961 AI offerings. Start with the job, not the tool.

A list of 100 tools no longer helps anyone choose; a map of where AI does work, what data it handles and what technology it rests on does.

These figures describe the study’s research inventory, including entries outside this website’s 485-record candidate catalogue. They are not a count of verified products or website admissions.

In 2024 we published a guide to more than 100 AI tools, sorted into categories such as natural language processing, computer vision, machine-learning frameworks and chatbots. Many of the names in it were developer libraries; several have since been renamed, merged or retired. The market has also grown too large for a list. For the Good AI Prism Global Market Study 2026/27 we mapped 998 market entries, 961 of them AI offerings, each recorded with a dated source from the provider's own pages.

The market also changes faster than any list. Within a few weeks of research we found 34 renames and 26 ownership changes, for example Acrolinx becoming Markup AI and LivePerson moving to SoundHound AI. So this edition is a method, not a catalogue: how to move from a business need to a defensible shortlist.

The Good AI Prism maps the AI market from the outside in.

Outer ring: 24 business areas, where AI does work.

• 14 functions: Customer Service; Assistants & Copilots; Knowledge & Search; Documents, Workflow & Agents; Sales & CRM; Marketing & Content; HR & Recruiting; Finance & Accounting; Legal & Compliance; IT & Security Operations; Software Development; Data & Analytics; Design, Image & Video; Meetings, Voice & Language.
• 10 industries: Health & Life Sciences; Financial Services & Insurance; Manufacturing & Industrial; Retail & Consumer; Energy & Utilities; Mobility & Logistics; Public Sector; Education; Media & Entertainment; Science & Research.

Middle ring: 8 data types, the kind of data an offering handles: language, voice, audio, image, video, tables, code and multimodal.

Inner ring: 6 technology layers, the technology underneath: models, chips and compute, clouds, data, agents and trust.

An offering can sit in more than one area; the 998 entries hold 1,037 placements on the map.

Pick one process in one business area and measure it before you look at suppliers: volume, cost per case, quality and cycle time. Measured gains are real but task-specific. In customer support, a generative AI assistant raised issues resolved per hour by 14% on average and 34% for less experienced agents (NBER). In software work, measured effects range from 19% slower (METR) to 21% faster (Google RCT). A supplier's productivity figure is not your baseline.

The middle ring narrows the field quickly. A voice agent, a document workflow and a coding assistant handle different data, carry different risks and need different checks. Voice and meeting tools, for example, raise consent questions: 65% of US workers who met with an AI notetaker said consent was inconsistent or missing (survey via KRDO).

Most business buyers purchase applications, but each application rests on layers that change independently.

• Models. The frontier now has three closed flagships: OpenAI GPT-6 Astra, Anthropic Claude Fable 5.1 and Google Gemini 4 Argon. Open-weight options include Mistral Large 4, previewed in October 2026 with European deployment and on-premises options. Prices span roughly 100 times, from USD 0.10 to USD 10 per million input tokens, and 81% of large enterprises already use three or more model families (a16z). Models also retire: OpenAI shuts down its GPT-4o and GPT-4.1 snapshots on 14 January 2027.
• Chips and compute. Capacity stays tight. ASML expects its planned 2027–28 supply "will not fully meet customer demand", and big-tech capital spending is set to rise a further 75% in 2026 (IEA). At the same time, inference prices per unit keep falling steeply (Epoch AI).
• Clouds and data. Hosting region, subprocessors and data use decide whether an offering is usable for a European buyer at all.
• Agents. Spending on AI agents and assistants is forecast to grow from USD 29.2bn in 2026 to USD 65.5bn in 2027 (Gartner), yet Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027.
• Trust. Documentation, certifications and their scope, oversight and incident handling.

Source-documented relations between offerings show applications resting on a few hubs: Zendesk, Salesforce and the Model Context Protocol on the application side; OpenAI, Anthropic, Amazon Bedrock and Hugging Face underneath. A single model or platform change can move many products at once, so ask every supplier which model and cloud they depend on.

