictory documents and reporting-period changes, then observe the corrections needed. A successful first example is a starting point for another test.
ERP access, sensitive information and the obligations of a regulated institution require an assessment specific to the organisation. This page presents uses and educational preparation. It does not provide accounting, tax or legal advice.
Results described at an event do not predict results for your team. Public references, sharing conditions and tool versions need another check before publication or practical use.
