The wrong question is: Which AI is best?
The useful question is: Which AI system is most effective for this workload, data boundary, failure tolerance, deployment model, and level of permitted autonomy?
There is no universal winner. A model that is excellent at long-horizon software engineering may be the wrong component for clinical documentation, credit analysis, tactical autonomy, industrial control, or legal research. In production, the model is only one layer. The inference engine determines how it runs. The harness determines what it can see, remember, call, and change.
This catalog is a 17 August 2026 market snapshot. It covers major globally relevant platforms and representative domain leaders. It does not claim to enumerate every AI product in every country. Product names link to primary vendor documentation or official product pages. Managed products sometimes do not disclose their exact model or serving stack; those fields are marked not publicly specified rather than inferred.
The fuzzy scores are my decision-support assessment for each product's stated best-fit use case. They are not laboratory benchmarks, market-share rankings, clinical validation, financial advice, or certification evidence.
TL;DR
- There is no universally "best" AI. Select a complete system for a defined workload, data boundary, failure cost, deployment model, and autonomy limit.
- Generative AI creates content. Agentic AI adds goals, state, tools, and an action loop. AGI remains a disputed target concept, not a verified commercial product category.
- Evaluate three layers: the model produces outputs, the engine runs inference, and the harness supplies data, memory, tools, policies, approvals, and observability.
- Use the catalog's fuzzy scores as shortlist guidance, not as benchmark results. Replace the analyst inputs with evidence from your own evaluation set before procurement or production use.
- In regulated, safety-critical, or security-sensitive domains, workflow controls, evidence provenance, human accountability, and failure containment matter more than a small difference in model capability.
Contents
- Generative AI vs. agentic AI vs. AGI
- Model, engine, and harness
- Fuzzy-effectiveness scoring
- Executive selection guide
- Foundation-model and AI-platform catalog
- Domain catalogs
- Production evaluation
- Final recommendation
- References
Generative AI vs. agentic AI vs. AGI
These terms describe different things and should not be used interchangeably.
| Term | What it is | What it does | What it is not |
|---|---|---|---|
| Generative AI | A class of models that produces derived synthetic content from learned patterns in data. NIST includes text, images, video, audio, and other digital content in this category. | Drafts, transforms, summarizes, predicts, synthesizes, or generates content. | It is not automatically an agent and does not inherently have permission to act. |
| Agentic AI | A system pattern in which a model operates inside a loop with goals, state, planning, tools, and action controls. NIST describes AI agents as capable of planning multi-step tasks and taking actions such as using tools and searching databases. | Observes state, selects a next step, calls tools, evaluates results, and continues until a stop condition or approval boundary is reached. | It is not a single model architecture. Adding function calling alone does not create a production-safe agent. |
| Artificial general intelligence (AGI) | A contested target concept for broadly general capability. Definitions vary: OpenAI describes highly autonomous systems outperforming humans at most economically valuable work, while Google DeepMind proposes levels based on performance, generality, and autonomy. | In theory, generalizes across a broad range of intellectual work at or above human level rather than excelling only within bounded tasks. | It is not a synonym for a strong chatbot, multimodal model, coding agent, or multi-agent workflow. |
Is AGI commercially available?
No system can be identified as verified AGI under a generally accepted, independently measurable standard as of this catalog date. Vendors and researchers disagree on the definition and threshold. Current systems can be highly capable, multimodal, tool-using, and autonomous over bounded workflows while still failing on reliability, transfer, factuality, long-horizon control, and unfamiliar real-world conditions.
The practical procurement position is simple: treat every current product as a bounded AI system with workload-specific evidence, not as general intelligence.
Primary definitions: NIST Generative AI Profile, NIST agentic AI evaluation work, Google DeepMind Levels of AGI, and the OpenAI Charter.
Model, engine, and harness: the stack that buyers should evaluate
These are normalized architecture terms. Vendors do not always use them consistently.
| Layer | Technical meaning | Examples | Procurement question |
|---|---|---|---|
| Model | The trained parameterized artifact that maps inputs to outputs. | GPT-5.6 Sol, Claude Fable 5, Gemini 3.7 Flash, Command A+, Llama 4 Maverick, MedGemma 1.5. | Does it perform the required task on our data, languages, modalities, and error cases? |
| Engine | The runtime that loads or hosts the model, schedules inference, manages accelerators, exposes APIs, and controls latency, throughput, scaling, and often isolation. | OpenAI managed inference, Amazon Bedrock Runtime, Vertex AI, Microsoft Foundry Models, NVIDIA NIM, vLLM, Hugging Face TGI. | Where does inference occur, what is retained, how is it isolated, and what are the availability and cost controls? |
| Harness | The application and orchestration layer around the model: prompts, retrieval, memory, tools, workflow state, policy enforcement, approvals, evaluation, and observability. | Codex, Claude Code, Bedrock AgentCore, Foundry Agent Service, Vertex Agent Engine, Palantir AIP, Security Copilot. | What can the AI read and change, under whose identity, with which approval and audit controls? |
This distinction matters because an enterprise may use the same model through several engines and harnesses with materially different risk. Claude through its native API, Amazon Bedrock, Microsoft Foundry, and a coding agent is not one control environment. The data path, retention terms, identity plane, tool permissions, logs, rate limits, and failure behavior differ.
