Glossary of AI Automation Terms

The 213 terms readers of this research set on AI-based automation will meet, explained in plain English, with a focus on India. All facts, prices and dates are as of 12 October 2026, re-checked against primary sources that day.

Terms are grouped by theme and listed alphabetically within each group, and each entry links to the document of the set that covers it in most depth; the quick index at the end lists every term from A to Z.

Core concepts

• Agent (AI agent, agentic AI): A system in which the model, not the developer, chooses the next step, calling tools in a loop until done or stopped. "Agentic AI" is the wider label. See: AI automation explained
• Agent washing: Gartner's term for relabelling chatbots, assistants or RPA tools as "agentic". It reportedly estimated in June 2025 that only about 130 vendors selling agentic AI were real (headline-level source; Gartner's release body was not read). See: AI automation explained
• AGI (artificial general intelligence): Loosely, AI that matches people across most cognitive work. There is no agreed test, so the set tracks measured capability rather than AGI claims. See: Outlook 2027-2030
• AI-based automation: Any production system in which a generative model reads input or chooses actions inside a business process, plus the tools, data, evaluation and controls around it. See: AI automation explained
• Automation versus augmentation: The split used by Anthropic's Economic Index: in automation the model does the task, in augmentation it helps a person do it. Chat use leans to augmentation, API use to automation. See: Claude (Anthropic)
• Autonomy levels (the automation ladder): The set's six-step scale from rules and RPA, through workflows with LLM steps and tool-using agents, to multi-agent and computer-use agents. Most small-business work should start at level 3. See: AI automation explained
• Context engineering: Anthropic's term for choosing, turn by turn, what goes into a model's limited context: instructions, retrieved facts, tool results and memory. See: Skills and implementation playbook
• Context window: The most text, counted in tokens, a model can consider in one request. Windows of 1M tokens are common in 2026, but some vendors charge more above a threshold. See: AI automation explained
• Distillation: Training a smaller model on a larger model's outputs so it inherits some of its skill cheaply. Unauthorised distillation through rival labs' APIs has drawn public accusations, a provenance risk for buyers. See: DeepSeek
• Evidence labels: Tags the set uses to show how far to trust a figure: vendor claim, secondary (news, aggregator or Wikipedia), single source, snippet or headline only, and calc (the set's own arithmetic). See: AI automation explained
• Fine-tuning: Further training a model on an organisation's own examples so it follows a format or domain more closely. OpenAI has closed hosted fine-tuning to new organisations (since 7 May 2026) and allows no new jobs from 6 January 2027, pushing this work to self-hosted models. See: OpenAI (GPT)
• Human-in-the-loop (HITL): A design in which a person approves consequential agent actions such as payments, writes and outgoing messages. It is the most recommended control, but weak alone because of "approval fatigue". See: Regulation and security
• Large language model (LLM): A model trained on vast text collections to predict the next token, which lets it read, write, classify and plan in natural language. See: AI automation explained
• Physical AI: AI that acts in the physical world through robots, vehicles and machines. Robotaxis are scaling, but the set treats humanoid robots as a 2028-or-later bet for most small firms. See: Outlook 2027-2030
• Robotic process automation (RPA): Software robots that replay recorded screen steps, such as copying data between applications. Predictable and auditable, but it breaks when a screen changes; RPA vendors are now adding agents. See: AI automation explained
• Token: The unit of text a model reads and writes, roughly a word fragment; usage and prices are counted per million tokens. Output tokens cost several times more than input. See: Costs and economics
• Tool use (function calling): A model's ability to request a named function, such as a database query or API call, which the application runs and returns. "Structured outputs" use it to return schema-valid JSON. See: Claude (Anthropic)
• Workflow (LLM workflow): A system in which the developer fixes the path and a model classifies, extracts or drafts at set points. Each step can be tested, so it is the usual first project. See: AI automation explained

Agents and architectures

• Agent framework: A code library that runs the agent loop, tools, state, retries and approvals. LangChain and its LangGraph runtime, CrewAI and Microsoft Agent Framework lead the field. See: Agent frameworks and protocols
• Agent loop and stopping conditions: The cycle in which a model calls a tool, reads the result and decides again. It needs stopping rules (final answer, error, turn or cost limit) or it can loop and overspend. See: AI automation explained
• AgentOps (LLMOps): Operations for agents in production: versioned prompts, CI/CD, telemetry, cost caps and model migrations. AWS recommends dedicated "DevOps for agents" teams. See: Skills and implementation playbook
• Claude Agent SDK: Anthropic's library that gives custom agents the same loop, tools and context handling as Claude Code. Its licence has two layers: the LICENSE file says MIT, but the README puts use, including in products sold to customers, under Anthropic's Commercial Terms. Claude Code itself is proprietary. See: Agent frameworks and protocols
• Coding agent: An agent that reads a codebase, edits files, runs tests and proposes changes, such as Claude Code, Codex, Gemini CLI, Copilot, Cursor or Devin. The leading agent category. See: Agent frameworks and protocols
• Computer-use agent: An agent that operates a desktop or browser from screenshots, needing no API; AI browsers such as Perplexity Comet apply this to the web. Reliability is low, and page content can hijack it. See: Agent frameworks and protocols
• Evaluator-optimizer: A pattern in which one model drafts and another critiques against set criteria, repeating until they are met. It suits drafting and translation but needs a stop rule. See: AI automation explained
• Google ADK (Agent Development Kit): Google's open-source, code-first agent framework (Apache-2.0) with graph-based sub-agents, and the export target of Google's visual Agent Studio. See: Google Gemini
• Long-running harness and background agent: A harness is the scaffolding that lets an agent work across many sessions, using progress logs and task lists. Background agents pursue a user's goals continuously, such as OpenAI's Dots. See: AI automation explained
• Low-code agent builder: A visual tool for assembling agents and RAG apps with little code, such as Dify or Langflow. Check licence and upkeep first: Flowise was archived in August 2026. See: Production tool stack
• Managed agent runtime: A vendor-hosted service that runs the agent loop, sandbox and state, such as Claude Managed Agents, OpenAI's Agents API or Copilot Studio. Convenient but less portable, and sometimes outside zero data retention. See: Agent frameworks and protocols
• Multi-agent system: A lead agent delegating to subagents, or agents handing a conversation to specialists ("handoffs"). It suits broad parallel research but uses about 15 times a chat's tokens. See: AI automation explained
• n8n: A widely used self-hostable workflow-automation tool with AI and MCP features. Its Sustainable Use License allows use only "for your own internal business purposes or for non-commercial or personal use", and distribution to others only "free of charge for non-commercial purposes". Running it as a paid managed service for third parties therefore needs a commercial agreement with n8n (inference from the licence text). See: Production tool stack
• OpenAI Agents SDK and Agent Builder: OpenAI's agent routes are the open-source Agents SDK, the managed Agents API and the low-level Responses API. Its visual Agent Builder shuts down on 30 November 2026. See: OpenAI (GPT)
• Orchestrator-workers: A pattern in which a lead model splits a task that cannot be planned in advance and hands parts to worker models. Flexible, but hard to test and variable in cost. See: AI automation explained
• Parallelisation: Running subtasks side by side, or one task several times with a vote, sometimes with a guardrail model screening alongside. It buys speed or confidence at a multiple of the tokens. See: AI automation explained
• Prompt chaining: A fixed sequence of model calls with code checks between them, such as extract, validate, then draft. The simplest pattern; errors pass along if the checks are weak. See: AI automation explained
• Routing: A classifier step that sends each input to a specialised prompt or a cheaper model. It suits mixed queues such as support tickets, with a fallback route for misroutes. See: AI automation explained
• Voice agent: A real-time loop of speech-to-text, a model, text-to-speech and telephony that handles calls. Cost it per minute all-in; full costs often run several times the platform's headline rate. See: Production tool stack
• Workflow engine: Software that triggers and connects steps across apps, such as Zapier, Workato, Activepieces or n8n (the iPaaS category). Durable-execution engines such as Temporal keep long workflows running through crashes and retries. See: Production tool stack

