Anyone starting an AI project today first discusses models: which language model delivers the best answers, which provider is reliable, what the cost per request looks like.A look at documented real-world projects significantly reorders these priorities.In eleven AI implementations in companies described in detail – from technical architecture to rollout – the actual bottleneck was more often trust, acceptance and the last few percent of implementation than the model itself.
This is an unspectacular but consequential insight.Because it shifts the focus of an AI project from technology selection to questions of adoption, responsibilities and data quality.And it explains why prototypes often convince, yet production operation stalls.
Why the model rarely limits
The available language models – i.e. AI systems that understand and generate text – have reached a performance level at which many typical tasks in everyday business are fundamentally solvable: summarising, classifying, creating drafts, researching documents.For these standard cases, choosing between two strong models is usually not a question of success but a trade-off between cost, data protection and operating model.
The delta arises elsewhere: in the question of whether the results are robust in the specific domain context, whether the underlying data is cleanly connected and whether employees trust the system enough to actually use it.An assistant that delivers 90 percent good answers but shows no traceable origin for its statements will be bypassed when in doubt.
Trust is a design task
Trust is not created through communication but through properties of the system.These include traceable sources for generated answers, clear labelling of uncertain results, the ability to correct suggestions and a visible handling of errors.
In practice this means: anyone integrating AI functions into existing applications should plan for traceability from the outset.Which documents were consulted?What state of the data was available?Who may see which content?These points are not additional features but a prerequisite for a system to make it out of the test phase.
Acceptance is decided in everyday work
The second hurdle is acceptance.An AI tool does not compete with the ideal state but with the established way users work.If the new solution requires an additional login, a context switch or an unfamiliar operating logic, it loses against routine.
That is why integration into existing systems – editorial interface, CRM, ticket system, intranet – is usually more important than the range of functions.AI support that appears where work is already being done needs no launch campaign.A separate standalone solution needs one and often fails anyway.
The last few percent cost the most time
The third factor, underestimated in practice, is the last few percent: special cases, exceptions, poor data quality in legacy holdings, permission logic, logging, monitoring, fallbacks during outages.A working prototype leaves these topics out – a production system does not.
For project planning this means: the effort is not distributed evenly.The first visible function comes quickly, robust operation takes time.Anyone who aligns the schedule with the prototype is planning too optimistically.
What can be derived from this for project design
- Use case before technology: First describe the process that is to be relieved, then decide on models and architecture.
- Check the data situation: AI results reflect the quality of the connected content.Unstructured or outdated holdings are the real preliminary work.
- Plan for interchangeability: Models change faster than processes.An abstraction layer keeps the application independent of the provider.
- Build in traceability: Provide source references, logs and correction options from the start.
- Involve users early: Acceptance cannot be created after the fact.
- Think about operations: Define responsibilities, costs and quality control after go-live.
Clarifying expectations instead of competing models
For decision-makers this is a helpful simplification: the question of the right language model is important but secondary.What is decisive is whether a project addresses a clearly defined process, whether the data basis holds up and whether the organisation is ready to adapt workflows.
For web and software projects this means a shift in effort: less energy into model selection, more into integration, data quality, permission and role concepts as well as into operations that make errors visible and keep them correctable.Anyone who plans AI functions as part of the existing system and not as an additional tool alongside it shortens the path to productive use considerably.And those who budget the final few per cent from the outset avoid the most common disappointment of such projects: a convincing prototype that never makes it into everyday use.
Sources
- The model isn't the problem: 6 lessons from 11 AI projects — t3n.de - News