Companies Doing AI Automation

Key takeaways

All facts in this document are as of 12 October 2026; deal terms, prices and results were re-checked against primary sources (SEC filings, vendor pages) on 12 October 2026 where readable, and facts not re-read keep their original read dates. Figures marked secondary, tracker, single-source or headline-only were not confirmed against a primary document. The open-questions section near the end lists the main ones.

• Most of the value sits with the model labs. Anthropic's own pages report a run-rate of $14B (February 2026), above $30B (April 2026) and above $47B (May 2026), and a $65B round at $965B (28 May 2026). OpenAI closed $122B at $852B on 31 March 2026 (secondary: several outlets agree, one aggregator cites $830B; company post not read). Everyone else in this map builds on these platforms, resells them or competes with them.
• Enterprise suites now report agents as recurring revenue. Salesforce's Agentforce ARR went from above $0.5B (October 2025) and $0.8B (January 2026) to above $1.5B (July 2026), partly through a wider definition. Microsoft's AI business passed a $37B run-rate (quarter to March 2026). UiPath, the RPA incumbent, grew ARR 12% to $1.901B.
• Agent startups trade at roughly 20x to 130x revenue (calculated here from mixed revenue measures; see the multiples table). Coding leads (SpaceX bought Cursor for about $60B; Cognition raised at $48B on almost $900M run-rate), then customer service (Sierra above $15B, Decagon $4.5B), legal (Harvey, Legora) and voice (ElevenLabs).
• Incumbents are buying the integration and front-end layers. Salesforce bought Informatica (about $8B) and Fin (about $3.6B), Workday agreed to buy Sana (about $1.1B) and Pipedream, ServiceNow closed Moveworks ($2.85B), and SAP invested in n8n at $5.2B. Larger vendors absorbed observability and AI-security tools.
• Indian IT reports AI in incompatible units. TCS reports a run-rate (above $3B, more than 10% of revenue, in Q2 FY27 to September 2026; about $2.6B in Q1), Infosys a revenue share (figure not confirmed), HCLTech a growth rate (+62%), Wipro nothing, and Accenture bookings (Indian figures via secondary reports of Q1 and Q2 FY27 results). The top five Indian IT firms shed a net 7,389 staff in FY26 (secondary).
• Market forecasts measure different things. Gartner's $201.9B agentic-AI spending figure for 2026 (secondary write-up) counts agent features built into enterprise software. Gartner also expects more than 40% of agentic AI projects to be cancelled by end-2027 (headline of a June 2025 Gartner press release; body not read).
• Buyers pay for the platform, not the model, and must plan for vendor churn. Outcome prices of $0.99 to $2.00 per AI resolution are about 20x or more the model cost (about $0.003 to $0.05). Shutdowns, model retirements and acquisitions make portability and exit terms part of every purchase. Indian SMEs face fixed deadlines: DPDP on 13 November 2026 and 13 May 2027, and WhatsApp INR billing by 31 December 2026.

Five units of scale that do not add up

Companies in this map report scale in five units. ARR (annual recurring revenue) is contracted subscription revenue at a point in time; Salesforce, UiPath, Sierra and ElevenLabs use it, with company-specific definitions (UiPath's release expands it as "annualized renewal run-rate"). Run-rate is the latest period's revenue multiplied out to a year; Anthropic, Cognition and TCS use it, and it moves quickly when usage-billed revenue rises or falls. Bookings are the value of contracts signed in a period and are recognised over several years; Accenture and Wipro report them. A private valuation is the price of the latest round or employee tender, not a measure of revenue. Usage metrics (monthly active users, tokens per minute, Salesforce's "agentic work units") show adoption but not money. This document never adds figures across these units. The one table that sets them side by side (the valuation multiples) labels each figure's unit and gives orders of magnitude only. The Glossary defines each term.

Each figure carries a source label. "Primary" means a company release, regulatory filing or the company's own page, read by the researchers. "Secondary" means Wikipedia, a news outlet or an aggregator. "Tracker" means a third-party estimate such as Sacra or Latka. "Single-source" means only one outlet reports the figure, and "headline-only" means researchers saw only a news headline. Where notes conflict, this document prefers the primary source and says why. The wider market picture is in Landscape, and the dated sequence of events is in Timeline.

Frontier labs sell the platform everyone else automates on

Anthropic has the best-documented revenue climb of any lab. With its $30B Series G at $380B post-money (12 February 2026), it reported a $14B run-rate, plus a Claude Code run-rate above $2.5B, more than half of it from enterprises (Anthropic Series G). By 6 April run-rate was above $30B, up from about $9B at the end of 2025, and more than 1,000 business customers were spending over $1M a year (Anthropic, Google and Broadcom compute). The $65B Series H at $965B post-money (28 May 2026), led by Altimeter, Dragoneer, Greenoaks and Sequoia, is on Anthropic's own page, which also says run-rate revenue "crossed $47 billion earlier this month" (re-read 11 October 2026; Anthropic Series H). A confidential draft S-1 followed on 1 June (Anthropic S-1 notice). The reported run-rate above $65B in July rests on news headlines only (HN listing of CNBC headline).

Anthropic also illustrates platform risk. A US export-control directive forced it to switch Fable 5 and Mythos 5 off for every customer, Indian ones included, from 12 June to 1 July 2026 (Anthropic access statement). The BBC reported that the US Department of Defense had stopped using its products (BBC). Anthropic opened a Bengaluru office on 16 February 2026 with partners including Air India, CRED, Razorpay and Cognizant, which it said was "deploying Claude to 350,000 employees" (Anthropic Bengaluru). Details are in Claude.

