What You Need for AI Automation: The Complete Checklist

Facts in this document are as of 12 October 2026. Dollar amounts are converted at ₹96 per US dollar, the rate used across this research set (X-Rates' October 2026 average to 11 October was 95.95) (X-Rates). Rupee prices from Indian vendors are given as quoted. It synthesises the other documents in the set into one checklist for a business starting or scaling AI automation, whether an Indian SME or a US or global company, and whether it builds in-house or with an implementation partner; the links lead to the evidence. Labels: vendor-reported is a company's claim about its own product; secondary is news, a blog, an aggregator or a law-firm summary; single-source has one source; assumption is an input chosen for this document's arithmetic. List prices read on a vendor's own page carry no label. This is research, not legal, tax or financial advice.

Key takeaways

• People, evaluation and compliance evidence are the scarce inputs, not tools or compute. In the worked budgets below, a small pilot run by two people costs $223-243 (₹21,446-23,366) a month in tools. An in-house automation team of about ten costs ₹12.8-18.4 lakh a month, 87-90% of it pay. One production WhatsApp-plus-voice support line for an Indian SME runs at ₹27,631-29,891 ($288-311) a month.
• For models, use two vendors, three tiers and one open-weight path. List prices run from $0.10 / $0.50 per million input / output tokens (Claude Haiku 5.5, GPT-6 Luna) to $10 / $50 (Claude Fable 5.1, GPT-6 Astra). A US export-control directive took two frontier models offline for every customer from 12 June to 1 July 2026, so keep a tested fallback.
• No vendor's own API processes requests in India. AWS Bedrock routes the previous Claude generation (Opus 5, Sonnet 5, Haiku 4.5), Kimi K3 and GPT-5.6 Terra/Luna only to Mumbai and Hyderabad. The newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud read.
• Every software layer has a free, permissive default, such as vLLM, LiteLLM, pgvector, Docling, Langfuse's MIT core, Activepieces or Temporal, and LiveKit. The traps are licences that restrict running a tool as a service for others, which matters when a partner hosts it for you or you offer it to your own customers: n8n, Dify, Arize Phoenix and LM Studio. Flowise is archived.
• Indian-language work needs its own test sets. No independent benchmark compares current frontier and Indian models. Older tokenizers need 4.1x (Hindi) to 13.0x (Malayalam) the tokens of English (preprint). Sarvam is the only vendor with a full rupee price list.
• The calendar is fixed. OpenAI's legacy snapshots shut on 23 October 2026, and Agent Builder and Evals on 30 November. India-billed WhatsApp accounts must move to INR by 31 December. Colorado's and California's automated-decision rules, and a doubling of Gemini Flash prices, arrive on 1 January 2027. India's main DPDP duties start on 13 May 2027.
• Deterministic controls beat classifiers. Classifier defences were bypassed in recorded incidents; one researcher beat a vendor's auto-approval classifier about 80% of the time. No bypass of least privilege, sandboxes, egress allowlists or human approval is recorded in the set.
• Staff the work with forward-deployed-style generalists first, then fund an evaluator, whether the people are yours or a partner's. Median AI/ML engineer pay is $17,436 (₹16.7 lakh) in India against $189,500 in the US, about an elevenfold gap (Stack Overflow 2025; the Indian figure rests on 25 respondents). Certifications cost $75-200 each.
• Method beats model choice. Score use cases, pilot the hardest step against go or stop KPIs signed before the build, and start with 20-50 test tasks drawn from real failures.

The checklist at a glance

1. Models and API access

Why. The model is the cheapest line in most text workflows. A five-turn support resolution costs $0.0027 (₹0.26) on Haiku 5.5 and $0.052 (₹5.01) on Sonnet 5.5, against $0.99 for one Fin outcome (Costs). But the model choice decides where your data and your customers' data go, which licence applies and how often code must be migrated. Sending 80% of cases to a value model and escalating 20% to a workhorse model cuts the blended cost to about $0.013 (illustrative, Top models).