Protocols and standards

• A2A (Agent2Agent protocol): An open protocol, created by Google and now under the Linux Foundation, for one agent to delegate work to another across vendors. Version 1.0 adds signed Agent Cards that identify each agent. See: Agent frameworks and protocols
• AAIF (Agentic AI Foundation): The Linux Foundation fund that has governed MCP since 9 December 2025. Anthropic (MCP), Block (goose) and OpenAI (AGENTS.md) contributed its founding projects; its Platinum members are AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. It also stewards AGENTS.md. See: Agent frameworks and protocols
• ACP (Agentic Commerce Protocol): OpenAI and Stripe's protocol for checkout by agents, the kind of agent-driven buying called "agentic commerce". Its repository marks it beta (Apache-2.0; latest spec 2026-04-17), maintained by OpenAI and Stripe per the README (MAINTAINERS.md also lists Meta), so treat it as early-stage. See: Agent frameworks and protocols
• Agent Skills: Anthropic's packages of instructions, metadata and optional scripts that a model loads only when relevant; OpenAI and Google now package skills similarly. Review skills like code, since they can run scripts. See: Agent frameworks and protocols
• AP2 (Agent Payments Protocol): Google's protocol for proving that a user authorised an agent's payment, extending A2A and MCP. Donated to the FIDO Alliance in April 2026, it is still at v0.2. See: Agent frameworks and protocols
• MCP (Model Context Protocol): The de facto open standard for connecting models and agents to tools and data, created by Anthropic and governed by the AAIF. Widely supported, so a connector outlives a change of model. The official MCP Registry listed 42,172 distinct servers on 12 October 2026 (registry entries, not a measure of production use). See: Agent frameworks and protocols
• MCP host, client and server: An MCP server exposes a system's tools and data; the client inside a host app, such as an IDE or agent, connects to it. Vet third-party servers like any code. See: Agent frameworks and protocols
• MCP specification 2026-07-28: The largest MCP revision so far: a stateless core without sessions, with Tasks, MCP Apps and Enterprise-Managed Authorization (single sign-on through an identity provider) as extensions. Existing servers need rebuilding. See: Agent frameworks and protocols
• OpenAI-compatible API: An endpoint that accepts OpenAI's request format, so code can switch provider by changing the base URL and model name. Test it: some silently remap models or ignore cache settings. See: OpenAI (GPT)
• OpenTelemetry GenAI conventions: Vendor-neutral names for recording model calls, tool calls and agent steps as traces, so observability tools can share data. The conventions have moved to the open-telemetry/semantic-conventions-genai repository and are marked "Development", not stable. See: Production tool stack