OpenAI has the largest user base and the least primary revenue data. Several outlets report that OpenAI closed a $122B round at an $852B post-money valuation on 31 March 2026 (one aggregator cites $830B; the company post was not read) (Bloomberg Law), and Wikipedia gives 2025 revenue of $13.1B (Wikipedia: OpenAI). A $40B run-rate in July 2026, an enterprise share above 40% and 9 million paying business users all come from secondary summaries of investor material (Shopifreaks). ChatGPT had about 900 million weekly users in February 2026 (Wikipedia: ChatGPT).

OpenAI's builder stack has four layers: the managed Agents API (a Codex harness run by OpenAI), the Agents SDK, the low-level Responses API and the ChatKit embed (OpenAI agents guide). Agent Builder shuts on 30 November 2026 (OpenAI Agent Builder). Two moves change who delivers OpenAI deployments, which matters when you choose an implementation partner. The first is DeployCo, a majority-owned deployment and integration company launched in May 2026 with more than $4B from a TPG-led group (secondary, via Wikipedia). The second is the February 2026 Tata deal, under which TCS's HyperVault supplies 100 MW of data-centre capacity, with potential to reach 1 GW, and Tata rolls ChatGPT Enterprise out across its workforce. The same article puts India's weekly ChatGPT users above 100 million (TechCrunch on Tata deal). Details are in OpenAI.

Google reports AI scale in its SEC filings. In Q2 2026 Google Cloud revenue rose 82% to $24.8B, with operating income of $8.8B. Gemini models processed 22 billion API tokens per minute, the Gemini app had 950 million monthly users, and "nearly 90% of the Fortune 100" used Gemini Enterprise (Alphabet Q2 2026 release). Google Cloud's share of the revenue backlog was $513.9B (Alphabet 10-Q). The production model, Gemini 3.8 Flash, doubles in price on 1 January 2027, from $0.75/$3.75 to $1.50/$7.50, with cached input rising from $0.075 to $0.15; Gemini 3.6 and 3.7 Flash rise to the same $1.50/$7.50 (Google generative AI pricing). Details are in Gemini.

SpaceXAI (formerly xAI) is the only lab with segment figures in a securities filing. It reported 2025 revenue of $3,201M against an operating loss of $6,355M, Q1 2026 revenue of $818M, and 117 million Grok-feature users at 31 March 2026 (SpaceX S-1/A). The merger closed on 2 February 2026, all stock, at a fixed 0.1433 SpaceX shares per xAI share (pre-2026 split), with a $75.46-per-share cash election for eligible service providers; SpaceX's prospectus states no dollar value (SpaceX 424B4). Wikipedia gives $250B for xAI (another Wikipedia article $125B) and a combined $1.25T (Wikipedia: SpaceXAI); none of these figures is in the filings read. Details are in Grok.

Chinese and other labs compete on price and open weights. In the week beginning 28 September 2026, DeepSeek had 21.8% of OpenRouter text requests, against 21.5% for Google, 18.2% for OpenAI and 2.5% for Anthropic. The four Chinese authors listed together had at least 37%. These shares count routed requests, not spend (OpenRouter rankings). Zhipu (listed January 2026, worth about $62B by August) and MiniMax (2025 revenue $79.0M, operating loss $1.87B) both trade in Hong Kong (Wikipedia: Zhipu AI; Wikipedia: MiniMax). Mistral raised €3B at above €21B on 8 September 2026 (Mistral news). India's Sarvam AI raised $234M at $1.5B in June 2026, with HCLTech investing $150M (Wikipedia: Sarvam AI). See DeepSeek and Other models.

Two inferences follow from the table. First, lab valuations and losses are large relative to revenue: Anthropic's February valuation was about 27x its run-rate and its May valuation about 21x, and SpaceXAI's 2025 operating loss was about twice its revenue. For a business choosing a vendor, the risk that the vendor changes course is as real as the risk that a model underperforms. Second, the labs are moving down the stack into services and up it into finished agents (OpenAI's DeployCo and Dots, Anthropic's Managed Agents). A third-party agent built on a lab's models may end up competing with the lab's own product, so buyers should ask how it differs from what the lab now sells directly.

Enterprise suites now report agents as ARR and price them per outcome

Salesforce gives the most complete agent-revenue series of any incumbent, from its own releases. The releases, re-read on 11 October 2026, correct one widely repeated figure and clarify another. The "$2.9B" figure circulating on blogs is the combined Agentforce and Data 360 ARR, which includes $1.1B of Informatica Cloud ARR; it is not Agentforce alone. And the Q3 FY26 release gives two deal counts that are easily confused: "over 18,500 Agentforce deals since launch" and "over 9,500 paid deals".

The definition change in Q2 FY27 means part of the jump from $1.2B to $1.5B is definitional, and the like-for-like growth rate cannot be calculated from the releases. A secondary summary puts an earlier point at about $100M of Agentforce ARR and about 4,000 paid customers in Q1 FY26 (No Jitter). Salesforce sells Agentforce through Flex Credits, which its worked examples price at about $0.005 per credit (Agentforce pricing). On 10 September 2026 it closed the roughly $3.6B purchase of Fin, formerly Intercom. Salesforce says Fin serves more than 30,000 companies and resolves an average of 76% of support volume end to end (Salesforce signs Fin agreement; Salesforce completes Fin). Fin's own homepage gives a lower count: "Fin customers average a 76% resolution rate, across 12,000+ businesses" (vendor-reported; resolution is not defined; the two counts may measure different things) (fin.ai).

