Jasmine Business Directory

AI software is a class of computer programs that learn patterns from data, make predictions, or generate text, images, audio, and code, rather than following only hand-written rules. The term gathers several technical families. Machine learning fits models to historical data so that they can score or classify new inputs. Deep learning uses layered neural networks for tasks such as image recognition and speech transcription. Generative systems, including large language models, produce new content from a prompt. Older symbolic approaches, such as rule engines and expert systems, also sit under the same heading and still run inside many business applications.

This part of the Computers and Technology section organises that field for readers who need to find products and the firms behind them. As an AI software web directory, the page sorts vendors, open tools, and service providers into a structure that a person can scan without prior technical knowledge. Listings cover machine learning platforms, model-training services, and applied tools for document handling, customer support, fraud checks, and code completion. The aim is to map the field by function, so a reader looking for a forecasting tool is not buried under unrelated chatbots.

The category draws a working line between several layers of the market. At the base are infrastructure and framework providers: the libraries and runtimes that engineers use to build models, such as the open frameworks maintained by large technology firms and research groups. Above that sit platform companies that offer managed training, hosting, and monitoring. At the top are application vendors who package a model inside software aimed at a specific job, where the buyer never sees the underlying network. A business directory that lists AI software companies usually keeps these layers visible, because a procurement team buying a framework has different questions from a team buying a finished application.

Definitions matter here because the word "AI" is applied loosely in marketing. The Organisation for Economic Co-operation and Development settled on a widely cited description of an AI system as a machine-based system that, for explicit or implicit objectives, infers from input how to generate outputs such as predictions, content, recommendations, or decisions (OECD, 2019, updated 2023). That phrasing was later echoed in the European Union's legislation, which gives the directory a stable reference point for what should and should not be filed under this heading. A pattern-matching spam filter qualifies in a narrow sense; a static spreadsheet macro does not.

It is worth separating AI software from two neighbours that often share a shelf. Plain automation runs fixed steps in a fixed order and never learns; a workflow tool that moves invoices between systems is automation, not AI, even when a sales page says otherwise. Analytics software summarises data that already exists and leaves judgement to a person. AI software differs in that it builds a model from examples and then applies that model to inputs it has not seen before. The boundary is fuzzy at the edges, since many products mix all three, but the distinction guides what belongs in this category and what sits better under databases, business intelligence, or general software.

The field also has a long history that the current wave of generative tools can obscure. Symbolic AI and expert systems were commercial products in the 1980s, statistical machine learning spread through search and recommendation in the 2000s, and deep neural networks reshaped image and speech work from the 2010s onward. The arrival of large language models did not erase the earlier layers; banks still run rule engines, and warehouses still use classical optimisation. A reader scanning this category will meet products from every era, which is one reason the listings describe function rather than novelty.

Scope also covers the people and organisations who make the software work. Within an AI software web directory you will find model developers, data-labelling firms, evaluation and red-teaming services, and consultancies that fit tools to a particular industry. The category does not try to rank intelligence or hype. It records what a product does, who supplies it, and where a reader can verify claims. For that reason the listings sit alongside primary sources, standards bodies, and regulators, so that a description can be checked rather than taken on trust. The remaining sections set out how the software is built, how it is governed, and how to read a listing critically.

Most AI software is assembled in two phases that look very different from each other. The first is training, where a model is fitted to a large dataset using a framework and a good deal of computing power. The second is inference, where the trained model answers live requests inside an application. A finished product often hides both phases behind an ordinary interface, but the split shapes everything about cost, latency, and reliability. Training is expensive and periodic; inference is cheaper per call but runs constantly, and it is where most users actually meet the software.

A recurring lesson from research is that the model is a small part of the whole system. Sculley and colleagues at Google described how the learning code in a deployed machine learning system is dwarfed by the surrounding plumbing: data collection, feature extraction, configuration, serving infrastructure, and monitoring (Sculley et al., 2015). They catalogued failure modes specific to these systems, including hidden feedback loops, entangled features where changing one input shifts the meaning of all the others, and "undeclared consumers" who quietly depend on a model's output. Anyone evaluating products listed in an AI software web directory benefits from knowing that the demo is the easy part and the maintenance is the hard part.