Models and model types

• Chinese open-weight models (Qwen, Kimi, GLM, MiniMax): Model families from Alibaba, Moonshot, Zhipu and MiniMax that lead on breadth and price. Licences range from MIT to custom terms with revenue thresholds or attribution duties. See: Other leading models
• Claude models (Haiku, Sonnet, Opus, Fable, Mythos): Anthropic's tiers from cheapest to most capable: Haiku, Sonnet, Opus and Fable. Mythos is restricted to vetted organisations. See: Claude (Anthropic)
• DeepSeek V4: DeepSeek's MIT-licensed open-weight family (V4-Pro and V4-Flash), among the cheapest capable models, with off-peak discounts. For client data, self-hosting the weights manages the risk better than using its API. See: DeepSeek
• Frontier model: The most capable general-purpose models at a given moment, in October 2026 led by Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1. Access to the newest is increasingly staged or restricted. See: Top 10 models compared
• Gemini models: Google's family: Gemini 3.8 Flash is the production default, the Pro tier is still in preview, and Gemini 4 Argon and Deep Think are gated. Gemma is Google's open-weight line. See: Google Gemini
• GPT-6 models: OpenAI's lineup: GPT-6 Astra (frontier), GPT-6.1 Sol (mid tier) and GPT-6 Luna (small). Astra is rated "Critical" for cybersecurity capability and is released in gated stages. See: OpenAI (GPT)
• gpt-oss: OpenAI's open-weight models under Apache 2.0. gpt-oss-120b fits one 80 GB GPU or a 128 GB desktop, making it the reference model for private, single-box deployments. See: Hardware
• Grok: xAI's models, now part of SpaceX. Cheap per token but priced up for prompts of 200K tokens or more, and the chatbot carries regulatory baggage from its January 2026 deepfake scandal. See: Grok (xAI)
• Language identification (LID): Detecting which language or script a speaker or text uses, measured as the share of utterances transcribed in the wrong language or script. Sarvam reports a 5.22% LID error across 22 languages (vendor-reported). See: Indian languages
• Mixture of experts (MoE): A model made of many expert sub-networks, of which only a few run for each token. Total parameters set the memory needed, active parameters the speed. See: Hardware
• Model lifecycle (preview, GA, deprecation): Models move from preview to general availability (GA), then to deprecation and retirement, after which calls fail. Releases now come about monthly, so pin model IDs and keep a migration calendar. See: OpenAI (GPT)
• Model tier: The price-and-capability band a model sits in: small, mid, flagship or frontier. List prices span roughly 300x, so budget by tier and move routine steps to cheaper tiers. See: Costs and economics
• Open-weight model: A model whose trained weights can be downloaded and self-hosted under a licence. Not necessarily open source: training data is usually withheld and licences may restrict commercial use. See: Other leading models
• Reasoning model and effort settings: A model that spends extra "thinking" tokens before answering, which helps on hard multi-step work. Anthropic and OpenAI bill those tokens as output even when the thinking text is not returned, and Google prices them in its output rate. Effort settings are a cost dial: maximum effort can cost several times more for a few points. See: Top 10 models compared
• Small language model (SLM): A model of a few billion parameters, such as Phi-4-mini or small Qwen and Granite models, that runs on a laptop, phone or edge box; suited to classification and extraction. See: Other leading models
• Sovereign AI: National control over the models, compute and data a country depends on, through domestic models, local data centres or data-location rules. In practice it often means not depending on one supplier. See: China and Asia-Pacific
• Sparse attention: An attention design in which each token looks at a selected subset of earlier tokens, cutting the memory and cost of long contexts. DeepSeek V3.2 and V4 use it. See: DeepSeek
• Speaker diarisation: Labelling who spoke when in a recording. Sarvam charges ₹45 an hour for batch transcription with diarisation against ₹30 an hour otherwise (undated price page, GST not stated). See: Indian languages
• Speech-to-text and text-to-speech (STT, TTS): Models that transcribe speech and synthesise voice, the two ends of every voice agent. Indian-language providers such as Sarvam price them in rupees. See: Production tool stack
• System card (model card): The document a lab publishes with a model, covering capabilities, evaluations, safety testing and limits. It often holds facts that launch pages omit. See: OpenAI (GPT)
• Tokenizer fertility: How many billable tokens a tokenizer needs for the same content. On the GPT-4-era tokenizer Hindi needs 4.1 times and Tamil 9.9 times the tokens of English; the newer o200k tokenizer cuts the average across ten Indian languages to 2.1 times (preprint). See: Indian languages
• Western open-weight models (Llama, Mistral, Phi, Granite, Nemotron): Downloadable families from Meta, Mistral, Microsoft, IBM and NVIDIA. Phi is MIT and Granite Apache 2.0, while Llama requires "Built with Llama" attribution and Mistral's commercial terms conflict between sources. See: Other leading models

Benchmarks and evaluation

• Artificial Analysis Intelligence Index: An independent composite score from Artificial Analysis, which also measures speed and cost per task. The index is at version v4.3.2, and Claude Opus 5.5 led it in October 2026 (57.6). See: Top 10 models compared
• AutomationBench: Zapier's benchmark of business-workflow automation across SaaS apps, the closest public proxy for workflow agents. Safety interventions count as failures. See: Top 10 models compared
• Benchmark: A fixed public test used to compare models. Most 2026 scores are vendor-reported on the vendor's own settings, so use benchmarks to shortlist and your own evals to decide. See: Top 10 models compared
• Containment rate: The share of conversations a bot resolves without a human, judged against a rubric and reported with the escalation rate. No figures are published for Indian small-business bots. See: Indian languages
• Epoch Capabilities Index (ECI): Epoch AI's independent index combining many benchmarks on one scale, calibrated so Claude 3.5 Sonnet scores 130 and GPT-5 150. Claude Opus 5.5 led at 167 on 1 October 2026. See: Top 10 models compared
• Eval (evaluation suite): A firm's own set of real tasks and graders, run after every change of prompt, model or tool. Start with 20 to 50 tasks from real failures and grade outcomes, not claims. See: Skills and implementation playbook
• GDPval: OpenAI's benchmark of real knowledge-work deliverables across 44 occupations, graded against work by industry experts. A proxy for back-office documents and analysis. See: Top 10 models compared
• Indian speech test sets (IndicVoices, Vistaar, Kathbath): Public test sets on which Indian vendors report speech-to-text error rates, each on a different subset, so vendor figures cannot be compared. Run a fixed set that includes your own recordings. See: Indian languages
• Indic language-model benchmarks (MILU, IndicGenBench, Indi-RomCoM): MILU tests 11 Indic languages across 41 subjects, IndicGenBench covers generation in 29 languages, and Indi-RomCoM (a June 2026 preprint) tests romanised code-mixed instructions, where every model lost accuracy as mixing rose. See: Indian languages
• LLM-as-judge: Using a model to grade another model's outputs so evals can scale. Calibrate the judge against human graders before trusting it. See: Skills and implementation playbook
• LMArena (Arena): A public leaderboard that ranks models by human preference votes (Elo), with text, WebDev and agent boards. It measures preference rather than task success, so use it to shortlist only. See: Top 10 models compared
• METR time horizon: METR's measure of the length of task, in human-expert time, that a model completes half the time. Horizons doubled every 3 to 7 months until 2026, but METR now calls measurements above 16 hours unreliable. See: Outlook 2027-2030
• OSWorld: A benchmark of operating a real desktop by screenshot, mouse and keyboard. In the OSWorld 2.0 paper (June 2026) the best model, Claude Opus 4.8 at max thinking, completed 20.6% of 108 long workflows under binary scoring (54.8% partial). Later leaderboard rows (Epoch's copy, secondary) show Claude Opus 5 at 31.4% binary. Vendor "OSWorld 2.1" partial-credit scores are a different version and run far higher. See: Agent frameworks and protocols
• pass@k: The share of tasks on which at least one of k attempts succeeds. Optimistic; it fits cases where a failed try is cheap and detectable, such as code with tests. See: Skills and implementation playbook
• pass^k: The share of tasks passed on every one of k trials, the measure that matters for customer-facing agents. At 75% success per trial, only about 42% of tasks pass three in a row. See: Skills and implementation playbook
• Randomised controlled trial (RCT) and field study: Studies that compare people working with and without AI, the strongest evidence of productivity effects. Gains are largest for less experienced staff and can turn negative outside AI's capabilities. See: Use cases and ROI
• SWE-bench: A benchmark of fixing real GitHub issues; SWE-bench Verified is the human-checked subset and SWE-bench Pro a harder variant. The official Verified board tops out at 79.2% and has not been updated since February 2026. Epoch's own Verified runs reach 83.5% (Claude Opus 4.7, April 2026), and no independent source supports the 95%+ claims. See: Top 10 models compared
• Word error rate (WER) and character error rate (CER): The share of words, or characters, that a speech-to-text or OCR system gets wrong against a reference text. CER suits scripts where word splits are ambiguous; report both per language and per level of code-mixing. See: Indian languages
• τ2-bench (tau-bench): A benchmark of multi-turn customer-service conversations in which an agent must use tools and follow policies. A proxy for support agents, but few current scores exist. See: Top 10 models compared
• Terminal-Bench: A benchmark of multi-step work in a command-line terminal, the closest public proxy for operations scripting and coding agents. The third-party tbench.ai board (each run uses an agent harness, usually the vendor's own CLI) put Claude Opus 5.5 first at 64.85% in October 2026. Vendor figures, at differing effort settings, can exceed it: Sonnet 5.5 claims 70.6% but scores 61.82% on the board. See: Top 10 models compared