Outcome pricing is now published, which gives buyers reference prices. Fin charges $0.99 per outcome, with a $49 monthly minimum that includes 50 outcomes (re-read 11 October 2026; Fin pricing). Zendesk includes AI agents in every plan and charges $1.50 per automated resolution on commitment or $2.00 pay-as-you-go, per the researchers' earlier read and secondary sources (the US page now redirects by region) (Zendesk pricing). On 11 October 2026 Zendesk's India page confirmed US$2.00 pay-as-you-go but showed the committed rate as £1.30 (Zendesk India pricing). Zendesk's German page (12 Oct 2026) shows €1.50 committed and €2.00 pay-as-you-go (Zendesk pricing, DE). Microsoft prices seats and credits rather than outcomes. Secondary sources put the Microsoft 365 Copilot Business add-on at $18 per user per month (promotional, annual; list $21) or $25.20 month-to-month, with conflicting promotion dates (Office Watch; microsoft.com did not load on 12 October 2026). Copilot Studio is sold as tenant-wide packs of 25,000 Copilot Credits at $200 a pack a month (about $0.008 per credit, calculated here), with a pay-as-you-go meter also available (Microsoft Learn). Under credit pricing, the bill depends on how many steps an agent takes, which is why the Costs page models agent loops rather than single calls.

The main point for buyers: incumbents with distribution (Salesforce, Microsoft, ServiceNow, Workday, SAP) are turning agents into line items on existing contracts. Most businesses will buy their first agents from a vendor they already pay, and the work then lies in configuring, integrating and governing them.

Workflow and iPaaS vendors gain value, but licences limit resale

n8n is the best-sourced company in this category. Its press page states $240M raised in total and a $5.2B valuation from SAP's strategic investment in May 2026, with n8n to be embedded in SAP Joule Studio. That valuation doubled the $2.5B of the October 2025 Series C, which raised $180M led by Accel (n8n press). The company's $240M figure supersedes tracker estimates of $193.5M to $313.5M. A press item dates the SAP deal to 12 May 2026 (Bolsamania). For buyers, the licence matters more than the valuation. n8n's Sustainable Use License permits use "only for your own internal business purposes or for non-commercial or personal use" (n8n licence). The licence also allows distribution or provision to others only "free of charge for non-commercial purposes", so a provider that runs n8n as a paid managed service for third parties needs a commercial agreement with n8n (an inference from the licence text, read 12 October 2026; no n8n FAQ was found) (n8n LICENSE.md). If a provider proposes to host n8n for you, ask which agreement covers it. On annual billing n8n lists Starter €20 a month (2,500 executions), Pro €50 and Business €667 (40,000 executions; "for companies with < 100 employees"; the page's hosting label is ambiguous) (n8n pricing).

Zapier has products but no verified 2026 revenue. Zapier MCP connects more than 9,000 apps (Zapier MCP), and the press page lists an Agents line (Zapier press). Zapier also built AutomationBench, which labs now quote; Anthropic reports 40.0% for Opus 5.5, a vendor-reported figure (Opus 5.5 page). Its last disclosed valuation is $5B, from a January 2021 secondary sale (Wikipedia: Zapier). Sacra's base case of $800M+ ARR by 2026 is a tracker forecast (Sacra on Zapier).

The claim that Make is owned by Celonis is not supported: the Celonis newsroom shows an Ikigai Labs deal but nothing about Make (Celonis news). Workato lists Agent Studio and Enterprise MCP (Workato press), but its $5.7B Series E valuation rests on a single, undated secondary source (Wikipedia: Workato).

Two things stand out in this category. Consolidation is real: Pipedream and Flowise went to Workday, n8n gained SAP as investor and channel, and Flowise's code was frozen within a year of its acquisition. Licences also differ in ways that matter when someone else runs the platform for you. A provider that runs one automation platform for many customers needs MIT or Apache cores (Activepieces, Temporal, Windmill's open parts) or a commercial agreement, not a source-available licence written for internal use. Running a tool yourself for your own internal business purposes is a different case, which n8n's licence permits. The Frameworks and Tool stack pages compare these tools in more depth.

Agent startups: coding and customer service take most of the capital

Agent startups have the highest valuations relative to revenue, and the evidence for them is the most uneven. Checking the widely repeated claims changed several of them. The SpaceX purchase of Cursor is confirmed and closed: Bloomberg Law reported the $60B deal "became effective on Aug. 14" (Bloomberg Law), and SpaceX's filings confirm it: the option agreement (April 2026, in the S-1) set an implied equity value of $60.0B, with a $1.5B termination fee and an $8.5B deferred services fee under specified conditions; the merger agreement followed on 16 Jun 2026, and SpaceX's 8-K reports the merger effective 14 Aug 2026, with 389,289,254 Class A shares issued to Cursor holders (SpaceX 8-K). Cognition's $48B valuation is confirmed by the company's own Series E post (8 September 2026, re-read 11 October 2026; Cognition Series E). Cursor's best-sourced revenue figure is $3B ARR in May 2026, from Wikipedia citing Bloomberg (Wikipedia: Anysphere). The "$4B+" figure rests on one anonymous source (Dealroom) and is not used here.

Customer service

Coding and software delivery

Knowledge, legal and go-to-market

General-purpose, voice and agent infrastructure

The multiples below are calculations from figures reported above, not company ratios. The revenue figures are in different units (ARR, run-rate, calendar-year revenue, a tracker estimate) and from different dates, so the multiples are orders of magnitude, not a ranking. They show what investors are paying for, and how far revenue must grow to justify it.