• A primary and a fallback vendor from different companies, both passing the same regression suite. Anthropic disabled Fable 5 and Mythos 5 for all customers from 12 June to 1 July 2026 under a US export-control directive (Anthropic).
• A router across tiers: plain code first, then the value, workhorse and frontier tiers. Escalate on schema failures, failed checks, policy-sensitive cases or irreversible actions.
• One open-weight path: gpt-oss-120b (Apache-2.0, fits one 80 GB GPU), Qwen3.8-27B (Apache-2.0) or GLM-5.3-Flash (MIT). For Indian languages, Sarvam 30B/105B or Gnani evon 30B (all Apache-2.0) (Indian languages).
• A processing location per workflow, written into vendor and partner contracts (table below).
• Retention settled before any commitment about data. OpenAI and Anthropic grant zero data retention (ZDR) only on approval. Anthropic's Fable and Mythos models require 30-day retention. On Anthropic, Batch, the Files API, code execution, the MCP connector, Agent Skills and Managed Agents are never ZDR-eligible (Anthropic retention). OpenAI's Agents API and vector stores keep data until deleted (OpenAI data controls).
• Reasoning tokens budgeted as output. Anthropic, OpenAI and Google say they bill them that way (Claude extended thinking; OpenAI reasoning).
• Per-project or per-team keys or workspaces with spend alerts. Anthropic's Start tier stops at a $500 monthly cap.
• Pinned model IDs, a deprecation calendar, and evals re-run at every change. Google re-routed gemini-3.7-flash to 3.8 Flash on 8 October 2026 (Gemini changelog).
• Staff seats. Claude Team Standard is $25 a seat monthly or $20 annually (2-150 seats); Pro is $20 a month (Claude pricing). Rupee billing for Team Standard is ₹2,999 monthly or ₹2,399 annually, GST included (single-source, Techstory). ChatGPT Business is $25 or $20 (secondary).

Sources: Claude pricing; OpenAI pricing; Google Cloud pricing; xAI models; DeepSeek pricing; Moonshot; Z.ai; Xiaomi; Sarvam. AA is the Artificial Analysis Intelligence Index v4.3.2 at the effort it lists. Detail is in Claude, OpenAI, Gemini, Grok, DeepSeek and Other models.

Where data can be processed in India.

AWS says not to rely on the in. prefix alone (call GetInferenceProfile), and that prompts may be stored in opt-in regions for abuse detection; Hyderabad is opt-in, so that stays in India.

Deprecation and price calendar.

Watch out for. Promotional prices are not floors. Long prompts reprice the whole request at OpenAI (above 272K tokens), xAI (at 200K) and Haiku 5.5 (above 100K). DeepSeek's terms let outputs train other models (DeepSeek terms). Kimi K3 and Qwen3.8-Max licences add revenue thresholds, and MiniMax M3's base grant is non-commercial (Other models).

2. Compute and hardware

Why. Most SME automation needs no owned GPU: APIs serve production and rented hours cover open-model tests. One H100 at Lambda's $3.99 an hour must sustain about 6,500 output tokens per second around the clock to beat a $0.17-per-million API price (Costs). Buy hardware for data control, offline work or a fine-tuned model no provider serves, not to save money.

• Developer laptops that run containers, so coding agents work in a sandbox. Models up to 10B need 0.4-5.4 GB at 4-bit, and gpt-oss-20b fits a 16 GB device (Hardware). The set holds no laptop prices.
• A GPU rental account in India, with GST noted. E2E Networks lists an H100 at ₹255.55 an hour on demand, ₹155.90 on a commitment of three months or more, and ₹70 on spot (vendor INR prices equal USD × 95; GST extra) (E2E H100).
• An IndiaAI eligibility check. The portal lists an H100 SXM at ₹153 per GPU-hour before a subsidy of "up to 40% subject to approval", or ₹91.80 at the maximum (IndiaAI calculator).
• A desk-side box only if data must stay on site, sized with the memory arithmetic in Hardware (about 0.55-0.6 bytes per parameter at 4-bit, plus KV cache).
• A rent-first rule. Measure throughput on your own prompts before committing to capacity.

IndiaAI eligibility. The portal's named categories include researchers, DPIIT-recognised startups and MSMEs with ₹50 lakh of annual revenue or ₹100 lakh of funding, students and government entities. Other businesses are not a named category, so a company qualifies only as a DPIIT-recognised startup or MSME (inference). Requests under 5,000 GPU-hours are auto-approved, submissions run from the 1st to the 25th of a month, and approvals are valid for 30 days (IndiaAI eligibility).

