deepset built a finance brain on Light

Around the ledgers sat a second set of duplicates. Spendesk ran expenses and AP in Europe, Ramp did the same job in the US, Lexoffice handled AR invoicing, and each entity banked with its own institutions, so two entities meant two versions of the same stack.

Month-end took nearly four weeks. For a company that builds AI agents for a living, that pace no longer fit the business it served.

The team signed in August 2025, which makes deepset one of Light’s earliest customers. They launched the US first, the smaller entity with fewer customers and employees, and went live in under two months with a hard stop on QuickBooks. Germany followed in January 2026 with a hard stop on DATEV, so the old and new systems never ran side by side.

A shared Slack channel kept both teams in regular contact through implementation. When deepset flagged something missing, or a setup it didn’t understand, Light turned it around fast.

DATEV dominates German accounting, but it has no natural partner tool in the other markets a growing company enters. Every new entity meant new data formats to stitch together and maintain. On Light, deepset runs every entity in one ledger, then exports a DATEV file at year-end so its tax advisor can handle the annual filing, which is now the only point in the year the advisor touches the accounts.

That first step grew into a full bookkeeping workflow. Once a week it reads every new card transaction and reimbursement in Light and assigns the account and cost centre with fixed rules, the way any bookkeeping process works. When the rules can’t decide, an AI layer reads the description and the invoice, recommends a posting and sends it back to Light, where Maren checks and approves it.

The workflow works because the data underneath it holds up. deepset set up its cost centres in Light and linked them to its HR tool, wrote a description for every account, and loaded its expense and travel policies. Light’s own agents use that context to run the first checks on reimbursements, and deepset’s external workflows use the same context through the API.

Employees get 24 hours to upload a receipt. After that, an automatic Slack message asks for it, and the employee uploads it through Light’s Slack connection or snaps it on the mobile app, where OCR reads and categorises it. deepset also moved expenses and AP from Spendesk and Ramp into Light, and the team expected pushback that never came, because the new flow felt familiar.

Two countries mean two sets of travel rules. deepset runs the Per Diem and Mileage apps from the Light App Store, so every expense claim uses the right local rate without a custom build. The team is now looking at the German Audit Export app for tax audits, and the Remote and Personio apps to bring people data into the ledger.

For software spend, a workflow calculates the variance against thresholds deepset sets itself. Anything inside the threshold passes with no comment, and anything outside it triggers an AI review of the invoices in Light, which explains the jump, such as ten extra seats on one tool. The team spends its time on the decision the analysis points to.

deepset sells to enterprise and public sector customers, so its revenue carries nuances, including services it can only recognise on delivery. A signed contract flows from Salesforce into Light, which creates the invoice and builds the deferred revenue schedule. The team reconciles that against ARR, tracks professional services each month with its customer engineers, and layers renewals, late-stage deals and draft invoices in Light on top to produce one revenue forecast.

The controls stay where they were. No AI posts to Light, a person approves every entry, and the monthly balance sheet and cash reconciliations catch anything that goes wrong. deepset builds one workflow at a time, so nothing turns into a black box.

Alex and Maren each built workflows in Claude Code, and at a team offsite they found the problem: neither could reach the other’s work. They now keep a shared repo in GitHub that holds every finance skill and agent in one structure, mirrors the context in Light, and adds context from the CRM, HR and product data. Month-end has become a list of tasks, and the team goes through it one task at a time, asking which ones AI can take.

Alex had never touched GitHub until two months before this interview. What sold the repo was the audit trail and the versioning, and the test it sets for the team: if anyone left, a replacement could read the repo and run finance within a week.

The next target is a close of two to three days in 2027, and Alex sees a path to one. After the close comes forecasting that updates in real time, built on more connected data sources, so finance can sit inside the decisions on hiring, new markets and big enterprise deals. The finance brain in GitHub will also meet a company-wide effort at deepset to build a shared context layer, so every team reads the same definitions, starting with what “revenue” means.

For finance leaders who feel cautious about AI, Alex’s advice is to start with the task that eats the most team time, run a small experiment beside the current process, and write down how finance works along the way.