Broadly, the multiples fall as revenue rises, though not strictly: Legora sits below Sierra and Harvey despite smaller revenue, and Cursor's acquisition multiple sits below both Anthropic round multiples. Investors pay most for small, early revenue bases, which leaves the smaller vendors most exposed if growth slows. Support and legal startups sit at the top of the range. The Use cases page examines whether deployments justify that pricing.

AI services firms: bookings at the top, published price lists at the bottom

At the top of the services market, AI is reported inside general bookings rather than as its own line. Accenture's Q4 FY26 (quarter to 31 August 2026) brought new bookings of $22.2B, of which managed services were $12.8B (book-to-bill 1.4). Full-year FY26 bookings were $84.5B and quarterly revenue was $18.7B (re-read 11 October 2026; Accenture 8-K exhibit). An earnings-call summary says more than 400 clients began their first advanced-AI work with Accenture in FY26, and gives FY27 guidance of 3% to 6% growth in local currency, of which 2% to 2.5% is inorganic (Motley Fool call transcript). The release has no FY26 generative-AI bookings line, and the research found none elsewhere. The same summaries mention about 110,000 AI and data professionals; the release gives no AI workforce count (re-read 12 October 2026). On 18 September 2026 Anthropic announced a partnership with Accenture on embedded evaluation (newsroom title only; Anthropic news).

The labs are also creating their own services capacity and channels. OpenAI's DeployCo is the clearest case. The lab partnerships with TCS, HCLTech (named as an OpenAI partner), Cognizant (Claude for 350,000 employees) and Accenture show the large integrators becoming resale and deployment channels for the labs. McKinsey/QuantumBlack, Deloitte, Capgemini (including WNS), IBM Consulting and Tech Mahindra yielded no usable AI figures in this research.

Below the large firms, US boutiques publish three-step pricing that buyers can use as a benchmark. Main & Machine, whose page was read on 11 October 2026, charges $3,500 to $8,500 for a two-to-four-week AI readiness audit and $18,000 to $60,000 for a 4-to-12-week implementation sprint quoted fixed in writing. Managed services start at $1,500 a month with no lock-in, and a full back-office build starts at $95,000. Model usage is budgeted separately (Main & Machine pricing). The remaining boutique items below come from an earlier background scan and were not re-verified.

Indian firms rarely publish package prices, and US boutiques usually do. The published hourly gap between India and the US is roughly 4x to 8x in these directory sources, which is a comparison of listed rates, not a survey of transactions. For a buyer these figures are reference points for comparing quotes, not measures of quality or results; the Playbook covers delivery method.

Indian IT reports AI in four metrics that cannot be compared directly

The large Indian IT firms now disclose AI revenue, but each uses a different metric. Their Q1 FY27 figures (quarter to 30 June 2026) therefore need careful reading before any comparison. TCS's Q2 FY27 figures (quarter to 30 September 2026) are added below; Infosys, HCLTech, Wipro and LTIMindtree Q2 FY27 figures were not checked.

Three cautions apply. First, a run-rate (TCS), a share of revenue (Infosys), a growth rate (HCLTech) and bookings (Accenture) are different quantities, so the firms cannot be ranked on "AI revenue" from these disclosures. Second, the TCS figures are internally consistent. Q1 FY27 revenue of ₹72,275 crore, at ₹96 per US dollar (close to the June 2026 average of about ₹95 (X-Rates monthly average)), is about $7.5B a quarter, or about $30B a year, and 8.5% of that is about $2.6B. Q2 revenue of ₹73,188 crore gives about $30.5B a year, so "more than 10%" fits an AI run-rate just above $3B. These are cross-check calculations, not company reconciliations. Third, AI growth is not yet lifting the overall business: Infosys trimmed its guidance in the same quarter its AI revenue climbed (TechMarketView headline), and Wipro guided for a sequential decline.

Headcount points the same way, on secondary reports. The top five Indian IT firms (TCS, Infosys, Wipro, HCLTech, Tech Mahindra) shed a net 7,389 employees in FY26 after adding 12,718 the year before (Forkast). IANS reports a fall of more than 7,000 (IANS). The TCS figures conflict: one source cites about 12,000 layoffs, while another reports a net reduction of more than 23,000 to 584,519 in FY26, followed by 9,279 net additions in Q1 FY27 (Entrepreneur India) and 4,258 in Q2 FY27, to 598,056 at 30 Sep 2026 (Q2 FY27 results, secondary) (Outlook Business). Sources give both AI efficiency and weak demand as causes. Meanwhile the large firms are becoming channels for the labs (see the services section) and chasing enterprise megadeals, and HCLTech is Sarvam's lead strategic investor. The India page covers the country context.

Deals of 2025-2026: incumbents buy front ends, labs buy developer tooling

The deal record shows a clear pattern. Platform incumbents bought the layers where customers meet automation: the support front end (Fin, Moveworks, Aisera), the data and integration plumbing (Informatica, Pipedream) and the agent builders (Flowise, Sana). Labs bought developer and agent tooling (Bun, Vercept, Stainless, promptfoo, Statsig). Security and observability vendors were absorbed by larger infrastructure firms. Disclosed consideration is now mostly on file: Informatica (about $8B equity value), Moveworks ($2.85B), Weights & Biases ($1.0B), Protect AI ($634.5M) and Prompt Security (about $180M) come from SEC filings, and Fin (about $3.6B) and Sana (about $1.1B) are company-confirmed.

