AI server for loss adjusters and appraisers: damage photos, report drafts, case files

What an AI server does for appraisers and surveyors

An AI server for loss adjusters and appraisers runs the language, image and speech models of the office on its own hardware. It sorts the photos of a claim and suggests the type of damage, types up the voice notes from the site visit, drafts the standard sections of the report from the case file and earlier reports, and checks the finished report against the client's requirements before it is sent; the appraiser decides, edits and signs. Because the models run in the office, photos, notes and reports never go to an outside AI service.

This guide is for the owner or managing partner of a firm of loss adjusters, vehicle appraisers, building surveyors or court-appointed experts, and for whoever looks after the office IT. It explains the work first, then the data questions, then the server in one section with a table. The notes at the end are for the IT team or service provider. Legal points are general information on EU law as of October 2026; your data-protection officer or legal counsel confirms what applies to your firm.

The steps of a case where AI helps

KI-Syndikat, a German-language catalogue of AI use cases for small and mid-sized companies, lists 20 use cases for appraisers and experts. Read as a sequence, they follow a case from the first call to the invoice. The time figures in this section are the catalogue's estimates, not measurements of ours, and the catalogue itself marks some effects as unproven.

Use cases from KI-Syndikat's catalogue for appraisers and experts (ki-syndikat.de, read 25 October 2026), arranged by case step; the division of decisions is our reading.

The catalogue estimates that drafting from CRM data and similar earlier cases cuts writing time per report from 2 to 3 hours to 20 to 45 minutes, and that photo analysis saves 15 to 20 minutes of sorting and first assignment per case. For a counter-expert report, it expects 60 to 80 per cent less preparation when a model extracts every objection and maps it to the passage of the original report it attacks. Treat these as the catalogue's planning figures; a pilot on your own cases gives yours.

Some tasks on the list need no GPU at all. Residual-value requests for total-loss vehicles and diminished-value calculations are data lookups and arithmetic with a short text at the end. The server earns its place with the steps that read images, listen to speech or write long text, and with the fact that one machine does all of them for every appraiser in the office.

Why case files stay confidential in the office

A claim file holds names, addresses, policy numbers and photos of homes, vehicles and sometimes people. Personal-injury and medical expert reports go further: they contain data concerning health, which the GDPR defines in Article 4(15) as personal data "related to the physical or mental health of a natural person". Article 9(1) prohibits processing it unless one of the exceptions in Article 9(2) applies. Which exception covers your reports is for your data-protection officer to decide.

When an office sends a case file to a cloud AI service, the provider processes the data on the firm's behalf. Article 28(1) then allows the controller to use only processors that give sufficient assurance of appropriate technical and organisational measures, which means a contract, a list of subprocessors and a check of where the data goes. Our guide to GDPR, cloud hosting and subprocessors goes through that contract. On a server in your own office, the AI step adds no provider. Article 32(1) still applies to the server itself, with measures such as "the pseudonymisation and encryption of personal data", and our guide on what local AI solves under the GDPR explains which duties remain.

Insurers and courts add their own expectations. An instruction from an insurer or a court can carry confidentiality terms of its own, which the office checks before any case data goes to an outside service. The insurer's side of the same claim, with its own regulatory frame, is covered in our page on GPU servers for banks and insurers, and the courts' own use of AI in our guide to an AI server for courts and the judiciary.

On the AI Act: point 5(c) of Annex III lists as high-risk "AI systems intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance". A model that drafts a property or vehicle damage report for an appraiser to check and sign is, by our reading, not named in Annex III. Your legal counsel confirms this for your use, and our guide to the AI Act for companies using LLMs covers the duties that apply to every deployer.

Damage photo classification: what the model sees and what it stores

A vision-language model reads images as well as text. Qwen3.8-27B is one: NVIDIA's NVFP4 version of it lists text, image and video as inputs and is published under Apache 2.0, "ready for commercial or non-commercial use". The model cuts each photo into tiles of 32 × 32 pixels and turns each tile into one token, the unit the model reads. Its image settings accept up to 16,777,216 pixels per image, which is 16,384 tokens.

That makes resizing a planning decision. By our arithmetic, a 12-megapixel photo of 4,000 × 3,000 pixels becomes about 11,700 tokens, while the same photo reduced to 1,280 × 960 pixels becomes 1,200. For sorting photos by room and suggesting "water", "fire" or "storm", the reduced size usually carries enough detail; for reading a serial plate or a crack width, the office sends a crop of the relevant area at full resolution. A case with 20 reduced photos and a page of notes fits a context of 32,768 tokens.

Storage grows with the photo archive, not with the AI. As an example with values to replace by your own, 300 photos of 6 MB per case and 2,000 cases a year come to 3.6 TB a year before backups. The server keeps the working copies of open cases on fast NVMe drives; the archive stays on the office storage or document system, which the AI reads with the same access rights as the appraiser. Our page on NVMe storage for an LLM server explains the split.

We found no published rate of photos per second for this model on the cards we supply. The method of our guide to vision-language models for document AI estimates about 0.24 document pages a second, each page about 2,100 image tokens, on one RTX PRO 6000 Server Edition with the model in FP8. A reduced photo has fewer tokens, so a case of 200 photos takes minutes, not hours, by that estimate; a test with your own photos gives the figure for your server.

