“Rather than looking at 100 job sheets, they can look at 40, and those are the 40 job sheets that they can actually save some decent money on or are really worth challenging.”
That is Charlie Brooks, VP of strategy, data and growth at Epyx, summarising the company’s imminent Automatic Vehicle Authorisation (AVA) system. The name is very ‘does what it says on the tin’, which is refreshing for an AI-powered tech product, and it is designed to approve straightforward maintenance jobs – things a fleet maintenance controller would glance at and say ‘yeah, fine’ – with no input required on their part. From the fleet or leasing company’s perspective, it is a theoretical load off their plate; for garages, there should be less waiting for the thumbs-up to start a job.
“We got some funding at the back end of last year to do something in the AI/machine learning space and looked at which part of the process we would augment or improve,” explains Brooks. “Each fleet [Epyx works with] still has several – sometimes up to 20 or 30 – people doing maintenance controlling. They’re reviewing each job sheet submitted by the garage, combing through it, saying ‘we should authorise this, we should not authorise this’. It still feels like a hugely manual task.”
Brooks estimates the company’s 1Link platform has processed “probably 20 million job sheets with the best part of 40 million line items or pieces of work done to vehicles,” during the past five years, data from which underpins the new development. It officially has five million vehicles registered to its platform, though internal estimates put that nearer seven million, and the system is said to process 20,000 jobs a day.
“We basically took all of that learning and behaviour to see if we can develop something that says, ‘you should have authorised this or you shouldn’t’ – that was the premise of it,” adds Brooks.
AVA comprises four new functions: an authorisation propensity model, a prior pricing model, an upsell probability model and a wear rate model. In short, they give the fleet an indication of how likely it would be to approve a certain type of job, a barometer of the cost, a heads up if the garage is being cheeky by systematically adding things to the bill and an idea of when a part would typically be changed.
It cross-references 1Link’s dataset with the fleet’s own, the former (blue in the screenshots) to illustrate industry-standard metrics, the latter (purple) to show the user what their fleet has historically done.
“The authorisation propensity model is probably the most intelligent one,” says Brooks. “That looks at the history of job sheets that have been authorised and reviewed by both the internal fleet and the 1Link benchmark view. It’s based on the amount of time maintenance controllers spend looking at this kind of line item, for this type of vehicle, in this type of garage, the number of job sheet versions, the amount of time they request information, the length of time they review it and how often they end up rejecting it versus authorising it.”
Both metrics – 1Link’s and the fleet’s own – are rated from zero to 100 (the numerical format is consistent across all four models), and a strong correlation between the two means a user’s score aligns with the industry standard. A score of 100 for authorisation propensity means Epyx is ‘pseudo-certain’ a maintenance controller would approve the job, while zero is a big, fat ‘no way’. The system stops short of generating a score when it lacks sufficient data (the exact amount varies by model) so if, say, there has only been one wiring loom repair in the previous six months, the user sees ‘N/A’ in place of a number.
The second of the four is the prior pricing model, which, according to Brooks, “basically says ‘how does the price of this type of line item, at this type of garage and for this type of vehicle compare with what you’ve previously authorised over the last six months’. A higher score indicates you are very likely to, because it’s far cheaper than everything else. A much lower score indicates it’s far more expensive than everything else.”
Third is the upsell probability model, which Brooks says is the “one that people seem to be most interested in. This basically asks, ‘how often do we see this line item being fitted to job sheets?’. So maybe something like a windscreen wiper is fitted on, say, 5% of job sheets. If a garage has a higher score, that indicates we are more confident that the garage is not upselling or adding too many of these line items to your job sheets.”
Upsell probability is not applied to every type of job, because the practice generally does not go hand-in-hand with the likes of services and MOTs. Its objective is to spot repetitive and unnecessary charges, typically for parts, and according to Brooks, one of the company’s large fleet customers was subject to exactly that earlier in the year.
“I think they said they had 20 job sheets in a row where a key battery was fitted at the same dealer site. Of course, those [sheets] were all sent to different people to review, and in isolation, each one looked perfectly reasonable, but actually, when you looked at them as a whole, maybe one in 100 job sheets should have a key battery – not 20 out of 20.”
Fourth and final is the wear rate model, which tells the user whether a part would typically be replaced at the time.
“It looks at the vehicle’s history, and we know that a first brake change or a pollen filter change on, say, an Audi E-Tron, is normally at 32,000 miles – but the car’s only done 17,000 miles,” says Brooks. “It would have a far lower wear rate score, because we don’t normally expect that part to be changed at this time.”
All of the above can be toggled, weighted and, crucially, automatically approved. If, for example, a fleet is more concerned with its authorisation history than with pricing, then the emphasis can be tweaked accordingly, and if a certain type of job meets certain scoring criteria, the system’s automated responses – which can be switched on and off – can independently tell the garage to go ahead with the work.
The threshold for auto-approvals can be customised, so setting it at 90 for authorisation means the fleet would almost certainly approve that job, but it is a high bar to trigger the auto-approval. It can also be set up to automatically approve work that costs less than a specified sum.
The system will ask garages for more information about work that does not qualify for automatic authorisation, as Brooks explains: “When the garage gets the response, it might say ‘this [work] is authorised, but we need a bit more information here and here’. That means they can at least start on one [job], and it kind of minimises downtime, rather than waiting for someone to manually review the whole thing, which might take 10-15 minutes.
“Say there was a very high upsell score, when we respond to the garage we would say, ‘it looks like you have added this item to several job sheets in the recent past, are you sure this vehicle needs it? Please provide evidence/a photo’. We tailor the wording a little bit in the response to highlight what we’re not sure about… and when they resubmit it, it then gets picked up by a member of the maintenance control team.”
Brooks claims the time between a garage issuing a job sheet and receiving an automatic approval notice is about a minute but stresses any efficiency gains will be down to where fleets set the thresholds and the extent to which they use the service. With that caveat applied, he estimates that it could “cut the maintenance control of work by 30-40% – but that’s based on my feeling of what was an acceptable level of risk.”
He is adamant that the scope of the company’s data means AVA “couldn’t be done by anyone but 1Link… we’re looking at four million jobs, not just the 100,000 that a specific lease company might. Likewise, knowing how the garage behaves, how many wiper blades you’d expect to see on job sheets – if you’ve got 100,000 of them, you can start to get to a decent view and be confident about what’s going on.”
As for pricing, Epyx says there will be a monthly cost for the standalone AVA portal, which fleets can initially use for training and familiarity, along with a per-transaction fee for automatic responses. The company was putting the finishing touches on the system at the time of our interview in early September, and hoping to trial it with a leasing company “broadly to make sure it does what we believe it should be doing,” according to Brooks, with the aim of rolling it out as soon as possible, so watch this space in the coming months.