Source context: MCP · authorization specification · 2026-07-28 ↗

A useful shortlist needs product- and plan-specific facts. Ask where storage, inference and backups take place, which sub-processors are involved, and which deployment locations you can select. An address in Europe or Switzerland does not establish EU hosting. Training only after opt-in, exclusion by default and an opt-out requirement are different policies. Inspect the dated evidence on each profile, retain qualifications, and ask the provider which contractual terms apply to your account. The study’s research records can overlap and are not a verified count of distinct products. Compare the facts for your own shortlist instead of using an earlier edition’s combined coverage ratio.

Two further patterns run across the market. Four in five offerings we added in 2026 are described only on marketing pages, not in technical documentation. And prices come in at least seven units: per resolution, per seat, per conversation, per minute, credits, plan tier and quote only.

Because prices come in different units, convert every option to one measure: total monthly cost (usage or licence, amortised integration, monitoring and training, human review, handling of failed cases) divided by the outcomes your business accepts. In a customer-support scenario with 10,000 cases a month, €11,000 of total cost and a 40% acceptance rate give €2.75 per accepted outcome, which you can set against today's cost per resolved case.

Whether an offering is high-risk depends on what you use it for. In the EU, Article 50 transparency duties apply since 2 August 2026; high-risk duties for Annex III uses such as recruitment, credit and education apply from 2 December 2027 under Regulation (EU) 2026/1744. Colorado and California duties on automated decisions apply from 1 January 2027.

Source context: EU AI Act · consolidated 2026-07-27 ↗

The Prism helps identify where to look. It does not decide that an offering fits your organisation, and placement in an area is not an independent verification result. An application may appear in several areas because it supports several kinds of work. That is useful context, provided repeated placements are not counted as separate products.

Keep the research inventory, the website candidate catalogue and the evidence depth separate. The research figures above describe their stated edition. The website exposes its own admitted candidate records and qualifications. An identified offering, a substantive profile and a decision-ready research record are different levels of usefulness. The interface should help you inspect the difference rather than convert presence on a map into an assurance.

This is why the buyer's first question matters. “Find an AI tool” is too broad. “Find options for preparing support replies in German using our approved policy, with people sending the final answer” creates a tractable search. It names the task, language, source and authority boundary before a product enters the conversation.

Start in the Customer Service area and select an offering. Read the panel for its purpose and the material source qualifications. Then open the profile for the facts that matter to your intended use. If data region or training use is not established for the relevant plan, keep that as a provider question. Do not infer an answer from the supplier's headquarters.

Add two contrasting options to Compare. The point is not to assemble the largest shortlist. It is to see meaningful differences in the same fields: billing basis, relevant data conditions, scope and the work the product actually supports. A quote-only price can remain a useful option, but it cannot be represented as zero cost. A documented regional option remains conditional until its applicability is confirmed.

Inspect the sourced connections where they affect dependency. An application using a model platform may inherit some technical constraints while retaining its own contract and data processing. The relationship is evidence about a connection, not permission to transfer every claim from the platform to the application. If the source does not establish the downstream condition, ask the supplier.

Use a short written brief for every candidate. Include the intended task, expected volume, languages, data sensitivity, acceptable output and actions that remain human. Ask each provider the same deciding questions. Otherwise one supplier can look stronger simply because it answered a different interpretation of the request.

Keep an explicit distinction between included functionality, optional configuration and separately scoped integration. A product can have a relevant feature without providing the complete workflow you need. Identify who will maintain the source documents, who handles exceptions and which existing systems must change. Those responsibilities affect both feasibility and cost.

Record the answer date and the service arrangement. When a sales conversation resolves a question, ask for the applicable documentation or contractual wording where the decision requires it. Do not silently replace a conflicting published source with a favourable informal answer. Preserve enough context for another buyer or reviewer to understand the conclusion.

Compare a shared workload. A per-seat tool, a per-minute service and a per-resolution product can all be relevant, but their prices need translation into the same operating scenario. Include the cost of review, failed cases, integration and supporting systems. Do not compare one provider's annual commitment with another provider's flexible monthly price without saying so.

Use a range for uncertain inputs. If the acceptance rate is unknown, show how the result changes rather than substituting an optimistic value. If the use case releases time, describe how the organisation intends to use it. A capacity scenario is different from a realised cash saving. The economics tools can help make those assumptions visible.

