I'm looking for a rough estimate for building an AI model.

Data in tables are the easiest to use in data science projects.

Text is often quite hard to structure. Numeric data is often a lot easier to use in AI projects.

The uncertainty of planning a project that requires integration of multiple data sources makes it hard to plan your project.

Numbers can easily be analyzed, but large free form texts, video and audio require a trained eye and a lot of manual inspection.

No personal data and no business secrets makes our lives a lot easier.

APIs are generally a wet dream for developers to extract data from because of their ease of use.

Interaction with legacy systems require custom data collection approaches.

Dealing with certifications, licenses, standards and protocols all take a lot of time.

A data scientist typically spends 60% of their time on cleaning data. If your data comes clean, we save a lot of time already.

Fixing holes in data is always very manual work.

Fixing these issues require deep contextual knowledge, which we'll have to figure out on the go.

We'll have to omit the "garbage in, garbage out" principle of AI models here, which is quite difficult.

There are a lot of pre-trained models that can deliver the value that you're looking for.

If completely pre-trained doesn't cut it completely for you, fine-tuning a model can work wonders.

Predictive tasks that are not based on NLP or computer vision require custom modelling.

If multiple predictions need to be made that might even be dependent on each other, modelling gets challenging.

No new data? Then there's no need for retraining.

New data can shift predictions, even so slightly. If there's some new data coming in, we recommend touch-ups.

Retraining consistently requires a good technical infrastructure and good monitoring.

The more data, the merrier. Training is the most expensive part of cloud usage bills though.

AI always comes with uncertainty of outcomes, which is a very important principle of statistics.

You'd be surprised how powerful the combination of AI with human experts is.

Relying on AI always comes with caveats that we'll investigate here.

Selecting this option puts us in the moral (and sometimes legal) obligation to make sure everything works as expected.

Any computer can train AI, it will just take a bit longer to retrain, but is good for the wallet.

For lots of AI models, getting access to GPUs is a good idea to speed up training.

We'll set up a strong architecture with multiple GPUs to get your training to a higher level.

We'd have to keep TPUs spinning to power your processes, which are not cheap.

If an AI model is not mission-critical, we can create a lean and inexpensive setup.

The AI model is important, but not necessarily requiring high availability.

Fast incident solving is required (within a day).

There is need for 24/7 support and zero downtime. The model has to always be reachable.

Standalone services without integrations will always be the cheapest option.

Future integrations with established platforms of APIs requires some preparation work.

The AI model needs to be integrated from the get-go with one or two well-documented APIs, or need a small level of customization.

The AI model needs to be deeply integrated into existing (legacy) infrastructure.

We could get away with scaling down the operations of your model.

Internal business processes typically invoke AI models a couple of times a day at most.

Typically, when an AI model is used as a service with authentication, you can expect requests a couple of times an hour.

No limits, and public availability can result in very high traffic coming to the model.

We work with a Proof of Intelligence, which is an MVP-like process to get up and running quickly.

Defining parts of the project as prototypes can detect early whether AI makes sense.

We always recommend to start small, but if that's simply not possible, we have to accept it.

We'll only find out whether the AI is successful after data cleaning, modelling and bringing it in production.