What a photo estimate actually knows

Photo-based damage estimation reads a visible surface and prices it against a database. The gap between what the photographs contain and what the repair will cost is where an adjuster still earns the file.

Updated September 13, 2026 Advanced

An adjuster opens a file and the estimate is already there. The question worth asking is the narrow one: what does that number actually know?

It knows what the outside of the vehicle looked like in a small set of photographs, and it knows what the parts and labour it inferred from that appearance cost in the database it queried. Everything else in it is inference, and the quality of the inference falls off sharply as the damage moves inward from the paint.

The two systems behind one number

The interface is a phone and a form. Underneath there are two quite different things doing the work.

The first is a vision model. It takes the images, locates the vehicle, segments it into panels and components, and classifies what it sees on each one: a scratch, a dent with a rough depth class, a crack in plastic, a torn or displaced part, glass, lamp, wheel. Alongside it the vehicle is identified — a plate or a VIN read from a photograph, decoded to a year, model and trim, which is what determines whether the bumper in front of the camera has a radar behind it. The classification is then translated into work: a repair or replace decision per component, a refinish operation on the panels that have been disturbed, a blend into the adjacent panel, and the labour hours attached to each of those.

The second system is the one with real authority, and it is old. Collision estimating databases have been pricing this work for decades: part numbers by trim, published labour times per operation, paint material calculations, local rates. The model does not price anything. It selects lines, and the database prices them.

That division is the whole story of where photo estimation is reliable. Audit a photo estimate for arithmetic and you will almost always find it correct, because the arithmetic was a lookup. The errors are in the selection — which lines were pulled, which were not, and whether the operations chosen describe the repair a shop will actually perform.

Where the surface stops telling the truth

Hidden damage is the structural problem, not an edge case. A bumper cover absorbs an impact and looks scuffed; behind it the absorber is crushed, a bracket is broken, and the reinforcement bar has moved. None of that is in the photograph, and a model fitted on photographs cannot recover it. Experienced estimators handle this by reading the energy of the impact rather than the surface: a deformation pattern implying load travelling into the rail, a gap that has opened where two panels meet, a wheel that no longer sits where the other one does. Some of those cues survive into an image and some do not.

Capture conditions degrade what is left. Direct sun flattens a dent, overcast light hides a crease, a wet panel reflects the sky and reads as damage that is not there, and a low angle turns a shadow line into a dent while making an actual dent disappear. Photographs taken at the roadside, at night, on a dirty vehicle, by somebody who has just been in a collision, are a different input distribution from the training images, and the failure is quiet. The estimate comes back looking exactly like an estimate that is right.

Then there is the vehicle itself. Repair method is a function of construction, and construction is not visible through paint. An aluminium panel that a steel-era rule would have pulled has to be replaced instead; rivet-bonded and adhesive-bonded structures have manufacturer procedures that forbid what a shop would once have done; high-strength steel sections carry sectioning restrictions. A model can learn the association between a vehicle identifier and the usual method, which holds until the trim decode is wrong or the vehicle is a variant with little history in the data.

Recalibration deserves its own paragraph, because it is the scope item most often missing on current vehicles and it goes missing for a structural reason rather than a careless one. Driver-assistance sensors sit behind bumpers, in mirror housings, in grilles, and behind the windscreen. Nothing about a scuffed bumper in a photograph indicates that the radar behind it must be aimed again once the cover is refitted, or that removing and reinstalling a windscreen triggers a camera calibration with its own procedure, its own equipment and its own labour line. The operation is invisible on the surface and it is expensive. Whether it appears on the estimate depends on the vehicle decode and on whether the estimating platform’s rules attach calibration operations to the components involved, which is a configuration question inside the insurer and not something a photograph can settle.

Supplements are the system’s own confession. A photo estimate is a first number that everyone in the chain expects to be revised once the vehicle is on a lift, and the supplement is where the revision happens. That is not in itself a failure; a staged estimate with an honest supplement process is a reasonable way to run a claim. It becomes a failure when the first number is treated as the estimate rather than as the opening of one, and when the supplement cycle carries friction the first number did not: a second authorisation, a second wait, a repairer with the vehicle already disassembled and no instruction.

The borderline that moves the whole file

A low first number does the most damage at the total-loss margin. Whether a vehicle is repaired or written off is a comparison between a repair figure and a value, and both sides of that comparison are uncertain early. An estimate built from visible damage, missing the reinforcement and the calibration, keeps a vehicle on the repairable side of the line on paper. The consequence is not a wrong number; it is a repair that is authorised, begun, supplemented twice and then written off anyway, with the vehicle disassembled, the customer in a hire car and weeks gone.

