How AI Is Changing Repair Quoting: From Photo Identification to Instant Pricing

By Hassan, Technical Lead at Cellbot

Published: 23 January 2025 · Fully reviewed: 27 August 2026

AI repair quoting is reliable only when the system can identify the device, understand the requested repair, retrieve a current shop price and present the price on the right terms. A fast answer is not automatically a sound quote.

Use AI for common, well-defined jobs with controlled pricebook coverage. Route ambiguous models, hidden damage, liquid exposure, board faults, data recovery and unsupported repairs to diagnosis or staff review. The best system is not the one that answers every enquiry; it is the one that knows when it should not quote.

The five gates in one view

!AI repair quote route showing five gates from device identity to monitored customer outcome

Download the AI repair quote test set. It includes routine, ambiguous, unsupported and adversarial cases so a sales demonstration cannot choose only easy examples.

What AI quoting should actually do

The useful workflow is retrieval and control, not imaginative price generation:

1. collect a customer's description, selection or image;
2. identify a device candidate and confidence or ambiguity state;
3. classify the requested service;
4. retrieve a shop-approved pricebook entry;
5. apply explicit location, tax, call-out, delivery or option rules;
6. present the device, service, price, assumptions and next step; and
7. preserve the accepted price and inputs with the booking or enquiry.

An AI model may help at steps two and three. It should not invent a missing price at step four. A configured rule engine is usually better for the final amount because the shop can inspect and test it.

Gate 1: prove the device identity

Text can be more reliable than a photograph

A clear model selection, settings screenshot or model number may identify a device better than an exterior photograph. Many phones share housings; cases hide distinguishing details; regional variants can affect parts; and a cracked or dark screen removes useful clues.

Treat a photo as evidence, not certainty. Ask the system to return a catalogue match, the input that supports it and an ambiguity state. Do not expose a raw model probability as if customers understand how it was calibrated.

Test the difficult neighbours

Build pairs that look or sound similar:

• base, Plus, Pro and Pro Max models;
• devices across consecutive years with similar housings;
• regional or network variants where parts differ;
• a device inside an opaque case;
• a partial model name such as “Galaxy A15”; and
• an image containing two devices.

The test passes when the system requests the missing information or routes the case safely. Guessing the most common model is a failure even if it is often right.

The ICO's AI accuracy guidance distinguishes personal-data accuracy from the statistical accuracy of an AI system and recommends minimising error risk for the intended purpose. Its broader AI risk toolkit is under review following UK legal changes, so check its status before relying on it for compliance decisions.

Gate 2: define the repair scope

“Broken screen” can mean cracked glass with a working display, no image, touch failure, frame damage or damage that needs inspection before a safe price exists. The quote route must distinguish a defined service from a symptom.

For a routine screen replacement, state what the price assumes. For water damage, intermittent faults, no-power devices, board work and data recovery, offer a diagnostic route with clear terms instead of predicting the final repair.

Use these scope states:

This is where generic chatbots often fail: fluent language hides an undefined job.

Gate 3: retrieve a controlled price

The pricebook entry should contain at least:

• catalogue device and repair identifiers;
• price and currency;
• shop or location scope;
• effective date and last reviewer;
• tax treatment;
• part or service tier where relevant;
• included labour and warranty description; and
• active, unavailable or review status.

Do not let the language model calculate a sale price from general web knowledge. If no approved row exists, return a no-price exception.

Version the price accepted by the customer

If the shop changes its pricebook after a customer accepts, the repair record should preserve the quoted amount, assumptions and acceptance time. Otherwise staff cannot explain the difference between the current price and the agreed one.

For shops maintaining thousands of combinations, the phone repair pricing guide explains price construction and the inventory guide covers stock evidence. The AI layer cannot rescue a pricebook nobody owns.

Gate 4: present the customer price clearly

Show the matched device, service, price, what is included, material conditions and whether the amount is a fixed price, estimate or diagnostic starting point. Give the customer a simple way to correct the device or request review.

The CMA's January 2026 price-transparency summary says consumer prices should be clear, complete and accurate and that unavoidable charges should normally be included up front. Apply the full CMA209 guidance to your particular invitation to purchase and obtain legal advice where needed.

Operationally, test:

• VAT and any mandatory booking, call-out or delivery charge;
• optional express service shown separately;
• deposits distinguished from the total price;
• variants with different part or service tiers;
• price changes during an open conversation; and
• the text shown when the total cannot yet be calculated.

Do not bury “from” or “subject to inspection” beneath a visually dominant number. If inspection can change the scope, describe what is provisional and what happens next.

Gate 5: monitor the real outcome

Pre-release accuracy is not enough. Compare the quoted record with the booked and completed job.

Track these measures separately:

A single “accuracy” percentage can hide the most expensive mistakes. Weight failures by consequence and publish the definition internally.

The NIST AI Risk Management Framework core recommends documented testing before deployment and regular measurement in operation. Its August 2026 draft TEVV-Athlon framework is designed to assess real-world AI impacts across different applications. Cellbot's five-gate test adapts that discipline to a repair quote; it is not a certification.

