Excel stays: when BI actually pays off in an SME

Monday morning, management meeting, item 2: August figures. The CFO sent "Mgmt-Report_2026-08_v3_final.xlsx" on Friday evening. The head of sales pulled his own CRM extract on Sunday, and revenue differs by 4 %. The first twenty minutes go to which number is right; that leaves half an hour for what to do. Searching for "replace Excel reporting" in this situation turns up tool comparisons. We are interested in a different question: when does BI pay off in an SME, and when does it not?

Excel stays in SMEs for good reasons

The Swiss figures are clear. In the IFZ study by the Lucerne University of Applied Sciences, covering 223 Swiss companies, 52 % of small and 40 % of mid-sized companies use no BI in controlling; Excel and Office are the standard tool for over 90 % (Egle/Keimer 2018, p. 15 ff.). 37 % of large companies rate the digitalisation of their controlling as fairly high to very high; among small companies it is 14 % (Egle/Keimer 2018, p. 3). Two years later the same authors confirm the picture: 65 % of Swiss companies put the digitalisation of their controlling below 50 %, only 8 % above 75 % (Keimer/Egle 2020, p. 3). For the German Mittelstand, Schmid-Gundram puts Excel-based controlling systems at over 80 % (Schmid-Gundram 2020, p. V).

These surveys date from 2016 to 2019, and the BI landscape has moved since. The finding still holds, and it is not negligence. Excel stays for three good reasons. It is flexible: the column the board asks for on Thursday is in on Friday. Everyone can use it: nobody needs training. And it costs nothing that has not already been paid for. Schmid-Gundram makes the point plainly: BI has not replaced Excel not because of the cost of the systems, but because many owners see little additional benefit and the people who run controlling on the side – often the owner – lack IT affinity (Schmid-Gundram 2020, p. V). To convince an SME of BI, you have to show the benefit.

Where Excel tips over

The benefit becomes visible as soon as Excel tips over. That happens in four places that keep turning up in our reviews.

First, reconciliation. Revenue in the management report comes from the ledger, revenue in the sales report from the CRM, and the two meet for the first time in the meeting. Cindi Howson describes the pattern from dozens of BI projects: executive meetings start with a debate about how the numbers were compiled and whose are correct, rather than with a discussion of the insights (Howson 2014, p. 111). That argument at the start of the management meeting is the surest sign that Excel has tipped.

Second, versions. "v3_final" is not the last version, and everyone in the room knows it. Three people on the same file make three truths, and "which file counts?" crowds out "what do we do?".

Third, effort. Langmann cites analyses of the reporting process according to which around 70 % of total effort goes into the first four steps: collecting data, preparing it, building the report, checking plausibility. In SMEs in particular, manual work dominates, leaving little time for root-cause analysis, commentary and actions (Langmann 2019, p. 12). A controller who spends two days a month copying files together has two days less for why the service margin is falling.

Fourth, detail. As long as the board wants revenue by month, a table will do. Once it wants revenue by month, region and product group – and by customer next month – the file grows in three dimensions. Schmid-Gundram calls this the real limit: Excel-based controlling systems quickly reach their boundaries when the level of detail changes often or keeps growing (Schmid-Gundram 2020, p. 12).

The cost fits into one scene: a service SME with 80 staff, three source systems and a monthly report ready on working day twelve. Two person-days of collecting, half a day of reconciling, and the meeting on day fifteen discusses figures three weeks old. A tool can solve that, but only if it is clear beforehand which decision the report is meant to carry.

First the decision, then the KPI, then the tool

In our approach the tool question comes fourth. The framework "From decision to KPI" starts with the decision the management team takes regularly, derives question and KPI from it, and only then turns to data, report and tool. Before every tile stands the purpose question: what are these numbers supposed to say, and what is the purpose? Whoever can answer it knows which sources are needed, who reads the report and how soon after the close it must be ready. Those are precisely the criteria for choosing the tool.

Howson's finding shows why the order matters: BI success and business impact are not the same thing. The two diverge in her survey, and she notes that it is possible to build a perfectly architected BI solution that has little impact on the business (Howson 2014, pp. 76–79). A data model with thirty tables is no success if the meeting still opens with the argument over numbers.