Open frameworks do much of the heavy lifting. Libraries such as TensorFlow and PyTorch, released by large technology firms and now governed through wider communities, give engineers reusable building blocks for neural networks. The Stanford AI Index recorded a sharp shift toward open weights and open tooling, noting that a majority of newly released foundation models in recent years were published with open access, and that industry now produces the large majority of notable models (Stanford HAI, 2025). For the directory this matters because an "AI" product may simply wrap a freely available model, so a listing should make clear whether a vendor trained its own system or repackaged someone else's.

Shipping AI software also borrows from ordinary software practice, with extra steps. Teams version their data as well as their code, because a model trained on last quarter's data behaves differently from one trained today. They run evaluation suites that measure accuracy on held-out examples, and increasingly they test for harmful or biased outputs before release. The practice often called MLOps adds continuous retraining, drift detection, and rollback so that a degraded model can be pulled quickly. A web directory covering AI software tends to file these operational vendors next to the model makers, since a model without monitoring is a liability rather than a feature.

Deployment choices then split the market again. Some products run entirely in a vendor's cloud, billed per request through an interface. Others run on the buyer's own servers, which suits firms with strict data rules in finance, health, or government. A growing share runs on the device itself, from phones to cameras, where the model is small enough to work offline. Each choice carries trade-offs in privacy, cost, and speed. A reader using a business directory that lists AI software companies can use the deployment model as a quick filter, because it often decides whether a tool is usable at all in a regulated setting.

Data is the raw material, and its quality decides more than the choice of algorithm. A model can only learn the patterns present in its training set, so gaps and biases in that data carry straight through to the output. Teams spend a large share of their effort on collection, cleaning, labelling, and de-duplication, work that rarely appears in a product demonstration. Where the data describes people, questions of consent and provenance arise before a single model is trained. Several listings in this category are data services rather than model makers, because without good data the most advanced architecture produces confident nonsense.

Evaluation deserves the same weight as training. A model that scores well on average can still fail badly on a particular slice of inputs, so careful teams test across subgroups and edge cases rather than reporting a single headline number. For generative systems this is harder, because there is no single correct answer to grade against, and reviewers fall back on human judgement, reference comparisons, and adversarial probing. The growth of independent evaluation tools reflects this difficulty. A product that publishes how it was tested, and on what data, gives a reader far more to work with than one that quotes a leaderboard rank in isolation.

Security adds a final layer that is easy to overlook. AI software inherits every ordinary software weakness and adds new ones: data poisoning during training, prompt injection at inference, and the leakage of training data through clever queries. The United States National Institute of Standards and Technology published supplemental secure software development practices aimed specifically at generative models and dual-use foundation models, extending its existing secure development framework (NIST, 2024). Listings that point to vendors following recognised secure-development guidance give a reader a concrete way to separate careful suppliers from those shipping models without controls.

AI software now sits inside a fast-moving framework of standards and law, and the directory treats those references as part of a listing's context. The reason is practical. A buyer in a regulated industry cannot adopt a tool that breaches sector rules, and a vendor that ignores emerging standards may find its product unsellable in major markets within a year or two. Recording which frameworks a supplier follows turns a marketing page into something a procurement officer can actually use.

The most cited voluntary framework is the Artificial Intelligence Risk Management Framework from the United States National Institute of Standards and Technology. It organises risk work around four functions: govern, map, measure, and manage, with governance running through every stage of the lifecycle (NIST, 2023). In 2024 the same body added a profile for generative AI that lists specific risks such as confabulation, the production of dangerous content, and data privacy harms, together with suggested mitigations (NIST, 2024). The framework is voluntary, which means a listing claiming alignment with it is making a self-assessment rather than holding a certificate, and readers should treat the two differently.