Serving, hardware and compute

• Capex (capital expenditure): Spending on long-lived assets such as data centres and chips. Big-tech AI capex rose about 80% in a year, far ahead of disclosed AI revenue, which suggests model prices may be subsidised. See: Outlook 2027-2030
• Custom accelerators: AI chips built as alternatives to NVIDIA GPUs, such as AWS Trainium, Microsoft Maia, Cerebras, Groq's LPU and Huawei Ascend. Most can be used only as cloud services, not bought. See: Hardware
• Edge AI and NPU: Running small models on laptops, phones or on-site devices, often on a neural processing unit (NPU) rated in TOPS. Suits offline OCR and classification; measure speed on the device first. See: Hardware
• FP8 and FP4: Number formats that store each weight in 8 or 4 bits instead of 16 (FP16 or BF16), roughly halving or quartering memory. Variants include MXFP4 and NVFP4. See: Hardware
• GPU (graphics processing unit): The parallel processor that trains and serves most AI models, sold mainly by NVIDIA and AMD. Memory capacity decides which model fits; memory bandwidth caps how fast it generates. See: Hardware
• GPU-hour: The rental unit for cloud GPUs: one GPU for one hour. An H100 hour ranges from ₹91.80 (IndiaAI's ₹153 list rate less its maximum 40% subsidy) through Indian commercial rates to several times more on hyperscalers. See: India deep dive
• HBM (high-bandwidth memory): Stacked memory packaged beside an AI chip to supply the terabytes per second that large models need. Memory is the binding supply constraint of 2026, and prices roughly doubled. See: Hardware
• Hyperscaler and neocloud: Hyperscalers are the giant general clouds (AWS, Azure, Google Cloud); neoclouds are GPU-specialist providers such as CoreWeave or Lambda, which usually rent GPUs for less. See: Hardware
• Inference: Running a trained model to produce output, which is what token prices pay for; training is the separate, far costlier process of creating the model. See: Production tool stack
• KV cache: A model's per-conversation working memory: stored attention keys and values for tokens already processed. It grows with context length and concurrent users, so size GPU memory for it. See: Hardware
• llama.cpp and Ollama: Free local runtimes for running quantised open models on laptops, desktops and CPUs; llama.cpp uses GGUF files. Good for development and pilots, not multi-tenant production. See: Hardware
• LLM gateway and OpenRouter: A gateway puts one API key in front of many model providers, enforcing budgets, rate limits, fallbacks and logs; LiteLLM is the open-source default and OpenRouter a fee-charging hosted router. See: Production tool stack
• MLPerf: An industry benchmark suite for AI hardware. The set counts vendor submissions made under MLPerf rules as measured figures, unlike a maker's own spec or claim. See: Hardware
• NVIDIA data-centre GPUs (Hopper, Blackwell, Rubin): NVIDIA's generations: Hopper (H100, H200), Blackwell (B200, and B300 or Blackwell Ultra, the volume part in October 2026) and Vera Rubin, which NVIDIA reported in production with partner racks running at several clouds in August 2026 (secondary). See: Hardware
• NVLink and NVL72: NVLink is NVIDIA's high-speed GPU-to-GPU link; an NVL72 rack joins 72 GPUs into one system. Liquid cooling and power needs make it a facility decision. See: Hardware
• Power capacity (gigawatts) and AI factories: Data-centre scale is quoted in megawatts and gigawatts of power; "AI factory" is the vendor and government term for a data centre built for AI. Announced gigawatts often far exceed operating capacity. See: Middle East and Africa
• Prefill and decode: The two phases of generation: prefill reads the prompt (compute-bound; sets time to first token), decode writes the answer (bandwidth-bound). Agents resend long contexts, so prefill speed matters. See: Hardware
• Quantisation: Storing weights at lower precision, typically 4 or 8 bits, to cut memory and raise speed at some quality cost; a 4-bit model needs just over half a byte per parameter. See: Hardware
• Serving engine (vLLM, SGLang): Software that runs models on GPUs for many users at once, with batching and KV-cache management. vLLM is the default to evaluate and SGLang its rival; Hugging Face's TGI is archived. See: Production tool stack
• TPU (tensor processing unit): Google's own AI chip, which runs Gemini and is rented through Google Cloud. TPUs are now sold outside Google, with Anthropic as the anchor buyer. See: Google Gemini
• Unified memory: A shared CPU-GPU memory pool, as in NVIDIA's DGX Spark, Apple's Mac Studio or AMD's Strix Halo, that fits big MoE models on a desktop but reads long prompts slowly. See: Hardware

Data, RAG and memory

• Agentic search: The agent version of RAG: the model calls search tools repeatedly, refining queries until it has enough evidence. It costs more tokens than a single retrieval. See: AI automation explained
• Document parsing and OCR: Turning PDFs, scans and office files into clean, chunked text and tables, using optical character recognition (OCR) for images. Parsing often sets the unit cost of document automation; Indic-script OCR is untested. See: Production tool stack
• Embedding: A list of numbers that captures a text's meaning, so similar texts can be found by similarity search. Open-weight embedding models avoid per-token fees. See: Production tool stack
• GraphRAG: Retrieval over an entity-and-relation graph built from the documents, which helps with questions spanning several sources. Indexing is costly and recurs as documents change. See: Production tool stack
• Hybrid search and reranking: Hybrid search combines keyword matching with vector similarity; a reranker then re-scores the top results. The set's researchers judge a reranker usually a cheaper quality gain than a new embedding model. See: Production tool stack
• Memory layer: Stored facts about users and past sessions that an agent can recall later, via engines such as Mem0, Zep or Letta or vendors' built-in memory. Admin controls differ widely. See: Production tool stack
• Prompt caching: Reusing an already-processed prompt prefix at a steep discount; cached input costs a small fraction of the normal price. Agent loops resend the same prefix, so caching cuts their cost sharply. See: Costs and economics
• Retrieval-augmented generation (RAG): Fetching relevant passages from a document store and having the model answer from them, ideally with citations. The 2026 evidence favours combining retrieval with long context rather than choosing one. See: Production tool stack
• Vector database: A store that indexes embeddings for similarity search, from pgvector inside an existing Postgres to managed services such as Pinecone or Qdrant. pgvector is the lowest-risk start. See: Production tool stack

