Autonomous Finance and AI in SAP Finance

What AI changes in finance

A new division of labor between people and AI is emerging in finance. People set goals and make decisions. AI assistants understand requests, consolidate information and coordinate the next steps. AI agents carry out individual tasks, such as proposing accruals, reconciling intercompany balances, clearing open items, preparing dispute cases in receivables management or drafting payment reminders.

The work thus shifts from data entry and reconciliation to reviewing, deciding and explaining.

One limitation is essential: language models estimate, they do not calculate. Robust solutions therefore have amounts calculated by system functions. The AI selects the appropriate step and explains the result.

Where companies stand today

A recent survey of SAP users shows that 43% of companies have already implemented AI use cases. Around three quarters of them rely on solutions outside SAP, such as Microsoft Copilot. Only around 3% use SAP’s own AI productively so far. In addition, a good half of the companies still work with SAP ERP or ECC.

Another study confirms the picture: high expectations of future benefits, but so far little measurable return. At the same time, many companies that already achieve tangible benefits with AI consider their own data quality or infrastructure in need of improvement. Perfect data is therefore not a prerequisite for getting started.

The SAP standard in detail

With the Autonomous Enterprise presented in May 2026, SAP aims for people and AI agents to carry out core processes together. For finance, assistants have been announced for financial planning, closing, invoicing, receivables, payables, cash and treasury, governance as well as tax and compliance. They are operated via the AI assistant Joule.

According to the SAP roadmap, the finance assistants are to be available by the end of 2026. Many of the related agents are currently in the early adopter program, some are generally available, others are in preview. For the financial close, SAP holds out the prospect of cycles up to 20% shorter. This is a vendor claim that still has to be confirmed in each company.

An example from tax: before posting, an agent checks the tax details of an invoice, justifies its proposal and submits doubtful cases for review. Only anomalies and uncertain cases still need to be handled.

SAP justifies this approach with the system context: its own agents work directly in the system, with business logic, authorizations and audit trail. If, by contrast, only tables are passed to an external model, the meaning behind the data is easily lost.

Access is closely tied to SAP Cloud ERP. For companies with SAP ECC or S/4HANA on-premise, SAP makes selected AI capabilities available only if they commit to a move to the cloud. Billing is tiered: basic functions are included at no extra charge, while advanced functions and agents are billed via a consumption credit, the so-called AI Units. From the fourth quarter of 2026, many functions that were previously chargeable are to become available at no extra charge.

Custom agents and external AI

SAP itself relies on open standards. In Joule Studio, SAP interfaces can be provided as tools for agents via the Model Context Protocol (MCP). Via the Generative AI Hub, custom agents can also use language models from other providers, and via the open A2A standard, Joule works together with custom agents.

Outside the SAP platform, AI assistants such as Claude or Microsoft Copilot can be connected to SAP via an MCP server, in principle also to ECC and S/4HANA on-premise.

The architecture is decisive. If data is copied into an external AI tool, the SAP authorizations no longer apply. A cleaner solution is an MCP server that runs via SAP Business Technology Platform with secure sign-on and only allows what the respective user is also permitted to do in SAP.

Since April 2026, SAP’s API policy has defined which interfaces are permitted as “Published APIs” for such integrations. This question belongs at the start of every project. Equally important: SAP’s commitments regarding the protection of intellectual property apply only to SAP’s own models, not automatically to custom agents using external models.

Five stages to Autonomous Finance

1. Foundation: SAP S/4HANA or SAP Cloud ERP.
2. Clean data and harmonized processes: master data, chart of accounts and a shared data foundation for accounting and controlling.
3. Automation: workflows, rules and e-invoicing without manual re-entry.
4. AI assistance: Joule or other AI assistants answer questions and suggest steps.
5. Autonomous processes: agents handle routine work, the business team manages the exceptions.

Each stage builds on the previous one. Project experience shows that AI does not scale on an unstable or fragmented foundation. First come S/4HANA, a clean core without unnecessary custom developments (Clean Core) and the SAP platform, then Joule. The S/4HANA Finance migration is therefore often the first step.

Nevertheless, the path does not have to be strictly sequential: initial use cases can be piloted in parallel, assessed for maturity and measured in use, even if not everything is perfect yet.

Staying in control when AI posts

Organizationally and legally, one distinction is decisive: an assistant prepares, an agent acts. As soon as agents prepare or execute postings, clear approvals, an adapted internal control system and a complete audit trail are required.

These rules should be part of the concept from the outset so that external auditors and internal audit can trace what has happened at any time. I support business and IT in this.

Impact on business and IT

In finance, work is shifting from data entry and reconciliation to reviewing, steering and explaining. IT takes on new tasks: it determines which agent may access which data and systems, secures interfaces and keeps an eye on usage and costs.

I support business and IT in clearly defining roles and responsibilities and in building the necessary AI literacy.

Free initial consultation

Whether CFO, head of IT or consulting firm: in a free initial consultation, we discuss the system landscape, objectives and possible first use cases. You receive an initial, non-binding assessment.

If desired, an assessment of the current situation follows: which systems and data are available, which processes are suitable, and which approach fits the contract? We then start with one or two clearly defined use cases, such as matching incoming payments or intercompany reconciliation, and measure the benefit. This results in a realistic roadmap.

Frequently asked questions

As of October 2026. AI offerings, licensing models and contract terms are currently evolving very quickly. Individual details on this page may therefore already be out of date.