How a production AI system actually behaves
A typical agentic request crosses these boundaries:
- The user or event enters the harness under an identity and tenant context.
- The harness loads system instructions, policy, conversation state, and retrieved enterprise data.
- A policy layer decides which model, tools, and data sources are eligible.
- The engine runs inference and returns text, structured output, or a requested tool call.
- A tool gateway validates the schema, authorization, target, and transaction risk.
- Read-only calls may execute automatically; material mutations should require deterministic policy and, where appropriate, human approval.
- Results return to the harness, which may repeat the loop, stop, escalate, or roll back.
- The system records evidence: actor, prompt version, retrieved sources, model, tool calls, authorization decisions, outputs, errors, cost, and latency.
The most important trust boundary is usually between model-generated intent and real-world execution. A model saying disable_account is not authorization. It is an untrusted proposal that must pass the same identity, policy, validation, separation-of-duty, and audit controls as any other privileged operation.
The fuzzy-effectiveness scoring model
The score estimates fitness for the specific best-fit use case listed in the catalog, not overall intelligence.
Each platform is assessed from 0 to 10 on five dimensions:
| Dimension | Weight | What it measures |
|---|---|---|
| Task fitness | 30% | Capability and workflow fit for the stated use case. |
| Ecosystem integration | 20% | Data connectors, APIs, identity integration, and workflow proximity. |
| Deployment control | 20% | Hosting choice, model choice, isolation, portability, and operational control. |
| Governance and safety | 20% | Access control, auditability, evaluation, approval, policy, and data protection. |
| Product maturity | 10% | Production evidence, support model, stability, and documentation. |
The defuzzified score is:
score = 0.30(task fitness)
+ 0.20(ecosystem integration)
+ 0.20(deployment control)
+ 0.20(governance and safety)
+ 0.10(product maturity)
The score is then fuzzified through overlapping membership bands. Using tri(a,b,c) for a triangular membership function and trap(a,b,c,d) for a trapezoidal one:
Conditional = trap(0.0, 0.0, 6.2, 7.2)
Capable = tri (6.2, 7.4, 8.3)
Strong = tri (7.5, 8.5, 9.3)
Leading = trap(8.7, 9.3, 10.0, 10.0)
The label with the highest membership is reported. A dual label, such as Strong/Leading, is used when the score sits materially in the overlap. This avoids pretending that an 8.4 and an 8.5 are meaningfully different without workload testing.
| Fuzzy label | Approximate center | Interpretation |
|---|---|---|
| Leading | 9.2 | Strong candidate for the stated workload; still requires local evaluation. |
| Strong | 8.5 | Production-capable with clear advantages and identifiable constraints. |
| Capable | 7.6 | Suitable when its specific strengths match the requirement. |
| Conditional | 6.5 | Use only with a narrow business case or compensating controls. |
For example, a coding agent assessed at 9.6 for task fitness, 9.3 for integration, 9.1 for deployment control, 9.2 for governance, and 9.4 for maturity produces:
(0.30 × 9.6) + (0.20 × 9.3) + (0.20 × 9.1)
+ (0.20 × 9.2) + (0.10 × 9.4) = 9.34 → 9.3 Leading
Those inputs remain analyst judgments. A buyer should replace them with scores derived from its own test set, architecture review, contractual evidence and operating requirements.
Evidence confidence is separate from effectiveness:
- High: detailed official technical documentation and clear component boundaries.
- Medium: official capabilities are documented, but model or runtime internals are partly opaque.
- Low: public evidence is mostly high-level product material. Low confidence does not mean poor capability; it means more buyer validation is required.
Executive answer: which AI should I use?
