Proprietary Algorithms: Valuing AI & ML Models

Proprietary Algorithms: Valuing AI and Machine Learning Models

Algorithms: The Intellectual Core of AI Businesses

Proprietary algorithms — the mathematical methods, optimisation techniques, and decision logic that power software systems — are classified as technology-based intangible assets under IFRS 3. In the AI era, this category has expanded dramatically to encompass machine learning models, neural network architectures, training pipelines, and the inference systems that deploy them.

The valuation of AI and machine learning assets is one of the most challenging areas in intangible asset accounting. The technology evolves faster than the accounting standards can adapt. The boundary between the algorithm (the logic), the model (the trained parameters), and the data (the training input) is often blurred. And the value can shift dramatically with each model generation — a state-of-the-art model today may be obsolete within 18 months.

$200B+ global AI market (2024)

Cost / Income primary valuation approaches

2-5 yrs typical useful life for AI models

The Algorithm-Model-Data Distinction

In an AI-powered business, three distinct intangible assets interact:

Asset	What It Is	IFRS 3 Classification	Separability
Algorithm	The mathematical method or architecture	Technology-based (proprietary algorithm)	High — can be described, published, or licensed
Trained model	The algorithm plus learned parameters from training data	Technology-based (computer software/algorithm)	Moderate — can be transferred but requires infrastructure
Training data	The dataset used to train the model	Technology-based (database)	High — can be sold, licensed, or transferred

Each may need to be separately identified and valued. A trained recommendation model, for instance, combines a neural network architecture (algorithm), learned user preference patterns (model weights), and the historical interaction data that produced them (training data).

★ Key Takeaway

In AI acquisitions, do not treat "the AI" as a single intangible asset. Decompose it into algorithm, model, and data — each has different useful lives, different risk profiles, and different values. The algorithm may be replicable; the trained model captures embedded knowledge; the data is the irreplaceable foundation.

Valuation Approaches

Cost Approach

The cost approach estimates the investment required to develop an equivalent algorithm and train an equivalent model:

Research and architecture design

The cost of the ML engineering team that designed the algorithm — researcher salaries, experimentation time, failed approaches. AI research is highly iterative; the successful architecture may represent 10% of the total experimentation effort.

Model training compute costs

GPU/TPU compute costs for training the model. For large language models, training can cost $10-100+ million. For domain-specific models, training costs may be $100,000-$5 million.

Data preparation and labelling

If the training data requires labelling (supervised learning), the annotation cost can be substantial — $1-50 per labelled example depending on domain complexity.

Deployment and optimisation

The engineering effort to deploy the model in production, optimise inference performance, and integrate with business systems.

Income Approach

Where the algorithm/model generates measurable economic benefit, the income approach captures its contribution:

Revenue attribution: What revenue does the AI enable? A recommendation engine that drives 30% of e-commerce revenue has a quantifiable revenue contribution. A fraud detection model that prevents £10 million in annual losses has measurable cost savings.

Incremental benefit: The comparison point is the business without the AI — using manual processes, simple rules, or no capability at all. The incremental benefit is the difference in revenue or cost.

✔ Example

A fintech company is acquired whose core product is a credit scoring algorithm that processes 2 million applications annually. The algorithm enables the company to approve 15% more applications at the same default rate as competitors using traditional scoring — generating £8 million in incremental annual revenue. Using the income approach with a 4-year useful life and 18% discount rate (reflecting AI technology risk), the algorithm is valued at approximately £20 million. The underlying training data (5 years of loan performance data) is valued separately at approximately £6 million using the cost approach.

AI-Specific Valuation Challenges

Rapid Obsolescence

AI models depreciate faster than almost any other technology asset. The state of the art advances continuously — a model that was competitive 18 months ago may now be significantly outperformed by newer approaches. Useful lives of 2-5 years are typical, and even this may be generous for models in rapidly advancing domains.

Reproducibility vs Uniqueness

Many AI architectures are based on publicly available research (transformers, diffusion models, reinforcement learning algorithms). The algorithm itself may not be unique. The value often resides in:

Proprietary training data that produces a uniquely capable model
Fine-tuning and domain adaptation that makes a general model effective for a specific use case
Production engineering that makes the model performant and reliable at scale
Feedback loops that continuously improve the model based on production data

High-Value AI Assets

Trained on proprietary, scarce data
Domain-specific with proven performance
Embedded in revenue-generating products
Continuous improvement through feedback loops
Difficult to replicate without equivalent data

Lower-Value AI Assets

Based on public data and architectures
General-purpose without domain specialisation
Proof of concept, not production-deployed
No feedback loop or continuous learning
Easily replicated by any competent team

⚠ Warning

The AI valuation landscape is rife with overvaluation. A machine learning model with impressive demo performance but no production deployment, no revenue attribution, and an architecture based on publicly available research has limited intangible asset value — regardless of the hype around AI. Valuation must be grounded in demonstrated, measurable economic benefit.

Useful Life

AI model useful lives are among the shortest of any technology asset:

Foundation models: 1-3 years (rapid architectural advancement)
Domain-specific production models: 2-5 years (depends on domain stability)
Traditional ML models (regression, classification): 3-7 years (more stable)
Optimisation algorithms: 3-5 years (mathematical foundations are more durable)

The Moat Question

For investors evaluating AI businesses, the key question is not "how good is the model?" but "how defensible is the advantage?" The most durable AI moat is proprietary data that continuously improves the model through a flywheel effect — more users generate more data, which trains a better model, which attracts more users. This compounding advantage is where the real intangible value lies, and it is best captured by valuing the data and the feedback system, not just the current model snapshot.

Proprietary algorithms are one of ten technology-based intangible assets under IFRS 3. For the full taxonomy, see 35 types of intangible assets. For AI valuation in M&A, read valuing AI in M&A deals.

Ivan Gowan is the Founder and CEO of Opagio. He brings 25 years of experience building and scaling technology platforms in financial services. Meet the team.

Share: