VIDEO TRANSCRIPT
The final thing I'd like to go through is just some custom case studies that I've personally been involved with or have led or delivered, either as part of my former employer or as part of Pathfinder.
So the first one is a case study for doing credit risk analysis for a bank. It used to take about a week for the credit risk team to get an answer to a question like, "Can you show me all of the lines that are maturing within the next two years, within these six postcodes, dropping from a high risk to a low-interest rate, and they are high risk, and they're probably gonna mature to a higher interest rate? Can you show me that list?" That would typically require the data team to run a whole lot of SQL queries, compile a report, and send it back. We put all that into a chat interface on the Google Cloud platform. They were already at Google shops, which made it quite a lot easier, but now you can just ask that question.
You can say, "Can you show me all the lines that are maturing in these three postcodes?" and it'll do that. You can ask to see it as a time series. You can then ask to export it as a CSV and interrogate it further.
So that basically took something that took a week and dropped it to a minute.
The next one is one that I think several CIOs could relate to, which is being able to modernize thirty years' worth of undocumented dirty Delphi code in six weeks instead of twelve months. Now, the modernization effort didn't take six weeks, but the documentation, we were able to get down to six weeks. The initial plan was to get a team of engineers to review thirty years' worth of code over the next year, which would have been a horrible process that they would have hated doing, and to document functional and technical specifications.
We were able to get this to work initially with one function, with one part of the code, but then eventually able to automate most of it, which was then allowed a technical team to review the outputs.
They did still have to critique it. That's why it wasn't an instant thing, but I was able to create beautiful, functional, technical architecture documents showing things like pseudocode, all of your business rules, all of your interfaces, your dependent systems, and even tried to describe how a user would use the code.
The next one was being able to go from three weeks to one day to assess reimbursements from a payer for a private hospital group. This private hospital group has about seventy-five hospitals in Australia.
This payer process is very manual. It requires that somebody looks at all of the doctor's notes, all of the blood tests, all of the procedures that happen during somebody's stay, then coding it to specific codes for reimbursement.
It was a process that had some OCR in it in the past, but it had a very low success rate. We were able to actually automate all of that by using GPT-4, where now most of that has an acceptance rate of, I think, over ninety percent where you now basically put all of this documentation through this model and it will automatically assign all of the different codes and describe all of the reimbursements that are required. So that used to be a three-week process and, or it didn't take somebody three weeks to do, but from the moment it started. It took three weeks before you got a result because they had such a load to do.
The next one is around reducing IT regular IT incidents.
Now this is. We've been able to achieve a ninety percent reduction at one of Australia's major alcoholic beverage companies.
And this is using all sorts of interesting combinations of tools with GPT-4 with simple things like if there is an outage, it will go through all of the log files automatically to try to highlight exactly where the errors are. So it'll look at all the comments of the log files. Before this, it had to get done manually. Somebody essentially ran macros across a whole lot of CSVs, even though they had a very mature monitoring stack.
This was still required. It also does lots of other things like if somebody raises an IT support request, it will ping them automatically on Microsoft Teams with possible resolution suggestions based on looking at what is not only in the support documentation but also it goes through all of the comments and resolutions of all previous support tickets and suggests things to that person that way. Also, if they haven't really given some great information on the save of the ticket, it will ping them to ask for more information. Maybe they haven't really described what the problem with the incident is.
So we've been able to automate all of that, and it's reduced the load for these common things by ninety percent.
The next one was really interesting. So this is where we had They needed to simplify a thousand pages' worth of loan and credit documentation for a bank in New Zealand.
They wanted to pressure from the regulator to simplify these because it was verbose. It was duplicative. It was legalese.
And they were migrating from one system to another and this was the time to simplify it. They realized that it wasn't possible to do this in the timeframe. It's just they don't have the capacity to get their legal team to do this. So what we did was we were able to make the model assess all of their documentation with two key criteria. The first was make it readable by an eighteen-year-old with a second language with English as a second language. The second was make it readable by an eighty-year-old with English as a second language.
But in both of those situations, it still needed to be legally enforceable.
So the first step was to get it to an audit. It went through the entire set of documentation. So first, it was just the loan agreement.
And it gave a few-page report of where things were for both duplicative, and suggestions of changes. The second step was where we started playing with actually automating, creating the documents themselves and playing with that, but essentially we're able to really scale up the impact there.
And the final one is with a major telco here in Australia. We have ninety-four percent of all calls on now being automatically summarized. And this is saving the call center on average five hundred hours a day, and that's a conservative figure.
It's because it would typically take two minutes to summarize a call now that two minutes is gone.
Ninety-four percent of the time it's been completely accepted by the agent. And, once you multiply it out, it's conservatively five hundred hours a day. In fact, it can actually be quite a lot more when you take into account the size of the telco.