Posted on: Tuesday 23rd of January 2018
Automating activities is far harder than it may seem. It works well when there is a high volume of repetitive tasks. But what looks repetitive at a high level can be much more variable when looked at granularly, step-by-step.
Typically the 80:20 rule applies so the logical approach is to automate the 20% of activities that account for 80% of the work. That works when there is an end-to-end process down which 80% of the volume flows. But often the high-volume activities are spread out in different areas, not connected to each other. With the result that to automate some steps, you have to automate others where the volume is insufficient for it to be economic or create islands of automation.
Also, most businesses aren’t starting from paper ledgers. The Y2K threat forced most large businesses to implement expensive Enterprise Resource Planning solutions which incorporate automated work-flow. So when evaluating the impact automation will have in the future, the scope is limited to activities not automated by existing systems. There are obvious examples – banks and life insurance companies remain heavily reliant on legacy mainframe systems. This is where the providers of robotic process automation tools are focusing their efforts, with some success through what I have observed suggests there is a limit to how transformational these initiatives will be.
Automating an end-to-end solution requires the integration of four elements – historic data, predictions (when uncertainty exists), judgment about the best action to take and execution of that action. That requires integrating a data warehouse with an AI solution with a decision engine with an execution tool such as a website, CRM system or RPA tool. All of which requires a significant investment. The business case has to stack up against replacing humans who are phenomenal multi-taskers – we intake data, use our experience to predict and judge as required, then use our hands to complete the task.
In the physical world, such integration is even harder, though not impossible. Two examples of automation in agriculture highlight the dichotomy. The Lettuce Bot can recognise weeds in a field of closely planted lettuces, also deformed and too closely planted lettuces, then precisely spray the offending plants with weed killer. This works because the classification is simple and the judgement required is minor.
Contrast this with the attempt by Makoto Koike to sort cucumbers grown on his parents’ farm into nine different categories. Training the solution required Makoto to spend three months taking photographs of 7,000 pre-sorted cucumbers. This is a small training sample in AI terms, and while it was 95% accurate in lab conditions, in production accuracy dropped to 70%. Even then it was only able to classify according to shape, length and distortion but not colour, texture, prickles and scratches, all dimensions being necessary for the classification required
Makoto will at some point develop a solution that is economic. And in the economy as a whole, levels of automation are bound to increase. But before believing the next disaster scenario you read about, ask yourself whether the author has ever tried to implement an automation solution. My guess would be no.