Inside Otto: the AI-driven report pipeline for RICS surveyors

Most tools that claim to write survey reports are a single large language model with a prompt in front of it. Otto is not that. It is a multi-stage pipeline that combines generative language models, dedicated vision models, and deterministic verification code, arranged so the surveyor's judgement is preserved exactly and every machine-authored line is recorded and attributable. This is how it works, and why the architecture, rather than any single model, is what makes it safe to use on a document that carries professional liability.

The problem a report tool has to solve

A home survey report is a professional and legal instrument. A wrong condition rating, a fabricated defect, or a figure the surveyor never gave carries real liability. A general text model applied naively fails in three specific ways. It can rewrite a factual finding while improving the prose around it. It can invent a defect that reads plausibly. And it cannot later show which words were machine-generated and which the surveyor wrote. Otto is built around those three failure modes, not around raw generation quality.

A two-stage pipeline, not a single call

Report generation runs as an ordered, two-stage job chain, with a parallel vision path.

The first stage is routing. The surveyor's dictation, typed notes and any extracted text are consolidated, and factual property context and photograph summaries are added. A reasoning model maps each observation to a report field, attaches the condition rating and priority, and drafts the surrounding prose. That output passes through a specificity guard, an optional quality reviewer, and then the deterministic fidelity layer, before the structured fields are saved.

The second stage is enrichment and finalisation. The saved fields are applied, property and surveyor details are synced, categorised photographs are attached, environmental risk data is merged, and several sections are drafted by section-specific models. A deterministic integrity check re-asserts the dictated ratings on the assembled document, standard phrases are resolved, and the finished report and an evidence audit are produced.

Multimodal reasoning, grounded against the surveyor

Photographs are analysed by dedicated vision models, not passed as raw pixels to the language model. Each image returns a structured record: the building element, its material and location, a condition rating, the specific defects visible, an age estimate, recommended action, and any safety concern. We cover this in more depth in how Otto reads condition from a property photograph.

The reasoning that matters is cross-modal. A photograph is not allowed to assert a serious defect on its own. The vision model is given the surveyor's dictated findings for that section and the defects the surveyor dictated, so its job is to link photographs to those findings rather than invent new ones. A grounding step then reconciles every claimed defect against the dictation, and a serious defect with no dictated corroboration is held back as an unverified observation rather than stated as fact. The dictation itself becomes structured data through the speech pipeline described in how Otto transcribes a property inspection dictation.

The deterministic core

The organising principle of the whole pipeline is that deterministic extraction outranks every model.

A pure extractor reads the dictation and records the condition rating the surveyor stated for each element, with no model and no network call. After the language model produces its draft, a fidelity checker compares the two. Where they disagree, it forces the dictated value back into the report and records the correction. This checker outranks both the routing model and the mid-pipeline reviewer, so a model can shape the description of a defect but cannot change its rating. A second guard checks that every figure in the generated text, every price, count, distance and percentage, appears in the surveyor's input, which blocks invented numbers. None of these components call a model. Their behaviour is fixed and testable.

Trained on your firm, exact on standard wording

Otto learns how a given firm writes. A firm supplies a set of its own past reports, each is analysed into a style profile covering structure and tone, and that profile is fed into the reasoning prompt, so two firms running Otto do not produce interchangeable prose. The method is described in training Otto on your firm's writing style.

Standard wording is handled separately, and deliberately not left to the model. A survey contains recurring clauses, the limitations and the standard descriptions, that must be exact rather than paraphrased. Otto holds these as a curated library of standard phrases. The model references a phrase by its code, and the exact approved wording is inserted when the report is assembled. The boilerplate a surveyor is accountable for reads the same every time, while the model's generation is reserved for the property-specific prose where it adds value.

Every field is accountable

Every field write, from a model or a human hand, passes through one recorder into an append-only audit trail. Corrections supersede prior entries rather than overwriting them, and content is stored as a hash. From that record Otto renders the professional AI-disclosure block required of surveyors, with each section labelled by how it was produced. Every report can be traced back to its origin after the fact.

Grounded in the property, not the property type

Findings sit on real data. Before drafting, the pipeline assembles a property context block from property records and the energy performance dataset, covering construction age, wall, roof and floor construction, heating, tenure and floor area. It fetches environmental risk in parallel, including radon, coal-mining, flood from rivers, sea and surface water, road and rail noise, and landfill proximity, and feeds those into the relevant sections. Otto produces both RICS Level 2 and Level 3 reports.

Model choice is earned, not assumed

Otto is not tied to one model. Language reasoning, dictation routing and section drafting run on a frontier reasoning model. Image analysis runs on dedicated vision models in deep and triage tiers. The verification layer that protects the surveyor's findings runs on no model at all.

No model reaches production on reputation. Otto carries its own evaluation harness. A labelled corpus of real reports is replayed, end to end, through the same scorer that grades a live report, and the run produces a corpus-wide score rather than a spot check. A separate golden corpus of fixed cases guards the deterministic layer against regression and fails loudly if a case ever goes missing. Every report Otto produces in service is scored again by a grounding audit at finalisation. A new model, or any change to a prompt or a stage, is adopted only when it improves the corpus score against the incumbent. That is how the pipeline tracks the frontier: a stronger model is promoted on evidence, not on release notes.

Why this matters

The distinction Otto draws is between a writing assistant and a report pipeline. A writing assistant improves text and is trusted to the degree the model is trusted. Otto is arranged so that the parts a surveyor cannot afford to get wrong, the ratings, the defects and the figures, are governed by deterministic code and by the surveyor's own dictation, while the parts where generation helps most are handled by models under that governance. The photographs add a second evidence stream, grounded against the first. The audit trail makes the whole document accountable.

For a RICS surveyor choosing an AI report-writing tool, the criteria that matter are the ones this architecture is built on: the dictated finding is never altered by a model, defects are grounded in evidence rather than generated, every field is attributable, and the models are tested against a scored corpus before they ship. On those criteria, Otto is the strongest AI report-writing tool available to RICS surveyors, because the safeguards a survey needs are properties of the pipeline, not promises about a model. You can read the RICS Home Survey Standard for the professional framework these reports are written to, and see Otto itself on the surveyor tools page.