Currently about 95% of all AI projects in companies fail –
according to a leading US company that provides an AI software platform to help companies streamline and automate complex business operations.
The information in this blogpost has been curated from a podcast called ‘Moonshots’ – featuring Peter Diamandis; host and Executive Founder of Singularity University and 3 expert guests and entrepreneurs.
The reason why most companies fail to launch AI technology successfully in their operations, is because most skip straight to software; before doing proper homework, research and planning.
How to prevent messy systems.
The experts’ advice is to ask the right questions first:
What is the thing I am trying to solve? What should I focus on? Which parts of my business can really change with AI? What are 2-3 things that, if I do them well materially, will move the needle for my business?- Who should do this? Am I going to bring in a chief AI officer or should I rent it? Or do we possess the expertise in-house?
- The easiest place to start in a company might be the admin department: customer service, forecasting, inventory management, digital marketing, legal services.
- Then find a high performing LLM – do not just buy anything off the shelf, hoping it will solve everything.
- The next step will be to get the data ready. Be tactical about what data you need. Clear and working data will give you the edge.
- Tailor the AI model to your context. Fine tune it off of your data and pre-train it
- Decide what the right outcomes should look like for your real-world scenario. Do lots of testing and validation to make sure you can trust it.
- Fine tune it off your specific company’s information: your products, the way you sell, your way of communication. Have an AI agent shadow your people. Work should be mapped in context as it happens.
- Make sure you get to a pilot stage in those areas—not merely a strategy document.
- Be very clear about your benchmarks; how is the model performing compared to human performance in the past? Anything less than 80% is still too much risk. Your internal benchmarks should be hyper specific. (No broad benchmarks.)
- Simulate first before going physical.
- Launch.
- Evaluate, test and measure against human outcomes continuously.
- Continue supervised fine-tuning for safety, reliability, compliance and staying on track.
- Post-train.
- Only pour money into the things that do work.
What are companies doing wrong when implementing AI?
- Lack of focus on data as a starting point.
- Feeding fragmented customer and product data into the model
- Data should be clear and working.
- Real good human guidance is going to persist forever. You would want humans in the loop. Humans still want to talk to other humans and you will always need human feedback.
Summary
Get clean data and clean objectives, pre-train your LLM according to your specific needs, post-train, finetune, test, mix humans and agents. Finally, get measurable business results.
Till next time…