Acquisitions and mergers

Lab rounds, strategic investments and partnerships

Startup rounds (Sierra, Decagon, Cursor, Cognition, Lovable, Factory, Harvey, Legora, Clay, ElevenLabs, n8n) appear with dates in the category tables above and are not repeated here.

Two lessons for buyers follow. First, owners of the tools a business depends on change fast. Langfuse, Helicone, promptfoo, Lakera, Invariant Labs, Pipedream and Flowise all changed hands between June 2025 and March 2026. Contracts and architecture should assume a component may be re-licensed, re-priced or frozen. Second, most of the startup money in the rounds listed here went to coding and support, so buyers in those functions can choose among well-funded vendors with outcome pricing. In verticals that funded startups do not serve, expect to rely more on implementation and integration work than on a packaged product. The People page profiles the founders and executives behind these deals.

Market forecasts: three scopes that must not be combined

The research found only one family of AI-market forecasts with figures and a URL: Gartner's, mostly through secondary write-ups. Those forecasts cover three different scopes: agent features built into enterprise software, total AI spending, and project outcomes. They are kept apart below. Standalone "AI agent market" estimates of about $7B to $8.5B circulate without a traceable source. That figure, a Gartner AI-spending figure for 2029 and an LLM-observability market estimate are excluded here for that reason.

Infrastructure forecasts are a separate class: they size the compute and energy that platforms buy, not automation demand.

Actual capex is the best-sourced number in the whole market, because the four company capex rows below come from SEC filings (the NVIDIA revenue row does not). They show the four largest US platforms spending at a rate far above the AI revenue anyone has disclosed.

Microsoft's AI run-rate of about $37B is roughly a third of its FY2026 capex of $116B. That does not mean the money is wasted, since not all capex is for AI and a run-rate is not annual revenue. It is consistent with model prices being subsidised today, and with some now rising: Gemini 3.8 Flash doubles in January 2027, and OpenAI's flagship list prices rose between 2025 and 2026. Company guidance points the same way: Salesforce guides FY27 revenue up 11% to 12% ($46.1B-$46.4B), while services firms guide lower (Accenture 3% to 6% in local currency; Infosys trimmed its guidance, per a headline). The Hardware, Costs and Future outlook pages develop these threads.

What this means for your business

Most businesses will meet AI agents through a vendor they already pay or through an implementation partner, not by building directly on a lab's platform. The evidence in this map points to the decisions below.

Know what an AI resolution really costs

Published per-outcome prices set a reference band of about $1 to $2 per AI-resolved ticket: Fin at $0.99 and Zendesk at $1.50 to $2.00. Computed model cost for a five-turn support resolution at list prices is about $0.003 on Claude Haiku 5.5 or GPT-6 Luna, and about $0.05 on Sonnet 5.5 or GPT-6.1 Sol (Costs, using Claude pricing and OpenAI pricing). The vendor's margin, not the model cost, sets the market price. If you already pay for Fin, Zendesk AI, Agentforce or Copilot Studio, the agent in that suite is usually the quickest place to start. If you compare it with a custom-built resolution agent, check that the custom quote covers tooling, monitoring, retries and human escalation, which the model cost excludes. Under credit pricing the bill depends on how many steps an agent takes, so budget per task, not per call.

Design for vendor churn from day one

The calendar is full of forced changes. OpenAI's Agent Builder shuts on 30 November 2026, and Claude Sonnet 4.5 retires from the Claude API the same day (Claude release notes). xAI retired a batch of Grok models on 15 May 2026 (xAI retirement notice). Google restricted Gemini 2.5 access to recent users from 18 September 2026 (Gemini API changelog), and Flowise reached end of life on 31 August 2026 (repository archived 13 Aug; npm packages and Docker images deprecated) (Flowise discussion). Helicone has run in maintenance mode since Mintlify bought it (3 Mar 2026) (Helicone). Beyond scheduled changes there are shocks: a US directive took Anthropic's top models offline for 19 days, three Asian regulators briefly blocked Grok in January 2026 (Wikipedia: Grok deepfake scandal), and OpenAI reportedly moved to cut off its models to SpaceX-owned Cursor. Check now whether anything you run depends on the dated shutdowns above. Design for model portability, with gateways such as the MIT-licensed LiteLLM (LiteLLM pricing) or OpenRouter's routed access, which charges a 5.5% platform fee (OpenRouter pricing), and keep an evaluation suite that lets you switch models in days, not months. Before go-live, test the automation on a second model.

Put the fixed Indian deadlines in your plan

The Digital Personal Data Protection regime has two dated phases. Consent-manager provisions start on 13 November 2026, and the core duties (notice, consent, security safeguards, breach intimation, rights, cross-border transfers) start on 13 May 2027 (Gazette G.S.R. 843(E)). The DPDP Rules require a fuller breach report to the Data Protection Board within 72 hours (DPDP Rules 2025), and penalties for failing to keep reasonable security safeguards can reach ₹250 crore (DPDP Act Schedule). Separately, eligible WhatsApp Business customers (partners and directly integrated businesses whose Sold-To country is India) must move to INR billing by 31 December 2026, or non-INR accounts stop delivering messages from 1 January 2027 (WhatsApp Business pricing). For budgeting, Indian per-message rates excluding 18% GST, as vendors report them ("effective 1 Oct 2026"; Meta's rate card not read): marketing ₹0.8631, utility ₹0.1150, authentication ₹0.1150 (ChatMaxima). Automations that touch customer data, WhatsApp order flows and CRM records should include consent, retention and breach-response workflows from the start. For chat or voice agents that face EU users, the EU AI Act's transparency rules (Article 50) apply from 2 August 2026 (EU AI regulatory framework). The Governance page details the obligations.