| Requirement | Best starting shortlist | Why | Non-negotiable gate |
|---|---|---|---|
| Complex professional reasoning and research | OpenAI GPT-5.6 Sol, Anthropic Claude Fable 5 or Opus 5, Google Gemini 3.1 Pro | Frontier reasoning, multimodal input, large contexts, and tool ecosystems. | Evaluate factuality, citations, retention, and tool behavior on your own corpus. |
| High-volume, lower-cost workflows | GPT-5.6 Terra/Luna, Gemini 3.7 Flash, Claude Sonnet 5/Haiku 4.5, Cohere Command A+ | Better latency and unit economics than always using the largest model. | Route by task complexity; do not silently downgrade high-risk decisions. |
| AWS-native enterprise AI | Amazon Bedrock with AgentCore or custom orchestration | Broad model choice, IAM integration, managed inference, and AWS-native operations. | Confirm model/Region availability, Marketplace access, logging, and cross-Region behavior. |
| Azure/Microsoft-native enterprise AI | Microsoft Foundry plus Agent Service or Copilot Studio | Broad catalog and close integration with Microsoft identity, data, and business applications. | Separate preview from GA services and verify model-specific data handling. |
| GCP-native multimodal and agent workloads | Gemini on Vertex AI plus Agent Engine/ADK | Strong multimodality, long context, grounding, and managed agent runtime. | Validate Region, data residency, search grounding, and tool authorization. |
| Sovereign or self-hosted AI | Llama, Mistral, Qwen, DeepSeek, or Granite on vLLM/TGI/NIM | Greater control over weights, network path, inference, and customization. | Budget for GPU capacity, patching, safety layers, evaluation, abuse controls, and lifecycle ownership. |
| Software engineering | Codex, Claude Code, GitHub Copilot, Gemini Code Assist | Repository-aware tools, command execution, edit/test loops, and developer workflow integration. | Sandbox execution, protect secrets, restrict network and CI credentials, and require reviewed diffs. |
| Security operations | The AI embedded in the telemetry platform you actually operate | Native security context usually matters more than a small difference in base-model capability. | Preserve evidence, require analyst validation, and never let a natural-language conclusion become an unreviewed containment action. |
| Healthcare | Dragon Copilot for documentation; MedGemma for validated application development; AlphaFold for molecular structure research | These tools solve different clinical and scientific problems. | Clinical validation, intended-use definition, privacy, regulatory review, and human accountability. |
| Finance | Bloomberg ASKB for Terminal workflows; Moody's for credit/risk; FactSet for investment data workflows | Domain data and provenance matter more than generic fluency. | No autonomous trading, lending, or risk acceptance without deterministic controls and accountable approval. |
| Defense | Palantir AIP for governed operational workflows; Anduril Lattice for C2 and mission autonomy; Shield AI Hivemind for AI-pilot autonomy | Each occupies a different layer of the mission stack. | Certification, rules of engagement, safety cases, degraded-mode testing, human command authority, and supply-chain assurance. |
| Creative production | Adobe Firefly for governed enterprise workflows; Midjourney for aesthetic exploration; Runway for generative video | Different strengths in governance, image aesthetics, and video production. | Rights, consent, provenance, brand controls, and human review. |
| Industrial and robotics | Siemens Industrial Copilot for automation engineering; NVIDIA Omniverse/Isaac for digital twins and robot learning | Domain context, simulation, and hardware/runtime integration. | Keep generated code and learned policies outside safety-critical control paths until verified and accepted. |
Global foundation-model and AI-platform catalog
Scores below apply to the stated best use, not every possible workload.
| Platform | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| OpenAI | Complex professional work, reasoning, coding, multimodal agents | GPT-5.6 Sol, Terra, Luna; specialized image, audio, and realtime models | OpenAI managed inference through the Responses and related APIs | ChatGPT, Codex, Agents SDK, tool/MCP integrations | 9.4 Leading | High |
| Anthropic Claude | Long-horizon reasoning, coding, document-heavy knowledge work | Claude Fable 5, Opus 5, Sonnet 5, Haiku 4.5 | Claude API or managed availability through major clouds | Claude, Claude Code, tool use, MCP, skills and subagents | 9.4 Leading | High |
| Google Gemini | Multimodal, long-context, grounded and agentic workflows | Gemini 3.1 Pro Preview, Gemini 3.7 Flash, Deep Research and specialized media/robotics models | Gemini API and Vertex AI managed inference | Gemini applications, Vertex Agent Engine, Agent Development Kit, tool and grounding services | 9.3 Leading | High |
| Microsoft Foundry | Model choice and governed agent deployment in Azure/Microsoft estates | Microsoft, OpenAI, Anthropic, Meta, Mistral, DeepSeek and partner catalog models | Foundry Models deployment and managed serving | Foundry Agent Service, Copilot Studio and application frameworks | 9.2 Leading | High |
| Amazon Bedrock | Multi-model generative AI in AWS | Amazon Nova plus Anthropic, OpenAI, Meta, DeepSeek, Mistral, Cohere and other supported models | Bedrock Runtime, Converse API and managed inference | Bedrock AgentCore, Bedrock Agents, Knowledge Bases, Guardrails and custom AWS orchestration | 9.2 Leading | High |