Security and risk

• Agent goal hijack: OWASP's top agent risk (ASI01): untrusted text in a web page, email or ticket redirects an agent to an attacker's goal. Separate untrusted content from instructions and block exfiltration paths. See: AI automation explained
• AI bill of materials (AIBOM) and AI inventory: An AIBOM lists an AI system's parts: models, data sources, licences, intended use and known weaknesses. An AI inventory is the organisation-wide register with one row per deployed component. See: Regulation and security
• AI security institutes (UK AISI, US CAISI): Government bodies that test frontier models: the UK AI Security Institute and NIST's Center for AI Standards and Innovation (CAISI), which NIST's site now calls the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI). Its evaluations are still published under the CAISI name. See: DeepSeek
• Bedrock India (geographic) inference profile: An AWS Bedrock profile with the "in." prefix that routes requests only to Mumbai and Hyderabad. It exists for Claude Opus 5, Sonnet 5, Haiku 4.5, Kimi K3 and GPT-5.6 Terra and Luna; AWS says to confirm each profile with GetInferenceProfile rather than rely on the prefix. See: Checklist
• CAISSI: The name NIST's site now uses for the centre previously called CAISI. Its evaluations are still published under the CAISI name. See: Regulation and security
• Copyright indemnity: A vendor's promise to defend customers against IP claims over outputs or training data. It typically excludes customer inputs and modifications, so it does not cover a whole solution. See: Regulation and security
• CSA AI Controls Matrix (AICM): The Cloud Security Alliance's catalogue of 243 AI control objectives, mapped to ISO/IEC 42001. A practical way to answer client security questionnaires without certification. See: Regulation and security
• Data processing agreement (DPA): The contract with a client that names sub-processors, models, regions, retention and ZDR status. A template ready before the first project saves negotiation time. See: Checklist
• Data residency: Where data is stored and processed, which clients and regulators may restrict. No vendor's own API processes requests in India: OpenAI offers India residency for stored data only, and in-region processing in the US, EU and UAE. AWS Bedrock does offer India-only routing (Mumbai and Hyderabad) for the previous Claude generation (Opus 5, Sonnet 5, Haiku 4.5), Kimi K3 and GPT-5.6 Terra and Luna; the newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud. Regional endpoints cost extra, for example 1.1x at xAI (US only) and Anthropic (US). See: Top 10 models compared
• Data Zone deployment: An Azure Foundry deployment type that keeps processing within a group of countries, such as an APAC Data Zone, rather than in one. It is narrower than a Global deployment (processing anywhere) but wider than a single geography. See: Regulation and security
• Excessive agency and least agency: Excessive agency (OWASP LLM06) is giving an AI system more tools, permissions or autonomy than its task needs. Least agency is the cure: narrow tools, read-only credentials, one identity per agent. See: Regulation and security
• Fallback model: A tested second model, ideally from another vendor, for when the first is withdrawn, throttled or changed; in June 2026 a US directive took two Anthropic models offline for 19 days. See: Claude (Anthropic)
• Global endpoint: A model endpoint that may route requests between regions, so the processing location cannot be fixed. From Bedrock's India regions the current generation (Claude 5.5, GPT-6) is available only this way, and Google processes Gemini 3.6 and later only in US and EU multi-regions. See: Indian languages
• Guardrails and guard models: Checks that screen inputs, outputs and actions for injection, personal data or unsafe content. Open guard models include Meta's Llama Guard; commercial guardrail start-ups are being bought by security firms. See: Production tool stack
• In-Region inference: Processing in the single named cloud region where the request lands, as opposed to a geographic profile or a global endpoint. AWS lists many open-weight models In-Region in Mumbai, such as gpt-oss and Qwen3. See: Indian languages
• ISO/IEC 42001: The certifiable international standard for an AI management system. A cloud provider's certificate does not pass to its customers, who must be certified in their own right. See: Regulation and security
• MITRE ATLAS: MITRE's knowledge base of attack techniques against AI systems, used to label red-team findings (results of deliberately attacking a system) and incidents. It now covers many agent-specific techniques. See: Regulation and security
• NIST AI RMF: The US National Institute of Standards and Technology's voluntary AI Risk Management Framework and its Generative AI Profile. The risk vocabulary US companies expect from AI vendors and teams. See: Regulation and security
• Observability and tracing: Recording each model call, tool call, cost and eval result so failures can be traced. Langfuse (MIT core) is the set's default example; Phoenix bars hosted service to third parties. See: Checklist
• OWASP Top 10 for Agentic Applications: OWASP's December 2025 list of the ten main agent risks, ASI01 Agent Goal Hijack to ASI10 Rogue Agents, built on the principle of least agency. The standard checklist for agent design reviews. See: Regulation and security
• OWASP Top 10 for LLM Applications: OWASP's 2025 list of risks for LLM apps, from LLM01 Prompt Injection to LLM10 Unbounded Consumption. The baseline application-security checklist. See: Regulation and security
• Prompt injection: Instructions hidden in content an AI reads, such as a web page, email or tool result, that override the user's intent; a jailbreak is the user's own bypass attempt. It remains unsolved. See: Regulation and security
• Sandbox: An isolated container or virtual machine in which an agent runs code, with limited network access and no keys inside. Approval classifiers such as auto modes do not replace it. See: Regulation and security
• Shadow AI: Staff using unapproved AI tools for work, often with company data. Surveys find it widespread; a written AI-use policy and approved business accounts are the usual response. See: Use cases and ROI
• SOC 2, ISO/IEC 27001 and HIPAA BAA: Security attestations and certifications enterprise buyers expect across the stack; a HIPAA business associate agreement (BAA) is required for US health data. Coverage varies by product, so check each one. See: Production tool stack
• Tool poisoning: An MCP attack in which a malicious or altered server hides instructions in its tool descriptions, which the model then follows. Pin server versions and review tool descriptions. See: Agent frameworks and protocols
• Zero data retention (ZDR): A contract option under which the vendor does not keep prompts and outputs after processing. It needs vendor approval, has carve-outs, and some models and managed services are excluded. At Anthropic, Batch, the Files API, code execution, the MCP connector, Agent Skills and Managed Agents are never ZDR-eligible, and the Fable and Mythos models require 30-day retention. See: Regulation and security

Regulation and compliance

These summaries are planning aids, not legal advice; the linked documents give the sources and caveats.

