Engineering consultancy puts two AI agents to work in its back office

Situation

A building services engineering consultancy in Eschborn, 40 staff and about €4.5 million in annual revenue, had a back office of three people that could not keep up. They handled about 250 supplier invoices a month, chased project timesheets, and assembled tender documents for the partners. Invoices sat in a shared mailbox for four working days on average before they were booked, month-end close ran late, and the three staff were doing regular overtime. Hiring a fourth person would have cost about €55,000 a year.

Approach

We mapped the department’s week task by task and picked the two workflows that ate the most hours. Then we built one AI agent for each. This is the same pattern we use to run our own holding company, where software does most of the administration for one founder.

The first agent handles supplier invoice intake. It reads each PDF that arrives in the invoice mailbox, pulls out the supplier, amount, VAT, and project number, matches the invoice to the purchase order in the company’s ERP system, and prepares the booking proposal. A person approves each one on a single screen. The second agent assembles tender responses. Given a tender request, it pulls the standard sections, project references, and staff CVs from the document library and drafts a first version, with every place that needs a human decision marked.

Both agents run on a model hosted in the EU under a data processing agreement, so no invoice or CV leaves European servers. Weeks one and two went on design and data access. In weeks three and four the agents ran in shadow mode: they made their proposals while the staff kept working the old way, and we compared the two. In weeks five and six the agents went live, the three staff were trained on the approval screens, and we left a written runbook for the day something goes wrong.

Result

In this scenario two agents were live in the finance and administration department after six weeks.

Supplier invoice turnaround fell from about four working days to one. A tender first draft took about 40 minutes to review instead of half a day to write. Month-end close finished two working days earlier than before. Across the three staff about 14 hours a week were freed, which they moved to chasing overdue receivables, and overdue days fell in the following quarter.

In month two under 2% of agent proposals needed a correction, and every one was caught at the approval step. The overtime stopped, and the fourth hire was not made.

What it cost

The AI-native operations engagement cost €11,000, quoted as a fixed price before work started. That bought the department mapping, two agents designed and built, four weeks of shadow and live operation, staff training and the runbook. The model usage of about €60 to €90 per month is a third-party cost and is excluded. The work was covered by our 100% money-back guarantee: full refund on request within 14 days of delivery.

Every fixed-price engagement carries the 100% money-back guarantee. How the guarantee works →