| NVIDIA AI platform | Accelerated self-hosted or hybrid AI, optimized inference, custom agent stacks | NVIDIA and supported open/partner models | NIM, Triton Inference Server and TensorRT-LLM on NVIDIA infrastructure | NeMo Agent Toolkit and ecosystem frameworks | 9.0 Strong/Leading | High |
| Cohere | Enterprise retrieval, multilingual RAG, private deployment and tool-using agents | Command A+, Command A family, Embed and Rerank models | Cohere API or supported private/cloud deployments | Cohere enterprise applications and custom RAG/agent orchestration | 8.8 Strong | High |
| Meta Llama | Custom and self-hosted multimodal applications where weight access matters | Llama 4 Maverick and Scout | Customer/partner-selected serving such as vLLM, TGI, NIM or cloud endpoints | Custom harnesses including LangGraph, Llama Stack and NeMo Agent Toolkit | 8.7 Strong | High |
| Mistral AI | European/sovereign deployment, efficient open and commercial models, document and speech workloads | Mistral Medium 3.5, Small 4, OCR 4, Voxtral and other current models | Mistral API/Compute or self-hosted inference for eligible models | Mistral Studio, Vibe and custom application frameworks | 8.7 Strong | High |
| Alibaba Qwen | Multilingual, coding, agentic and open-model deployments, particularly in Asian ecosystems | Qwen 3.8-Max and open Qwen families | Qwen/Alibaba Cloud APIs or self-hosted serving for released weights | Qwen applications, API tools and custom agent frameworks | 8.7 Strong | Medium |
| IBM watsonx | Regulated enterprise AI, hybrid deployment and governance-heavy workflows | IBM Granite family and supported external models | watsonx.ai inference on IBM Cloud or Software Hub | watsonx Orchestrate and governed enterprise workflows | 8.6 Strong | High |
| DeepSeek | Cost-conscious reasoning, coding and OpenAI/Anthropic-compatible API workflows | DeepSeek V4-Pro and V4-Flash | DeepSeek API; self-hosting where weights and compatible runtimes are available | DeepSeek applications, API integrations, Codex integration and custom harnesses | 8.6 Strong | High |
| xAI Grok | General reasoning, coding, realtime-search-connected applications and media generation | Grok 4.6 plus dedicated image, video and voice models | xAI API | Grok, Grok Build and custom Responses/API harnesses | 8.5 Strong | High |
The open-weight caveat
Open weights can improve deployment control, inspection, customization, offline operation, and sovereignty. They do not make the resulting system secure by default.
A self-hosted stack transfers responsibility to the operator for:
- model provenance and license compliance;
- artifact integrity and supply-chain scanning;
- GPU isolation, patching, capacity and denial-of-service controls;
- API authentication, tenancy, rate limits and abuse monitoring;
- prompt-injection resistance and tool authorization;
- safety tuning, evaluation, rollback and model lifecycle management;
- telemetry without sensitive prompt or completion leakage.
vLLM and Hugging Face Text Generation Inference are engines, not models or complete agent platforms. They serve models. A separate harness still owns retrieval, tool use, policy, state and approvals.
Software engineering AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| OpenAI Codex | Long-running repository work, complex implementation, testing, review and secure sandboxed workflows | GPT-5.6 Sol, Terra and Luna | OpenAI managed inference | Codex app, CLI and IDE extension with repository tools, sandboxing, skills, MCP and approvals | 9.5 Leading | High |
| Claude Code | Codebase reasoning, multi-file implementation, terminal/IDE workflows and extensibility | Current Claude models selected by product or configuration | Claude API or supported cloud engine | Claude Code with file, search, command, web, MCP, skills, hooks and subagents | 9.4 Leading | High |
| GitHub Copilot | Enterprise developer adoption inside GitHub, IDE and pull-request workflows | Selectable OpenAI, Anthropic, Google, xAI, Microsoft and open-weight models | GitHub-managed routing across documented hosting providers | Copilot Chat, coding agent, code review, extensions and GitHub context | 9.3 Leading | High |
| Gemini Code Assist and Gemini CLI | Google Cloud development, IDE assistance and open-source terminal agent workflows | Gemini 3.1 Pro Preview and Gemini 3.x Flash models, subject to product selection | Google-managed Gemini inference | Code Assist agent mode and Gemini CLI using a ReAct loop, tools and MCP | 9.0 Strong/Leading | High |
Coding-agent failure modes
Coding agents can increase throughput while also increasing the velocity of mistakes. Common production failures include:
- modifying generated or vendored code instead of the source of truth;
- passing unit tests while violating integration, concurrency or security assumptions;
- exposing secrets through prompts, logs, command output or remote tools;
- executing dependency-install hooks or repository instructions from untrusted branches;
- using broad cloud, GitHub or Kubernetes credentials inherited from the developer shell;
- creating plausible but unreviewed migrations, IAM policies or CI workflows;
- optimizing for a green test suite by weakening assertions or removing controls.
The control pattern is: isolated workspace, least-privilege credentials, egress restrictions, protected secrets, diff review, independent tests, software-composition and security scanning, and an accountable human merge decision.
