AI development: from task to verified use
AI development starts with a business problem and verifiable requirements. Data, oversight and operation affect effort alongside model choice. A useful quote defines deliverables, checks and remaining limitations.
Define the task before the technology
Should the application classify documents, extract information, forecast or draft text? Define who uses the outcome and what makes it useful. Compare a simple rule-based alternative too. A chatbot provides conversation; an agent may also choose steps and tools. Include those capabilities only when needed and explicitly agreed.
Clarify data and usage rights
Inventory data types, quality, currency and available examples. Establish usage rights and access for every source. Missing or conflicting information can impair results. A publicly accessible file is not automatically unrestricted for reuse. Descriptions suffice initially; protected raw data belongs in an agreed secure process.
Choose appropriate development
An existing model, targeted instructions, retrieval from approved knowledge and a specially trained model are different approaches. Ask why the proposed adaptation is necessary. Not every AI application requires a proprietary model. Integrated sources must remain current and accessible within permissions; using them does not guarantee error-free answers.
| Approach | Check first |
|---|---|
| Defined rules | Are explicit conditions sufficient? |
| Existing model | Suitability for task and languages |
| Additional adaptation | Data, effort and measurable need |
Bound actions and responsibility
Drafts and recommendations differ from automatic changes or binding decisions. Define human checks and approvals. Swiss data-protection requirements also apply to AI processing of personal data; plan transparency and risk assessment early. Specific legal needs depend on the use. A technical solution does not replace that assessment.
Measure quality on your own cases
Agree a test set with ordinary, difficult and disallowed inputs. Assess business correctness, missing information, languages and required response times. A general model ranking says little about your use case. Have domain experts review results and document limitations. Define when the application should decline to answer or hand over to a person.
Agree integration and ongoing operation
Describe the destination application and required data flows. Clarify access, retention, deletion, failure behaviour and ownership. New models, data or instructions can change results. Agree version records, retesting and rollback. Handover should include documentation and support for responsible staff.
Compare costs beyond development
Separate analysis, preparation, pilot, integration and acceptance from usage and maintenance. Consumption may depend on inputs, outputs, model calls, storage and additional tools. Compare a bounded usage scenario with explicit assumptions and overages. Limited retries and suitable reuse can help. Ask about effective usage controls; budget alerts alone do not automatically stop spending.
| Cost item | Compare |
|---|---|
| Development | Pilot, preparation and integration |
| Usage | Units, included volume and overages |
| Maintenance | New data, versions and retesting |
Checklist and project procedure
Note the task, users, desired result, languages, available data types and applications. Add expected usage, required oversight and timing. Send no credentials or sensitive datasets initially. Proceed through feasibility assessment, a bounded pilot, domain evaluation, a quote for further work and documented approval before real-world use.
Frequently asked questions
Do I need a proprietary model?
Not necessarily. The task and verified quality determine appropriate adaptation.
Is AI development the same as a chatbot?
No. AI can also extract, classify or forecast. A conversational interface is a separate scope item.
Can AI make mistakes?
Yes. Test your use case with suitable examples and define checks and clarification paths.
Does data protection apply to AI?
Yes. Clarify permitted data use, transparency and safeguards for the specific application.
Does a pilot guarantee future quality?
No. It shows results within the tested scope. Changes and new inputs may require further checks.
Must I upload training data?
No. Describe types and availability first. Secure transfer is agreed later if needed.