Watch out for. Memory shortages raised hardware prices in 2026, so treat quotes as short-lived and check that each states its FX date and GST. Dense 70B models fit desktops badly, while 120B mixture-of-experts models fit well. No measured single-H100 throughput for gpt-oss-120b was found, so treat the break-even as a threshold to test. Indian electricity and colocation prices have no source.

3. Software stack

Why. A model API is not a production automation; the table lists the other layers. Each has a permissive default, and hosted options mostly add cost at volume or at enterprise tiers (Tool stack).

Licences and prices are from Tool stack and Frameworks, read from each repository or pricing page, including the n8n licence, LM Studio terms, Llama 4 licence, ChatMaxima and Plivo India.

• One framework matched to your cloud: Microsoft Agent Framework for Azure, ADK for Google Cloud, the Agents SDK for OpenAI-only work, LangGraph as the neutral default.
• A reusable MCP server library for your core business systems, each scanned and version-pinned.
• A licence register, re-checked at upgrades. LiteLLM, Langfuse, Activepieces and Mastra change licence by folder.
• Evaluation suites in a repository you own, including when a partner builds them, not on a vendor dashboard.
• A billing-currency check for every WhatsApp Business account you run. Two vendors report that from 1 October 2026 each number gets 1,000 free service messages a month, then pays the utility rate (vendor-reported; Meta's file not read) (Gallabox).

Watch out for. Ownership churn changes roadmaps. Langfuse joined ClickHouse, Helicone joined Mintlify and OpenAI agreed to buy promptfoo. Keep hosted layers swappable behind a gateway and OpenTelemetry traces.

4. Data

Why. Retrieval finds only what the parser extracted, and agents act only on data their tools can reach. India's DPDP Act needs a lawful basis for each purpose and has no "legitimate interests" ground, so reusing chats for prompts, evals or fine-tuning beyond the notice probably needs fresh consent (inference in Governance).

• A data map per workflow: sources, owners, categories (personal, children's, financial, health), storage, the model and region that see the data, retention and outbound channels.
• A consent basis on record. Consent must be free, specific, informed, unambiguous and unconditional, and as easy to withdraw as to give (DPDP Act s.6).
• Retention and log schedules. DPDP Rules 6(1)(e) and 8(3) require a year of logs from 13 May 2027. CERT-In already requires logs kept for 180 days within India (CERT-In).
• A quality check: structured or scanned, which scripts, how much handwriting.
• A baseline before the build: handling time, error rate, volumes, repeat contacts.
• Test sets: 20-50 tasks drawn from real failures, a regression suite near 100%, an injection test set and a reference solution per task (Anthropic evals). Use masked or consented samples for identity documents.
• Indian-language sets per language and script, reported separately (table below).
• Measured tokens per language. On the older cl100k tokenizer, Hindi takes 4.1x and Malayalam 13.0x the tokens of English; on o200k the ten-language mean falls from 8.0x to 2.1x (preprint). Current Claude, Gemini and GPT-6 tokenizers were not measured.
• Memory hygiene: provenance on memory writes, isolation between customers and tenants, expiry.

Source: Indian languages. In the only independent 2026 test of code-mixed input, every model lost accuracy as mixing rose: Claude Opus 4.6 from 68.7 to 61.2, Sarvam-30B from 64.2 to 56.1 (Indi-RomCoM, preprint).

Watch out for. Most Indian speech and OCR scores are vendor-run, such as Saaras V3's reported 19.31% word error rate. Score your own audio and scans before relying on any accuracy figure.

5. People and skills

Why. The forward-deployed engineer (FDE) role describes the core automation job, whether your own staff or a partner's do it: scope a proof of concept, build it on the business's own systems, and turn fixes into reusable patterns. The evaluator turns a pilot into evidence, yet no posted pay for that role was found anywhere (Playbook).

Sources: Stack Overflow 2025 medians of total compensation (survey; small Indian samples), PayScale, BLS and job boards, via Playbook.

