Hi all,

As we all are experiencing, AI is IBM's thing now. However, leadership and management are just telling us "infuse AI into your daily workflow" with not much more context and it will be on this year's OKR's (Objective and Key Results). Many of us are starting with the Watsonx Assistant, since it's the easiest to start with.

Delving into this new field is very challenging. I've made some progress. Like if I was talking to a stranger, I can chat all about AI. But creating, implementing, how to run an AI Project, what architecture to choose from etc.... my biggest blockers. I know IBM must have a whole bunch of AI experts.

My question is, creating and running an AI project is very different. Resources are very hard to find. I have spent months searching every where, reading IBM documents, reading guidelines here and there, and it is like drinking from a fire hose.

How would I go about to get help or study a more methodical way of doing things. I randomly come up with architectures based on my readings. I don't really know if I need Discovery or ML or NL etc.... to achieve my goals. I've submitted my project to AI Infrastructure Guild Sangam.AI requesting help, but they seem backed up. I've attended every office hour related to AI.... but is there a starting point, a mentor, an expert that I can use (or get my department to pay for), to give me a clearer path on how to start and what to do?

I feel like I'm doing all sorts of random poc's. I managed to connect Watson Assistant to various API's. I've created a custom extension to watsonx.ai. I've set up intents, entities, created look up tables in cloudant, trying to design dialogs etc.... But can't seem to tie everything together.

My gut is uncomfortable, because I know this isn't the best way, but I don't know the best way, and management is giving time constraints. Right now it's just me. I don't need to be handheld, but it would really good to have someone provide a framework from start to finish, the forms i have to fillout, the governance, the process. someone that can tell me that there is a much better way to do xyz etc...... I've pinged random people from the AI Guild, but no responses and no time to help. 

Sorry for the long post, but I wonder if anyone has suggestions so I might do this in a more systematic way and learn the correct things. Thank you!!

Thank to share.

Regards

Muchos estamos en el mismo camino y con sensaciones similares.

Ahora estoy empezando a utilizar LLM para ayudar a valorar las ideas y que me asesore en el mejor camino 

Puede no ser lo mejor, pero hay ingenieria de Prompting  que puede ayudar a encontrar un camino, tanto para validar que idea desarrollar, como también a evaluar que camino recorrer.

Utilizar la IAG como parte de los recursos para lograr identificar los objetivos y ayudarnos a encontrar el camino.

------------------------------
Anibal Gonzalez
------------------------------

100% agree.

Prompt engineering should be learnt at school. This is such a powerful skill for programming and learning overall.

You are not stuck on tools. You are stuck on sequence.

A practical way to scope these projects is:

1. Start with one business outcome. Example: reduce support handle time by 20% or automate one internal lookup workflow.

2. Define the exact user input, system action, and expected output.

3. List what must be deterministic vs probabilistic. API calls, approvals, and record updates should be deterministic. Drafting, summarizing, and classification can use AI.

4. Pick the smallest architecture that fits. For many first projects: Assistant + retrieval/search + 1 or 2 API actions is enough. You usually do not need a full ML pipeline first.

5. Create a fallback path. If confidence is low or a policy check fails, send to a human or return a safe next step.

6. Add 3 evaluation sets before launch: happy path, ambiguous input, and failure cases.

7. Track only a few metrics at first: task completion rate, escalation rate, latency, and user correction rate.

A simple decision rule:

- Need structured answers from approved knowledge: retrieval.

- Need action in a system: workflow/API orchestration.

- Need prediction from historical labeled data: ML.

- Need open-ended generation: LLM with guardrails.

If you already have Assistant, intents/entities, APIs, and Cloudant in place, I would build one narrow end-to-end flow first instead of adding more components. One production-grade use case teaches more than five disconnected PoCs.

If helpful, I can also share a 1-page template for turning an AI idea into a build sequence and evaluation checklist.