Voice notes from the site visit

Notes from a site visit often arrive as voice memos, which the catalogue's first use case names. A speech recognition model turns the recording into text on the server. NVIDIA's Parakeet TDT 0.6B v3 has 600 million parameters, transcribes 25 European languages with punctuation and timestamps, detects the language by itself and is published under CC BY 4.0, a licence that allows commercial use. Its model card states up to 24 minutes of audio in one pass with full attention. Whisper large-v3, the alternative for languages outside that list, covers 99 languages, and its model card warns that "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)", so the appraiser reads the transcript before it becomes part of the file.

Speech recognition takes little of the server. Our guide to speech-to-text servers for Whisper and Parakeet gives the published throughput figures; an office's voice notes of a day are a small fraction of one card.

Expert report drafting, checks and research

The draft is the step with the largest saving in the catalogue and the one that needs the most care. The model receives the case data, the transcribed notes, the photo tags and two or three earlier reports of the same type as examples, and writes the standard sections: the description of the object, the findings, the list of damage items. It does not set values, assess causes or reach conclusions for the appraiser; those sections stay empty for the expert to write. A check before sending compares the finished report with the client's requirements, flags missing photo references and inconsistent figures, and leaves every correction to the appraiser.

StorageReview measured NVIDIA's NVFP4 Qwen3.8-27B on one RTX PRO 4500 Server Edition at 320 output tokens per second with 32 parallel requests of 512 tokens in and 512 out, and 153 tokens per second with 8,192 in and 1,024 out, on vLLM 0.27.1 (review of 14 October 2026). By our arithmetic, 153 tokens per second is about 540 requests of that size an hour, far more drafts and checks than an office of 30 appraisers writes.

Research in case law and technical standards is retrieval: the model searches a collection of documents and answers with the passage and its source. Two limits apply. The appraiser verifies every citation, as the catalogue itself notes, and standards bought from a standards body come with licence terms that may not allow their use in AI, which our guide to standards in your company AI explains. Reports for foreign clients go to a translation model: EuroLLM-9B-Instruct covers all 24 official EU languages under Apache 2.0, and our guide to on-premise machine translation compares it with others.

We supply the RTX PRO 4500 Server Edition and build AI servers for appraisal offices to order; send us the number of appraisers, cases per year and the languages of your reports, and our team replies with a configuration.

Which server, how many appraisers, and what it needs in the building

The work of an appraisal office needs a modest model. It pays to run it on one server that the whole office shares, for photos, reports, speech and the general assistant, and to size that server by the number of people working at the same time. The capacity column follows our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card, less the model weights, which leaves 25.7 GiB per RTX PRO 4500 Server Edition and 83.04 GiB per RTX PRO 6000 Server Edition.

Usable memory and conversations by our users-per-card rule with an FP8 cache of 1.14 GiB per 32K conversation; weights 20.42 GiB (NVIDIA NVFP4 build, 21.9 GB) and 28.75 GiB (Qwen FP8, 30.9 GB) from the Hugging Face file lists; power by our estimate (cards, processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

One card. One RTX PRO 4500 Server Edition with NVIDIA's NVFP4 build leaves 5.28 GiB for conversations, enough for four cases of full length at once. That suits a pilot or a small office where two or three appraisers write at the same time; speech and translation then need a second card.

Four cards in 2U. The NVIDIA RTX PRO 4500 Blackwell Server Edition is a passive, single-slot data-centre card with 32 GB of GDDR7 at 800 GB/s and 165 W. Four of them run Qwen3.8-27B in FP8 as a pair with 19 cases at once, the translation model on the third card and speech recognition and search on the fourth, at about 1.4 kW, within one 230 V 16 A feed (2.94 kW at 80 per cent). This is the configuration we would size first for an office of 10 to 40 appraisers; four RTX PRO 4500 Server Edition cards in 2U lists the chassis.

Two RTX PRO 6000 Server Edition cards. Our entry server, with two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, holds 47 cases of full length on one 96 GB card and leaves the second for the other models. It suits a larger firm, or one that also wants a bigger general model for the rest of the office.

A server of this kind is loud and gives off its full power as heat, so it belongs in a server room or a cooled rack, not under a desk; our page on a GPU server in an office sets out the limits and our server-room checklist the preparation. Without a server room, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU. A firm without its own IT staff finds the division of tasks in our guide to an AI server without an IT department.

Notes for the IT team or service provider

1. Integration. The case management system stays the system of record. The AI reads case data and photos with the appraiser's rights and writes drafts, tags and check results back as suggestions; vLLM's structured outputs return them in a fixed JSON schema.
2. Photos. Reduce photos to a set size before the model (for example 1,280 pixels on the long side) and send crops for detail. Keep the original files untouched in the archive.
3. Thinking mode. Qwen3.8-27B reasons before it answers by default. Switch thinking off per request for tagging and standard sections, and leave it on for the report check.
4. Versions and security. Load models from a pinned revision in safetensors format, run vLLM 0.31.0 or later on an internal network behind a gateway with single sign-on, and keep the NVIDIA driver at a release that NVIDIA's bulletin of 30 September 2026 lists as fixed, as our guide to securing a GPU server explains. The gateway log holds case data, so its access and retention belong in the DPIA where one is required.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 4500 Server Edition and RTX PRO 6000 Server Edition and build AI servers to order with the number of cards your work needs, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the models, the search over your earlier reports and the connection to your case management system, under an NDA and with a data processing agreement on request; describe how your office handles a case today, and we propose a pilot on one report type.

FAQ