Choose the workflow before the tool

Demand in India is broad but unguided. In Anthropic's February 2026 India brief, India was the second-largest Claude.ai market but 101st of 116 countries per capita, and use was concentrated in Maharashtra, Tamil Nadu, Karnataka and Delhi (Anthropic India brief); in the June 2026 release (consumer plans only) India had 7.12% of Claude.ai usage in May 2026 and a per-capita usage index of 0.30 (1.0 = proportional to working-age population) (Anthropic Economic Index data). In a survey of 10,000 retailers and MSMEs, 71% of retailers use AI tools daily, but 26% are unsure which AI features would help (PayNearby via Newswav). The first decisions are therefore which workflow to automate, how it connects to the systems you already use, and how staff will be trained. If you run SAP, n8n's place inside SAP Joule Studio gives a natural integration path, subject to n8n's licence terms.

For Indian-language voice agents, add up the full cost per minute

Funded voice-agent platforms price in US dollars per minute: Retell's own range is $0.07 to $0.31 per minute before Indian telephony. INR prices exist on other legs (vendor-reported): Plivo India lists ₹0.38 and ₹0.25 a minute (rows not labelled by direction) and transcription at ₹0.81 a minute; Bolna charges ₹5.52 (6.00 cents) a minute on its $15 pilot plan; Vapi's platform fee is $0.05 a minute plus components (Plivo; Bolna). Sarvam publishes INR speech rates (₹30 per hour for speech-to-text, ₹3 per 1,000 characters for text-to-speech) (Sarvam pricing), which removes currency exposure on the speech leg. Ask for a quote that adds up every leg of the cost per minute (platform, LLM, speech-to-text, text-to-speech, telephony), plus WhatsApp messaging costs if voice and WhatsApp run together. Outbound automated calls also fall under TRAI rules. Its Direction of 4 May 2024 moved voice calling onto DLT-based number series (140-series within 60 days; a 160-series plan within 30 days, implementation in 90) (TRAI Direction). The draft TCCCPR Third Amendment Regulations 2026 add an A2P (application-to-person) call definition, pre-declaration and termination charges, and a lower complaint threshold before a sender is suspended; a fully automated voice agent appears to fall within A2P as drafted (inference from comments hosted on TRAI's site) (Exotel comments to TRAI). Final rules were reported as notified on 18 Sep 2026 (secondary; official text not read), and no explicit AI-voice rule is confirmed. Synthetic voice and video for Indian audiences should also carry the disclosures that the IT Rules amendment, in force since 20 February 2026, requires for synthetically generated content (Business Today).

Put guardrails around coding agents

The fastest-growing revenue in the market is coding agents: Claude Code above $2.5B of run-rate (February 2026), Cursor at $3B ARR (May 2026) and Cognition at almost $900M of run-rate (September 2026). Even UiPath now builds automations through coding agents. If your engineering team adopts these tools, plan for repository setup, review gates, test pipelines, security scanning (Invariant's MCP Scan is open source and now maintained by Snyk) and usage-cost controls.

Budget in stages

US boutiques publish a sequence of fixed-fee assessment, fixed-fee build and monthly retainer: Main & Machine lists a $3,500 to $8,500 audit, an $18,000 to $60,000 implementation sprint and managed services from $1,500 a month, with model usage budgeted separately (Main & Machine pricing). At ₹96 per US dollar, that audit is about ₹3.4 to 8.2 lakh, above what most Indian SMEs can spend on a first engagement. Indian firms rarely publish package prices, so ask for a written quote broken down the same way: assessment, build, ongoing operations and model usage. Consider outcome-based pricing only once a pilot has produced logged costs.

Risks to weigh

Agent projects fail often. A June 2025 Gartner press-release headline says more than 40% of agentic AI projects will be cancelled by the end of 2027 (the release body, including its "agent washing" warning, was not read), so ask for measured outcomes and production references, not demos. Vendor continuity is uncertain: SpaceXAI's AI segment lost $6.4B on $3.2B of revenue in 2025, MiniMax lost $1.87B on $79M, and agent startups trade at up to about 130x revenue. Licences can bind whoever runs a managed service for you. n8n's licence covers internal use, Kimi K3 and Qwen3.8-Max need separate agreements once a model-as-a-service operator's revenue, counting affiliates, passes a threshold, and Mistral's commercial terms conflict across its own pages. Introductory prices also end: Gemini 3.8 Flash doubles on 1 January 2027, and OpenAI marks GPT-5.6 Sol's price as promotional. The Playbook turns these risks into delivery practices.

Questions to ask any vendor or implementation partner

• What is the total cost per resolved ticket, call minute or task, including platform fees, retries, monitoring and human escalation?
• Which models does it use, how fast can it switch if one is retired, suspended or repriced, and has it been tested on a second model?
• Which licence covers each platform when it is run for you as a paid service, and who holds any commercial agreement?
• Which prices are introductory, and what applies after they end?
• What happens to your workflows, data and logs if a component is acquired, re-licensed or frozen?
• How does the design handle DPDP consent, retention and 72-hour breach reporting, and WhatsApp INR billing?