Then choose the next evidence step. It may be a written price, a deployment confirmation, a small evaluation or a decision to use an existing tool. A good shortlist is not a promise to purchase. It is a manageable set of options with clear reasons for inclusion and concrete conditions for proceeding.

For every retained candidate, write one reason to consider it and one fact that could rule it out. This prevents a familiar brand from staying on the list without a relevant advantage. It also prevents a less familiar offering from being rejected merely because its documentation uses different language.

Ask another person to reconstruct the comparison from the sources. Can they find the price period, the deployment condition and the unresolved question? If not, improve the record before adding more products. A shortlist with four understandable options is more useful than a longer table whose cells conceal incompatible assumptions.

When the comparison changes, preserve the reason. A new contract condition, a corrected identity or a better-defined task may explain the update. The original selection was not necessarily careless; it relied on a different state of knowledge. Keeping that history makes the process reusable without presenting yesterday's answer as today's certainty.

For a focused comparison of up to three offerings, inspect the AI Decision Brief. If the business question and implementation assumptions are still unclear, the AI Business Blueprint may be a better starting point. A deeper investment case is separately scoped through the Enterprise AI Opportunity Case.

Carry the selected products, language and unresolved questions into the enquiry. Check the draft before sending it yourself. A prepared message is not a submitted order, and a provider question is not a negative verdict. The map has done its job when it helps you move from an unbounded market to a useful decision with the right evidence and a responsible owner.

What to decide now

• Choose one process, place it on the map (business area, data type, layer) and measure its baseline before any demo.
• Send every shortlisted supplier the same written questions: hosting region, training use, billing unit, model and cloud dependencies, exit terms.
• Compare options on cost per accepted outcome at your own volume, with a stop condition agreed in advance.

The Good AI Prism, the full taxonomy and the October 2026 research findings are available at ethority.ai/ai-prism-study, where providers and readers can also submit corrections.

• ETHORITY, Good AI Prism Global Market Study 2026/27 (research edition 20 Oct 2026): https://ethority.ai/ai-prism-study
• NBER Working Paper 31161: https://www.nber.org/papers/w31161
• METR study, via Simon Willison: https://simonwillison.net/2025/Jul/12/ai-open-source-productivity/
• Google RCT on developer productivity: https://arxiv.org/html/2410.12944v2
• KRDO / Stacker, AI notetaker survey (Jul 2026): https://krdo.com/stacker-business-economy/2026/07/16/ai-notetakers-have-sat-in-on-1-in-3-us-workers-meetings-but-only-a-third-say-they-were-asked-first/
• OpenAI, GPT-6 Astra: https://openai.com/index/gpt-6-astra/
• Anthropic, models overview: https://platform.claude.com/docs/en/about-claude/models/overview
• Google Cloud, Gemini 4 Argon: https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud
• Mistral AI, Mistral Large 4: https://mistral.ai/news/mistral-large-4/
• OpenAI, GPT-6.1 Sol and Luna pricing: https://openai.com/index/introducing-gpt-6-1-sol/
• a16z, enterprise AI survey (Jan 2026): https://a16z.com/leaders-gainers-and-unexpected-winners-in-the-enterprise-ai-arms-race/
• OpenAI, deprecations: https://developers.openai.com/api/docs/deprecations
• ASML, Q3 2026 results (6-K): https://www.sec.gov/Archives/edgar/data/0000937966/000162828026065813/pressreleasefinancialresul.htm
• IEA, data-centre electricity 2025: https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
• Epoch AI, inference price trends: https://epoch.ai/data-insights/ai-inference-price-trends
• Gartner, worldwide AI spending forecast (16 Sep 2026): https://gartner.com/en/newsroom/press-releases/2026-09-16-gartner-forecasts-worldwide-ai-spending-to-grow-49-point-5-percent-in-2026
• Gartner, agentic AI project cancellations: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
• Regulation (EU) 2026/1744: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
• Colorado SB 26-189: https://leg.colorado.gov/bills/sb26-189
• California Privacy Protection Agency, ADMT regulations: https://privacy.ca.gov/2025/09/california-finalizes-regulations-to-strengthen-consumers-privacy/

Updated edition of the article first published on ethority.net in 2024; the original remains archived.