Cybersecurity and security-operations AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Microsoft Security Copilot | Microsoft Defender, Sentinel, Entra, Intune and Purview investigations and security workflows | Azure OpenAI models from Foundry; exact selection can vary | Microsoft-managed security and model services | Security Copilot orchestration, grounding, plugins and embedded security experiences | 9.2 Leading in Microsoft estates | High |
| Gemini in Google Security Operations | Google SecOps investigation, summaries, detection creation and playbook assistance | Gemini family; exact production selection is vendor-managed | Google-managed inference integrated with SecOps | Google SecOps investigation assistant, SIEM/SOAR context and response workflows | 9.1 Leading in Google SecOps | Medium |
| CrowdStrike Charlotte AI | Falcon-native triage, investigation, custom security agents and agentic SOAR | Proprietary/vendor-managed model mix; exact model not publicly fixed | CrowdStrike-managed Falcon AI services | Charlotte AI, AgentWorks, mission-ready agents and Falcon data/action plane | 9.1 Leading in Falcon | Medium |
| SentinelOne Purple AI | Singularity-native threat hunting, OCSF-grounded investigation and response guidance | Proprietary/vendor-managed model mix; exact model not publicly fixed | SentinelOne-managed inference and Singularity data plane | Purple AI investigation workflows over native and third-party OCSF-normalized data | 9.0 Strong/Leading in Singularity | Medium |
There is no credible cross-platform claim that one of these products is the best SOC AI in isolation. The winning factor is usually telemetry coverage and action context. Security Copilot without relevant Microsoft signals, Charlotte without Falcon depth, or Purple AI without the right normalized data will underperform a supposedly weaker model embedded in a better-instrumented environment.
For production SecOps, require:
- links to original events, alerts, entities and queries;
- reproducible query generation and visible time windows;
- separation between an AI verdict and a containment decision;
- explicit authority for account disablement, host isolation, firewall change and token revocation;
- immutable logs of model, prompt, retrieved evidence, tool calls and analyst approval;
- false-positive, false-negative, mean-time-to-decision and analyst-override measurement;
- prompt-injection testing against logs, tickets, threat intelligence and retrieved documents.
Enterprise workflow, CRM, ERP and data AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Salesforce Agentforce | CRM-centric sales, service, commerce and business workflow agents | Supported OpenAI, Anthropic and Google models; selection varies | Salesforce-managed model access and Atlas Reasoning Engine | Agentforce Studio/Builder, Agent Graph, Data Cloud, actions and Salesforce trust controls | 9.1 Leading in Salesforce | High |
| SAP Joule | SAP-centered finance, procurement, supply chain, HR and business-process automation | SAP-managed and partner models; exact model depends on scenario | SAP Business AI Platform and managed model services | Joule Assistants, Joule Agents and Joule Studio across SAP and approved external systems | 9.1 Leading in SAP | Medium |
| Databricks AI and agents | Governed custom AI applications over lakehouse data, MLflow evaluation and multi-agent systems | Databricks-hosted, external and customer-selected models | Databricks Model Serving, compute and AI Gateway | MLflow ResponsesAgent, Agent Framework, Apps, Unity Catalog tools, MCP and Supervisor Agent | 9.2 Leading for data/ML teams | High |
| Snowflake Cortex Agents | Governed agents over structured and unstructured Snowflake data | Customer-selected supported orchestration model plus Cortex-managed models | Snowflake Cortex inference and Snowflake compute | Cortex Agents, Analyst, Search, code execution, skills and custom UDF/procedure tools | 9.1 Leading in Snowflake | High |
Agentforce, SAP Joule, Databricks and Snowflake are not substitutes for one another. They are strongest where the authoritative data, permissions and business actions already live. Moving a workflow to a marginally stronger base model while weakening data lineage or authorization is usually a net loss.
Healthcare and life-science AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Google MedGemma | Building and validating medical text/image applications under a defined intended use | MedGemma 1.5 4B multimodal; MedGemma 1 4B and 27B variants | Local or customer-selected serving, including documented Vertex/model-garden options | Customer-built clinical application, retrieval, validation and workflow controls | 8.7 Strong for developers | High |
| AlphaFold 3 and AlphaFold Server | Biomolecular structure and interaction prediction for research | AlphaFold 3 | AlphaFold Server or approved local/research deployment paths | AlphaFold Server and scientific analysis workflow | 9.5 Leading for its narrow scientific task | High |
| Microsoft Dragon Copilot | Ambient clinical documentation, information surfacing and workflow assistance | Microsoft/Nuance clinical speech and language models; exact stack not publicly specified | Microsoft-managed healthcare AI services | Dragon Copilot applications and supported EHR integrations | 9.2 Leading for clinical documentation | Medium |
| NVIDIA BioNeMo | Training, adapting and deploying biomolecular models for computational drug discovery | AMPLIFY, ESM-2, Evo2, Geneformer and supported biomolecular models | NVIDIA GPU runtime and BioNeMo-optimized framework components | BioNeMo Framework, recipes, containers and customer scientific pipelines | 9.0 Strong/Leading for biotech engineering | High |
Do not compare AlphaFold's 9.5 with a general-purpose model's 9.4. AlphaFold is exceptionally effective at a narrow scientific task; it is not a general assistant. MedGemma is a developer foundation, not an approved diagnosis engine. Dragon Copilot assists clinical workflow; it does not transfer clinical accountability to Microsoft or the model.
