Building an AI Valuation Engine for Heterogeneous Assets: What We Built for SAM
The Engineering Problem
Most valuation systems are built for one asset class. A mortgage AVM values residential property. A vehicle pricing tool values cars. The methodologies work because the asset universe is relatively uniform — the comparable data is abundant, the adjustment factors are well-understood, and the model can be trained on millions of similar examples.
The problem becomes structurally different when the asset universe is heterogeneous. An industrial surplus auction platform might value a Vector Network Analyzer on Monday, a fleet of forklifts on Tuesday, a data centre decommission on Wednesday, and a food processing line on Thursday. Each has a different market, a different buyer base, different condition variables, and different discount logic for liquidation channel dynamics. A single model trained on one category generalises poorly to the next.
SAM runs auctions across industrial and surplus categories at volume. Their appraisal process was manual, inconsistent across categories, and slow. The direct consequence was financial: assets undervalued meant opening bids set too low and recovery left on the table. Assets overvalued meant stalled auctions and unsold inventory. Neither outcome was acceptable at their scale.
They needed a system that could take an asset’s details — make, model, year, manufacturer, specs, condition, and photos — and produce a structured valuation with an auction strategy attached, automatically, before the listing went live. We built it. This post describes the design.
Inputs: What the System Works From
The valuation engine works from the inputs that are already available when an asset is catalogued for auction. No additional data collection is required.
Structured asset data. Manufacturer, model, year, category, quantity, and condition status. For industrial equipment, manufacturer and model alone carry significant information — a Rohde & Schwarz Vector Network Analyzer occupies a specific position in a specific market, with known buyer demand, known typical sale ranges, and known condition adjustment patterns.
Asset specifications. Technical attributes, accessories included, calibration status where relevant, operational history if documented. These details move the valuation within the range that manufacturer and model establish.
Condition classification. The condition input is one of the highest-weight variables in the model. Tested and working, untested as-is, cosmetically damaged, and parts-only represent meaningfully different asset states. Each carries a different discount profile relative to retail market value.
Photos. Listing photos serve two purposes. The first is identity verification — confirming that the asset matches the catalogued description. The second is visible condition assessment — surface damage, completeness, missing components, and storage condition signals that may not appear in the written description.
The combination of structured data and photos means the engine has both the market context to find comparables and the asset-specific detail to adjust them correctly.
Stage One: Asset Identification and Classification
Before any valuation work begins, the engine needs to confirm it understands what it is valuing.
This sounds straightforward, but at auction scale it is one of the most consequential steps in the pipeline. Assets are catalogued by humans under time pressure, and catalogue entries vary in quality. A listing might say “Rohde & Schwarz Network Analyzer” without specifying the model. It might misspell the manufacturer. It might categorise a specialist piece of test equipment under a generic “electronics” label that returns irrelevant comparables.
A language model processes the listing description, asset specs, and any available documentation to extract and confirm the manufacturer, model, and category. Where the input is ambiguous, the model flags it rather than proceeding on a low-confidence identification. An asset incorrectly identified at this stage will be compared to the wrong market, producing a valuation that is wrong for reasons that have nothing to do with the downstream model.
The photo analysis layer runs in parallel. A computer vision model checks the visible asset against the catalogued description — confirming that what is in the photos is consistent with what was entered. It also extracts visible condition signals: surface wear, physical completeness, evidence of damage or heavy use.
The output of this stage is a confirmed asset identity, a category classification, and a condition supplement from photo analysis that adds to or modifies the written condition input.
Stage Two: Market Research and Comparable Finding
With a confirmed asset identity, the engine searches for comparable sale data across the relevant channels.
For industrial surplus and specialist equipment, comparables come from multiple sources: specialist test equipment marketplaces, industrial auction platforms, reseller databases, and historical auction results. The appropriate source set depends on the asset category. A Vector Network Analyzer’s comparables come from specialist test equipment channels. A fleet of forklifts has a different comparable universe entirely.
The research layer retrieves recent sale data for the same or closely related models, filters for relevance — same manufacturer, same model family, similar specifications — and extracts the sale prices and the conditions under which those sales occurred. A working, calibrated unit sold through a specialist dealer is a different comparable from an untested unit sold in a liquidation auction. Both are relevant, but they anchor different points in the valuation range.
The model synthesises this comparable data into a base market range: the price range a known-working unit of this type typically achieves in the open market. That base range is the starting point for adjustments, not the final valuation.
Stage Three: Condition and Channel Adjustment
This is where the valuation becomes specific to the actual asset and the actual sale context.
Condition adjustment. The gap between a working, calibrated unit and an untested, as-is unit is significant and varies by asset category. For specialist test equipment, an untested unit with unknown calibration status typically trades at a 40 to 60 percent discount to retail market value. For heavy machinery, the discount range is different. For office furniture, condition adjustment works on an entirely different scale.
The adjustment model applies a discount derived from historical outcomes for that asset category under comparable condition inputs. It is not a fixed percentage — it reflects what the market has historically accepted for this type of asset in this condition state.
Channel adjustment. A liquidation auction is not the same channel as a specialist dealer or an end-user direct sale. Auction channel buyers expect a discount for the risk they are absorbing: no warranty, no recourse, as-is terms, the possibility that the asset underperforms its description. The channel adjustment accounts for the expected discount that auction buyers apply relative to alternative acquisition channels.
Lot structure adjustment. A single unit of a specialist asset and a lot of ten identical units have different valuation dynamics. Single units are easier for end-user buyers to absorb but lose the volume discount that dealer buyers expect. Lot pricing requires a different adjustment to the per-unit value.
The output of this stage is the estimated value range for this specific asset, in this specific condition, in this specific sale channel. Not a theoretical market value — an expected auction realisation range.
Stage Four: Valuation Report Construction
The valuation output is a structured report, not just a number.
The report contains a low estimate, a high estimate, and a most likely sale range — with the most likely range representing the realistic central outcome given current market conditions and the asset’s specific inputs. It documents the key assumptions behind the valuation, the market research that anchored the base range, the value drivers that support stronger realisation, and the risk factors and discounts that were applied.
The confidence level is explicit. A well-documented asset with abundant recent comparables and clear photos produces a high-confidence valuation. An unusual asset in a thin market with limited comparable data produces a medium or low-confidence valuation. The distinction matters for how the auction team treats the output. High confidence means the recommended ranges can be applied directly. Lower confidence means the output is a starting point for human review, not a final answer.
The plain-language market context section explains who the likely buyers are, which channels they typically operate in, and what they prioritise in their purchasing decision. For specialist equipment, buyer type significantly affects realisation. A maintenance department buying for operational use will pay closer to working-unit market value than a surplus reseller buying for parts or refurbishment.
Stage Five: Auction Strategy Recommendation
Valuation alone is not sufficient for an auction platform. The output needs to translate into operational decisions: what opening bid to set, what reserve to apply, how to structure the lot, and how to position the listing.
The auction strategy module takes the valuation range and produces a set of recommendations.
Opening bid. Set to generate early momentum without anchoring the final price. Too high and the lot receives no early bids, which suppresses later bidding. Too low and recovery depends entirely on competitive pressure that may not materialise. The opening bid recommendation is calculated as a percentage of the reserve, calibrated by category — the appropriate relationship between opening bid and reserve varies across asset types and typical buyer behaviour.
Reserve price. The floor below which the asset should not be sold. Derived from the low estimate with an adjustment for the seller’s recovery expectations and the cost of relisting an unsold asset. The reserve recommendation is explicitly distinguished from the opening bid to prevent the common error of setting both at the same level.
Lot structure. Where the asset intake involves multiple units, the system recommends whether to lot them individually, in groups, or as a single combined lot, based on likely buyer type and the historical performance of comparable lots.
Marketing positioning. The listing should lead with the asset attributes that drive buyer interest for this category. For specialist equipment, that means manufacturer, model, and technical specifications before any condition or logistics information. For heavy equipment, operating hours and service history lead. The recommendation reflects what buyers in that category search for and respond to.
What the Output Looks Like in Practice
To make this concrete: a Rohde & Schwarz ZNB20 Vector Network Analyzer, catalogued as untested and sold as-is, produces the following from the engine.
The base market range for working, calibrated units of this model sits between $20,000 and $35,000 on specialist channels. Applying the untested, as-is condition discount of 40 to 60 percent produces an adjusted auction estimate of $7,000 on the low end and $28,000 on the high end, with a most likely sale range of $12,000 to $18,000. The confidence level is medium, reflecting the photo-based assessment without operational verification.
The auction strategy recommends an opening bid of $5,600 and a reserve of $6,300, with a single-unit lot structure. The marketing guidance is to lead with manufacturer and model rather than the category label, since specialist buyers in this segment search by model number and the brand recognition is a primary value driver.
Expected realisation under a well-structured listing: approximately $15,000. If the listing is vague or miscategorised, recovery trends toward the low estimate.
That is what the system produces. Not an opinion. A structured output with documented reasoning that the auction team can act on directly or review before the listing goes live.
Where the Same Architecture Applies
The architecture described here is not category-specific. The same system applies wherever you need to value heterogeneous assets under auction or liquidation conditions.
Vehicle auctions. Make, model, year, mileage, condition, and photos feed the same pipeline. Comparable channels are different. Adjustment factors are different. The structure is identical.
Real estate auction. Property type, location, condition, and comparable transactions anchor the base range. Auction channel discount and as-is adjustment modify it. The output is the same structure: estimated range, reserve recommendation, opening bid, marketing notes.
Government surplus. Often the most heterogeneous asset mix — furniture, vehicles, equipment, electronics within a single auction. The asset identification layer matters most here, where catalogue quality is typically lowest.
Bankruptcy and liquidation. Time pressure and creditor recovery expectations add constraints to the auction strategy output. The valuation logic is the same. The strategy layer accounts for the additional context.
What to Think Through Before Building This
The asset identification stage is where most teams underestimate the work. A valuation model that operates on correctly identified assets is straightforward. A system that handles messy, inconsistent, human-entered catalogue data requires meaningful investment in the extraction and verification layer before any valuation work can be reliable.
The condition adjustment model is the other area that needs careful calibration. The discount applied for untested, as-is condition is not universal. It varies by category, by typical buyer behaviour in that category, and by current market conditions. Building the adjustment model on historical auction outcomes — what did assets in this condition actually achieve relative to working-unit market value? — produces better calibration than applying a fixed percentage across categories.
The photo analysis component adds real value and adds real complexity. Deciding how much weight to give photo-derived condition signals relative to the written condition input, and how to handle cases where the two contradict each other, requires deliberate design decisions before the build begins.
The return on investment for an auction platform is direct and measurable. Better opening bids produce more early engagement. Accurate reserves reduce unsold inventory. Correct lot structuring maximises recovery per asset. Every element of the system output has a traceable effect on auction performance.
Next in the series: Intelligent Sales Representative Assignment — matching inbound leads to the right rep based on intent signals and performance history.
GTC builds AI-powered systems across auctions, real estate, SaaS, and enterprise software. If you are thinking through an asset valuation or pricing engine, let’s talk.