That is the case worth building an exception queue around. Where the estimate lands inside a defined band of the threshold, the cheapest intervention available is an inspection before authority, and it is cheap only while the vehicle is still assembled.

What to check, in order

Read the operation lines before the total. Every component should carry a repair or replace decision, and the ones replaced where a photograph showed a scuff are the ones to ask about. Look for what is absent rather than what is wrong: calibration operations on a vehicle equipped for them, corrosion protection after a panel is replaced, seam sealer, the diagnostic scan before and after. Absence is harder to see than error, and it is where the money is.

Check whether the labour times came from the database or were entered by hand, because a hand-entered time on an automated estimate is a human overriding a published figure and the reason belongs in the file. Check the parts sourcing — original, aftermarket, recycled — against what the policy wording and the local rules permit, which is a legal question the model does not model.

Then go back to the photographs and count them. An estimate built on a handful of images taken from two positions is a weaker document than the same estimate built on a full set with the required angles, and the difference is visible in seconds. Missing angles are the cheapest thing to fix in the life of a claim: asking for four more photographs costs a message and a day, and it costs the same message whether it is sent on day one or after the repairer has phoned.

Whose number it is

The last thing to hold about a photo estimate is where it comes from socially rather than technically. The images were taken by the policyholder, usually on the day, in whatever conditions existed, to satisfy a form they were told to complete. Those images then become the evidentiary basis of the first offer on their claim, and the first offer is the anchor everything downstream negotiates against: the reserve that gets set, the authority that gets granted, the settlement conversation, the complaint if there is one.

There is an asymmetry in that arrangement worth naming plainly. The claimant supplied the input and has no view of how it was read; the insurer holds the reading and can revise it at a cost that rises the longer the file runs. The cheap correction sits at the start, and it belongs to the side that knows the input was thin. An adjuster who asks for better photographs on day one is not being generous. They are buying back the part of this process that money cannot fix later.

The authority in the number comes from the database. What the photographs decided was which lines the database was asked to price, and nothing in the file records how well they decided it.

Ariski's take

The part of estimating that automation has taken is the part that was already mechanical: matching visible damage to operation codes and pulling prices from a database a human was reading anyway. That work was never the adjuster's contribution, and defending it as though it were concedes the argument. What the model cannot do is decide what the photographs failed to show, and that decision has always been the job. Our position is that an estimating model earns its place when it makes the first number arrive early enough for a reinspection to still be cheap, and loses it the moment the first number becomes the only number because nobody was resourced to challenge it. That is an operating choice, not a property of the technology.

Rules in your jurisdiction

Deadlines, fault rules and minimum coverage differ by state and country. Pick yours to see the rules that apply to this topic.

Select a jurisdiction to see its rules.

Frequently asked questions

The photo estimate and my reinspection differ by a lot. Which number do I defend?

Neither, until you know which of the two saw the vehicle. A photo estimate is evidence about the visible surface at the moment of capture; a reinspection is evidence about the vehicle in the state the repairer found it. If the difference is scope — operations present in one and absent in the other — the reinspection is almost always the better document, because it was produced after panels came off. If the difference is price on identical operations, the estimate is usually right and the reinspection has used a different parts sourcing assumption or a different labour rate, which is a question about the schedule rather than about the damage. Write the difference down as scope or as price before arguing about either, because the two are settled by different evidence and confusing them is how a file spends a month going nowhere.

Can I rely on the model's confidence score to decide what to inspect?

Only as a queue, never as a conclusion. A confidence score reports how familiar the images looked to the model, which tracks accuracy where the vehicle and the damage resemble what the model was fitted on and stops tracking it where they do not. The predictable consequence is that the cases most likely to be wrong are cases the model has seen little of, and some of those come back with a comfortable score attached. The practical use is inverted: treat a low score as a routing instruction, and treat a high score as no information at all about hidden damage, because hidden damage is invisible to a system whose input is photographs of the outside of the car.

Does an automated estimate change what we owe the policyholder procedurally?

The obligations do not move because the estimating tool changed. Where a regulator sets acknowledgement, investigation and payment periods, they run from the same events they always ran from, and an estimate produced in minutes neither shortens a policyholder's right to question it nor lengthens your window to answer. Two things do become more live. The first is the reasons you give, since a number the handler cannot explain is hard to defend at a complaint stage. The second is documentation of the images: the photographs are now the evidentiary basis of the first offer, so how they were captured, what was requested and what was missing belongs in the file. Where periods have been recorded for your jurisdiction, they are in the rules below.