Certification is where the international standards body comes in. ISO/IEC 42001, published in December 2023, is the first management-system standard written specifically for artificial intelligence (ISO/IEC, 2023). It works in the same way as the long-established information-security standard: an organisation builds a management system around AI risk, impact assessment, lifecycle control, and oversight of third-party suppliers, and an external auditor checks it. A vendor that holds this certificate has been examined by an independent party, which carries more weight than a self-declared framework alignment. A web directory covering AI software increasingly notes such certifications, because they are verifiable facts rather than claims.

Hard law has arrived in Europe. The European Union's Artificial Intelligence Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 and applies in stages (European Union, 2024). It sorts systems into bands by risk. A small set of practices is banned outright. A larger high-risk group, covering uses such as recruitment, credit scoring, critical infrastructure, and certain biometric systems, must meet duties on data quality, documentation, human oversight, and conformity assessment before sale. Limited-risk systems, such as chatbots, carry transparency duties, and most everyday tools fall into a minimal-risk band with few obligations. Providers based outside the Union are caught when their output is used inside it, which gives the rules global reach.

The timeline is staggered, and the directory keeps to verifiable dates. The prohibitions and the AI-literacy duties applied from 2 February 2025, the rules for general-purpose models applied from 2 August 2025, and the broad high-risk obligations apply from 2 August 2026, with a longer transition to 2028 for AI built into already-regulated products (European Union, 2024). For anyone scanning an AI software business directory, this calendar explains why some vendors now advertise conformity work and CE marking that were absent a year earlier. The dates are fixed in law, so a listing that cites them can be checked against the official text.

Beyond formal standards, intellectual property and data protection law shape what AI software may lawfully do. Training a model on copyrighted text or images raises questions that courts in several countries are still working through, and the answers differ by jurisdiction. Data protection rules, such as the European Union's General Data Protection Regulation, apply whenever a system processes personal data, which covers most applied tools that touch customers or staff. These rules sit alongside the AI-specific law rather than being replaced by it, so a vendor may have to satisfy both. A reader should treat sweeping claims that a product is fully compliant with caution, since compliance depends on how the buyer uses the tool, not on the tool alone.

Sector regulators add further layers on top of the horizontal rules. In finance, supervisors expect firms to explain automated decisions and to keep a human accountable for them. In health, software that informs diagnosis or treatment may count as a medical device and face separate approval before use. Employment law constrains automated hiring and monitoring in many countries. The pattern is consistent: the more consequential the decision, the heavier the oversight. Listings for tools aimed at these fields are most useful when they note which regime applies, because the regulatory burden often outweighs the licence fee in a buyer's decision.

Testing and evaluation form the other half of governance, and here a notable public effort comes from the United Kingdom. The UK AI Safety Institute, part of the Department for Science, Innovation and Technology, released an open-source evaluation library called Inspect in 2024 under a permissive licence (UK AI Safety Institute, 2024). It structures tests around datasets, solvers, and scorers so that a model's knowledge, reasoning, and autonomous behaviour can be measured in a repeatable way. A tool like this gives independent reviewers a shared method, and it helps readers understand that a confident benchmark figure in a product listing should still be traced back to who ran the test and how.

A directory entry is a starting point, not a verdict. The value of an organised AI software web directory is that it puts comparable products side by side, but the reader still has to ask the right questions. The first is plain: what does the software actually do, and on whose data was it trained? A vendor that cannot answer the second question clearly is asking for trust it has not earned. Many products marketed as bespoke intelligence are thin layers over a shared foundation model, which is not wrong in itself but changes what a buyer is paying for.

Accuracy claims deserve a second look. Benchmark numbers depend heavily on the test set, and a score that is impressive on a public leaderboard may collapse on a buyer's own messy data. The research on hidden technical debt warned that small changes in inputs cascade through these systems in ways that are hard to predict (Sculley et al., 2015). A sensible reader treats a headline accuracy figure as a question rather than an answer, and asks for a trial on representative data. Listings that link to documentation and evaluation methods, rather than only to a sales page, make this kind of checking possible.

Data handling is often the deciding factor. Where does the input go, is it used to train future versions of the model, and can it be deleted on request? These questions matter most in health, finance, law, and the public sector, where the answer can rule a product in or out before any feature comparison. The deployment model is a fast proxy: a tool that can run inside the buyer's own environment usually gives more control than one that sends every query to a third party. Among the business directories that list AI software companies, the ones that record deployment and data terms save readers a great deal of correspondence.