European Union

• AI literacy (Article 4): The EU AI Act duty to ensure that staff who use AI understand it well enough, in force since 2 February 2025 and enforced by national authorities from 2 August 2026. Internal training records suffice; no certificate is needed. See: Regulation and security
• AI Omnibus: The EU law, in force from 27 July 2026, that moved most AI Act high-risk duties to December 2027 and August 2028 while leaving transparency duties largely on schedule. See: Americas and Europe
• Article 50 transparency duties: EU AI Act rules, applicable from 2 August 2026, that people must be told when they are dealing with an AI system, and that AI-generated content and deepfakes be marked or labelled. 2 December 2026 is only a grace period for generative systems already on the market before 2 August. They cover nearly every chatbot or voice agent serving EU users. See: Americas and Europe
• EU AI Act: The EU's risk-based AI law: bans on some practices, duties for high-risk systems and general-purpose models, and transparency rules. Applicable since 2 August 2026, with exceptions, it follows work done for EU users, wherever it is built. The Omnibus adds two prohibitions from 2 December 2026: AI systems that generate non-consensual sexual deepfakes or child sexual abuse material. See: Americas and Europe
• GDPR Article 22: The EU right not to be subject to solely automated decisions with legal or similarly significant effects. Where exceptions apply, people keep the right to human review and to contest. See: Regulation and security
• General-purpose AI (GPAI) model: The EU AI Act term for broadly capable models such as frontier LLMs, whose providers face AI Office enforcement from 2 August 2026 and can sign a voluntary Code of Practice. See: Americas and Europe
• High-risk AI system: The EU AI Act category with the heaviest duties. Stand-alone uses such as hiring, education and credit scoring apply from 2 December 2027; AI inside regulated products from 2 August 2028. See: Regulation and security
• Provider and deployer: Under the EU AI Act, the provider develops an AI system and places it on the market; the deployer uses it under its own authority. A business can be either, and so can the vendor or partner that builds its system. See: Regulation and security

United States

• ADMT (automated decision-making technology): US state-law term for systems that make or substantially shape consequential decisions about people, such as hiring or lending. California's CCPA rules add ADMT duties from 1 January 2027. See: Regulation and security
• Colorado SB 26-189: Colorado's law, signed 14 May 2026, that repeals and re-enacts its AI Act as an ADMT law. From 1 January 2027 developers must document systems, and deployers must give notice, explain adverse decisions and offer human review. Enforcement is clouded by X.AI LLC v. Weiser: a federal court order of 27 April 2026 bars the Attorney General from enforcing the old AI Act, or replacing legislation enacted that session, until 14 days after a ruling on xAI's preliminary-injunction motion, and whether this reaches SB 26-189 is arguable. Until 1 January 2030 the AG must give a 60-day cure notice, and there is no private right of action. See: Americas and Europe
• EO 14365: The US executive order of 11 December 2025 that seeks to pre-empt state AI laws through litigation and funding pressure. It puts no direct duties on private firms, and state laws still apply. See: Americas and Europe
• Model-level export control: Government restriction of access to specific AI models rather than chips. A US directive of 12 June 2026 made Anthropic switch off two models for all customers until 1 July. See: Timeline 2025-2026

India

• 140 and 160 number series: Indian number series for commercial voice calls: 140 for promotional, transactional and service calls, and 160 for service and transactional calls only. Both run on DLT-managed platforms under TRAI's May 2024 Direction. See: Indian languages
• A2P (application-to-person) calls: A telecom category for calls initiated by software rather than a person. TRAI's 2026 draft amendment defines it, and a fully automated voice agent appears to fall inside it; the final text is unverified. See: Indian languages
• Consent manager: Under India's DPDP regime, a registered intermediary through which people give, review and withdraw consent across services. Consent-manager registration starts on 13 November 2026, the only DPDP tranche due that day. See: India deep dive
• Data fiduciary: Under the DPDP Act, the organisation that decides why and how personal data is processed; the data principal is the person, and processors act for fiduciaries. Significant Data Fiduciaries face extra duties under Rule 13: an annual impact assessment and audit, due diligence on algorithmic software, and possible localisation of specified data. See: India deep dive
• DLT (distributed ledger technology) registration: The operator-managed platform on which commercial calls and messages in India are registered. TRAI's May 2024 Direction moved 140-series voice calling onto it, so a business must finish its operator's registration before a campaign. See: Indian languages
• DPDP Act and Rules: India's Digital Personal Data Protection Act 2023 and Rules of November 2025. Core duties, including purpose-specific consent and 72-hour breach reports, apply from 13 May 2027, with penalties up to ₹250 crore. See: India deep dive
• India AI Governance Guidelines: MeitY's principles-first guidelines of November 2025, which prefer existing laws and voluntary compliance to a new AI statute. See: India deep dive
• IT Rules on synthetic content: India's 2026 amendment to the IT intermediary rules, in force since 20 February 2026, requiring labels on AI-generated audio, images and video ("synthetically generated information") and much faster takedowns: 3 hours (down from 36) for court or government orders, and 2 hours for complaints about non-consensual intimate imagery (secondary). See: India deep dive
• RBI FREE-AI: The Reserve Bank of India's committee report on responsible and ethical AI in finance (13 August 2025: 7 Sutras and 26 recommendations under six pillars). It is a report, not a direction, and no RBI AI direction has been found. Banks and NBFCs are expected to pass its duties to tech vendors by contract (recommendation details from secondary sources). See: India deep dive
• Significant Data Fiduciary: A data fiduciary under the DPDP Act that carries extra duties. Under Rule 13 it must run an annual impact assessment and audit, check its algorithmic software, and may be barred from transferring specified data categories abroad. See: India deep dive
• TRAI: The Telecom Regulatory Authority of India, whose rules govern commercial calls and messages. The only primary TRAI text read on voice calling is the Direction of 4 May 2024; no explicit AI-voice rule has been confirmed. See: Indian languages

