How many small businesses use AI, and what holds them back
- 26%
- of companies in Germany used AI in 2025
- 23%
- with 10 to 49 employees
- 36%
- with 50 to 249 employees
- 57%
- with 250 or more employees
That is according to the Federal Statistical Office's survey for 2025, as of 24 November 2025; the next survey is announced for November or December 2026. Of the companies not using AI, one in five had already considered using it. According to the table of obstacles, the reasons these companies most often gave for not doing so were lack of knowledge (72%), uncertainty about the legal consequences (62%) and data protection concerns (60%); costs followed far behind at 32%.
The three most common reasons are exactly the ones a small, planned pilot resolves: it builds knowledge on a real task, it forces you to clarify the data route and legal questions beforehand, and it shows whether the benefit justifies the effort. The six steps below guide you through such a trial.
1. Choose a task whose result you can judge
“We want to use AI” is not yet a testable task. “Summarise our approved monthly data from CSV in an overview we can check” is one. Other suitable candidates are comparing quotes or drafting a presentation from an existing report. The Biwak demo and the article on AI agents and office tasks show concrete inputs, results and acceptance criteria.
- The task comes up regularly and its current process is known.
- It can be carried out with a manageable, approved copy of the files.
- A person with the relevant expertise can spot errors in the result and in the sources.
- A failed attempt ends the test without affecting customers, payments or original files.
2. Clarify responsibilities and the data route before the trial
| Subject-matter responsibility | Who checks the finished work?Record the name, a deputy and the acceptance criteria. The tool provider cannot confirm the professional accuracy of your result on your behalf. |
|---|---|
| Data and access | Which information may go into this trial?Take inputs, attachments, conversation context and tool output into account. For a first functional test, sample data you create yourself can make sense; real data requires the appropriate approval. |
| Technology and operations | Who sets up access, permissions and the way back?Document the working folder, account, permitted services, budget, backup and the procedure in case of an error. |
| Employees | Who needs to be involved and trained?The employees who will continue working with the result belong in the trial. If you have a works council, you inform it under § 90 BetrVG (German Works Constitution Act) while you are still planning; the Act expressly mentions the use of artificial intelligence there. Training is one of the measures you use to promote your staff's AI literacy under Art. 4 AI Act. |
At Biwak, the privacy policy describes the model route: for each task, your question, the context and the file contents used go via the Biwak relay in Frankfurt am Main to Microsoft Azure in the EU Data Zone; only if Azure declines do they go to PREM SA in Switzerland and from there to its compute partners in the EU. For businesses, Biwak's data processing agreement applies, with the TOMs in Annex III. As for the way back: the checkpoint only covers the working folder, and the folder is not a technical barrier, because the agent's tools reach as far as the user account does. Terminal commands and internet access can be switched off in the settings; a checklist for IT is on the security page. Which agreements your use requires depends on roles, data and processing. The guides on processing on behalf, works councils, data protection impact assessments and AI literacy training serve as orientation; the pilot plan is not legal clearance.
3. Record today's effort as a baseline
First, work through representative tasks the way you do today. Record active working time, waiting time, errors and necessary corrections separately. For the AI trial, use comparable tasks with the same acceptance criteria. If the same person works on the same file a second time, they may be faster because of the learning effect; do not count this difference as an AI effect without checking.
Record failed attempts as well. An evaluation that only contains successful demonstrations will not help you decide about everyday use. For each attempt, note the product version, model or performance level, inputs and relevant settings, so that you can re-check the result after a later change.
4. Properly sign off an Excel analysis
- Check the input: Are the file version, the period and the number of rows read in correct?
- Check the calculation: Are totals and subtotals correct? Were currencies, decimal separators, cancellations and empty values handled correctly?
- Check the exceptions: Are unclear or discarded rows visible, instead of silently disappearing from the analysis?
- Open the file: Do the relevant formulas and links work in your spreadsheet program? Can the result be understood without the conversation?