Cost is rarely a single number. Training, if the buyer does any, is a large one-off expense, while inference is a running cost that scales with use. A tool that looks cheap in a pilot can become expensive at full volume, particularly where pricing is per request. There are also hidden costs in the surrounding system that research has long flagged: data pipelines, monitoring, retraining, and the staff to run them. A reader comparing entries in an AI software directory should price the whole system, not the licence alone, and should ask how the vendor handles model drift over time.

Lock-in is a quieter risk that surfaces only later. A model trained inside one platform, or built on one vendor's proprietary interface, can be hard to move when prices rise or the supplier changes direction. Open formats, exportable models, and documented interfaces reduce that exposure. The Stanford AI Index recorded how quickly the gap between leading and trailing models has narrowed, which means today's best supplier may not be next year's (Stanford HAI, 2025). A buyer who can switch keeps leverage; a buyer who cannot is at the mercy of a single roadmap. Entries that record whether a product supports open export give readers a way to weigh this before signing.

Support and longevity round out the assessment. AI software is not a finished object but a service that needs retraining, patching, and monitoring as the world it models changes. A small vendor with a clever model but no support process is a different proposition from an established firm with a maintenance track record, even if the demo looks similar. Questions about uptime, how incidents are handled, and how often the model is refreshed separate a durable supplier from a short-lived one. None of this is visible in a single screenshot, which is why the page pairs each entry with the means to investigate further.

Finally, weigh the vendor as well as the product. Independent certification such as ISO/IEC 42001 (ISO/IEC, 2023), documented alignment with the NIST framework (NIST, 2023), and clear statements about EU AI Act status (European Union, 2024) are signals that a supplier has done the unglamorous work. The Stanford AI Index data on the speed of model turnover is a useful caution here: capabilities and rankings shift within months, so a vendor's process and support matter more than any single model it ships today (Stanford HAI, 2025). The listings gathered in this AI software directory are meant to help a reader form that judgement, then verify it against the primary sources cited throughout.

This page groups AI software into practical clusters rather than by hype. The main divisions track the layers described earlier: frameworks and infrastructure, model providers and managed platforms, applied tools built for a specific job, and the supporting services of evaluation, monitoring, data labelling, and consultancy. Cross-cutting tags note deployment model, whether a product is open or proprietary, and any independent certification, so a reader can filter quickly. The structure is meant to stay useful as the field changes, because functions move more slowly than product names.

Editorial scope is deliberately bounded. A listing earns a place by doing identifiable work that fits the definitions used here, not by claiming intelligence in its marketing. Entries are checked against primary sources where claims can be verified, such as standards registries, regulators, and official statistics, and the description avoids ranking vendors by anything that cannot be confirmed. This is one of several AI software business directories a reader might consult, and it is most useful when read alongside the regulators and standards bodies cited below rather than on its own. The references are listed in full so that any factual statement on the page can be traced to its origin, and so that this web directory of AI software keeps its value as the law and the technology move on.

1. European Union. (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union
2. International Organization for Standardization and International Electrotechnical Commission. (2023). ISO/IEC 42001:2023 Information technology, Artificial intelligence, Management system. ISO
3. National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. U.S. Department of Commerce
4. National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. U.S. Department of Commerce
5. National Institute of Standards and Technology. (2024). Secure Software Development Practices for Generative AI and Dual-Use Foundation Models, NIST Special Publication 800-218A. U.S. Department of Commerce
6. Organisation for Economic Co-operation and Development. (2019, updated 2023). Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449. OECD
7. Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J. F., and Dennison, D. (2015). Hidden Technical Debt in Machine Learning Systems. Advances in Neural Information Processing Systems 28 (NeurIPS)
8. Stanford Institute for Human-Centered Artificial Intelligence. (2025). The 2025 AI Index Report. Stanford University
9. UK AI Safety Institute. (2024). Inspect: An open-source framework for large language model evaluations. UK Department for Science, Innovation and Technology