A worked test case

This is an illustrative test, not a customer result.

Input: “How much to replace the screen on this?” plus a photograph of a cased phone with two rear cameras.

Bad outcome: the system guesses a model from appearance and presents a fixed price.

1. image analysis returns two plausible model families;
2. the identity gate fails because the exact catalogue device is not known;
3. the system asks the customer to select the model from Settings or supply the model number;
4. the confirmed identifier maps to one active screen-replacement row for the customer's shop;
5. the price route displays the total and included service terms; and
6. the accepted quote stores the device, pricebook version, amount and timestamp.

The extra question makes the process slower by one step and safer by several orders of consequence. Optimise for a correct decision, not the shortest chat transcript.

The minimum useful test set

Use at least these categories before launch:

• ten common supported device and service pairs;
• five near-neighbour models;
• five unclear customer descriptions;
• five unsupported devices or repairs;
• five diagnosis-only symptoms;
• a stale price, missing price and location-specific price;
• tax, deposit and optional-service cases;
• duplicate submissions and a price change mid-session;
• blurred, cased, partial and multi-device images; and
• hostile instructions asking the assistant to ignore the pricebook.

Record the expected route before running the test. If the expected answer is written afterwards, the evaluation can be bent around the output.

Use the downloadable CSV as a starter set, then replace its illustrative rows with the shop's own catalogue and policies.

When AI quoting is the wrong choice

Do not deploy a customer-facing quote route when:

• the shop has no maintained pricebook;
• staff routinely change prices without recording why;
• most work needs bench diagnosis;
• nobody owns exceptions or monitoring;
• the software cannot preserve the displayed and accepted terms; or
• a simple service menu already handles the enquiry reliably.

A structured booking form may outperform AI for a narrow service list. AI earns its place when natural-language or image input reduces customer effort without weakening price control.

How Cellbot handles the route

Cellbot's customer-facing AI checks the shop's configured pricebook and asks for more detail or routes the enquiry when it cannot make a reliable match. Shops remain responsible for reviewing prices and policies. Pricebook coverage depends on the paid tier: Starter includes 1,000+ iPhone and iPad prices, Pro includes 3,000+ iPhone, iPad and Samsung prices, and Pro Plus includes up to 10,000 prices across those families plus other Android phones and MacBook. Custom prices remain available on every tier. Current prices, limits and feature availability are on the Cellbot pricing page.

Cellbot publishes this guide and sells the product being discussed. Do not accept these statements as test results. Run the same labelled cases against Cellbot and every shortlisted supplier, preserve the output and compare the final route.

The wider repair shop automation guide explains triggers, fallbacks and rollout. The software evaluation plan provides a controlled 30-day buyer test.

AI repair quote FAQs

Can AI identify every phone from a photo?

No. Similar housings, cases, image quality and regional variants create genuine ambiguity. A responsible route asks for another identifier or sends the case to a person.

Should an AI quote be fixed or estimated?

That depends on the scope and the shop's terms. A supported standard repair with a current price may support a clear fixed price. Hidden damage and uncertain faults usually need an estimate or diagnostic route. Label the amount accurately and seek appropriate advice on customer terms.

What should happen when the quoted model is wrong?

Stop the automated route, explain the mismatch, preserve the original input and quote, and have an authorised person review the price and customer remedy. Use the incident to add or strengthen a test case.

Is a high confidence score enough to show a price?

No. Confidence must be calibrated for the shop's catalogue and failure cost. Device identity is also only one gate; job scope, pricebook coverage and price presentation must pass.

Does AI quoting increase conversion?

It may reduce response delay, but no universal conversion uplift should be assumed. Measure comparable enquiries from input through accepted booking and completed repair, and disclose other changes that could affect the result.

How often should the system be retested?

Run the critical set after changes to the model, catalogue, prompt, price rules or customer flow. Monitor production exceptions continuously and schedule a full review at a frequency proportionate to quote volume and failure risk.

Sources, search evidence and update note

This article was fully rebuilt on 26 August 2026. It removes unsupported speed, conversion, identification-accuracy, after-hours-demand and Cellbot plan claims that were not tied to current first-party evidence.

Fresh UK DataForSEO data recorded 10 monthly searches for “AI quoting software”, commercial intent and £19.77 cost per click, with keyword data last updated 14 July 2026. The live desktop query “AI quoting software repair shop” triggered an AI Overview. Cellbot's software-and-AI guide ranked fourth on the screen and second among standard organic listings after a trade quoting product; the dedicated article did not appear in the sampled results. These are dated observations, not a traffic or future-rank forecast. Exa supported semantic competitor and source discovery but was not used as rank evidence.

Primary references:

• ICO guidance on AI accuracy, checked 26 August 2026
• ICO AI and data protection risk toolkit, checked 26 August 2026 and currently flagged by the ICO as under review
• CMA price-transparency guidance, updated 7 January 2026
• NIST AI RMF core, checked 26 August 2026
• NIST TEVV-Athlon evaluation framework, draft announced 7 August 2026

Continue with the repair automation control guide, the pricebook guide or the software buyer test.