Healthcare deployment gates should include intended-use documentation, representative clinical validation, subgroup and edge-case analysis, privacy and residency review, human factors, monitoring for distribution shift, incident reporting, rollback, and regulatory assessment in every deployment jurisdiction.
Financial-services AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Bloomberg ASKB | Conversational institutional research and discovery inside Bloomberg Terminal workflows | Bloomberg/vendor-managed models; exact production model not publicly specified | Bloomberg-managed financial data and inference services | ASKB in the Bloomberg Terminal with Bloomberg content and workflow context | 9.3 Leading for Terminal users | Medium |
| Moody's AI solutions | Credit analysis, lending, insurance, KYC, third-party risk and research | Moody's/vendor-managed models; exact production model varies | Moody's managed data, analytics and AI services | Research Assistant, coordinated agents, credit memo and risk workflow products | 9.2 Leading for credit/risk | Medium |
| FactSet AI | Investment research, portfolio and financial-data workflows in FactSet | FactSet/vendor-managed models; exact production model not publicly fixed | FactSet-managed data and AI services | FactSet workstation, APIs and workflow-specific AI experiences | 9.0 Strong for investment analytics | Medium |
Financial AI should accelerate evidence gathering, comparison, drafting and exception identification. It should not silently become the system of record for prices, credit decisions, sanctions screening, regulatory reporting or trade execution. Material decisions need traceable source data, deterministic calculations where possible, maker-checker controls, model-risk governance and post-decision surveillance.
Defense, national security and mission AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Palantir AIP | Governed operational data, decision support and AI-enabled workflows across enterprise or mission contexts | Customer-selected and platform-supported LLMs plus conventional ML; exact choice varies | Palantir AIP/Foundry managed model and data services | AIP Chatbot Studio, AIP Logic, Workflow Builder, Ontology, actions and granular access control | 9.2 Leading for governed operational workflows | High |
| Anduril Lattice | Sensor fusion, command and control, tactical networking and multi-system mission autonomy | Proprietary perception, tracking and autonomy models; details not publicly specified | Lattice distributed edge/mission runtime | Lattice Command & Control, Mission Autonomy and Lattice Mesh | 9.1 Leading for integrated C2/autonomy | Medium |
| Shield AI Hivemind | Platform-agnostic AI-pilot and multi-agent autonomy in GPS/comms-denied environments | Mission-specific perception, cognition and control models; details not publicly specified | Hivemind Edge | Hivemind Design for development/test and Hivemind Commander for operator interaction | 9.2 Leading for unmanned mission autonomy | Medium |
These scores do not certify suitability for a weapon, aircraft, ship, C2 network or safety-critical mission. Public product material is not a safety case.
For defense use, the evaluation boundary must include:
- digital-twin, software-in-the-loop, hardware-in-the-loop and live-range evidence;
- contested PNT, degraded communications, sensor deception and intermittent data;
- identification confidence, track provenance and uncertainty propagation;
- deterministic command authorization and rules-of-engagement enforcement outside the generative model;
- human command authority, abort paths and loss-of-link behavior;
- adversarial ML, prompt injection, data poisoning, supply-chain and update-channel threats;
- classification, cross-domain, crypto-boundary and coalition-release controls;
- configuration control, signed artifacts, reproducible builds and rollback;
- operational test and evaluation against mission-specific acceptance criteria.
Creative image and video AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Adobe Firefly | Governed enterprise creative production, Adobe workflows and brand customization | Adobe Firefly image/video models plus approved partner models in eligible workflows | Adobe managed generative services and APIs | Firefly web/Creative Cloud, Creative Production, workflow authoring, APIs and Custom Models | 9.2 Leading for enterprise creative governance | High |
| Midjourney | High-quality aesthetic exploration and image ideation | Midjourney V8.2 default as of 24 July 2026 | Midjourney-managed GPU inference | Midjourney web/Discord workflow, parameters, personalization and reference controls | 9.4 Leading for image aesthetics | High |
| Runway | Generative video, image/audio APIs and production-oriented media workflows | Runway Gen-4.5 and supported first/third-party media models | Runway managed inference and Model Router | Runway application, API, recipes and model-routing configuration | 9.3 Leading for generative video workflows | High |
The operational differentiator is not only visual quality. Enterprise buyers should examine training-data and indemnity terms, consent, likeness and voice controls, content credentials/provenance, geographic restrictions, model-change behavior, retention, private asset handling, and whether outputs are acceptable for the intended commercial channel.
