AI bank reconciliation automation for a group in Berlin: cost in 2026

The verdict in three sentences

For a Berlin group of 9 companies processing 14,000 bank lines a month, manual reconciliation eats most of the time of 3 accountants and pushes the monthly close to day 12. An off-the-shelf module (EUR 300 to 1,200 a month) is enough while flows stay simple, but a custom rules-plus-AI automation connected to the ERP (EUR 15,000 to 35,000 excl. VAT) pays off once bulk payments, multi-entity direct debits and incomplete bank references dominate. The realistic 2026 target: 85 to 95% of lines matched automatically and a close brought down to day 5, in a 6 to 10 week project.

Why reconciliation blocks a group close

In a multi-company group, the difficulty is not raw volume but heterogeneity. The same bank sends CAMT.053 statements for some entities and MT940 for others. Customers pay several invoices in a single transfer, sometimes with a reference truncated to 35 characters. Intercompany flows (recharges, cash pooling agreements, management fees) create mirror entries that must be matched on both sides.

The typical pattern in mid-sized finance departments: 60 to 70% of lines are trivial (recurring debits, bank fees, payroll), but the remaining 30 to 40% absorb 80% of the time. That is exactly where AI helps, matching by similarity of amount, date, counterparty and description, then proposing a match the accountant approves in one click.

Off-the-shelf module or custom automation: comparing 2026 costs

Reconciliation modules built into accounting software or sold as SaaS work well on one or two entities with standard flows. Their limit shows up on group-specific rules: intercompany allocation keys, cost centres, factoring, escrow accounts. Custom work means building a business rules engine, a probabilistic matching model and an approval interface, plugged into the existing ERP (SAP Business One, DATEV, Microsoft Dynamics, Sage) and into EBICS statements.

The tipping point usually sits around 5,000 lines a month and 4 entities: beyond that, the cumulative 3-year subscription approaches the cost of custom work without covering complex cases. To scope this kind of project, our custom web applications page describes the method, and the automation and AI page covers the technical building blocks.

What the specification must include

Three elements drive the final match rate. First, input data quality: automatic statement retrieval via EBICS T or TS, normalised descriptions, a deduplicated counterparty master. Second, a two-layer engine: deterministic rules for certain cases, then a model that outputs a confidence score for ambiguous ones. Third, full traceability for the auditor: every automatic match must keep the rule or score that triggered it, the user who approved it and the timestamp.

Bank data stays hosted in Europe, and the AI model only receives pseudonymised descriptions where possible. A configurable confidence threshold (for example 0.92) decides what goes to automatic matching and what goes to the approval queue.

Mini case study

Sarah, head of accounting at a technical distribution group of 9 companies in Berlin, manages 3 accountants who spend about 310 hours a month on reconciliation. At a fully loaded cost of EUR 48 per hour, that is close to EUR 14,900 a month. After deploying a custom automation at EUR 28,000 excl. VAT, the automatic match rate reaches 91% and time drops to 90 hours a month, freeing 220 hours, about EUR 10,560 a month. Adding EUR 600 a month for hosting and AI calls, the net gain is about EUR 9,960 a month: the project pays back in under 3 months, and the close moves from day 12 to day 5, so the executive committee gets its reporting a week earlier.

FAQ

What automatic match rate can we realistically target?

On standard group flows, 85 to 95% after 2 to 3 months of rule tuning. In the first month you often see 70 to 80% while thresholds and the counterparty master are adjusted.

Can AI match lines without human review?

Yes for lines above the confidence threshold, usually 0.90 to 0.95. The rest, 5 to 15% of volume, goes to approval, and every decision stays logged for the audit.

Do we need to change ERP to automate reconciliation?

No. The engine connects to the existing ERP through API, file import or a database connector. Integration often represents 25 to 35% of the total budget.

What does it cost to run after go-live?

Count EUR 150 to 450 a month for hosting and EUR 80 to 250 for model calls on 14,000 lines. Plan ongoing maintenance for new banks or entities.

How long does delivery take?

Between 6 and 10 weeks: 2 weeks of scoping and statement collection, 3 to 5 weeks of development, then 1 to 3 weeks of testing in parallel with a real close.

Let's scope your project. Send us your number of entities, banks and monthly lines, and we will price the scope, an indicative budget and a rollout plan ahead of your next close. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.