Open questions and unverified items

Sources

1. Anthropic: $30B Series G at $380B post-money (12 Feb 2026)
2. Anthropic: Google and Broadcom compute partnership (6 Apr 2026)
3. Anthropic: Series H announcement (28 May 2026)
4. Anthropic: confidential draft S-1 submitted to the SEC (1 Jun 2026)
5. Hacker News listing: CNBC headline on Anthropic's $65B run-rate (17 Aug 2026)
6. Anthropic: statement on Fable 5 and Mythos 5 access (12 Jun 2026)
7. BBC: US Department of Defense ceases use of Anthropic products
8. Anthropic: Bengaluru office and partnerships across India (16 Feb 2026)
9. Wikipedia: OpenAI
10. Shopifreaks: OpenAI says enterprise revenue has crossed 40% of total sales
11. Wikipedia: ChatGPT
12. OpenAI agents guide
13. OpenAI Agent Builder
14. TechCrunch: OpenAI taps Tata for 100 MW AI data-centre capacity in India (18 Feb 2026)
15. Alphabet Q2 2026 release
16. Alphabet 10-Q
17. Google generative AI pricing
18. SpaceX S-1/A amendment no. 2 (SEC EDGAR, 3 Jun 2026)
19. Wikipedia: SpaceXAI
20. OpenRouter rankings
21. Wikipedia: Zhipu AI
22. Wikipedia: MiniMax
23. Mistral news
24. Wikipedia: Sarvam AI
25. Claude plans and pricing
26. Anthropic Economic Index: India country brief (16 Feb 2026)
27. TechCrunch: OpenAI launches Dots (29 Sep 2026)
28. OpenAI API pricing
29. Constellation Research: OpenAI touts broadening Codex usage, 5 million weekly active users (2 Jun 2026)
30. Alphabet Q1 2026
31. Agent Platform pricing
32. Google Argon
33. Next '26 wrap-up
34. xAI models
35. Wikipedia: SpaceX IPO
36. DeepSeek pricing
37. DeepSeek releases
38. Wikipedia: Qwen
39. Qwen3.8-Max licence
40. Coze Studio
41. Wikipedia: Doubao
42. MiniMax pricing
43. Kimi K3 licence
44. Mistral Large 4
45. Mistral pricing FAQ
46. Cohere blog
47. Sarvam pricing
48. Salesforce Q3 FY26 results (3 Dec 2025)
49. Salesforce Q4 FY26 results (25 Feb 2026)
50. Salesforce Q1 FY27 results (May 2026)
51. Salesforce Q2 FY27 results (26 Aug 2026)
52. No Jitter: Salesforce highlights Agentforce momentum
53. Agentforce pricing
54. Salesforce signs Fin agreement
55. Salesforce completes Fin
56. Fin pricing
57. Zendesk pricing
58. Office Watch: Microsoft 365 Copilot Business AI add-on (secondary)
59. Microsoft Learn: Copilot Studio licensing and subscriptions
60. Salesforce Informatica
61. Salesforce Claudeforce
62. Microsoft FY26 Q3
63. Microsoft FY26 Q4
64. Wikipedia: ServiceNow
65. ServiceNow release
66. ServiceNow: Moveworks acquisition release, 8-K exhibit 99.1 (10 Mar 2025)
67. UiPath Q1 FY27
68. UiPath Q1 FY27 results, SEC 8-K exhibit 99.1 (28 May 2026)
69. Wikipedia: Workday
70. Workday Sana
71. Workday Pipedream
72. n8n press
73. Wikipedia: Automation Anywhere
74. Pega news
75. Wikipedia: Pegasystems
76. Intercom pricing
77. Bolsamania: n8n valuation doubles to $5.2bn as SAP makes strategic investment
78. n8n: Sustainable Use License (LICENSE.md)
79. n8n pricing
80. Zapier MCP
81. Zapier press
82. Anthropic: Claude Opus 5.5 (22 Sep 2026)
83. Wikipedia: Zapier
84. Sacra on Zapier
85. Celonis news
86. Workato press
87. Wikipedia: Workato
88. Zapier pricing
89. Make pricing
90. Tray.ai news
91. Sacra on Tray.io
92. Pipedream pricing
93. Activepieces pricing
94. Windmill pricing
95. Temporal pricing
96. Dify pricing
97. Flowise discussion
98. Botpress pricing
99. Bloomberg Law: SpaceX completes its $60 billion Cursor acquisition
100. Cognition Series E
101. Wikipedia: Anysphere
102. Dealroom: Cursor tops $4B annualized revenue (single anonymous source)
103. VKTR: Sierra raises $950M at $15B valuation
104. PYMNTS: OpenAI veteran raises $950 million for AI customer-service agents
105. Sierra blog
106. Bloomberg Law: AI customer-support startup Decagon valued at $4.5 billion
107. Sacra on Decagon
108. Crescendo news
109. Cursor pricing
110. Hacker News search: SpaceX and Cursor headlines (incl. Reuters, 27 and 29 Aug 2026)
111. Wikipedia: Cognition
112. Wikipedia: Lovable
113. Lovable blog
114. Factory news
115. Glean press
116. Wikipedia: Hebbia
117. Harvey: raises $550M at a $15.5B valuation
118. Wikipedia: Harvey
119. Wikipedia: Legora
120. Clay Series D
121. 11x blog
122. Wikipedia: Manus
123. Genspark blog
124. ElevenLabs Series D
125. ElevenLabs blog
126. ElevenLabs tender
127. Retell pricing
128. Vapi pricing
129. Bland pricing
130. Synthflow pricing
131. TechCrunch: Mem0 raises $24M (28 Oct 2025)
132. Accenture Q4 FY26 earnings 8-K exhibit (SEC)
133. Motley Fool: Accenture Q4 FY26 earnings-call transcript (1 Oct 2026)