Legal and education AI
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Harvey | Legal research, drafting, contract analysis, due diligence and source-grounded legal workflows | Harvey/vendor-selected frontier and specialized models; exact routing is not publicly fixed | Harvey-managed legal AI services | Harvey platform, Agents, Vault, Knowledge, Contract Intelligence and integrations | 9.2 Leading for legal workflows | Medium |
| Khanmigo | Guided tutoring and teacher support grounded in Khan Academy content | Current underlying model not publicly specified on the cited product page | Khan Academy-managed service | Khanmigo tutoring/teaching workflow, Khan content and education safety design | 9.0 Strong/Leading for guided education | Medium |
| Duolingo Max | Conversational language practice and explanation inside a structured learning product | OpenAI GPT-4 is the officially documented launch model; current routing should be confirmed before procurement | Duolingo/OpenAI managed services | Duolingo curriculum, roleplay, video-call and feedback experiences | 8.8 Strong for language practice | Medium |
Legal citations must be opened and checked. A source-grounded interface reduces, but does not eliminate, fabricated authority, stale law, jurisdiction mismatch, privilege leakage or incorrect application of facts. Education systems need age-appropriate design, teacher/guardian oversight, privacy controls, pedagogical evaluation and safeguards against shortcutting the learning objective.
Manufacturing, industrial AI, digital twins and robotics
| Product | Best use | Model | Engine | Harness | Fuzzy score | Evidence |
|---|---|---|---|---|---|---|
| Siemens Industrial Copilot | Automation engineering, industrial code assistance and maintenance workflows | Azure OpenAI models enriched with Siemens domain and process context; exact version varies | Microsoft Azure OpenAI plus Siemens industrial services | Siemens Industrial Copilot integrated with Xcelerator, TIA Portal and supported industrial workflows | 9.2 Leading for Siemens automation estates | High |
| NVIDIA Omniverse | Industrial digital twins, physically based simulation and synthetic-data pipelines | Simulation, rendering and workload-specific AI models rather than one foundation model | Omniverse libraries/microservices on NVIDIA-accelerated infrastructure | OpenUSD-based applications, digital-twin workflows and integration pipelines | 9.3 Leading for industrial digital twins | High |
| NVIDIA Isaac GR00T, Sim and Lab | Robot foundation models, simulation, synthetic data, policy training and deployment | GR00T robot foundation models plus customer perception/control policies | Isaac/Jetson accelerated runtime, simulation and inference libraries | Isaac Sim, Isaac Lab, ROS packages, data pipelines and robot-development workflow | 9.3 Leading for robotics development | High |
Generated PLC code, robot policies and maintenance recommendations belong outside the trusted safety envelope until independently verified. Simulation helps expose faults but cannot prove that the real environment, sensors, actuators, timing, wear, operators and adversaries match the model. Keep certified safety interlocks deterministic and independent from generative output.
Domain selection map
This map covers the major use domains without pretending that a single vendor owns each category.
| Domain | Primary candidates | Use them for | Do not use them as |
|---|---|---|---|
| Knowledge work and research | OpenAI, Anthropic, Gemini, Cohere | Synthesis, drafting, analysis, grounded research and document workflows | Unverified authority or a substitute for accountable expert judgment |
| Software engineering | Codex, Claude Code, GitHub Copilot, Gemini Code Assist | Codebase exploration, implementation, testing, review and documentation | An autonomous production deployer with inherited admin credentials |
| Cybersecurity | Security Copilot, Gemini in SecOps, Charlotte AI, Purple AI | Triage, hunting, query generation, evidence synthesis and response planning | Sole incident commander or unreviewed containment authority |
| Cloud and platform engineering | Foundry, Bedrock, Vertex AI, Databricks, Snowflake | Building governed AI applications close to enterprise data and identity | A reason to bypass architecture, IAM, data classification or change control |
| Healthcare | Dragon Copilot, MedGemma, specialized regulated products | Documentation and validated narrow clinical applications | A general diagnostic authority without intended-use validation |
| Life science and drug discovery | AlphaFold, BioNeMo, domain models | Structure prediction, representation learning and computational discovery | Clinical proof or experimental validation |
| Finance and insurance | Bloomberg, Moody's, FactSet, enterprise model platforms | Research, credit/risk evidence, document analysis and decision support | Uncontrolled trading, lending, pricing or regulatory filing authority |
| Defense and public safety | Palantir AIP, Anduril Lattice, Shield AI Hivemind | Governed operational workflows, C2, sensor fusion and bounded autonomy | AGI or a replacement for command responsibility, ROE and safety certification |
| Legal | Harvey and other source-grounded legal platforms | Research, drafting, review, due diligence and matter workflows | Final legal judgment or unchecked citation authority |
| Education | Khanmigo, Duolingo and institution-governed general models | Guided tutoring, teacher support and structured practice | A mechanism for bypassing assessment or unsupervised child profiling |
| Customer service and sales | Agentforce, Foundry/Copilot Studio, custom cloud agents | Case handling, summarization, next-best action and governed workflow automation | An unrestricted refund, pricing or account-change authority |
| ERP, procurement and supply chain | SAP Joule, Palantir AIP, enterprise agents | Cross-process analysis, planning and workflow coordination | An unreviewed approver for payments, suppliers or inventory movements |
| Manufacturing and OT | Siemens Industrial Copilot, Omniverse, domain ML | Engineering assistance, digital twins, maintenance and optimization | Direct safety-controller logic without verification and acceptance |
| Robotics and autonomy | NVIDIA Isaac, Shield Hivemind, Anduril Mission Autonomy | Simulation, policy development and bounded autonomous behavior | A safety case based only on benchmark or simulated success |
| Creative and marketing | Adobe Firefly, Midjourney, Runway | Ideation, brand assets, image and video generation | Proof of rights, consent, truth or provenance |
| Sovereign/offline deployments | Llama, Mistral, Qwen, DeepSeek, Granite with vLLM/TGI/NIM | Controlled network paths, local inference and customization | A lower-effort security or operations model |
Why benchmark leaderboards are not enough
Public benchmarks are useful for model research and initial screening. They rarely reproduce the production system:
- the prompt template and retrieved context differ;
- tool schemas, retry logic and stop conditions differ;
- quantization and serving configuration change behavior;
- latency and concurrency affect user and agent decisions;
- the harness may summarize or truncate history;
- enterprise permissions restrict available evidence;
- model aliases may move to new snapshots;
- vendors apply different safety filters, post-processing and data-retention terms;
- domain errors carry radically different impact.
