AI in Business and Economics

AI in business and finance denotes the use of statistical machine learning and, more recently, generative language models to make or support decisions about credit, risk, fraud, trading, customer service and knowledge work inside firms. It covers two technically distinct families: predictive models trained on structured records, which have been used in finance for well over a decade, and generative assistants for text and code, which entered workplaces widely after 2022.

Firms adopt AI to reduce the cost of prediction (who will default, which transaction is fraudulent, which ticket is urgent) and, with generative models, the cost of producing and summarising text. The evidence base differs by family. Predictive models in credit and fraud have peer-reviewed accuracy comparisons and decades of production use. Generative assistants have a smaller but growing set of field experiments measuring productivity and quality. Claims of firm-level profit gains, by contrast, rest mostly on vendor and consultancy material that does not meet the Tier 1-2 standard of this encyclopedia and are not used here.

This article separates three evidence classes: deployed products (in production use at firms), validated results (peer-reviewed or regulator-published measurements), and research demonstrations (back-tests, preprints, and prototypes).

Statistical credit scoring predates modern machine learning. Machine-learning methods entered consumer credit forecasting in the late 2000s (Khandani et al., 2010). Academic finance then compared methods systematically; a 2020 study evaluated a broad set of machine-learning methods on predicting stock returns (Gu et al., 2020). Fraud and anomaly detection moved in parallel, with ensemble methods applied to accounting fraud (Bao et al., 2020). After 2022, generative models were tested in call centres and consulting tasks (Brynjolfsson et al., 2025; Noy & Zhang, 2023; Dell'Acqua et al., 2023) and domain-specific language models for finance were reported (Wu et al., 2023).

Predictive risk models. A lender estimates the probability of default $p = P(\text{default} \mid x)$ from applicant and account features $x$. Flexible methods (gradient-boosted trees, neural networks) can capture nonlinear interactions that logistic regression omits. In asset pricing, the target is the expected excess return $r_{i,t+1}$, modelled as $\mathbb{E}[r_{i,t+1}] = g(z_{i,t})$ for a flexible function $g$ of firm characteristics $z$; Gu et al. (2020) found that the best-performing methods were trees and neural networks and traced gains to nonlinear predictor interactions.

Evaluation pitfalls. Financial data are non-stationary, labels (defaults, frauds) are rare, and back-tests are vulnerable to look-ahead bias and data snooping. Out-of-sample, time-ordered evaluation is the minimum standard. For language models, a further hazard is training-data leakage: a model may have seen the period it is asked to forecast, which is why one study explicitly used headlines after the model's knowledge cutoff (Lopez-Lira & Tang, 2023).

Generative assistants. In customer-support and writing tasks, a language model drafts or suggests text, and a human accepts, edits or rejects it. The gains are therefore a property of the human-plus-tool workflow.

• Credit and underwriting: scoring, limit setting, pricing, insurance risk assessment.
• Fraud, anti-money-laundering and accounting-fraud detection: anomaly and classification models with heavy class imbalance.
• Investment research and trading: return forecasting, text analysis of filings and news, execution algorithms.
• Customer operations: chat and agent-assist systems for support, collections and advice.
• Back-office and knowledge work: drafting, summarisation, coding assistance (see APP-SW-001), document review.
• Agentic systems: models that take multi-step actions with tools; their use in finance is largely at pilot stage in the sources reviewed here (see TECH-AGENT-001).

#6.1 Deployed products

• Adoption in UK financial services. A joint Bank of England and Financial Conduct Authority survey published on 2024-11-21, with 118 responding firms, reported that 75% of firms were already using AI and a further 10% planned to within three years; insurers reported the highest use (95%) and financial market infrastructures the lowest (57%) (Bank of England & FCA, 2024). The survey also reported that most use cases involve some automation of decisions but only a small share are fully autonomous. These are self-reported firm statements for one jurisdiction.
• Customer-support assistant. A generative AI conversational assistant was introduced in a staggered rollout to 5,172 customer-support agents at one company (Brynjolfsson et al., 2025).

