Integrating AI into existing finance processes

Integrating AI into existing finance processes means plugging AI into the accounting package, MS365, banking portal, and Excel models already in place — not replacing the landscape, but picking connection points where manual work sits. For finance teams, the build-vs-buy choice (custom agents on MCP, or a platform like Saldus) determines whether the integration is audit-traceable or becomes a shadow-IT layer.

Most finance functions don't have a greenfield. There's an accounting package (Exact, Twinfield, AFAS), an MS365 environment, a banking portal, a reporting tool, and a set of Excel models where the real work happens. Integrating AI doesn't mean replacing this landscape — it means plugging AI into what's already there, at exactly the points where manual work sits.

Anyone starting with "we'll install a new AI platform and migrate our finance stack" gets stuck on change management, IT objections, and cost. Anyone starting with "where in our current close or AR flow is the manual work AI can take, and how do I connect AI to that step" is productive within weeks.

The three connection types

There are roughly three connection types between AI and existing finance systems. Anyone who understands these three understands any integration diagram.

1. Webhooks — "let me know when something happens"

A webhook is an HTTP callback: system A sends a message to system B as soon as something happens. "A new purchase invoice has come in," "a payment has been approved," "a debtor balance has crossed a threshold." Event-driven: the source system pushes, the receiving system reacts.

Typical finance use: a banking portal sends a webhook as soon as a new bank statement is available. The AI workflow picks it up, classifies transactions, and puts a match proposal in a queue. No polling, no delay.

2. APIs — "I fetch or write it"

An API is a set of rules letting software call other software. Unlike webhooks the initiative is reversed: your workflow actively asks the system for something ("give me the last 50 purchase invoices," "post this journal entry," "fetch the open balance").

Most modern finance tools (Exact, Twinfield, AFAS, Outlook, Teams, Slack) have public APIs. In tools like n8n, Zapier, or Make they're available as ready-made nodes; for custom work you write the calls yourself.

3. MCP — "every API as a tool for the model"

Model Context Protocol is an open standard that turns every API into a tool an LLM can call itself. Where with APIs and webhooks the workflow platform holds the steering wheel, with MCP the model can decide when to call a tool: "look up this customer in the books," "fetch the open invoices," "create a draft journal entry in the approval queue."

For finance, MCP is the run-up to agentic workflows. For deterministic close steps, webhooks and APIs remain the workhorse; for more flexible tasks (the Q&A agent answering questions about the books), MCP is the right tool. See What MCP is and why it matters for finance.

Where to hook AI in — the four integration points in finance

In order of implementation ease.

Mail — universal trigger and output in finance

Almost every finance process starts or ends with mail. Inbound purchase invoices, payment reminders going out, VAT correspondence with the tax authority, board updates to the supervisory board. And almost every reply goes out by mail too.

Pattern: mail trigger → AI processing → draft back in the Drafts folder. Outlook supports both triggers and draft creation via the Graph API. Manual sending stays with the human (HITL in the right place). Often the first productive AI workflow in finance.

Accounting — extraction, matching, and draft postings

The heart of finance AI. Inbound invoices, bank transactions, open items. AI is strong at structured extraction from unstructured documents and at pattern matching against historical postings.

Patterns:

• Inbound invoice in mail or folder → extract → draft posting: PDF invoice → vision + LLM pulls supplier, date, amount, VAT, and ledger category → entry in the accounting package via API, in an approval queue. HITL for amounts above a threshold.
• Bank-transaction matching: AI links unknown bank transactions to open items based on description, amount, and historical patterns.
• Periodic reconciliation: AI compares balance-sheet totals with returns (VAT, payroll) and flags discrepancies.

MS365 — knowledge, calendar, and reporting

A rich substrate of finance-relevant context: the close manual in SharePoint, board decks in Word, models in Excel, calendars for MT meetings, meeting notes in Teams.

Patterns:

• Meeting → CRM/Notion update: transcript from Teams or Fireflies → AI extracts finance-relevant actions and next steps → written to the right file.
• Excel model → analysis: file drop in SharePoint triggers a workflow that reads the model, runs an analysis, and generates commentary.
• Board-pack template → completed board pack: Word template + monthly numbers + draft commentary → assembled board pack via file creation.

Banking and payments — the most sensitive layer

Pulling bank statements, preparing payment batches, monitoring payment statuses. In the Netherlands: integrations via PSD2 or via the bank's own API. AI here always via HITL — no autonomous payments, not even at high confidence.

Patterns:

• Daily bank statement → match-proposal queue (see above).
• AP payment proposal: AI proposes a payment batch based on open payables, payment terms, and cash position; finance manager approves.
• Status monitoring: after a payment, AI checks whether the offset matched and flags discrepancies.

The integration canvas for finance

For every AI integration it pays to walk through these questions:

1. Where does the process start now? → trigger.
2. Which systems hold the data? → API calls or MCP tools.
3. Which systems should receive the result? → action nodes.
4. Where does a human need to sit? → HITL step (see Human in the loop in finance workflows).
5. Which credentials are already in place, and which do I need to request? → drives the lead time.
6. What happens if a step fails? → retry, notification, fallback. For finance: never fail silently. Failure is allowed, but must be visible.
7. Which audit log do you need? → design this from day one, not as an afterthought.

The last two questions get forgotten often. Production workflows fail — an API is temporarily down, a credential has expired, a record has been deleted. For finance, an unnoticed error in an automated posting flow is an audit finding in the making.

Not everything needs connecting

Integration costs maintenance. Every extra connector is another place where credentials expire, APIs change, or rate limits get hit. Three rules:

• Only integrate where time or cash is gained. An integration that saves 30 minutes once a month isn't worth the maintenance.
• Reuse existing connectors before adding new ones. A Saldus building block already providing an Exact connection is cheaper than a custom API integration.
• Keep the workflow visible to the user. If AI output appears in Outlook drafts or the Exact approval queue, everyone sees what's happening. If it disappears into a tool nobody opens, it doesn't get used.

Build vs buy

For many finance integrations, building isn't necessary. Product categories like AP automation tools, invoice OCR, AR platforms, and cash-forecast tools are pre-built workflows with good integrations. Building deserves attention only when:

• The use case is business-specific (custom chart-of-accounts logic, custom processes, no standard product fits).
• Volumes are substantial (more than 200 invoices per month — below that SaaS is cheaper).
• Data governance requires a custom environment (clients with strict customer-owned infrastructure requirements).

In all other cases: buy a tool, integrate via the standard connectors, build only what's unique to your finance function.