Start with a task you know
Requests arrive by email, information is scattered across files and software, and someone has to check everything before replying. For a microbusiness or SME in western Switzerland, this is a concrete starting point for assessing AI.
First choose a recurring task, an expected result and the person who will check it. You can then compare a feature in your current software, an automation and specific development. Better organisation is sometimes enough.
The SECO SME portal recommends defining uses, preparing data and assessing the impact. Here is how to turn that approach into a first trial.
Three situations to examine in your business
The following examples are illustrative. They describe possible uses, without claiming client results or guaranteed time savings.
A building contractor prepares quotes
A request contains a description of the site, some photos and notes from a visit. AI could prepare a summary, identify missing information and suggest a description of the work based on approved templates.
The person responsible retains pricing, quantity checks and commitments on delivery dates. A photo alone cannot establish a reliable measurement. Rates must come from the company's price list, and calculations must follow its usual rules.
Start by testing preparation of the quote file. Check whether the description omits work, invents materials or needs as much correction as writing it normally. Sending the quote still requires approval.
A property manager handles tenants' requests
A shared inbox receives questions about invoices, reports of problems and requests for repairs. A first trial could extract the building concerned, summarise the request and draft a reply following the property manager's instructions.
The workflow would be explicit: receive the message, suggest a category, have a colleague check it, then reply or create a work order. Missing information stops preparation. Emergencies follow a dedicated procedure with human involvement, without relying solely on AI classification.
One request, a checked response
A request arrives
A tenant reports a problem in their home.
Authorised documents
The system retrieves case information the manager has permission to access.
AI prepares a draft
It suggests a response with the documents used. It flags missing information.
The manager approves
They check the facts and recipients, make any corrections and send the response.
Start with one common, low-risk category. The property manager decides who can view each record and who approves a reply. The first trial can use drafts before any connection to the business software.
A service company searches its documents
Someone needs the procedure that applies to a case or a clause in internal documentation. An assisted search could suggest an answer with the passages and documents that support it.
Limit the test to an up-to-date collection. The system must respect each person's reading permissions, including when producing a summary. A colleague without access to a record must not be able to obtain its contents by asking a question.
Require a link to the source, its date and a clear indication when information is missing. The person opens the document and checks their interpretation before acting. A well-written answer does not prove that its source is correct.
An existing feature, an automation or a custom tool?
Look at your email, office suite and business software first. A built-in feature deserves a trial if it covers the task, allows the necessary checks and processes data under acceptable conditions.
- Starting point
- Test an AI feature in an existing tool
- Example
- Prepare meeting minutes from notes you are authorised to use.
- What to check
- Data-use terms, draft quality and the time spent reviewing it.
These examples help you think through the options. The choice depends on your tools, your data and the consequences of an error.
Automation becomes relevant when the problem lies between tools. For example, copying an approved request into a record and notifying the person responsible. A conventional rule may be enough for that transfer. AI is most relevant for interpreting text or preparing content for review.
Custom development deserves assessment when your rules, access permissions or connections to existing software cannot be covered satisfactorily. Describe that difficulty and check that resolving it justifies implementation and maintenance. Our article on moving from Excel to a business application helps assess this decision.
What budget should you plan for?
The advertised subscription is only part of the cost. Ask for an estimate that distinguishes initial expenses from costs that recur during use.
| Cost item | What to clarify |
|---|---|
| Licences and usage | Required accounts, included features, limits and usage billing if an API connects the software |
| Setup | Defining the task, preparing documents, configuration and trials |
| Integrations | Software access, data exchanges, error handling and validation |
| The team's work | Training, review, corrections and responsibility for follow-up |
| Operation | Any hosting, monitoring, updates and maintenance |
Volume matters too. Occasional preparation and continuous document processing do not incur the same expenses. Specify who monitors consumption and what happens when a limit is reached.
Compare this full cost with the current work. Time freed up can help you handle requests better or absorb additional volume. It does not automatically become a financial saving. If review cancels out the gain, reduce the scope or change the approach.
Data and colleagues: decide before the trial
The FDPIC states that the Federal Act on Data Protection applies to AI processing. A customer's or colleague's personal data remain covered when they pass through an assistant.
For the first trial, prefer fictional examples or examples with identifying information removed. Removing a name is not enough if other information can identify the person. Avoid commercial secrets too until you have checked the processing conditions.
The FDPIC advises checking purposes, access, transfers to third parties and use for training. Ask the provider for precise answers on retention, processing locations, deletion and contractual commitments. Advertised Swiss hosting does not answer all these questions.
With the team, agree on authorised tools, permitted data and results that need checking. Assign someone to collect errors. Also show how to complete the task without the assistant when it is unavailable or produces an unusable result.
A first pilot with a decision at the end
Choose one workflow and a few representative cases, including incomplete or unusual ones. Measure the current work before the trial and agree on a test period with the people involved.
Then compare similar situations using these criteria:
- The total time to complete the task, including review and corrections.
- Omissions, errors and cases that need to be redone completely.
- The ability to find sources and respect the agreed access permissions.
- Usage costs and the team's ability to work without outside help.
Set acceptable thresholds before starting. A reduction in time does not compensate for an incorrect amount or exposed confidential information. At the end, decide to continue, adjust or stop. The trial serves to reach that decision before a wider rollout.
Prepare a discussion about your first use
Bring a task that recurs often, an anonymised example, the tools you use and the expected result. Add who prepares the work, who approves it and where it gets stuck today.
At Appik Studio in Lausanne, we can examine this workflow and the options that fit your tools. The first discussion and initial opinion are free. Detailed scoping has a scope and budget agreed before work starts.
Let's discuss the task you want to improve. This starting point allows us to assess a limited trial and determine whether an existing feature, an integration or a specific tool is justified.
