Event Recaps

What's in Your Stack? Carsten Gerger, VP Finance at LucaNet

Luc Hancock CFO Connect

Most finance leaders talk about optimising their tech stack. Carsten Gerger, VP Finance at LucaNet, actually built one — from scratch, across 20 entities and 15 countries, to support €180M in ARR growing at 30% per year. In this CFO Connect session, he opened the hood entirely: every tool, every integration, every AI workflow, and every mistake along the way.

TL;DR

One tool per category, all entities on the same stack: LucaNet uses the same ERP, CRM, spend management, HRIS, and banking platform across every entity — and it took years to get there
Integration is everything: the difference between a good tech stack and a great one is whether the tools talk to each other; Workato replaced Power Automate for Salesforce-to-Business Central and now drives 99% invoice automation
Spendesk replaced a scattered procure-to-pay landscape: procurement, AP, credit cards, and expenses are now unified — the clarity this gives the team on cash position and supplier payments is cited as one of the biggest wins
Data first, AI second: Carsten invested in Fivetran, Snowflake, dbt, and Lightdash years before using AI — and that foundation is what makes Claude Code and MCP actually work
AI must be embedded, not bolted on: the vision is AI inside specialist tools (Spendesk, LucaNet, Hibob) — not a standalone chatbot that produces numbers you cannot reconcile

Key Takeaways

A scalable CFO tech stack is built on standardisation: one ERP, one CRM, one spend management platform, applied consistently across all entities
99% invoice automation is achievable — but only once you replace brittle middleware with purpose-built integration tooling
Spendesk gave LucaNet's 40-person finance team visibility over every supplier payment and employee expense across 14 countries, with a daily sync to Business Central
A clean semantic data layer on top of Snowflake is what makes Claude Code finance querying via MCP actually reliable
Carsten's AI rule: if you cannot reconcile the number and explain it to an auditor, it is not ready to use — 80% accuracy is not acceptable in finance

The diagram shows every tool in the LucaNet finance stack, the integrations between them (file import vs API), and which workflows have an AI layer applied.

1. What does a scalable multi-entity CFO tech stack actually look like?

The foundation of LucaNet's finance stack is a single principle: every entity runs on the same tools. No local ERPs, no shadow systems, no regional workarounds.

The context that shapes the stack

Carsten joined LucaNet in 2016 and built the finance function from zero. Ten years later, the team manages:

20 entities across 15 countries
~€180M ARR growing at approximately 30% year-on-year
40 finance staff covering order to cash, procure to pay, general ledger, group accounting, data, and FP&A
A headquarter in Berlin with a fully centralised finance function

According to the CFO Connect Top CFO Tools Report 2025 (survey of 253 finance leaders across Europe and the US), finance teams at comparable scale typically run 8 to 12 distinct tools across their core categories. The challenge is not acquiring tools. It is making them work together.

The standardisation thesis

All of our entities are in Business Central. All of our entities are on Salesforce. We don't have any other accounting systems or any other CRM systems. — Carsten Gerger, VP Finance, LucaNet

This sounds simple. It is not. Maintaining standardisation across entities in China, Singapore, the US, France, Spain, and Germany — each with their own compliance requirements, banking relationships, and payroll complexity — requires constant discipline.

Carsten's framing: every tool decision is made for now, but constantly re-evaluated. The stack is not fixed. It is actively managed.

The stack is not a one-time build. It is an ongoing set of decisions that must be revisited as the business changes.

What to do this quarter

Map every tool your finance team currently uses across all entities — including local or shadow systems not centrally approved
Identify any category where different entities use different tools for the same function
Prioritise one category to standardise in the next 90 days — start with the one causing the most reconciliation work
Document your standardisation principle explicitly: it will save the argument every time a new entity wants to introduce a new tool

2. How do you automate order to cash across multiple entities?

LucaNet achieves 99% invoice automation by connecting Salesforce to Business Central via Workato — replacing a Power Automate integration that was not working reliably enough.