• FDE-style generalists first, in-house or from a partner, who can run discovery, build and hand over to the team that runs the workflow.
• An evaluator second, then juniors trained as evaluators and approvers. AI helps novices most: +34% issues resolved per hour for novice support agents (NBER w31161).
• A named part-time governance and security reviewer at design review and before release.
• A pay-growth allowance. Indian projections are 15-20% a year (secondary), and Naukri reported AI/ML hiring up 33% year on year in July 2026 (secondary) (India; Playbook).
• An AI-literacy training log. The EU AI Act's Article 4 needs no certificate; internal records suffice (Commission Q&A).

Watch out for. Indian pay data is thin: AmbitionBox, Glassdoor and Naukri could not be read. Treat ₹10-17 lakh as the mid-market engineer band; FDE-calibre seniors cost well above it.

6. Delivery process

Why. MIT NANDA's "95%" of pilots showed zero measurable P&L return six months on (interview-based; the authors call it "directionally accurate") (MIT NANDA). Gartner forecasts more than 40% of agentic projects cancelled by the end of 2027 (headline read only). Method is the variable you control.

Source: Playbook, built on Microsoft's Cloud Adoption Framework and OpenAI, Anthropic, AWS and Google guidance.

• A discovery questionnaire using CAF's rule: "If the steps are clear, repeatable, and follow strict rules, use regular code".
• KPIs measured together: pass rate (pass^k for customer-facing flows), throughput with quality (CSAT or error rate), volumes and repeat contacts, resolution rate, cost per completed task, time against baseline, and human edit rate.
• Gate rules. Go when the threshold is met, quality holds, cost is within budget and no high-severity finding is open. Change direction when only the hardest step fails. Stop when safeguards or consent cannot be secured, or the hardest step fails after two rounds of fixes.
• Eval-driven QA. At 75% success per run, an agent passes three runs in a row only about 42% of the time. OpenAI's GDPval grader agreed with experts 66% of the time, against 71% between experts, so check AI graders against people (GDPval).
• Runbooks: a kill switch (revoke tokens, disable tools, end sessions) tested quarterly; incident response with a 6-hour CERT-In step, and notice from vendors and partners well inside 72 hours; model migration with the regression suite re-run; rollback.
• The least autonomous design that passes. Agents use about 4 times the tokens of chat, multi-agent systems about 15 times (Anthropic).

Watch out for. Self-reported speed gains mislead: in METR's 2025 trial, developers took about 19% longer with AI tools yet believed they were 20% faster (METR). The best model in the June 2026 OSWorld 2.0 paper (Claude Opus 4.8) completed about 20.6% of long computer-use tasks under strict scoring (OSWorld 2.0). Keep a human in that loop.

7. Legal, compliance and contracts

Why. For a business operating in India, the nearest binding dates come from data protection, media labelling, incident reporting and calling rules, not from an AI statute. US and EU duties follow your customers and their data wherever the work is done, and processors, including vendors and implementation partners, take on many duties through contracts.

There is no federal US AI statute. EO 14365 directs challenges to state laws, but state duties keep binding while those proceed.

• Six short policies: acceptable use; build and release; customer data and vendors; human oversight; incident response; disclosure and labelling.
• Four named roles: an accountable owner, a data protection lead, a release approver who is not the builder, and a business owner per workflow.
• A data processing agreement (DPA) with each vendor and partner naming sub-processors, models, regions, retention and ZDR status.
• A vendor-terms file. Anthropic's IP indemnity excludes customer inputs, modifications and combinations and caps liability at 12 months' fees (Commercial Terms). OpenAI and xAI keep API data for up to 30 days by default.
• Vendor and partner contract clauses covering:
  • the parties' roles and named models, with notice before changes;
  • no secondary use of your data, and model substitution with re-testing;
  • an agent authority schedule, and human review for decisions about people;
  • incident notice well under 72 hours (for example 24 hours, as suggested in Governance);
  • an indemnity read against the model vendor's own terms, and disclosure duties for each channel;
  • a licence schedule, a price-review and FX clause, and a written outcome definition if you pay per outcome.
• Colorado and California packs for US operations automating consequential decisions: a model card, a human-review guide and adverse-decision explanation templates.
• A disclosure default. Every chatbot and AI call says it is AI, and generated media is labelled.
• A licence review in every statement of work. Check n8n, Dify, Phoenix, LM Studio, Skyvern (AGPL-3.0), the Llama licence, the Kimi K3 and Qwen3.8-Max thresholds, MiMo's metadata-only MIT tag, Sarvam-Translate (GPL-3.0), Claude Code (proprietary) and the Claude Agent SDK's commercial terms.

