IVS 105 and artificial intelligence: what a valuation model may do, and what it may not

A model does not sign. Wherever the analysis comes from, the value that leaves the report is the valuer's, and it is the valuer who answers for it.

1. IVS 105: the framework for automated models

IVS 105— the International Valuation Standard published by the IVSC that governs valuation approaches and methods — is where the use of automated models is framed. It does not displace the valuer's judgement. It defines the conditions under which a model may be relied on without the conclusion becoming unverifiable.

An automated valuation model (AVM) is a mathematical system that estimates the value of a property from statistical algorithms, machine learning or neural networks. It reads transaction databases, physical characteristics and macroeconomic indicators, and returns a figure instantly. The appeal is obvious. The discipline the standard imposes is the price of that appeal, and it rests on three requirements.

Methodological transparency. The model must be documented: what data goes in, what algorithm processes it, what assumptions it embeds. A black box whose output nobody can explain is not compliant, whatever its accuracy on a test set. Deep neural networks pose this problem in its sharpest form: a system that cannot say why it reached a figure gives the reader of the report nothing to examine.

Validation and back-testing. The model must be tested regularly against real transactions — the estimated value set against the price actually achieved — rather than presumed to work because it worked once. Validation is a recurring obligation, not a launch formality, and a model calibrated on one market segment does not transfer silently to another.

Human supervision. A qualified valuer must oversee the use of the model, test the relevance of its output against what he knows of the asset and its market, and assume final responsibility for the conclusion. The formula is short and it is the whole of the standard: the machine assists, the human decides.

2. What the Red Book adds: responsibility that cannot be delegated

The Red Book goes further than IVS 105. It does not merely frame the model — it defines the obligations of the valuer who uses one. A RICS member who brings an AVM or an AI tool into the valuation process remains personally responsible for the outcome. Four consequences follow, and each of them is checkable by the reader of a report.

• Disclosure. The report should state explicitly that automated tools were used, which ones, and to what extent they informed the value adopted. This is the point on which the reader has most to gain: a stated tool can be weighed, an unstated one cannot.
• Independent verification. The valuer tests the model's output against his own work — comparables gathered and adjusted by hand, physical inspection, market judgement. An automated tool relieves the process of no step; it feeds into it.
• Technological competence. The valuer must understand how the model he relies on actually works. Using a tool without grasping its limits is a competence failure, not a technical detail — and competence is a professional obligation before it is a preference.
• Non-delegable responsibility. The value adopted belongs to the signing valuer, not to the algorithm. If the figure is challenged, it is the valuer who answers for it — the software does not appear.

One clarification worth making, because it is now the most common confusion: general-purpose language models are not valuation tools. They can help draft, summarise or organise textual material. They do not produce a property value that meets professional standards, and a figure obtained that way carries no methodology a reader could examine.

3. Where the algorithm ends and judgement begins

Valuation is not an exact science. It combines quantitative data with qualitative judgement, and the boundary between the two is where the argument about AI actually sits. An algorithm is excellent at processing large volumes of structured data, detecting statistical trends, screening a portfolio quickly and flagging figures that look out of line with their neighbours. Those are real gains and there is no reason to refuse them.

What it does not do is perceive. It does not register the feel of a street, the quality of an outlook, the condition of a common stairwell, or the early signals of a market turning. It has no view on an atypical asset for which no comparable exists — and the moment a property stops resembling the population the model was trained on, the model is extrapolating rather than measuring.

A concrete illustration. An upper-floor apartment in a Casablanca residential block: a model will read the recent transactions in the district and return a figure consistent with them. Only an inspection establishes that this particular unit has an open sea view that the units below do not, and only local enquiry establishes that a consented development on the adjoining plot will remove that view within a few years. Both facts move the value in opposite directions; neither is in the database. This is the ordinary work of the visit and of the professional network, and it is precisely what the automated figure omits.

The corollary matters for anyone reading a valuation: what makes a conclusion trustworthy is not the sophistication of the tools behind it but the fact that it is documented and verifiable line by line — named sources, stated assumptions, an adjustment grid that can be reworked. The reading grid is set out in our guide to reading a valuation report.

4. Where these tools genuinely earn their place in Morocco

The limits are real; so are the uses. Technology belongs in the process provided professional judgement stays at the centre of every conclusion. Four applications are worth naming.

