Artificial Intelligence
Artificial intelligence (AI) is software that performs tasks which would otherwise need human judgement — recognising patterns, interpreting language and images, predicting outcomes, making or recommending decisions — mostly, since the 2010s, by learning from data rather than by following rules written by people. In fintech it runs fraud detection, credit decisions, identity verification, customer service and compliance; in M&A it is both a sector of its own and a capability every target claims and few can evidence.
- Also known as
- AI, Artificial intelligence (AI), Machine intelligence
- Used at
- fraud and transaction monitoring, credit underwriting, identity verification, document processing, customer service and assistants, personalisation, trading and risk, compliance screening, and software development itself; across the economy, every function that has data and decisions
- Binding
- the EU AI Act (in force August 2024, applying in stages to 2027) classifies systems by risk — prohibited practices, high-risk uses including creditworthiness assessment and insurance pricing, transparency duties, and obligations on general-purpose model providers; sector rules (model risk, consumer credit, anti-discrimination, data protection) apply in full; the US regulates through existing agencies and state laws; the UK through regulators' guidance
- Typical economics
- the labs' frontier models cost hundreds of millions to train and are sold by the token; the applications built on them are priced as software or as outcomes; enterprise AI spend grew from single-digit to double-digit billions a year through 2024–2025; AI companies took a majority of venture funding in 2024–2025 and the largest acquisitions and licence-and-hire deals of the period
- Typical distinction
- narrow AI (a system for one task — the fraud model, the document reader) is what runs in production; general-purpose models (language, vision) are the platforms narrow applications are now built on; artificial general intelligence is the labs' stated goal and nobody's product
- Parties
- the labs (OpenAI, Anthropic, Google DeepMind, Meta, and the open-weight community), the infrastructure (Nvidia, the clouds), the application companies, the enterprises deploying, the regulators, and the customers whose decisions the systems touch
How artificial intelligence works in practice
A payments company's AI is a portfolio. A fraud model — gradient-boosted trees over hundreds of features, retrained monthly — scores every transaction in milliseconds. A document model reads onboarding documents. An identity vendor's models verify faces and detect deepfakes. A language model, rented from a lab, drafts support replies from the knowledge base and summarises merchant conversations. A model in compliance prioritises monitoring alerts. None of them is "AI" in the sense of the films; each is a trained system for one task, with a training set, an evaluation, a monitoring dashboard and a model card, governed under model-risk rules and, for the credit and identity ones, the EU AI Act's high-risk obligations. Together they are why the company runs with a fifth of the staff its predecessors needed.
The 2020s changed what is possible and what is claimed. General-purpose models made language, vision and code tasks available through an API to any company, so a feature that needed a data-science team in 2019 needs a prompt and an evaluation set in 2025. That lowered the barrier and raised the noise: every fintech now says "AI-powered", and a buyer's diligence has to separate a rented model behind a prompt from a proprietary model on proprietary data — the first is a feature, the second is an asset. The labs' own economics — capital intensity, concentration, the race for compute and talent — produced the licence-and-hire deals of 2024, the infrastructure acquisitions, and the valuations that an applied company borrows when it says the word. For a fintech, AI is a set of production systems with governance obligations; for an investor, it is a sector whose value chain — chips, clouds, models, applications — captures value unevenly, and mostly not at the application layer that most fintechs occupy.
For a fintech founder, AI is what the models in production do and what the model cards can prove; for a buyer, it is a claim to be tested against data rights, evaluation results and the question of what would remain if the rented model were switched off.
Example
A buyer evaluates two "AI-native" lending fintechs. The first has a credit model trained on 400,000 of its own loans with 18 months of out-of-time validation, a fraud model on its own transaction data, a document pipeline built on a rented vision model, and a customer assistant on a rented language model — with model cards, monitoring and an EU AI Act classification for the credit model as high-risk with the required documentation. The second has the same features built on rented models with a prompt layer and no proprietary data of note. Both grow at similar rates. The buyer values the first at 2.5× the second's revenue multiple: its credit and fraud models are assets its data alone could produce, and its governance file makes them acquirable; the second's "AI" is a feature any competitor can rent, and its pitch was priced accordingly.
Common mistakes
- Calling a rented model behind a prompt a proprietary AI capability. - Deploying models on decisions about customers without the governance the rules require. - Reading the labs' valuations as the application layer's. - Treating AI as a project rather than a portfolio of production systems with lifecycles. - Diligence that reads the pitch and not the model cards, the data rights and the evaluations.