Every finance team has one. It tends to have a name like Group_Recon_MASTER_v14_USE_THIS.xlsx, it lives on a shared drive with permissions nobody remembers granting, and it costs one person two working days a month to keep alive. When that person takes leave, the board pack slips. When they resign, a colleague spends a fortnight reverse-engineering a formula written in 2021 by a contractor who no longer answers email.
Most companies call that file a workaround. It is more useful to think of it as a translation. The ERP knows what was invoiced; the CRM knows what was sold; the two cannot agree on what a customer is. One holds legal entities, the other holds relationships. One closes a month, the other closes a deal. The spreadsheet is where a human quietly settles that disagreement, applying rules that exist nowhere else in the organisation — not in a data warehouse, not in a policy document, not in anyone’s job description.
That file is not evidence of a failure. It is a specification. And it is increasingly where the money is.
Investors have started paying for the plumbing
On 3 September, Munich-based Zeit AI announced a seed round of €4.3 million — reported by several outlets at around €5 million — for a platform called ZeitMind. The founders, Leopold von Waldthausen and Marvin Bornstein, are Palantir alumni. The backers are an unusually political mix: Y Combinator, the University of Oxford’s seed fund, a Sequoia scout fund, Hasso Plattner VC, and angels including Germany’s former finance minister Christian Lindner.
The pitch is deliberately unglamorous. ZeitMind connects to ERP, CRM, HR and more than 600 other source systems, cleans what it finds and builds the report on top, with results traceable back to individual transactions and processing kept in German data centres. Von Waldthausen calls it “the autonomous data engineer.” The target is explicitly the European mid-market: companies that reach for a spreadsheet precisely because their ERP or FP&A system cannot answer the question being asked. The company is planning a London office, and its early customers — an automotive components maker, an agricultural machinery firm, a grain trader, a business aviation operator — are exactly the sort of firm that has never employed a data engineer and never expected to.
What makes the raise interesting is not the category but the admission underneath it. After three years of AI budgets, the binding constraint in most mid-sized companies is not intelligence. It is that nobody can say, in machine-readable terms, what the business does between its systems.
The UK numbers are blunt about this
Britain is not short of spreadsheets holding the line. A survey by Odoxa of 303 CFOs at private UK companies with 250 or more employees, published in May, found that 67% still use Excel for account analysis and reconciliation: 57% for accruals, 54% for manual journal entries, 53% to run the close calendar itself. More than a third of those finance chiefs said they lacked high confidence in the reliability of their own figures.
In FP&A more broadly, over half of teams juggle ten or more reporting tools each quarter and 82% of companies say they act on stale information. Technology spending rises; the spreadsheet layer thickens underneath it. Buying another system has mostly produced another thing to reconcile.
Set that against the Office for National Statistics’ latest read and the shape of the problem sharpens. AI use among UK businesses with ten or more employees has tripled since late 2023 to 35% by June 2026, and 49% among large firms. Yet the average adopter runs just 1.6 AI technologies, up from 1.4, and only 10% describe their use as extensive. Adoption has spread sideways rather than deepened. The barrier medium-sized firms name most often is not cost but a lack of in-house expertise, followed by difficulty identifying a use case worth doing.
That last point is the one worth sitting with. Companies are not short of AI use cases. The best ones are simply undocumented. A task that lives in a job description is easy to spot. A task that lives in a file, maintained by one person out of habit and loyalty, is invisible to every strategy workshop ever run.
What the spreadsheet actually costs
Two days a month of a qualified accountant’s time is real money, but it is not the interesting number. The interesting cost is behavioural. In the same UK survey, 93% of CFOs said the close harms employee motivation and 86% said they do not get enough time to analyse. That is what a manual reporting layer buys: a finance function that spends its most experienced people on assembly and its least available hours on judgement. Ask a controller why they stayed late and the answer is rarely “to interpret the variance.” It is to make two extracts line up.
Then there is trust, which behaves differently from cost. Reconciliation done in a spreadsheet is auditable in theory and unexplainable in practice: the logic is legible only to its author, and the number reaching the board carries a person’s reputation rather than a system’s evidence. Which is why traceability, not speed, keeps appearing in the pitch decks. A finance director does not want a faster number. They want one they can defend in a room.
And there is the risk mid-market firms rarely name out loud: institutional knowledge concentrated in individuals. A company can run SAP, Salesforce and a data warehouse and still have its revenue reporting depend on whether one person in Solihull picked up the phone.
Where the agents actually fit
The practical move is smaller and duller than most AI strategies: treat the manual reporting layer as a discovery exercise rather than an embarrassment. Inventory the recurring files — the ones rebuilt by hand every week or month. For each, note which systems it draws on, what the human does in between, and which decision depends on the output. That list is usually the highest-quality AI backlog the company owns: every item has a known input, a known output, a defined frequency and a person who knows whether the answer is right. Those are the four things a failed pilot lacks.
The sequencing is unromantic. Fix the identity problem first — a shared definition of a customer, an account and a period across CRM and ERP — then let an agent handle extraction, matching and exception-flagging, and keep the human on the exceptions and the interpretation. Success is not that the spreadsheet disappears, but that it stops being load-bearing.
Implementation, not model choice, decides the outcome here. Getting an agent to read two systems is a demo; getting it to hold through a quarter-end, a pricing change and a staff departure is an operating discipline — monitoring, guardrails, escalation paths, someone accountable when the output looks wrong. It is why firms such as Riophany help businesses build their work around CRM, ERP and finance-system integration with managed operations rather than one-off deployments. In the mid-market, the gap between an AI pilot and an AI capability is almost entirely what happens after go-live.
The last observation
Zeit AI’s raise will be read as a story about European software catching up. It is more interesting as a comment on how companies find their AI opportunities — not through a transformation programme, but by looking at what their own people have been doing by hand for years.
The spreadsheet nobody wants to maintain has been documenting the gap between the systems all along. It is the most honest map of how the business really runs. Most organisations treat it as a mess to be cleared up. The ones getting value from AI have started reading it as a brief.
Find Your First Spreadsheet-to-Workflow AI Setup