Watch out for. GST on foreign AI APIs was not read from any statute or vendor page; reverse charge on business imports of services is the usual position (background only), so get tax advice.

8. Security

Why. Most serious 2025-2026 incidents followed one pattern: an agent read untrusted text, then acted with real credentials or sent data out. In Anthropic's own test, only 13.6% of 1,053 testers refused a swapped dangerous command, while its auto mode would have blocked 89% (vendor-reported). A researcher then bypassed auto mode about 80% of the time (Simon Willison).

• A threat model per agent against the OWASP Agentic Top 10, ASI01 to ASI10 (OWASP).
• No toxic flow. No agent holds untrusted input, private data and an outbound channel at once without a deterministic gate.
• Least privilege: one identity per agent, narrow read-only tools by default, minimum OAuth scopes, and authorisation enforced downstream (OWASP LLM06).
• Human approval for writes, deletes, payments, external sends and account changes, showing the exact action.
• Sandboxes with no home directory, SSH keys or cloud credentials, and an egress allowlist.
• No production access for agents by default; destructive operations need a second person.
• Supply-chain discipline: pinned packages and MCP servers, MCP Scan on tool descriptions, no "trust all tools" flags.
• Exposure management: consoles and self-hosted workflow tools behind VPN or SSO, patched fast. MITRE ATLAS case studies cover attacks on n8n and Langflow (ATLAS).
• Output validation: never render untrusted Markdown, links or images.
• Spend and rate limits per agent, with automatic stops.
• Audit logs of every tool call (identity, inputs, outputs, approver), kept at least a year.
• Red-teaming: an injection test set in every release, a quarterly red-team for Tier 1, and findings labelled with MITRE ATLAS (120 techniques, 73 case studies).
• Security questionnaires, the ones you send vendors and the ones your customers send you, mapped to the CSA AI Controls Matrix (243 controls, 18 domains). Pursue ISO/IEC 42001 only when customers or regulators ask for it, because cloud certificates do not flow down (AWS FAQ).
• The same rules on your developers' laptops and CI. The s1ngularity attack used local AI command-line tools to hunt secrets, exposing 2,180 accounts.
• A risk tier per deployment, scoring impact, authority, data sensitivity, untrusted input and outbound channels from 1 to 3. A total of 11 or more, or a 3 on impact or authority, makes it Tier 1, which needs all 22 controls in Governance.

Watch out for. OpenAI's security chief calls prompt injection "a frontier, unsolved security problem" (Simon Willison). NIST's CAISI (which NIST's site now calls CAISSI) found DeepSeek R1-0528 agents 12 times likelier to follow hijacking instructions, so injection-test each model before adoption (CAISI).

9. Budget

Why. All three budgets are monthly at list prices, converted at ₹96. Replace every assumption with logged data after the first pilot.

Budget A: a small pilot run by two people (tools only).

Budget B: an in-house automation team of about ten.

Budget C: one production WhatsApp and voice support line. This is an Indian SME's Hindi/English support line handling 5,000 WhatsApp conversations and 2,000 three-minute inbound calls (6,000 minutes) a month on one number. It uses Sarvam speech and LLM, Plivo telephony and LiveKit Cloud. Lines C1-C9 are the worked example in Indian languages, where every volume and token count is an assumption.

Changing one leg at a time shows where the decisions lie (Indian languages). A Claude Haiku 5.5 LLM saves ₹1,938 a month, while Sonnet 5.5 adds ₹14,109. ElevenLabs voices add ₹24,653, and per-character translation of every chat reply adds ₹45,000. For a US business, a support agent handling 3,000 conversations a month on a workhorse model costs 3,000 × $0.0522 × 2 + $29 = $342 (₹32,851), or $45 on Haiku 5.5 (Costs). Bought per outcome at the $0.99 Fin benchmark, with an assumed 50% resolution rate, the same volume would cost $1,485 a month.

