Platform or Feature? Valuing AI in Your Next Acquisition

Platform or Feature? Valuing AI in Your Next Acquisition

Almost every pitch deck we review has 'AI' on the third slide. The claim is table stakes now. The critical question for an investor is no longer 'Are you using AI?' but 'How are you using AI?' The answer separates targets with a compounding, defensible moat from those with a thin feature layer that will be a commodity in 18 months.

The distinction we look for during technology due diligence (tech DD) is between a genuine AI platform and a simple AI feature. A feature uses AI to perform a discrete task. A platform uses AI to build a system where the value to every user grows as more users join. Understanding this distinction is the key to correctly valuing the asset you are about to acquire.

The Anatomy of a Platform vs. a Feature

An AI feature is often a veneer over a third-party service. Think of a simple chatbot on a website that uses the OpenAI API. It answers questions, but it operates in a silo. It does not learn from interactions across the entire customer base to improve its service for the next user. It is a bolt-on, easily replicated by a competitor with the same API key and a modest budget.

An AI platform is architecturally different. It is designed from the ground up for multi-tenancy and data network effects. Consider a supply chain optimisation platform that serves 500 logistics firms. It anonymises and aggregates data on shipping lanes, delays, and customs from all clients. Its predictive models, which forecast disruptions, become more accurate for every single client with each new data point from the network. This creates a powerful flywheel; new clients join for the superior predictions, and their data makes the predictions even better, widening the moat.

Reading the Signals During Tech DD

Discerning a platform from a feature requires looking past marketing claims and interrogating the architecture and organisation. During tech DD, we hunt for specific tells that reveal a company's true AI strategy.

Architectural Giveaways

The codebase and infrastructure diagrams do not lie. We look for common patterns that signal a feature masquerading as a platform:

• Thin API Wrappers: The system makes direct, unmodified calls to a public Large Language Model (LLM) or other AI-as-a-service provider. There is no proprietary logic, data processing, or model fine-tuning that adds unique value.
• Per-Tenant Data Silos: The AI engine operates strictly on a single client's data. There is no infrastructure for anonymising, aggregating, and learning from data across the entire customer base. This is a structural barrier to creating a network effect.
• No Feedback Loops: The system is a one-way street. It makes a prediction or generates content, but there is no mechanism to capture the outcome or user feedback to systematically retrain and improve the core models.

Organisational and Commercial Tells

How a company is structured and how it goes to market often reveal more than its technical documentation.

• The 'AI Lab' Trap: If the AI talent sits in an isolated 'labs' or 'innovation' unit, separate from the core product and engineering teams, their work is likely experimental. True AI platforms have machine learning engineers and data scientists embedded directly within product teams, working on the core offering.
• Feature-Based Pricing: When the AI capability is sold as an optional, per-seat add-on, it signals that the company itself views it as a non-essential feature, not the core engine of value. Platform value is typically tied to outcomes or consumption, not a simple feature flag.
• 'Powered by...' Marketing: A company that heavily markets the brand of its underlying AI provider (e.g., 'Powered by Anthropic') is implicitly admitting it brings little of its own intellectual property to the table. A true platform sells the power of its own proprietary insights.

Quantifying the Platform Advantage

Vanity metrics like 'petabytes of data processed' are meaningless. The proof of a platform is in the measurement of its compounding data advantage. We demand to see metrics that demonstrate a learning effect over time.

Key performance indicators we analyse include:

1. Cross-Tenant Model Lift: The measured improvement in model accuracy for Client A that is directly attributable to data ingested from Clients B, C, and D. A platform company can prove this; a feature company cannot.
2. Declining 'Cold Start' Time: The time it takes for a new client to receive high-quality, accurate output from the system. For a platform, this time-to-value should decrease as its aggregate models become more sophisticated.
3. Alpha Generation: In financial or operational tools, this is the demonstrable improvement in client outcomes—such as higher forecast accuracy or lower error rates—that correlates with the growth of the platform's dataset.

The Exception: When a Feature Creates Defensible Value

While rare, a non-platform AI feature can sometimes create significant, defensible value. This occurs under specific conditions where the moat is derived not from the technology itself, but from its context.

One such condition is deep workflow integration. An AI feature that automates a critical step inside an industry-specific ERP or a regulated compliance system has high switching costs. The defence is the workflow, not the AI. Another is a dominant distribution channel; a company with a captive audience can push a simple feature that becomes sticky simply because it is there. Finally, a feature trained on a truly unique and inaccessible proprietary dataset can also be defensible, even if the model architecture is simple.

These are exceptions. The default assumption for any investor should be that an AI feature is on a path to commoditisation. The strategic premium belongs to the platform.

Takeaway: The Blueprint Is the Moat

In your next technology due diligence, look beyond the 'AI-powered' label. Scrutinise the architecture for data feedback loops and multi-tenant learning. Ask the CTO to map, on a whiteboard, how data and insights from one customer systematically improve the core product for all customers. If they cannot clearly and convincingly draw that flywheel, you are likely buying a feature, not a platform, and your valuation model must reflect that reality.