• Portfolio screening. Where a holder has several hundred assets to revalue periodically, a model gives a first pass: it ranks assets by exposure and identifies those requiring full inspection, so that the effort goes where the value and the uncertainty actually are.
• Continuous market monitoring. Tools that read listing platforms and agency sites continuously pick up movements in asking prices by district, changes in the depth of supply, and listings that sit far from their neighbours. This is evidence to be verified, not conclusions — but it is a far better starting point than an annual survey.
• Yield analysis by area. Cross-referencing property, demographic and macroeconomic data supports a view of how returns are moving across a city. These readings are firmest where data is abundant and weakest in emerging districts — which is exactly where investors most want an answer, and where the valuer must say plainly how thin the evidence is.
• Quality control on reports. A model can test a report for internal consistency: are the comparables genuinely comparable, is the capitalisation rate within the range observed on the market, do the calculations reconcile? A useful additional safety net, and no substitute for review by a human.

5. Why the Moroccan market bounds what a model can conclude

Automated models are only as good as the record they learn from, and the Moroccan record has particular features that any honest account has to state.

• Transaction data is imperfect. Declared prices do not always reflect the price actually paid. A model trained on the declared record will inherit that gap and reproduce it systematically — not as random noise, but as a directional bias.
• The stock is heterogeneous. Traditional riads, modern villas, social housing, walk-up blocks without a lift: a single model cannot capture that range, and the segments where data is thinnest are often the ones where the asset is most singular.
• Part of the market sits outside the formal record. Whatever is not recorded is invisible to the model, which therefore generalises from a visible subset while presenting itself as reading the market.
• Micro-local dynamics are decisive. Two parallel streets can trade very differently on the strength of building quality, noise, aspect and reputation. Those differences are legible on foot and largely absent from any dataset.

A second, less discussed risk is algorithmic bias. A model trained on a historical record that embeds systematically depressed valuations for certain districts will reproduce and entrench them. Vigilance on this point is part of the valuer's job: identify the pattern, correct for it explicitly, and say in the report that the correction was made.

Finally there is the question of responsibility. If a decision is taken on the strength of a figure produced by a model and the figure proves wrong, who answers — the party that built the tool, or the party that relied on it? The professional standards give the plain answer for a compliant valuation: the signing valuer answers. That is a further reason to keep automated output as an input to the analysis rather than as the conclusion of it, and it is why an instant online estimate and a valuation are different objects — a distinction we set out in our note on free online property valuations and what they are worth.

6. What an investor should ask of a report that used these tools

For a foreign investor or an auditor, the practical question is not whether the valuer used technology — increasingly, everyone does — but whether the report allows the reliance placed on it to be tested. Four questions cover it:

• Is the use of automated tools disclosed? Named, with the extent of their influence on the value adopted. Silence here is itself informative.
• Was the output verified independently? By comparables gathered and adjusted by the valuer, and by an inspection — with the report showing that work rather than asserting it.
• Are the assumptions stated? Including those inherited from the model: the data source, its date, its coverage, and the segments where it is thin.
• Who signs? A named valuer, with a statement of compliance and of independence. That signature is what converts an analysis into a valuation.

These are the same questions an international auditor asks of any Moroccan valuation placed in a consolidation, and they are set out at length in our note on how an international auditor reads a RICS valuation of a Moroccan asset.

7. Where this leaves the practice

The profession is not being displaced; the balance of the work is shifting. Data handling that once consumed days now takes minutes, which frees time for the parts of the exercise a model cannot perform: the inspection, the legal characterisation of what is actually being valued, the reconciliation of methods, and the reasoning that a reader can follow. The valuations that will hold up are the ones where the tools are named and their output tested — not the ones where they are hidden behind a confident figure.

Our reports are prepared by RICS-certified expertsand comply with Red Book standards. The conclusions are built to be argued with: they rest on named assumptions, cited sources and a stated methodology, which makes them contestable point by point rather than defensible in a block. A private valuation informs a decision and an arm's-length negotiation.

Fees start at 3,500 MAD excl. tax for standard assets, with a firm quote within 24 hours and delivery in 5 to 8 days, 48-72 hours on the express service. ReaConsult has been advising owners, investors and institutional clients since 2019, with more than 5,000 valuations completed, offices in 6 cities and a rating of 4.9/5 across 47 reviews.

Note:this article sets out the framework that international valuation standards apply to automated models and the obligations the Red Book places on the valuer who uses them. It does not describe any particular commercial tool. A private valuation informs a decision and an arm's-length negotiation. To instruct us, see our contact page or the property blog.