134. Anthropic newsroom
135. Main & Machine pricing
136. Main & Machine services
137. OptiWork.ai
138. Antino Labs: artificial intelligence services
139. Infyom: n8n workflow automation services
140. Ideas2IT
141. The Tribune: TCS Q1 FY27 results, AI revenue run-rate
142. TechMarketView: Infosys trims guidance as AI revenue climbs (24 Jul 2026)
143. DQ India: HCLTech Q1 FY27 management on AI growth
144. Whalesbook: LTIMindtree Q1 FY27 results
145. Wipro Q1 FY27 release
146. X-Rates: USD to INR monthly averages, 2026
147. Forkast: India's IT giants are cutting jobs for the first time in 20 years
148. IANS: top IT firms' headcount falls over 7,000 in FY26 (25 Apr 2026)
149. Entrepreneur India: headcount declines from TCS to TechM
150. Wikipedia: 2025 in AI
151. Meta 10-Q, quarter to 30 Jun 2025 (SEC)
152. Snyk: acquires Invariant Labs
153. SentinelOne 8-K: Prompt Security acquisition completed (8 Sep 2025)
154. The Verge: OpenAI acquires Statsig
155. CXOToday: Check Point completes acquisition of Lakera
156. Lakera website (Check Point)
157. OpenAI: acquires Software Applications Incorporated (Sky)
158. Anthropic: acquires Bun as Claude Code reaches $1B milestone
159. Wikipedia: Groq
160. ClickHouse blog
161. SiliconANGLE: ClickHouse raises $400M, acquires Langfuse (16 Jan 2026)
162. Anthropic: acquires Vercept
163. Helicone: joining Mintlify (3 Mar 2026)
164. Promptfoo blog: agreement to be acquired by OpenAI (9 Mar 2026)
165. Axios (headline only): Meta acquires Moltbook
166. OpenAI: acquires TBPN
167. Anthropic Stainless
168. Salesforce Contentful
169. DuckLabs: to join AWS
170. Nvidia 8-K: agreement to acquire Hugging Face (filed 3 Sep 2026)
171. BBC: OpenAI scraps rollout of new model over safety concerns; reports Nvidia agreed to buy Hugging Face (29 Sep 2026)
172. Deno: joining Cloudflare
173. TechCrunch: OpenAI raises $110B (27 Feb 2026)
174. MediaNama: Sarvam raises $234 million (Jun 2026)
175. CNBC (headline only): Google to pay SpaceX $920 million a month for xAI compute
176. Software Strategies Blog: Gartner forecasts agentic AI overtakes chatbot spending in 2027 (16 Feb 2026)
177. Gartner: over 40% of agentic AI projects will be canceled by end of 2027 (25 Jun 2025)
178. OODA Loop: over 40% of agentic AI projects will be scrapped by 2027, Gartner says
179. Wikipedia: AI bubble
180. Wikipedia: Stargate
181. Wikipedia: environmental impact
182. Epoch data centres
183. IANS: India's data-centre capacity expected to reach 6 GW by 2029 (JLL, 2 Sep 2026)
184. Energetica India: India data-centre capacity to reach 12 GW by 2030 (Wood Mackenzie)
185. Fortune: India's AI Impact Summit closes with the New Delhi Declaration (23 Feb 2026)
186. SEC XBRL: Microsoft
187. SEC XBRL: Alphabet
188. SEC XBRL: Meta
189. SEC XBRL: Amazon
190. NVIDIA Q2 FY2027
191. Claude Platform API pricing
192. Claude release notes
193. xAI retirement notice
194. Gemini API changelog
195. Wikipedia: Grok deepfake scandal
196. LiteLLM pricing
197. OpenRouter pricing
198. Gazette G.S.R. 843(E)
199. DPDP Rules 2025
200. DPDP Act Schedule
201. WhatsApp Business pricing
202. EU AI regulatory framework
203. Newswav: PayNearby MSME Digital Index 2026
204. Business Today: government tightens rules on AI-generated content (10 Feb 2026)
205. usecarly: LLM market share 2026 (secondary summary of Menlo Ventures data)
206. Zendesk India pricing (read 11 Oct 2026)
207. Bloomberg Law: OpenAI valued at $852 billion after closing $122 billion round
208. SpaceX 424B4 prospectus (SEC EDGAR, 12 Jun 2026)
209. Google: AP2 donated to the FIDO Alliance (28 Apr 2026)
210. Moonshot: Kimi API pricing
211. Sarvam blog: Sarvam 30B and 105B open-sourced (6 Mar 2026)
212. Plivo: India voice pricing
213. fin.ai homepage
214. Zendesk pricing, Germany (read 12 Oct 2026)
215. Salesforce: Informatica agreement, 8-K exhibit 99.1 (27 May 2025)
216. n8n: LICENSE.md raw text (read 12 Oct 2026)
217. SpaceX 8-K: Cursor (Anysphere) merger effective (14 Aug 2026)
218. Outlook Business: TCS Q2 FY27 results (8 Oct 2026)
219. CoreWeave 10-Q, quarter to 30 Jun 2025 (SEC)
220. Palo Alto Networks 10-K, fiscal 2025 (SEC)
221. ChatMaxima: WhatsApp API pricing in India
222. Anthropic Economic Index data (Hugging Face; June 2026 release)
223. Bolna pricing
224. TRAI Direction of 4 May 2024 on DLT-based voice calls
225. Exotel comments to TRAI on the draft TCCCPR Third Amendment Regulations 2026
226. Gallabox: WhatsApp pricing overview