We are building new insurance products that can be responsive and relevant to our customers. Part of this is using photos taken from mobile phones at the start of the policy and during a claim. The big challenge here is how to look at large volumes of images each day.
Lets consider some numbers – high volume is say 50,000 or 100,000 images a day.
Images per day ~100,000. Number of humans needed to look, just look – open-look-close-next – say 10 seconds each – that’s 1,000,000 seconds = 35-man days per day!! This is just to LOOK, nothing else. Day in day out – add in hiring, churn, 24 hour operations and soon you will have over 100 people employed doing this simple task, just looking at the pictures!! Add in data collection, quality review, report writing, audits and reports and you can easily double that to 200 people.
How can a human look at all the pictures and collect a vast number of data points analysed in real time from a single image or multiple images, and then determine location, time, date, bike not bike, bounce not bounce, damaged or not and where.
Humans don’t have good memories when compared to a computer. How can a human retain the data and reproduce it in real time analytics for consumption and actionable reporting? Its no wonder that in the world of insurance doing sensible things like visual inspections of vehicles prior to taking out insurance is not done and the ownness of fraud detection etc is left to when a claim happens.
However in my world of micro insurance and high volume / low costs premiums and claims this is not an option.
Computers specialise in speed and volume and automation repays investment with data. Every photo consistently and accurately assed and the data put into actionable reports. Comparisons? Computers can easily compare large volumes of data providing information for the business to act upon.
Training a large human team is costly and time consuming. Once trained its difficult to update the process and even harder to completely change a system or process. Once trained humans leave, and churn gives rise to constant levels of training and support for core business operations. Training is easy for computers, once its understood the rule and command base they just keep going and going….
Biases and inconstancies are constant in human driven processes even after training. Computers have bias too but are consistent in this meaning that over time it can be driven out. Holiday’s, moods and sick days are prevalent in a human workforce, we all get sick and want days off, we don’t want to work 24/7, we love a good holiday and have moods which all effect the quality of work and the productivity, plus if we get a better offer – we leave!
The process of hiring a good member of staff on average takes 4 to 6 weeks this doesn’t include notice periods and gardening leave which drastically change from days to months and from industry to industry and location to location. If you want to double the capacity of a computer automation system it could all be done overnight, fully tested and ready for the new capacity challenge.
If your business relies on users taking photos to prove your business operation, then without an AI automation system you make huge compromises in your business that costs real money and real customer issues day in day out. A 1 in N approach to service is just not good enough in the personal digital world. 100% is the only way to guarantee success.
Use case comparison for simple vehicle inspection
Number plate extraction
Human: 10 to 15 seconds, 10 start of day 15 seconds towards end of day, higher possibility of missing and getting confused with characters. Max images per person in 8 hours = 2,400 – but more realistically 1,500 per day
GPU: 2 second all day every day lower probability of inconsistencies. = 21,000 in a typical day
Human: 10 to 15 seconds, 10 start of day 15 seconds towards end of day maybe longer with a possibility of missing damage. Humans lack consistency within processes and come with biases. Max images per person in 8 hours = 2,400 – but more realistically 1,500 per day
GPU: 4 seconds all day every day = 11,000 in a typical day
Is it a car or bike in the photo?
Human: 5 to 10 seconds, 5 start of day 10 seconds towards end of day maybe longer with a possibility of missing damage. Humans lack consistency within processes and come with biases. Max images per person in 8 hours = 4,000 – but more realistically 3,000 per day
GPU: 4 seconds all day every day = 11,000 in a typical 12 hour day
So to complete a task of say 50,000 images for all three tasks over a 12 hour work day:
Number Plate Team = 34 people
Damage Team = 34 people
Not Bike = 17
Reporting Team = 10 people
Management = 12 people
Churn = 30% = 30 people constantly in training and hiring
Simple Cost = $110,000 per month to process ONLY 50% of images!!
Number Plate Team = 3 GPU’s
Damage Team = 4 GPU’s
Not Bike = 4 GPUs
Reporting = Automated and VERY valuable delivered in real time
Management = 0 (included)
Churn = 0
Cost = $10,000 per month
Costs are all local costs and people costs vary around the world, however there is less variables when costing GUP’s to do the work.
Using the image data to our advantage must not be undervalued. An automated AI system will return huge value per vehicle or journey. For a human based solution, the process will always be a compromise and the insurance company will never achieve the quality service or operational efficiencies that is needed in its business.
Without technology we will not be able to maintain service standards because of poor image standards, data quality and lack of actionable reporting across every policy.
This not about computers taking jobs, this is about the ability to server millions of customers with insurance and sharing in their risk each day. It’s about making insurance available to everyone.
At MIC Global we are focused on changing the way business insurance is developed and processed. We are insurance with AI built in with API’s. We are in the forefront of that change; developing policies by the season, job, by the hour, by the day and by the Km, thus fitting our model to that of the platforms and the way small and micro businesses see risk. We are unbundling business policies so that the cover offered fits with peoples and business needs or the actual job or process being undertaken. Making Business Insurance transactional.