AI in the finance function: what changes now

OpenAI shipped ChatGPT for Financial Services, a product that takes on company research, model building and presentation work, the tasks usually done by junior bankers. A few weeks earlier Anthropic released ready-to-run agent templates for finance work, including closing the books at month-end. This is not another general-purpose chatbot. It is a signal that AI is moving into a specific function inside the company, finance, and doing it with data access, source citations and permission controls. For a board, the question is no longer „whether", but „which finance processes to hand to a machine, and under what control".

What happened

On 10 September 2026 OpenAI announced ChatGPT for Financial Services, a version of its enterprise product tailored to financial institutions. The tool researches companies, analyses financial data and generates presentations, the work that used to belong to entry-level bankers. Its built-in data comes from providers such as Daloopa, PitchBook, LSEG News and Crunchbase, indexed by OpenAI and marked with granular citations, so a figure can be traced back to its source. Morgan Stanley and Evercore were named as design partners.

Earlier, Anthropic released Claude for Financial Services and extended it with ten ready-to-run agent templates for the most time-consuming tasks: building pitchbooks, screening KYC documents and closing the books at month-end. The agents connect to FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG and Daloopa, as well as firms' internal systems, under governed access. They run inside Excel, PowerPoint, Word and Outlook through the Claude add-ins for Microsoft 365, and the whole thing sits on enterprise controls: SSO, SCIM, audit logs, configurable data retention and ISO/IEC 42001 certification.

Two launches in a short window point in the same direction. Model providers are moving away from selling a general assistant and toward selling a tool embedded in the data and processes of one function, finance.

Why it matters for business

Until now, AI in finance mostly meant an assistant you had to paste data into by hand, then check for anything it made up. Now the tool gets access to the sources and can carry a task through to the end, from pulling the numbers to a finished document with citations. That moves AI from an experiment to a part of the process.

The scale of the shift is measurable. Gartner predicts that by 2029 CFOs who deploy AI strategically will add ten margin points of growth. The same Gartner warns that around 40 percent of agentic AI projects will be scrapped by the end of 2027 over cost, unclear value and weak risk controls. Deloitte adds a governance maturity gap: companies want to deploy agents faster than they can control them.

For a company outside Wall Street, the takeaway is simple. Providers are building AI straight into the finance department, with data and controls, so technology stops being the bottleneck. The bottleneck becomes whether the company has clean data, documented processes and a person who decides what the machine may and may not do.

Business use cases

None of these are reserved for banks. They are processes any company with a finance team already runs, and the only difference is scale.

1. Month-end close and management reporting

Business problem: the close drags on for days, rests on manual reconciliations and copying data between spreadsheets, and the board report comes last, often late.

AI solution: an agent pulls data from the ERP and spreadsheets, reconciles balances, prepares draft statements and flags exceptions for a human to decide. Anthropic ships a ready template for month-end close.

Data and processes to put in place: a chart of accounts, reconciliation rules, defined materiality thresholds and a list of typical exceptions, plus controlled access to the ERP.

Potential effect: a shorter close and less manual work, with the finance team shifting from re-keying to analysis. Anthropic frames this as a deployment measured in days, not months.

Risk and limitation: if the data comes from several inconsistent systems, the agent produces output nobody trusts. A process owner and a quality check on results are both required.

2. Accounts payable and invoice processing

Business problem: matching invoices to orders and deliveries by hand, data errors and late payments.

AI solution: an agent reads invoices, matches them to purchase orders, checks amounts and terms, and routes exceptions to a person.

Data and processes to put in place: an order register, payment terms, approval rules and thresholds, access to the document workflow system.

Potential effect: fewer manual reconciliations, faster processing and fewer late payments.

Risk and limitation: a fraudulent or inflated invoice slipping through automation with no control on approval thresholds. Automation does not remove the need to authorise payments.

3. Forecasting and planning (FP&A)

Business problem: forecasts are built in Excel, slowly, and running a few scenarios takes an analyst a full day.

AI solution: an agent builds models, runs scenarios, states the assumptions and drafts commentary on the numbers. Both OpenAI and Anthropic advertise model and decision-material building.

Data and processes to put in place: historical financial data, explicit assumptions, metric definitions and access to the data warehouse.

Potential effect: faster scenario runs and more time to interpret results instead of building the spreadsheet.

Risk and limitation: a model can repeat a wrong assumption at scale. Without a controller verifying it, you get a bad forecast faster, not a better one.

4. Document review and counterparty checks

Business problem: reviewing contracts and counterparty documents is slow, and different people apply the criteria inconsistently.

AI solution: an agent gathers and summarises documents, catches discrepancies and packages a file for decision. Anthropic shows this as a KYC screener, but the same pattern works for vendor and customer checks.

Data and processes to put in place: a contract and document repository, a checklist, clear risk criteria and access permissions.

Potential effect: a faster review and consistent criteria regardless of who happens to handle the case.

Risk and limitation: the agent can miss context or assert a conclusion the documents do not support. Without an audit trail, that decision is hard to defend.

5. Research and materials for the board

Business problem: gathering market data and preparing a board pack takes an analyst several days.

AI solution: ChatGPT for Financial Services runs research on built-in data, marks it with citations and builds a deck in which every figure can be traced to its source.

Data and processes to put in place: access to reliable data sources, material templates and confidentiality rules.

Potential effect: less time to prepare the material and easier control over where each number came from.

Risk and limitation: a citation does not remove the need to verify, and feeding confidential data into a tool without a contract is a security problem, not a convenience.

What companies can do now

Steps that work for a small or mid-sized company, in this order:

1. Pick one process with repeatable steps and clear boundaries. Gartner advises CFOs to start with a pilot that has clear boundaries and to avoid high-consequence processes.
2. Measure the baseline. Find out what the process costs today in hours and money, because without that number you cannot judge whether the rollout paid off.
3. Clean up the data. Decide which system is the source of truth when the data disagrees, and assign that decision to a named person.
4. Set the agent's boundaries. Write down what data it can access, what actions it may take and where automation ends and a human decision begins.
5. Name a process owner. Someone has to be able to say „from Monday we do this differently" and own the result.
6. Measure the effect in numbers. Time, cost and error count, not just the fact that the tool works and people use it.
7. Handle security. SSO, audit logs, data retention and a clear ban on pasting financial data into consumer tools.

Risks

Data security: finance holds some of the most sensitive data in the company. Pasting it into a tool without a contract and access control is a potential leak, not an improvement.

Shadow AI: if the company does not provide a safe tool, employees will use their own. We covered this in our piece on shadow AI in Polish companies.

Cost: the agent bill grows with the number of calls and the volume of data. Without limits and monitoring it is easy to blow the budget, which we covered in AI cost control.

Wrong rollout: a tool added on top of a process instead of changing it increases the workload rather than cutting it.

No owner: when nobody owns the process, there is no one to decide what the agent may and may not do.

No quality control: a clean format and citations do not mean the numbers are right. Without a person verifying the result, the company reaches a bad decision faster, on the strength of a confident-looking answer.

Key takeaways

AI is entering the finance function with data access and controls. That is a shift from an assistant you dictate to, into a process operator you have to supervise. The value of that shift depends on clean data and processes, not on the choice of model. Start with one process, with an owner, clear boundaries and metrics, and put the controls in place before you extend agents to higher-stakes work. Governance before scaling, not the other way around.