Despite the noise around frontier AI models, most of the non-tech economy still runs on spreadsheets, manual lookups, and legacy systems.
This is even more so the case in automotive repair, where the gap between what AI promises and what it actually delivers costs the industry real money.
The global auto repair market is worth over $1 trillion but the US alone accounts for more than $180 billion of that annually, spread across more than 250,000 businesses, with no single operator holding more than 5% market share.
Latest Videos FromTechRadar
CEO and co-founder of Partly.
The average American vehicle is now 12.6 years old, the oldest fleet on record, meaning more frequent repairs, more complex parts, and more pressure on workshops already running on thin margins.
Since 2022, repair costs have risen by 25%, well above general inflation. Despite all of this, the industry still matches parts largely through manual catalogue lookups, experience, and guesswork. This results in wrong parts getting ordered, vehicles sitting in shops and workshops absorbing the cost of returns, rework and lost technician time.
That problem exists across every market, but the US is where it is most apparent; the market is enormous, deeply fragmented, and has operated without AI-native infrastructure designed specifically for it. That reflects a genuine technical challenge that general-purpose AI has not solved yet and cannot solve in its current form.
Why general-purpose AI models fall short
General-purpose models are built for optimizing language, reasoning and creativity, but not to be ruthlessly precise within one specific industry. In industries like automotive repair, tolerance for error is close to zero. For example, if a model recommends a headlight that isn’t the right fit, it holds up the technician repairing the car, delays the customer and costs the insurer paying for the repair considerably more.
When you benchmark leading general-purpose models against domain-specific automotive parts tasks, the results are stark. The best general models achieve around 5% precision on parts identification benchmarks. A well-trained human does considerably better, but is constrained by memory, catalogue complexity and time.
A model purpose-built for this problem, trained on proprietary original equipment manufacturer (OEM) data, normalized catalogue schemas, fitment rules and live transactional feedback, achieves over 90% precision on the same benchmarks – not just a marginal improvement but a different category of capability.
The data and architecture problem
There are structural reasons why the gap is that large. Automotive parts and repair data is not a larger version of generic text, it is a constantly evolving graph of relationships:
VINs, trims, sub-models, region-specific variants, supersessions, aftermarket substitutions and workshop-specific preferences. Much of this lives in fragmented OEM catalogues and proprietary formats not available on the open web.
The architecture required is different too; a parts decision in a real workshop needs to account for vehicle history, insurer regulations, supplier inventory, contractual pricing and technician preferences simultaneously, requiring a model built to ingest structured catalogues and enforce hard constraints, not one optimized for broad conversational usefulness.
Building the infrastructure that makes scale possible
Building a domain-specific foundation model is not a matter of fine-tuning a general base as the dataset has to be assembled through years of OEM agreements, schema normalization and continuous integration of repair, claims and inventory data.
The model needs to be wired into the systems that run the business, so that every accepted recommendation, return and job outcome feeds back in and compounds accuracy over time. The competitive advantage builds over time through the combination of proprietary data, deep integration and continuous learning. The longer the model runs, the better it becomes.
For the US market, this matters more than anywhere else; the scale of the opportunity, 250,000 repairers, a fleet getting older every year, costs rising faster than inflation, means the compounding value of getting parts right the first time is enormous.
The markets where this technology has been deployed in Europe and Asia-Pacific have demonstrated that the productivity gains are real and measurable. The US has, until now, had no equivalent infrastructure to access them.
What buyers and investors should be asking
For technology buyers in complex verticals, the right question to ask of any AI vendor is not how capable their model is in general, but how it performs on the specific failure modes that cost your industry money. In automotive repair, that means first-time-right rates, return rates and cycle time compression.
The same logic applies to investors: the size of a model is not a measure of how hard a business is to displace. The relevant questions are how hard the dataset is to assemble, how embedded the model is in operational workflows and how general the reasoning ability is within the domain.
General-purpose models will continue to improve and will remain valuable for a wide range of tasks. But the deepest, most durable value in AI will be created where models stop being generic assistants and start becoming invisible infrastructure, optimized for one hard problem at a time.
In automotive repair, that means fixing the parts problem that has taxed the industry for decades. In the US in particular, where scale amplifies both the cost of the problem and the value of solving it, that shift is overdue.
We’ve listed the best GPS fleet tracking systems.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
https://cdn.mos.cms.futurecdn.net/DVffQnnibMWmNpx2Wfb5Se-1920-80.jpg
Source link