The maturity path: S, M, L, XL

For the tool question itself we use the T-shirt sizes described after Branger in the TDWI self-service analytics working group: S is experimental, directly on source data; M a semantic model; L a staging layer or data lake; XL an organisation-wide data model (TDWI 2022, p. 37). The value of the scheme lies in the deliberate transition: you move up a size only when criteria are met that were fixed beforehand.

Maturity path S/M/L/XL: criteria, effort, use

Size	What it is	Typical criteria	Effort	When it makes sense
S	Excel or Power BI directly on source data, no model	1–2 sources, 1–3 readers, refreshed manually	Days; ongoing upkeep every month	Testing a new question, one-off analysis, prototype
M	Semantic model (Power BI), sources and budget combined	2–4 sources, management team of 5–10 readers, refresh after the close	Weeks; low upkeep	Monthly reporting for management, one cockpit, one truth
L	Staging layer / lakehouse (Fabric), history, several models	4–8 sources, department heads with their own views, daily refresh	Months; in-house or external support	Several entities or divisions, operational cockpits
XL	Organisation-wide platform with governance and data ownership	More than 8 sources, whole organisation, several refreshes a day	Continuous operation; team or partner	Groups, regulatory requirements, self-service for departments

Four criteria decide the switch. Sources: a ledger and a budget sheet are size S or M; once CRM, time tracking, payroll and a project system join in, the data needs a place to come together before anyone displays it. Readers: a report for three people may be a file; a report for twenty department heads with their own views needs a model, or you get twenty files. Days to close: if the monthly report is ready later than working day five – Schmid-Gundram's upper limit (2020, p. 8) – and the reason is collecting, automation pays off. Cost of errors: as long as a wrong value is caught on reading, Excel is good enough; once a pricing, hiring or investment decision rests on a wrong number, it is not.

The distribution is an estimate, but it shows where BI earns its keep: it shortens the first two bars. The time freed up flows into analysis and decisions, and the argument about whose numbers are right disappears from the start of the meeting.

Size S with AI: Excel lasts longer

Since the surveys of 2016 to 2019, something has shifted at size S. An AI assistant on a company licence costs about as much per person and month as a lunch, and it takes over part of the work hidden in the first two bars. Used properly, that means a project in the assistant holding the chart of accounts, the last few monthly reports and one page of definitions: what counts as revenue, margin or an active customer in your company. The AI then drafts the variance comments, builds formulas and Power Query steps, and helps track down the difference in a reconciliation. Meisser counts such low-code tasks among the best uses of AI in controlling, because they recur and their benefit can be measured in francs (Meisser 2024, pp. 32 f.).

The evidence is good, but it has a limit. In a field experiment with 758 consultants at Boston Consulting Group, those using AI completed 12 % more tasks, 25 % faster and at over 40 % higher quality, as long as the task lay within the AI's capabilities; on a task outside them, they were 19 percentage points less likely to get it right (Dell'Acqua et al. 2023). For bookkeeping there is early field data: in 79 small and mid-sized firms, accountants working with AI-based software closed the month 7.5 days faster (Choi/Xie 2026). Meisser draws the line for controlling clearly: KPIs need exact values according to the company's own definition, not an AI's approximation (Meisser 2024, p. 34). An assistant given the revenue from the ledger and the revenue from the CRM will calculate with both equally convincingly. It will not end Monday's argument about whose numbers are right.

What matters is how the assistant enters the company. According to KPMG and the University of Melbourne, 77 % of employees in Switzerland use AI at work, and more than half of them have already used it against company rules, for instance with confidential data in free tools (KPMG 2025). The assistant in controlling therefore belongs on a company licence, with four rules:

• The company takes out the licence. Business plans such as Microsoft 365 Copilot, ChatGPT Business or Claude Team do not use inputs for training by default and offer a data processing agreement; free private accounts offer neither.
• Personal data stays out, above all payroll and HR data. The Swiss Federal Act on Data Protection applies directly to AI (FDPIC 2023).
• Transfers abroad are settled. If the provider processes data in the US, it needs to be certified under the Swiss-U.S. Data Privacy Framework or bound by standard contractual clauses; whoever is responsible for data protection in the company checks this once.
• The AI drafts, a person signs off. In the same study, 74 % of respondents in Switzerland said they do not check AI output.

The biggest benefit of this size shows at the next step. The definitions you write down for the assistant are the same ones a semantic model at size M needs: what counts as revenue, when a customer is active, from which variance a comment is due. For Meisser, BI and AI are accordingly symbiotic concepts (Meisser 2024, p. 34). Set up size S with AI properly, and a good part of the BI rollout is already done when Excel does tip over at one of the four points.

