An explorable layered model of a smash-repair management platform evolving from a system of record to an AI-native operating system, with one repair job animated through every stage
This model is a smash-repair management platform like Clear21, drawn as eight layers and shown fully built: the AI-native destination. The bottom layer stores the business: customers, vehicles, estimates, parts, jobs, invoices and payments. The lane on top is one repair job in ten steps; its bars show who does the work: orange is people, blue is software, violet is AI.
Drag to turn the model. Click any step on the lane to see who does what there.
Today the software mainly remembers. Staff assess the damage, write the estimate, chase the insurer, order parts and plan the workshop, then type the results in. Only the data layer is built. The software is the filing cabinet; people are the engine.
What goes wrong: rekeying errors, lost photos, and jobs stuck waiting on a phone call nobody made.
Workflow and business rules are added. Screens guide staff through each step in order, track status and send reminders when an insurer deadline slips. People still do all the thinking, but the work now happens inside the software, so every step leaves a record.
This record of how work is really done becomes the raw material for AI later.
AI models and unstructured data join: damage photos, PDFs, emails. AI helps with single tasks. It drafts an estimate from photos, writes the customer update and checks an invoice against the approved estimate. A person reviews every suggestion and still moves the job along.
What goes wrong: dark photos or missing angles give a wrong estimate, and damage hidden behind the bumper is invisible to a camera.
Tools and agents are added. An agent can now act: send the estimate to the insurer, order the parts, book the paint booth. It completes multi-step work across the system, but stops for a person to approve important decisions such as the estimate, a supplement or the final invoice.
What goes wrong: the wrong part variant ordered, or an action taken on out-of-date information. Permissions and approvals limit the damage.
Evaluation and the human feedback loop complete the stack. The system watches every job in the shop, works out what should happen next, does the routine work and hands people only the exceptions: a disputed supplement, a structural repair, an unhappy customer. Software becomes the engine; people steer.
Orange that remains marks where people still decide or physically repair the car.
Bolting a chat window onto the old product leaves everything underneath unchanged. The chatbot cannot see the photos or the live job, cannot take actions, cannot check whether its answers are right and learns nothing from corrections. People still do the work. Switch to AI-native to see what must be built underneath.
Doing work safely takes more than a model: tools connected with permissions, guardrails on what AI may do alone, monitoring and an audit log of every action, and evaluation against real outcomes. The organisation changes too. Staff become reviewers and exception handlers, and someone owns the quality of each AI-run process.
Each correction a person makes is captured and flows back into evaluation. That learning loop, not the chat window, is what makes the system improve.
More work done by AI produces more observable outcomes and more human corrections. Those become better evaluation data, which make better agents, which can safely take on more work. Each turn widens what can be automated, and the data comes from real jobs on your platform, which a newcomer cannot copy.
The automation shares shown are illustrative, not measured.