- Count the effort: How much work was needed for the task, follow-up questions, checking and repair?
Our measurement with two test PDFs shows what such a sign-off against a fixed target file looks like: it was set in advance which 16 transactions with which seven values had to come out, and only results that were completely correct without rework were counted. The same rule works for your pilot: first the target values, then the trial, then the count.
For document comparisons, source references and a completeness check replace the reconciliation of totals. For presentations, the substantive message of each slide and a visual check come on top. A convincingly worded text does not replace any of these checks.
5. Compare benefits and costs without false precision
Compare the effort per accepted task. This includes human processing, checking and rework, a share of the set-up and ongoing tool costs. Technical waiting time only counts as working time gained if the person can do something useful in the meantime.
| Current process | 40 minutes per accepted task |
|---|---|
| With AI | 25 minutes5 minutes of preparation, 12 minutes of checking and 8 minutes of rework. |
| Time difference | 15 minutes per taskFor 20 similar tasks, that would be 5 hours on paper. Set-up and ongoing costs still have to be offset against that. Whether this happens for you is what the pilot will show. |
For a trial with Biwak, you get a one-off 100 trial credits without payment details; after that, the monthly subscription costs €19 (Base Camp) or €99 (Rope Team), cancellable monthly; see plans. Price and available usage alone do not settle the matter: a cheap trial with a lot of rework can cost more than the current process. We do not derive any promise of a particular time saving from this example.
6. Make a clear decision
| Expand | Quality and total effort are rightThe agreed criteria are met across the relevant mix of tasks, and responsibility and data route have been clarified. Expand to similar tasks first. |
|---|---|
| Narrow down and re-test | The benefit depends on certain conditionsExample: tables with clear columns work, scans do not. Document this limit and only allow the tested case. |
| Stop | The task or the tool does not fitNo usable time saving, insufficient quality or unresolved questions about data processing are reasons not to adopt the trial into everyday work. |
If Biwak fits your task, download and set-up lead to the first trial. To choose a suitable task together, you can request an appointment. The product comparisons help if collaboration, fully local models or a specialist application are decisive for you, such as patent software as in the comparison Biwak or DeepIP.
Where this checklist comes from
This checklist is Biwak's proposal for a traceable pilot, based on the linked legal sources, the figures from the Federal Statistical Office and our own measurement with test PDFs. The example tasks and the worked example are invented; they are not a customer study and not independently measured proof of productivity. The linked articles and product pages state their own sources and review dates.
Frequently asked questions
How long should an AI pilot last?
Until the relevant mix of tasks and its failure cases can be assessed. A fixed number of days proves no benefit. Set a limited scope and a decision date in advance, and document it if the data is not yet sufficient for a decision.
Which task is suitable for the first trial?
A recurring task with clear inputs, manageable consequences and a verifiable result, such as a monthly overview from an approved CSV export or a comparison of quotes with source references. Choose a task whose current handling you know.
How does an SME measure the time saved by AI?
Compare the total human effort per accepted result with the current process. Count set-up, input, checking and rework, and document failed attempts. A few quick demonstrations are not a reliable average.
Do we need to be able to program?
You do not need to write any code to give a task in the Biwak interface. You do, however, need a person who knows the task and checks the result. Technical set-up, access rights and special file formats may require additional support.
Sources
- Federal Statistical Office: Companies using artificial intelligence technologies by employee size class, 2025 (as of 24 November 2025)
- Federal Statistical Office: Reasons against using artificial intelligence technologies, 2025
- Anthropic: Building effective agents, 19 December 2024 (simplest solution first, tests in a sandboxed environment)
- § 90 BetrVG (rights to information and consultation)
- Regulation (EU) 2024/1689 (AI Act), Art. 4 as amended by Regulation (EU) 2026/1744
- Biwak: Privacy policy, section 14 (requests to a language model)
