How to decide what to use AI to do
If your business has written processes then they can be automated.
Humans Are Easily Impressed
Today’s AI is human intelligence on an artificial substrate and real artificial intelligence will be different. So says Danny Hillis, who pioneered parallel computers at a time when science said they were impossible. Today they are the engine of cloud computing, AI processors and all advanced chips.
Hillis says tasks we think of as hard are the easiest for machines to perform. These also the ones that impress us the most, such as an ability to play chess, write code and solve intricate maths problems. By contrast, routine human intelligence, such as jumping to conclusions, weighing up problems and even listening, are hardest for technologists to develop. But they’re getting there.
OpenAI’s reasoning models turn what we know about prompting on its head. Rather than breakdown problems step-by-step, as you do for free and earlier versions of large language models (LLMs), the models now reason for you. You choose the depth of thought you want and pay accordingly.
Achieving Human Intelligence
OpenAI’s o3 model, which is still in testing, beat the benchmark for human intelligence on the ARC-AGI test of general intelligence. This is a series of puzzles that humans find simple while machines struggle to complete. The model did this by reasoning, but with way more compute and cost than allowed to win the $1,000,000 ARC prize. Yet it will get cheaper.
Reasoning shifts the model’s workload of providing intelligence from training to inference, which is the term for running a model. This takes us a step closer to intelligence and is advantageous for model providers. While they must cover the cost of training up front, users pay for inference as they go. Shifting the workload from training to running costs changes the economics of LLMs.
With the rapid development of AI capabilities, businesses might be forgiven for adopting a wait and see approach to what the technology can do. They might also assume existing suppliers will incorporate AI and absolve them of responsibility. Yet this is far from certain.
Microsoft CEO Satya Nadella argues that interfaces will be replaced with AI capabilities. For example, you will not need a financial dashboard when you can ask an AI model to draw any chart you imagine. You don’t need to spend hours setting up reports in your CRM when, for instance, a model tracks your overdue accounts and emails the relevant clients.
While the jury is out on the degree to which SaaS will be replaced by AI models, what should businesses be doing today to determine where to use AI?
Document Your Processes
One starting point is to determine the business processes that do not change, and are time-consuming and tedious enough that staff make mistakes repeating them. These might be entering numbers in a spreadsheet, which is an example of a software service, or summarising market research. Then determine how these processes are performed.
The most cost effective way to automate depends on the complexity of tasks. For instance, entering a positive number in one column and negative in another may be automated without AI. Alternatively, research into the five largest markets for synthetic textiles requires a single prompt or API call to an LLM. Longer queries require a new way of working.
Agentic workflows are when multiple models combine on a task. These are a way to transfer thought into action. For example, using data from the web in a research report, then emailing it to interested parties and posting about it online. Agents can automate this. What matters is where the knowledge of the process resides – if it is written down it can be automated.
For many businesses, AI is a catalyst to think about processes that should be automated anyway. Thereafter, LLMs become efficient search engines. Once a decision must be made about a subsequent course of action, AI agents are a means of making that choice.
All of this may be achieved at low cost. It is only when outcomes are indeterminate, and the best course of action changes with time and circumstance, that agents require deep reasoning. That may be prohibitively expensive for now, but is worth thinking about because it won’t be forever.
Thought Exercises:
- What are the documented processes in my business?
- Which ones have automatic follow-on actions?
- Is a decision required to determine this action?