#6.2 Validated results

• Customer support. Access to the assistant raised issues resolved per hour by 15% on average; less experienced and lower-skilled workers gained in speed and quality, while the most experienced saw small speed gains and small quality declines (Brynjolfsson et al., 2025). Setting: one firm, one task type.
• Professional writing. In a preregistered online experiment with 453 college-educated professionals, exposure to ChatGPT cut average task time by 40% and raised output quality by 18% (Noy & Zhang, 2023). The tasks were short, incentivised writing tasks, not full jobs.
• Consulting tasks and the "jagged frontier". In a preregistered experiment with 758 consultants, those using GPT-4 completed 12.2% more tasks and 25.1% faster on tasks within the model's capability, but were 19% less likely to produce a correct answer on a task designed to lie outside it (Dell'Acqua et al., 2023). The study is a working paper; the final published version was not checked.
• Mortgage credit. Using US mortgage data, machine-learning default models gained less for Black and Hispanic borrowers than for others, and in an equilibrium model increased disparity in rates between and within groups (Fuster et al., 2022).
• Accounting fraud. An ensemble model using raw accounting numbers outperformed two benchmark fraud-prediction models on a ranking-based metric (Bao et al., 2020). This is an out-of-sample academic evaluation, not a statement about any firm's production system.
• Asset pricing. Machine-learning forecasts produced large economic gains for investors in a research setting, in some cases doubling the performance of leading regression-based strategies (Gu et al., 2020). Back-test gains are before the frictions of live trading.

#6.3 Research demonstrations

• Language models and returns. A preprint reported that GPT-4 scores for news headlines predicted initial market reactions and subsequent drift, and that strategy returns declined as LLM adoption rose (Lopez-Lira & Tang, 2023). It is a preprint, has not been independently replicated in this review, and may not survive transaction costs.
• Domain-specific language models. BloombergGPT was reported as a language model for finance (Wu et al., 2023); this is a developer report of a model trained for in-house use, with benchmark results that are not independent evaluations.

Advantages. Higher predictive accuracy in some tasks; measured productivity gains for routine writing and support; scalable monitoring.

Limitations.

• Distributional effects. Gains can be uneven across groups, as in the mortgage study (Fuster et al., 2022).
• Uneven capability boundary. Performance can improve inside and deteriorate outside the region where a model is reliable, and users may not see the boundary (Dell'Acqua et al., 2023).
• Skill heterogeneity. The support study found larger gains for novices and small quality declines for top performers (Brynjolfsson et al., 2025).
• Opacity and governance. The Bank of England and FCA survey reported that 46% of respondent firms had only a partial understanding of the AI technologies they use, against 34% reporting complete understanding (Bank of England & FCA, 2024).
• Regulatory scope. Under the EU AI Act, systems that evaluate the creditworthiness of natural persons or set credit scores, and certain life and health insurance risk assessments, are classified as high-risk (Regulation (EU) 2024/1689, Annex III).

Unresolved questions. How much of measured task-level gain converts into firm-level profit; how crowding of similar models affects market stability; liability for erroneous automated decisions; whether returns to LLM-based trading decay as adoption spreads, which the preprint above itself suggests.

As of 2026-10-15: the use of predictive machine learning in credit, fraud and risk is mature and widespread among regulated firms, with self-reported adoption above 70% in the UK survey (Well Supported for UK firms; other jurisdictions not verified here). Field experiments find positive average productivity effects of generative assistants on writing and support tasks, with heterogeneity by skill and task (Well Supported for the studied tasks; generalisation to whole occupations is Contested). Evidence that language-model trading signals are durable and tradable is Preliminary. For labour-market and macro effects see ECON-LABOR-001; for market structure see ECON-IND-001.

EU application dates for high-risk obligations were revised by the Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the Official Journal on 2026-07-24, moves Annex III obligations to 2027-12-02 and Annex I product-embedded obligations to 2028-08-02 (checked against the Official Journal text on EUR-Lex). See GOV-EU-001 for the authoritative treatment.

Near-term (0-3 years): wider use of assistants in regulated workflows with human review; supervisory expectations on model risk and explainability; more field experiments on whole-job effects. Medium-term (3-10 years): agentic workflows in operations, subject to liability and control rules. Long-term claims about the fraction of financial tasks automated are speculative and not rated here.

• Foundational Concepts:
  • TECH-ML-001 Machine Learning: Methods underlying predictive models.
  • TECH-LLM-001 Large Language Models: Basis of generative assistants.
• Related Methods & Technologies:
  • TECH-AGENT-001 Autonomous AI Agents: Agentic workflows in operations.
  • CAP-GENAI-001 Generative AI: Generative capabilities used in the field experiments.
  • APP-SW-001 AI in Software Engineering: The software-development use of the same assistants.
• Advanced Implications & Future Trajectories:
  • ECON-LABOR-001 AI, Productivity and Labor Markets: Aggregate labour effects.
  • ECON-IND-001 The AI Industry Structure: Vendors and market structure.
  • GOV-EU-001 The EU AI Act: High-risk classification of credit scoring.