Finance professionals need more than a new prompt list. They need to work from source material, build useful outputs and take responsibility for the checks between the two.
What changed on September 10, 2026
OpenAI introduced ChatGPT for Financial Services for eligible financial institutions. It describes a tailored ChatGPT Work experience using GPT-6 Astra, with its initial focus on investment banking and equity research. It combines financial data, research, modelling and client-material preparation.
The announcement names built-in datasets from Daloopa, PitchBook, LSEG News and Crunchbase, alongside connections that depend on a firm’s existing subscriptions. It also describes detailed citations, firm document templates, access controls and supported audit-log exports. Availability through this institutional product does not mean every individual ChatGPT plan includes the same data or features.
AI + Excel and Sheets: review the workbook itself
OpenAI’s separate spreadsheet announcement describes AI-assisted model creation, updates, scenarios and cell-linked explanations. Its May 5 update states that ChatGPT for Excel and Google Sheets became generally available across plans. Tool access and workplace administrator settings still need checking.
The practical skill is inspecting the workbook after an AI edit. Can you trace a formula to the correct period? Did a forecast assume that revenue and cash receipts arrive together? Did an extra row fall outside a SUM range? A professional must answer those questions before sharing the model.
A useful exercise is to give AI an inherited budget workbook with three intentional faults: one hard-coded value, one broken reference and one inconsistent unit. Ask it to explain the model and propose changes, then independently confirm which faults it found and what it missed.
Financial modelling still needs finance judgement
A three-statement model is a set of accounting relationships. Generating formulas quickly does not remove the need to understand those relationships. Cash must reconcile. The balance sheet must balance. A debt schedule needs a defensible link to interest expense. Forecast assumptions must have a business reason.
Build the historical model first, then ask AI to help extend the forecast. Keep assumptions separate from calculations. Test base, upside and downside cases. Recalculate selected outputs independently. For valuation practice, explain the cash-flow definition, discount assumptions and enterprise-to-equity bridge rather than accepting a single attractive number.
The learning goal is a model another analyst can inspect, update and challenge. Model appearance and a confident AI explanation are not evidence that the formulas are right.
Research needs citations that actually support the claim
A company review should distinguish reported results, management guidance, analyst assumptions and the learner’s own judgement. A transcript is useful evidence of what management said; it is not proof that a forecast will happen.
Collect the filing, earnings release, investor presentation and transcript. Record publication dates and reporting periods. Ask AI for a cited earnings summary, then open the referenced passages. Check whether a figure is consolidated or standalone, reported or adjusted, and whether currency and units match.
Connected financial data can reduce retrieval work, but permission to access a source, current entitlements and source quality remain separate checks. A public-data learning project can teach the same evidence discipline without promising access to costly databases.
Investment workflows: build the memo, not just the answer
A practical analyst assignment can produce a company profile, comparable-company table, valuation snapshot, catalyst and risk map, and a short research memo. Require a section that argues against the proposed thesis. This makes unsupported certainty harder to hide.
Investment banking exercises can use public information to prepare buyer screens, meeting briefs, diligence questions and pitch drafts. Record gaps in transaction information explicitly. Do not fill missing terms with plausible guesses or imply that a training exercise used a proprietary banking system.
These are research and learning workflows. They do not provide a promise of investment performance or replace authorised professional advice.
Finance operations and compliance need different labs
The launch’s initial product focus is investment banking and equity research. The following finance-operations examples are AI5’s teaching recommendations, rather than a claim that the new product automates every process.
For accounting, reconcile invoices, bank records and ledgers using synthetic data. For FP&A, build an actual-versus-budget bridge and explain the largest drivers. For MIS, reconcile dashboard totals before drafting management commentary. For audit, link review notes to evidence and record exceptions for the auditor.
For AML and KYC, check identity matches, adverse-media sources and transaction context. Assess false positives and missed suspicious activity. An AI-generated case narrative or STR/SAR draft remains subject to authorised human review and the organisation’s regulated process.
Agentic finance: make the approval boundary visible
A reusable research workflow can collect approved documents, extract figures, calculate ratios, assemble citations and prepare a review pack. A monthly-close assistant can identify unmatched items and draft commentary. These are useful learning targets because the output can be checked against known inputs.
Write down what the workflow may read, what it may change and when it must stop. Missing sources, stale figures and failed calculations should produce a review request. Test that behaviour deliberately. Keep action logs and make any external write or submission depend on explicit authority.
As AI systems take on longer sequences of work, finance professionals need to inspect intermediate results as well as the final memo. A polished final document can conceal an early source-selection error.
A practical skills plan for 2027
The following plan is AI5’s recommendation for professional development, not a prediction about every finance job. Pick one workflow from your own role and improve it through repeated builds and reviews.
- Create a source register with dates, units, periods and access permissions.
- Build a finance workbook and explain each important formula.
- Use AI to analyse a document pack, then verify the citations and calculations.
- Produce a role-specific output: forecast, research memo, audit paper or KYC assessment.
- Add a repeatable workflow with failure checks, logs and a human approval gate.
- Keep before-and-after examples showing where your judgement corrected AI.
Research sources
Primary publications used for this article:
Common questions
Is ChatGPT for Financial Services included in AI5 courses?
No. AI5 does not promise access to the institutional product or paid financial databases. Labs use public or approved data and the tools confirmed for each cohort.
Should a finance learner start by learning to code?
Not necessarily. Begin with finance knowledge, spreadsheets, source-based AI work and verification. Coding can be useful for some workflows, but it is not a prerequisite for these professional application courses.