The order to cash flow, step by step

Software is delivered to the customer
A workflow triggers automatically in Salesforce
The invoice dataset is pushed via Workato to Business Central
Business Central sends the invoice to the customer the next day
One human performs a final sense-check — the only manual step remaining

The result: 99% of invoices require no human intervention. The finance team has effectively removed itself from the invoice creation and sending process entirely.

The lesson from Power Automate

The original Salesforce-to-Business Central integration used Microsoft Power Automate. It did not work reliably. LucaNet switched to Workato two years ago — a purpose-built integration platform that is consumption-based, meaning cost rises with volume. Carsten is honest about the friction:

As a finance guy, I'm struggling always a bit — how do I predict future usage and cost? But essentially, it's still working. We have super good performance with no technical issues.

The e-invoicing challenge ahead

Business Central does not support e-invoicing natively in all markets. LucaNet uses B2B Router for Spain and China. Germany — home to 5,000 LucaNet customers — goes live with e-invoicing requirements next year. A significant integration project is already on the roadmap.

99% automation is achievable — but only once you stop accepting that a brittle integration is good enough.

What to do this quarter

Map every step in your current order to cash process that involves a human touch
Identify which of those steps is a genuine control check versus a workaround for a broken integration
Evaluate whether your Salesforce-to-ERP middleware is purpose-built or a compromise — if it is the latter, price Workato or a comparable tool
Begin your e-invoicing compliance assessment now if you operate in Germany, France, or Spain

3. How does Spendesk fit into a modern multi-entity finance stack?

Carsten replaced a scattered, disconnected procure-to-pay landscape with a single Spendesk implementation covering procurement, accounts payable, credit cards, and expenses across 14 countries.

⭐ Spendesk spotlight

Before Spendesk, LucaNet's AP and expense processes ran across disconnected tools with no single source of truth on what had been paid, what was pending, or where a given expense sat in the approval chain.

We implemented Spendesk just recently. Before, we had a bit scattered landscape — different tools, not connected. We decided last year to just have one tool. We're happy with it — procurement, invoices, credit cards, expenses — which gives us a lot of headspace and clarity for reporting. — Carsten Gerger, VP Finance, LucaNet

The full platform is in use. Spendesk handles the complete procure-to-pay workflow inside the tool. At the close of each day, transaction data is exported automatically to Business Central — keeping the balance sheet, creditor ledger, and cost centre reporting permanently current in the ERP.

The result is not just process efficiency. It is organisational clarity. Employees, suppliers, and the finance team all have visibility into where a payment or expense sits — something that did not exist before.

Spendesk is live across most LucaNet entities. China is excluded due to local requirements. For the US and other international entities, payment functionality is live where Spendesk holds a banking licence.

Carsten's forward-looking view was clear: the next step is AI embedded natively inside specialist tools like Spendesk — where the system already understands the procure-to-pay workflow — rather than relying on a standalone AI tool running alongside it. That embedded context is where the real productivity gain will sit.

For teams evaluating spend management as part of a broader stack build, Spendesk's Modern CFO Finance Tech Stack Map shows how spend management connects to ERP, banking, and data systems in practice. The CFO Connect Top CFO Tools Report 2025 covers what 250+ finance leaders are using across AP, spend, and ERP categories today.

What to do this quarter

List every tool currently touching procure-to-pay — including email approvals, shared inboxes, or spreadsheet trackers
Calculate how long it takes to answer: has this supplier been paid? — if more than 60 seconds, you have a visibility problem
Assess whether your current AP tooling integrates natively with your ERP, or whether data is moved manually
Map what full platform coverage would look like: procurement, AP, credit cards, expenses, and daily ERP sync — and score each current tool against those five

4. How do you manage payroll and HR across 15 countries?

LucaNet is replacing 10 separate local payroll providers with a single Lano implementation sitting natively on top of Hibob — cutting relationships, systems, and reporting cycles from 10 to one.