A 2% benchmark advantage is irrelevant if the system cannot honor tenant boundaries, cite the record, survive rate limits, integrate with your identity plane, or roll back a bad action.
A production evaluation that can survive audit
Use a staged evaluation rather than a vendor demo.
1. Define the decision and failure cost
Write the intended use, prohibited use, accountable owner, data classification, maximum autonomy, and impact of false positive, false negative, fabricated output, privacy breach and unavailable service.
2. Build a representative evaluation set
Include normal work, difficult work, multilingual inputs, malformed documents, stale data, missing permissions, adversarial retrieved content, conflicting sources, long contexts, rate limits and downstream tool failures. Preserve a holdout set.
3. Test the complete system
Evaluate model plus engine plus harness. Measure task success, citation correctness, tool-call precision, privilege adherence, recovery behavior, latency, cost, analyst overrides and unsafe-action attempts.
4. Exercise trust boundaries
Attempt indirect prompt injection through documents, tickets, logs, source code, web pages and email. Test cross-tenant retrieval, confused-deputy conditions, schema manipulation, tool-result poisoning, memory contamination and approval bypass.
5. Prove operational control
Verify SSO, least privilege, service identities, network egress, encryption, retention, regional processing, key management, audit logs, alerting, quotas, kill switch, fallback, version pinning where available, canary rollout and rollback.
6. Re-evaluate continuously
Managed models and harnesses change. Re-run the regression suite on model, prompt, retrieval, tool, policy or vendor changes. Monitor production drift and override rates. A passed pilot is not a permanent control attestation.
Common misconceptions
“The smartest model gives the best enterprise result”
Usually false. A slightly weaker model with authoritative data, narrow tools, deterministic policy and strong observability can outperform a stronger model operating with poor context and excessive authority.
“Agentic AI means autonomous AI”
Not necessarily. Agency is a spectrum. An agent can propose steps, perform read-only retrieval, execute reversible low-risk actions, or operate with high autonomy. The permitted level should follow impact and assurance evidence.
“A reasoning engine is the model”
Sometimes vendors use “reasoning engine” as a product term for an orchestration layer. Salesforce Atlas, for example, coordinates logic, state and model calls. In normalized architecture terms, that is closer to a harness/orchestration engine than a standalone foundation model.
“Self-hosted means private and secure”
Self-hosting gives control; it also creates operational responsibility. A publicly reachable unauthenticated vLLM endpoint with broad tool access is less secure than a well-governed managed service.
“RAG prevents hallucination”
Retrieval-Augmented Generation supplies context. It can retrieve the wrong record, omit the controlling record, expose unauthorized data, or ingest malicious instructions. Citation and authorization must be validated independently.
“No human in the loop means more advanced”
Removing human approval is an operating-model choice, not a maturity badge. High-volume reversible actions may justify automation. Irreversible, safety-critical, financial, legal, clinical or mission actions need stronger independent control.
Final recommendation
Choose the system, not the logo.
Start with the data and action plane you already trust. Shortlist two or three model/engine/harness combinations. Test them on representative work and adversarial cases. Score task success and failure cost, not conversational polish. Keep model-generated intent outside the authorization boundary. Preserve evidence. Design for model changes and service failure.
If forced to nominate broad starting points in August 2026:
- OpenAI, Anthropic and Google form the strongest general-purpose frontier shortlist.
- Microsoft Foundry, Amazon Bedrock and Vertex AI are the primary hyperscale enterprise control planes.
- Meta Llama, Mistral, Qwen, DeepSeek, Granite and NVIDIA tooling lead when self-hosting, sovereignty or customization drives the architecture.
- Domain platforms win when their proprietary data and workflow context matter more than general model capability.
The best AI is therefore not one product. It is the best-evidenced, least-overprivileged, operationally supportable stack for a defined decision.
Top comments (0)