Other jurisdictions

• China generative-AI filing and labelling: China requires public-facing generative-AI services to file with the Cyberspace Administration before launch, and AI-generated content to carry visible and metadata labels. See: China and Asia-Pacific
• Singapore agentic AI framework: IMDA's voluntary Model AI Governance Framework for Agentic AI (January 2026), which holds organisations accountable for the agents they build or buy. See: Regulation and security
• South Korea AI Basic Act: Korea's AI law, in force since 22 January 2026, which requires telling users about AI use and makes foreign firms above revenue or user thresholds appoint a local representative. Fines (up to KRW 30 million) are suspended for at least a year, so they start around January 2027 (secondary); corrective orders are not suspended. See: China and Asia-Pacific
• Vietnam Law on AI: Southeast Asia's only comprehensive AI statute in the research, in force since 1 March 2026 and covering foreign firms operating in Vietnam. See: China and Asia-Pacific

Business and pricing

• Agentforce ARR: Salesforce's reported annual recurring revenue from its Agentforce agents, above $1.5 billion in its second quarter of fiscal 2027 (to July 2026, reported 26 August 2026), partly after a wider definition. See: Companies
• ARR (annual recurring revenue): Contracted subscription revenue at a point in time, expressed per year. Never add it to run-rate, bookings or valuations, which measure different things. See: Companies
• Assessment, sprint and retainer ladder: The common way implementation partners price work: a fixed-fee assessment, a fixed-scope build sprint, then a monthly retainer, with model usage passed through. For Indian SME budgets, smaller fixed-scope packages are usually a better fit than the full ladder. See: Skills and implementation playbook
• Cost per task: The modelled cost of one unit of work from tokens and list prices: fractions of a cent for a text step, dollars for a long agent run, plus a margin for retries and evals. See: Costs and economics
• Credits (consumption pricing): Prepaid units that meter agent steps or messages, as in Microsoft Copilot Studio or Salesforce's Flex Credits. The bill grows with every step an agent takes. See: Costs and economics
• Custom model licence: A model licence with terms beyond Apache or MIT, such as revenue thresholds for model-as-a-service (MaaS) businesses, user caps or attribution duties. Keep a licence register per deployment. See: Other leading models
• Digital worker subscription: A monthly service in which a vendor supplies an AI "worker" for a fixed scope, such as invoices processed, with a cap, named human supervision and a KPI report. Customers should never be led to think the worker is a person. See: Skills and implementation playbook
• Discount and speed tiers (batch, flex, off-peak, fast): API price modifiers: batch and flex processing cost about half, DeepSeek halves prices off-peak, and fast modes cost double or more: OpenAI's Fast (renamed from Priority on 30 July 2026) is 2x and Ultrafast 6x; Anthropic's Fast mode is 2x on its own API only. Anthropic's Priority Tier is no longer sold for current models. Put overnight and bulk work on the cheap tiers. See: Costs and economics
• Forward-deployed engineer (FDE): An engineer who embeds with a client, scopes proofs of concept and production deployments, and codes "side-by-side" on the client's systems. The core role in an automation team, in-house or at a partner. See: Skills and implementation playbook
• Outcome-based pricing: Charging per result, such as a resolved ticket, rather than per seat or hour. Support platforms charge about $1 to $2 per AI resolution, so define exactly what counts as an outcome. See: Costs and economics
• Partner programme: A model or cloud vendor's scheme for implementation and consulting firms, offering credits, training, listings and referrals. Some vendors also sell services directly, so a buyer can often choose between a listed partner and the vendor itself. See: Skills and implementation playbook
• Permissive licence: An open-source licence such as Apache 2.0, MIT or BSD that allows commercial use, modification and hosting with few conditions. The safe default when a model or tool is shipped to others or hosted for them. See: Production tool stack
• Per-seat pricing: A fixed monthly price per named user, still the default for assistants and copilots; seats cluster at $20 to $30 a month. See: Costs and economics
• Pilot and go/stop gates: A time-boxed trial on real work, measured against a baseline taken first, with agreed KPIs that decide whether to continue, change course or stop. Most pilots never reach measured production. See: Skills and implementation playbook
• Run-rate: The latest period's revenue multiplied out to a year. It moves fast when usage-billed revenue rises or falls, so it is not comparable with ARR. See: Companies
• Self-hosting break-even: The volume at which running an open model on owned or rented GPUs beats paying an API. It needs sustained high utilisation, so data control is the usual reason to self-host. See: Costs and economics
• Source-available and copyleft licences: Source-available licences (such as n8n's) publish code but restrict use; copyleft licences (such as AGPL) carry duties when software is hosted for others. Review both before distributing such software or hosting it for others. See: Production tool stack
• Token pricing: Model APIs charge per million tokens, with separate rates for input, cached input and output. Some charge more once a prompt passes a size threshold, which can multiply long-agent costs. See: Costs and economics