The tools

Hibob (HRIS, 2.5 years in use): replaced Personio; chosen for stronger reporting functionality and team-level transparency across entities
Lano (payroll, currently implementing): consolidates payroll across all international entities; native Hibob integration; data flows from Hibob into Lano, which processes payments to local payroll accounts in each country

The problem this solves

Before Lano, changing anything cross-entity — a compensation structure, a headcount report, a benefit — required coordinating with 10 different advisers across 10 different systems.

If you want to have a cross-entity report, that was slowing us down because we needed to speak to 10 different tax advisers. Now we are implementing Lano as one payroll tool. The first entities are live and it is helping us massively.

What to do this quarter

Count the number of payroll providers currently active across your entities
Assess whether your HRIS and payroll systems share data natively or require manual export and re-entry
If you have 3 or more payroll providers, build the business case for consolidation — the admin saving alone typically justifies the switch
If you are on Personio and finding reporting functionality limited, evaluate Hibob as an alternative

5. How do you build a finance data platform that supports AI querying?

LucaNet built its data stack four to five years ago — and that foundational investment is the reason Claude Code finance workflows and MCP integrations now actually work for the team.

The data stack

The platform is built in layers, each with a distinct job:

Fivetran pulls data from all source systems into the warehouse
Snowflake serves as the central data warehouse for all non-financial data
dbt transforms and models raw data into clean, structured outputs
Lightdash provides a self-service BI front-end for non-technical users across sales, marketing, and revenue operations
A semantic layer built on top of Snowflake standardises how every key metric — ARR, churn, discount — is defined and calculated across the business

Lightdash replaced Tableau because Tableau required the data team to build every dashboard — creating a permanent bottleneck. With Lightdash, business teams build their own dashboards. The data team controls the environment and the metric definitions, but end users own their own reporting.

Why the semantic layer is the critical piece for Claude Code finance querying

The semantic layer is what most teams skip — and what separates plausible AI answers from trustworthy ones. Without it, connecting Claude Code to your data warehouse via MCP produces fast answers that cannot be verified. With it, every query maps to a defined, auditable metric.

We connected Lightdash and Claude via MCP. Users ask: what is our new ARR in Singapore in December 2025? And the system gives a reliable answer — because we invested in the data architecture and the semantic layer. The data is clean, trained, and easy to access.

Carsten's warning is direct:

If you haven't invested in clean data, you can't use a semantic layer. You can't use MCP. Everything doesn't work. But if you built that foundation years ago, then you can.

Only 35% of CFOs believe their current tech stack meets their needs, according to a 2025 industry survey — and the most commonly cited gap is data visibility and reporting quality. That is exactly the problem the LucaNet data stack was built to solve.

For a view of what Claude Code for finance looks like at the application layer — including a fractional CFO who built a full three-way FP&A portal in six weeks — the Claude Code FP&A application build recap covers the method and validation approach in detail.

What to do this quarter

Assess whether your current BI tool requires the data team to build every report — if yes, it is a bottleneck, not a resource
Check whether your key metrics have consistent, documented definitions across all systems
Begin building a semantic layer — even a simple shared document — that defines each KPI before investing in tooling
Prioritise data cleanliness over prompt design: Claude Code finance query quality is a direct function of input data quality

6. How do you use AI in a finance tech stack without sacrificing accuracy?

Carsten is LucaNet's heaviest Claude user outside of R&D. His rule: if you cannot reconcile the number and explain it to an auditor step by step, it does not go in the report.

Two AI workflows in production

Workflow 1: AI analyst agent for monthly close An AI agent checks consolidated numbers during the monthly closing process — flagging accounts that look odd, surfacing actuals-versus-budget variances, and highlighting items that need review. It is a first-pass filter that reduces manual review time, not a replacement for the financial controller. This is the kind of agentic AI finance workflow that operates autonomously in the background and surfaces only what needs human attention.