For some SMEs, M is the target

Here belongs an honest sentence you rarely read in tool comparisons: for many SMEs, size M is the target. A company with one ledger, one budget and a management team of five needs a semantic model that brings ledger and plan together cleanly, a cockpit in consistent notation and a process that refreshes the report after the close. It does not need a lakehouse. Langmann says so explicitly for his maturity model: the maximum level is not the best solution for every dimension; it depends on circumstances and business model (Langmann 2019, p. 51). Keimer and Egle add the second condition: data, technology, processes, methods and skills must be at a comparable maturity (Keimer/Egle 2020, p. 7). An XL platform on top of an M ledger and S skills is mainly a cost block.

Two rules from our practice apply at every size. First: reconcilability before beauty. Before a report gets IBCS® notation, its revenue must reconcile to the franc with the ledger – otherwise the beautiful report is the next occasion for the argument over numbers. Second: the report hangs on the close. An automated report is only as current as the bookkeeping behind it; if the fiduciary posts quarterly, no tool will deliver a monthly report. The close schedule therefore belongs in every reporting build.

What matters is the deviation from target. Before the switch the report sits around seven days above the baseline every month, afterwards within tolerance. That is a KPI by which a BI build can be measured before anyone talks about licences.

In his case study on a BI rollout at a Swiss ticketing company, Caviezel describes what decides acceptance: start with an application that solves an existing problem, not one that improves an existing solution (Caviezel in Keimer/Egle 2020, p. 116). Projects that simply rebuild existing solutions and processes in a new technology he considers doomed to fail (p. 119). For the Excel question this means: the first Power BI cockpit should end Monday's argument over numbers. A prettier copy of the management report does not.

What you can do tomorrow

Take the last monthly report and answer four questions. How many sources do the numbers come from? How many people read it, and do they need different views? On which working day after month-end was it ready – and why? And which decision in the last quarter rested on a number that later turned out to be wrong? The answers tell you which size you are at and whether a switch is due. The reporting check above asks the same in a structured way, free and in ten minutes.

If a switch is due: our scoping sprint clarifies in two to three weeks at a fixed price which decisions the report is meant to carry, which KPIs with which cadence belong to them, which sources are needed and which size – M, L or XL – is the target. The result is decision-ready: decision inventory, question catalogue, KPI profiles, measurement plan, notation concept and target picture. Only then is the tool decided – and it often turns out smaller than expected.

Sources

• Egle, U., Keimer, I. (2018): Digitaler Wandel im Controlling. Schriften aus dem IFZ, vol. 37. Verlag IFZ, Lucerne University of Applied Sciences.
• Keimer, I., Egle, U. (eds.) (2020): Die Digitalisierung der Controlling-Funktion. Anwendungsbeispiele aus Theorie und Praxis. Springer Gabler – including Caviezel: BI rollout (ch. 7).
• Langmann, C. (2019): Digitalisierung im Controlling. Springer Gabler essentials.
• Schmid-Gundram, R. (2020): Controlling-Praxis im Mittelstand. Aufbau eines Controllingsystems ausgehend von Lexware, DATEV oder SAP. 3rd edition, Springer Gabler.
• Howson, C. (2014): Successful Business Intelligence. Unlock the Value of BI & Big Data. 2nd edition, McGraw-Hill.
• TDWI e.V. (ed.) (2022): Self-Service Analytics. Best of TDWI Themenzirkel Self-Service & Analytics, e-book 2022.04. SIGS DATACOM – including the T-shirt sizes S/M/L/XL after Branger (p. 37).
• Choi, J. H., Xie, C. (2026): Human + AI in Accounting: Early Evidence from the Field. Journal of Accounting Research.
• Dell'Acqua, F. et al. (2023): Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013.
• FDPIC/EDÖB (2023): Geltendes Datenschutzgesetz ist auf KI direkt anwendbar. Statement of 9 November 2023.
• KPMG, University of Melbourne (2025): Trust, attitudes and use of artificial intelligence: A global study 2025. Swiss results.
• Meisser, F. (2024): KI im Rechnungswesen und Controlling: Was sind sinnvolle Anwendungen? SwissAccounting «Standard» 3/2024, pp. 32–35.