Indian context

• 22 scheduled languages: The 22 languages listed in India's Constitution. Vendor support ranges from one of them (Llama, Claude's interface) to all 22, and Hindi is the only one every frontier vendor lists. See: Indian languages
• AI Impact Summit: The global AI summit India hosted in New Delhi in February 2026. It produced the non-binding New Delhi Declaration and an expected "over $200 billion" of investment that is not a commitment. See: India deep dive
• Bhashini: The Indian government's language-translation platform under MeitY, which underpins services such as ONDC's Saarthi app. No usage figures for it were found. See: India deep dive
• CERT-In: The Indian Computer Emergency Response Team, the national cyber-security agency. Its binding 2022 Directions require reports of listed cyber incidents within 6 hours, including attacks on AI and ML systems, and logs kept for 180 days in India. Its bill-of-materials guidelines (v2.0, July 2025), which include an AI bill of materials, are non-binding guidance likely to shape procurement. See: Regulation and security
• Code-mixing, Hinglish and romanised text: Writing Hindi, Tamil or Bengali in Latin script, often mixed with English (Hinglish), as WhatsApp and SMS users commonly do. Accuracy falls as mixing rises, and retrieval on such text is untested. See: Indian languages
• DPIIT recognition: Recognition by India's industry-promotion department that, together with ₹50 lakh in revenue or ₹100 lakh in funding, qualifies start-ups and MSMEs for IndiaAI compute subsidies. Services firms are not a named category. See: Indian languages
• GCC (global capability centre): An in-house Indian centre run by a foreign company for technology, operations or AI work. GCCs compete with service firms for the same talent. See: India deep dive
• GST (goods and services tax): India's indirect tax, 18% on the Meta and GPU charges the set costs. Many Indian price lists exclude it, so state whether each rupee figure includes GST. See: Indian languages
• IndiaAI Compute Portal: The IndiaAI Mission's pool of subsidised GPUs from empanelled private providers, 45,081 GPUs across 14 bidder rows on an undated portal list. It is allocated by application to researchers, DPIIT-recognised start-ups and MSMEs, students and government entities. The portal lists an H100 at ₹153 per GPU-hour before a subsidy of up to 40% (GST extra), and requests under 5,000 GPU-hours are auto-approved. Listed capacity is far larger than the GPU-hours sanctioned so far. See: India deep dive
• IndiaAI Mission: India's national AI programme under MeitY, covering subsidised compute, foundation-model grants, the AIKosh dataset portal and skills training. See: India deep dive
• Indian IT services and BPO: India's outsourcing industry for software, operations and business processes, the front line of AI's labour effects. AI revenue is growing (TCS reported annualised AI revenue above $3 billion in its quarter to September 2026), while headcount trends differ by firm and quarter (TCS's rose by 4,258 to 598,056 that quarter; secondary). See: India deep dive
• Indian sovereign model builders: Domestic teams building foundation models with IndiaAI support, such as Sarvam, BharatGen, Soket AI, Gnani.ai and Gan.ai, alongside Ola's Krutrim. Most have modest uptake so far. See: India deep dive
• Indic-language AI: Models and tools for India's languages, scripts and code-mixed text in speech, translation, search and OCR. Vendor claims are largely untested, so pilot on the client's own data. See: India deep dive
• IndicTrans2: AI4Bharat's MIT-licensed translation models, the first family to cover all 22 scheduled languages, and still the open reference for Indian-language translation. See: Indian languages
• Lakh, crore and the ₹96 convention: One lakh is 100,000 and one crore is 10 million. The set converts at ₹96 per US dollar; Indian GPU rental quotes usually exclude 18% GST. See: India deep dive
• MeitY: India's Ministry of Electronics and Information Technology, which runs the IndiaAI Mission and issues the DPDP Rules, the IT Rules and the AI Governance Guidelines. See: India deep dive
• MSME: Micro, small and medium enterprises, India's legal category for smaller businesses. Their AI use is widely claimed but poorly measured. See: India deep dive
• NASSCOM: India's technology-industry association, whose reports supply most figures on the IT sector, GCCs and AI talent. See: India deep dive
• Sarvam AI: The IndiaAI Mission's first foundation-model grantee, and the only domestic option with both open weights and a commercial API priced in rupees. See: India deep dive
• WhatsApp Business Platform: Meta's business-messaging API, central to Indian order and support automations. Eligible accounts billed with India as the Sold-To country must move to INR billing by 31 December 2026, or Meta stops delivering their messages from 1 January 2027. See: India deep dive
• WhatsApp message categories: Meta bills marketing, utility, authentication and service messages differently. Vendors quote ₹0.8631 for marketing and ₹0.1150 for utility (excluding GST, vendor-reported), and say service messages beyond 1,000 a number a month are charged from 1 October 2026; Meta has not confirmed it. See: Indian languages

Quick index

Every term from A to Z, with its group and the document that covers it in most depth.

Sources

Every definition draws on the documents of this set, which carry the full citations. The quoted phrases and the standards, protocols and laws defined above come from these primary pages:

• Anthropic, Building effective agents (agent and workflow definitions; workflow patterns)
• Anthropic, Demystifying evals for AI agents (eval suites; pass^k)
• AWS, Operationalizing agentic AI (AgentOps)
• OpenAI, forward-deployed engineer posting (FDE role)
• Claude Agent SDK for Python, README and LICENSE (MIT file; Commercial Terms govern use)
• Claude Code repository, LICENSE.md (proprietary)
• OpenAI, deprecations (fine-tuning dates)
• Artificial Analysis, Grok 4.7 (Intelligence Index version)
• n8n licence (Sustainable Use License)
• MCP blog, 2026-07-28 specification
• A2A specification
• AP2 documentation
• Agentic Commerce Protocol repository (beta status; maintainers)
• OpenAI, GPT-6 Astra system card ("Critical" cybersecurity rating)
• Llama 4 Community License ("Built with Llama")
• OpenTelemetry GenAI semantic conventions
• OpenTelemetry GenAI span conventions repository ("Development" status)
• OWASP Top 10 for Agentic Applications 2026
• OWASP Top 10 for LLM Applications
• MITRE ATLAS changelog
• NIST AI RMF Generative AI Profile
• CSA AI Controls Matrix
• AWS, ISO/IEC 42001 FAQ (certification does not pass to customers)
• Simon Willison, OpenAI CISO on Atlas (prompt injection as an unsolved problem)
• METR, time horizons
• Epoch AI, Epoch Capabilities Index
• European Commission, AI Act regulatory framework
• European Commission, AI literacy Q&A
• Federal Register, EO 14365
• Colorado General Assembly, SB26-189
• Gazette of India, DPDP commencement notification G.S.R. 843(E)
• MeitY, DPDP Rules 2025 (G.S.R. 846(E))
• Hogan Lovells, India's DPDP Act brought into force (consent grounds; Significant Data Fiduciary duties)
• Hogan Lovells, India introduces mandatory labelling for AI ("synthetically generated information")
• Fortune, India AI Impact Summit closes ("over $200 billion")
• Linux Foundation, Agentic AI Foundation announcement and AGENTS.md (founding projects; members)
• Official MCP Registry API (server count, 12 October 2026)
• Anthropic, extended thinking and OpenAI, reasoning (reasoning tokens billed as output)
• OSWorld 2.0 paper, Epoch AI benchmark data and tbench.ai (OSWorld, SWE-bench Verified and Terminal-Bench figures)
• IndiaAI price calculator, eligibility criteria and empanelled GPU list (GPU rates; allocation)
• NIST, CAISSI
• AWS Bedrock, regional availability (India profiles and In-Region models)
• Anthropic, API and data retention (ZDR carve-outs)
• EU AI Act Service Desk, implementation timeline (Omnibus prohibitions)
• CourtListener, X.AI LLC v. Weiser (Colorado enforcement)
• Gazette of India, DPDP Rules (Rule 13 text)
• iPleaders, IT Rules 2026 (takedown times; secondary)
• RBI, FREE-AI committee press release
• CERT-In, Directions of 28 April 2022
• OpenAI pricing and Anthropic service tiers (fast modes)
• TRAI, Direction of 4 May 2024 (DLT; 140 and 160 series)
• Indi-RomCoM, arXiv 2606.30790 and MILU, arXiv 2411.02538 (preprints; Indic benchmarks)