Workflow 2: Claude Code with Lightdash via MCP Users connect Claude to Lightdash via MCP and query live ARR and operational data in natural language — a Claude Code finance use case that works because the semantic layer ensures every question is interpreted against a consistent, auditable metric definition. The system does not guess at what ARR means. It knows.

The accuracy standard Carsten will not compromise

Finance needs to be precise. If I hand over to my auditors numbers where the answer is I think they are 80% correct — no auditor will accept that. And I'm also not trusting numbers that are only 80% correct.

His reconciliation protocol for every AI-generated number:

Ask the AI to show the calculation — method, not just result
Ask for multiple examples so the approach can be understood, not just the output
Cross-check against a known anchor point you already trust
If you cannot explain to an auditor exactly how the number was produced, it does not go in the report

In the end, I'm still responsible for the numbers. No auditor will accept Claude did it. It's mine.

The bigger vision: AI embedded in specialist tools

The most important observation Carsten made about AI in finance is not about what tools he currently uses. It is about where he wants AI to live going forward: inside the specialist tools that already understand the workflow context, not alongside them as a separate layer.

A standalone Claude Code finance assistant that requires context to be re-explained every session is not the end state. An ERP, a spend management platform, or an HRIS that surfaces AI-driven insight directly inside its own workflow — with full knowledge of how that workflow operates — is.

The State of AI in Finance 2026 report supports this: the highest satisfaction rates from AI in finance come from teams using it embedded inside specialist tools, not via standalone AI assistants.

The most dangerous AI output is the one that is wrong but sounds exactly right. Build your reconciliation process before you build your prompts.

What to do this quarter

Define your AI accuracy standard explicitly: what confidence level is required before an AI-generated number enters a report?
For any AI workflow in use, document the reconciliation step in writing — how does a human verify the output?
Test Claude Code finance querying on historical data where you already know the answers before using it for forward-looking decisions
Ask your ERP, spend management, and HRIS vendors directly: what agentic AI finance features are on the roadmap, and when?

7. How should a finance tech stack evolve as the business scales?

Every tool decision LucaNet makes is provisional. The principle is not get it right once. It is make the best decision now, and check it again in six months.

The three forcing functions that drive stack changes

1. Entity growth Adding a new country is not just an HR and legal task. It is a question of whether existing tools support that market — banking licences, payroll compliance, e-invoicing, local tax requirements.

2. Volume growth At a certain invoice volume, Power Automate breaks. At a certain data volume, Tableau becomes a bottleneck. At a certain payroll complexity, local providers become unmanageable. The stack that works at €50M ARR may not work at €180M.

3. AI capability shifts The tools that existed five years ago did not have AI features. The tools today do. Claude Code finance automation capabilities that required engineering effort in 2024 are available to finance teams directly in 2026. A stack decision made today is partly a bet on which vendors are embedding these capabilities fast enough to stay relevant.

Every decision we make, we make it for now — but we constantly check: is that still valid, or do we need to change something?

What to do this quarter

For each core tool, ask: at what point would this stop working as we grow — in volume, entities, or complexity?
Identify the most fragile integration in your stack — the one that breaks most often or requires the most manual intervention. That is where to invest next.
Map where your current vendors are on agentic AI finance feature development — if a vendor cannot give you a clear roadmap, treat that as a risk to manage
Set a six-month review cadence for the stack: a structured check on whether each tool is still the right tool for its job

A simple 90-day finance tech stack audit plan

Days 1 to 15 — Map the current state

List every tool in use across all entities, including unofficial or local tools
Document every integration: which tools send data to which, and how (API, file export, manual)
Identify every category where different entities use different tools for the same function
Flag every manual step that exists because an integration does not work properly

Days 16 to 45 — Prioritise and plan

Rank integration gaps by the manual work they create
Build the business case for fixing the top two — time saved, error rate reduction, reporting quality improvement
Engage your top two vendors on their AI roadmap — what Claude Code finance features or MCP integrations are coming, and when?
Define your data quality standard: what does clean enough for AI look like for your most important metrics?