Watch out for. No measured production token profile was found. Budget USD lines at a stated FX rate, because the rupee moved from 90.73 to 95.95 per dollar between January and October 2026. Ask for GST to be stated on every rupee quote, and keep WhatsApp and telephony as separate usage lines billed at cost.

30-60-90 day setup order

What this means for your business

Start with method. Score use cases, pick one to three, record a baseline and sign go and stop KPIs before building (section 6). Write 20-50 test tasks from real failures, and run two model vendors against one regression suite.

Budget for people and evidence, not hardware. A two-person pilot needs about ₹21,000-23,000 a month in tools. Pay is 87-90% of an in-house team of about ten, and one support line costs ₹27,631-29,891 a month to run.

Decide early where each workflow's data may be processed (in India, only on the verified routes in section 1, confirmed in writing), which licences you accept, and who signs off evaluation and security.

Diary the deadlines: 23 Oct and 30 Nov 2026 (OpenAI shutdowns), 31 Dec (WhatsApp INR billing), 1 Jan 2027 (Colorado, California, Gemini Flash prices) and 13 May 2027 (DPDP).

Test before you trust. No published benchmark settles which model, voice or OCR engine suits Indian-language traffic, so test your own calls, chats and scans. Rely on deterministic controls rather than classifiers, and keep a tested open-weight fallback, because prices, models and access all changed in 2026.

Ask any vendor or implementation partner:

• Which models and regions will process your data, with what retention, and what notice comes before a model change?
• Does each component's licence allow them to host it for you?
• Will the evaluation suites live in a repository you own?
• How fast will they report an incident, and how are price and FX changes handled?

Open questions and unverified items

Sources

Documents in this set: Index, Top models, Claude, OpenAI, Gemini, Grok, DeepSeek, Other models, Hardware, Frameworks, Tool stack, Costs, Governance, Playbook, India, Indian languages, Americas and Europe, Use cases, Timeline and Glossary.

Key primary sources:

1. Models and data terms: Claude pricing; Claude data residency; Claude data retention; Claude models overview; Claude release notes; Claude extended thinking; OpenAI pricing; OpenAI deprecations; OpenAI data controls; OpenAI reasoning; Google Cloud pricing; Gemini changelog; xAI models; DeepSeek pricing; DeepSeek privacy policy; DeepSeek platform terms; Moonshot pricing; Z.ai pricing; Xiaomi MiMo pricing; Artificial Analysis
2. India processing: Bedrock regional availability; Bedrock Sonnet 5 card; Google Cloud data residency; Azure region availability; OCI regions; E2E inference endpoints; Yotta Shakti endpoints
3. Compute: IndiaAI calculator; IndiaAI eligibility; IndiaAI empanelled GPUs; E2E H100; Lambda; AWS Price List, Mumbai; llama.cpp DGX Spark benchmarks; Apple US Mac Studio; Apple India Mac Studio
4. Software and licences: MCP registry; Claude Agent SDK; n8n licence; LM Studio terms; Llama 4 licence; Flowise discussion; Langfuse pricing
5. Speech and messaging: Sarvam pricing; Plivo India; ChatMaxima; Gallabox; Meta WhatsApp pricing
6. Seats and people: claude.com pricing; Techstory, rupee billing; GitHub Copilot plans; Cursor pricing; Stack Overflow 2025; PayScale India; BLS data scientists; NBER w31161
7. Method and evidence: Microsoft CAF; Anthropic evals; Anthropic multi-agent research; GDPval; METR; OSWorld 2.0; Indi-RomCoM; MIT NANDA
8. India law: DPDP Act; G.S.R. 843(E); DPDP Rules, G.S.R. 846(E); CERT-In directions; TRAI Direction; TRAI-hosted draft comments; RBI FREE-AI; iPleaders on the IT Rules
9. EU and US law: European Commission AI Act; AI literacy Q&A; GDPR Art. 22; Colorado SB 26-189; X.AI v. Weiser docket; CPPA ADMT rules
10. Security: OWASP Agentic Top 10; OWASP LLM06; MITRE ATLAS changelog; Simon Willison; Simon Willison on OpenAI's CISO; NIST CAISI; AWS ISO/IEC 42001 FAQ; Anthropic Commercial Terms; Anthropic, Fable and Mythos access
11. Exchange rate: X-Rates