I am working on supplier evaluations. This often means extracting data from invoices, proposals, rate cards or other sources and getting them into a format that makes them comparable and allows analysis and prediction. The data sources are quite varied, pdfs, excels, csv files and it can take a long time to extract meaningful data. As an example, freight providers often list a base price, but then add extra charges like a variable fuel levy, charges for exceeding size restrictions, surcharges for remote locations etc.
An ideal task for AI, I thought.
I threw the data at ChatGPT and asked it to compare rates in a new proposal with a big chunk of data from existing invoices. AI ingested, read and analysed the data. It produced beautiful comparison tables and a bullet point analysis within 30 seconds. The presentation was amazing.
I was startled, this task would have taken me hours. I knew all the stats on how AI increased productivity, but every now and then the experience of how efficiently AI replaces office work(ers) is a revelation, and prompts (again) serious thoughts about the future of work.
Then I checked the output and realised that AI had not considered some of the surcharges, affecting the comparison. I asked “Have you considered the fuel levy and the minimum charges in the proposal?” and it immediately revised the output.
Perfect outcome! AI had saved me hours of work, but it had not been completely right, so I was still relevant and needed to check and correct the results. It felt good. However, a further check a little later showed that AI had not used all the right numbers from the information and delivered some very distorted results where the data structure was not recognised or misread. It had done so without a blink, presenting with conviction and authority – while hallucinating and showing absolutely wrong or partly inaccurate numbers.
So the result was not immediately usable, lesson learned. I had to ask follow up questions, re-iterate and double check output. AI still saved me so many hours. I was able to process additional, newer data and re-submissions almost instantly. But, so far at least, it still needs expertise to get results.