Days 46 to 90 — Execute one thing well

Fix the most fragile integration in your stack — not the most interesting one, the most painful one
Implement one AI workflow on historical data where you can validate the output before going live
Document the reconciliation process for that workflow before running it on anything forward-looking
Present the result to your team: one integration fixed, one AI workflow validated, one tool decision reviewed

FAQ: CFO tech stack building

What tools does LucaNet use in its finance tech stack? As of 2026: Salesforce (CRM), Business Central (ERP), Workato (integration middleware), Spendesk (spend management), Hibob (HRIS), Lano (payroll), JPMorgan (banking), Fivetran, Snowflake, dbt, and Lightdash (data platform), plus LucaNet's own software for financial consolidation and operational planning. Claude is used as an AI layer for data querying and monthly close analysis — including Claude Code finance workflows via MCP.

How does Spendesk integrate with Business Central? LucaNet uses Spendesk for the full procure-to-pay workflow: procurement, invoice approval, payment, credit cards, and expenses. Transaction data is exported from Spendesk to Business Central daily, keeping the balance sheet, creditor ledger, and cost centre reporting permanently current in the ERP.

What is the difference between Lightdash and LucaNet in the reporting stack? Lightdash handles all non-financial reporting — ARR, pipeline, people data, revenue operations. LucaNet handles all financial reporting — consolidated P&L, balance sheet, cash flow, IFRS group accounts, and operational planning. The boundary is financial versus non-financial data.

How does LucaNet use Claude Code and MCP in finance workflows? Claude is connected to Lightdash via MCP, enabling natural language queries against live ARR and operational data — a Claude Code finance use case where the reliability comes from the semantic layer underneath, not just the AI model. A separate AI analyst agent assists with monthly financial close by flagging variances and unusual accounts. All outputs are reconciled against known data before use.

How do you ensure AI accuracy in financial reporting? Carsten's protocol: ask the AI to show the calculation method, not just the result; cross-reference against a known anchor; test Claude Code finance queries on historical data before using for forward-looking decisions; only use outputs that can be explained step by step to an auditor. The standard is full reconcilability — not 80%.

What should a finance leader look for in a spend management platform? Based on LucaNet's experience: full platform coverage (procurement through to payment, expenses, and credit cards); daily automatic ERP integration; multi-entity and multi-currency support; and a vendor AI roadmap that embeds agentic AI finance capabilities into the workflow. See Spendesk's overview of the best spend management tools for 2026 for a structured comparison.

How do you manage a finance tech stack across 15 countries? Standardise wherever possible: all entities on the same ERP, CRM, and spend management platform. Handle local exceptions with targeted add-ons rather than separate systems. Centralise the finance function and review every tool decision against a consistent standardisation principle.

What is Claude Code for finance? Claude Code is Anthropic's agentic coding tool, increasingly used by finance teams to build custom FP&A applications, automate reconciliation workflows, and connect data sources via MCP without engineering support. LucaNet's use of Claude connected to Lightdash via MCP is an example of this applied in a live finance stack. For a deeper look at Claude Code finance use cases, see the CFO Connect Claude Code for finance recap series.

Closing thought: The best finance tech stack is the one your whole team actually uses

Carsten Gerger did not build LucaNet's finance stack by chasing the best-reviewed tools on G2. He built it by asking one question repeatedly: does this tool solve the problem, integrate with what we already have, and work for every entity on the same basis?

The result — 10 years in the making — is a 40-person finance team managing 20 entities across 15 countries, with 99% invoice automation, a fully self-service data layer, and Claude Code finance workflows that are live and reconcilable.

The tools are not the point. The principle behind them is: one tool per category, all entities on the same stack, every integration reliable enough that a human only touches it when something genuinely needs judgment.

That is not a technology problem. It is a decision-making standard applied consistently over a long time.

And now, with agentic AI finance capabilities beginning to embed themselves inside specialist tools rather than sitting alongside them, the finance leaders who have already done the hard work of standardisation and data cleanliness are the ones who will extract the most value from what comes next.