Skills, Team and Implementation Playbook

Key takeaways

• A small, senior team does the work, whether you hire it or bring in an implementation partner. Forward-deployed engineers (FDEs) scope a proof of concept, then build it into production on the business's own systems. Posted US base pay for individual-contributor FDE roles runs from US$135,000 (Palantir) to US$325,000 (OpenAI); manager and security FDE posts go higher. An agent engineer, an integration engineer, an evaluator and a part-time architect complete the team.
• Pay for AI engineers differs about elevenfold between the US and India, and Indian supply is tight. Stack Overflow's 2025 survey puts median AI/ML engineer pay at US$189,500 in the US and US$17,436 (₹16.7 lakh) in India, though the Indian figure rests on only 25 respondents. India's AI talent pool of 6 to 6.5 lakh people is reported to grow about 13% a year against 25% to 35% demand growth, and Naukri reports AI/ML hiring up 33% year on year in July 2026 (secondary), so allow time to hire.
• Certifications cost US$75 to US$200 each, and three vendor tracks have changed. Microsoft has retired AI-102; its current AI track is AI-103, AI-500 and an AB- series. Anthropic's four Claude certifications (US$99 to US$175), launched in 2026, are open only to staff of Claude Partner Network firms and last 12 months. AWS moved its ML Engineer exam to a US$75 beta (MLA-C02) after English delivery of the old exam ended on 28 Sep 2026. Databricks and NVIDIA credentials last two years, AWS's three. Check that a partner's certificates are current, and treat them as a signal, not proof.
• Vendor guidance now covers every stage of a project. Microsoft's Cloud Adoption Framework (CAF) screens out agents where ordinary code will do. It scores use cases from 1 to 5 on impact, feasibility and desirability, and says to pilot the hardest step against baseline KPIs that act as go or stop gates. Anthropic says 20 to 50 tasks drawn from real failures are enough to start an evaluation suite. The same steps are a yardstick for judging a partner's method.
• Choose the least autonomous design that passes the tests. At list prices, a workflow with LLM steps or a retrieval (RAG) assistant costs US$0.0001 to US$0.015 per task, and a five-turn tool-using support agent US$0.003 to US$0.05 per resolution (small to mid-tier model). Multi-agent systems use about 15 times the tokens of a chat.
• Published prices give buyers a benchmark for implementation work. One US boutique, Main & Machine, charges US$3,500 to US$8,500 for an assessment and US$18,000 to US$60,000 for a build sprint, with managed service from US$1,500 a month; another US firm starts lower. Indian firms rarely publish prices. Outcome fees of US$0.99 (Fin) to US$2.00 (Zendesk) per resolution are 19 to 38 times the mid-tier model cost of that resolution, so compare them with your fully loaded cost per resolution, not with model cost alone.
• AI-assisted delivery saves money only if it is measured and contained. Measured effects range from 19% longer task times (METR's 2025 trial) to 26% more completed tasks (Cui et al.'s published abstract, 4,867 developers; the authors' slides report 5,114 developers and about 25% to 31% more pull requests). One researcher bypassed a vendor's agent auto-approval classifier about 80% of the time. Whether your own developers or a partner use AI agents, the controls that hold are human review, sandboxes and written disclosure.
• Fixed dates drive the first year. OpenAI Evals turns read-only on 31 Oct 2026, and Evals and Agent Builder shut down on 30 Nov 2026. Businesses whose Meta "Sold-To" (billing) country is India must move their WhatsApp Business accounts to INR billing by 31 Dec 2026. Colorado's laws on automated decisions and chatbots start on 1 Jan 2027. India's main data-protection (DPDP) duties start on 13 May 2027.

Scope, date and evidence labels

All facts here are stated as of 12 October 2026. Most pages were read on 11 October; facts that changed or were disputed were re-checked against primary sources on 12 October 2026. Amounts are in US dollars, with rupee equivalents at ₹96 per US dollar. That rate is the X-Rates average for October 2026 and is used across this research set (X-Rates monthly average). One lakh is ₹100,000; one crore is 100 lakh.

Three labels appear in the text:

• [Synthesis] marks a recommendation, template or roadmap built from the cited facts. It is not itself a sourced fact.
• Vendor-reported marks a company's claim about itself or its customers.
• Secondary marks a news, blog or aggregator source whose original was not read.

The playbook is written for businesses building AI automation capability, in-house or with an implementation partner, with most detail on Indian SMEs and US companies. It covers the roles and skills needed, how to run and gate a project, reference architectures, what implementation work costs, how to evaluate a partner, contract terms, risks and a first-year plan. It describes no particular firm. The most decision-critical fees, pay ranges and vendor terms were re-read on their primary pages on 11 October 2026. Related pages: Costs and economics, Agent frameworks, Tool stack, Governance and security, Use cases and ROI, Companies, People, India, Indian languages, Claude, OpenAI and the Glossary.

Demand is broad, but most projects still fail to reach measured production

Adoption is wide but shallow. In May 2026, 52% of US employees used AI at work, but only 15% used it daily (Gallup AI indicator). 37% of US firms with 250 or more employees used AI in their operations, against under 20% of firms with four or fewer staff. The Census Bureau changed its question in late 2025, so trends cannot be compared across that change (US Census BTOS). In a representative US survey from late 2024, the time people said they saved came to only 1.4% of work hours (Bick, Blandin and Deming, NBER w32966).

India has the same pattern. It ranks second in total Claude.ai consumer use but 101st of 116 countries per head of population (Anthropic India brief, February 2026). In Anthropic's June 2026 Economic Index release, India had 7.12% of Claude.ai usage in May 2026 but a per-capita usage index of 0.30, where 1.0 means usage in proportion to working-age population (Anthropic Economic Index data). It also has more than 100 million weekly ChatGPT users (TechCrunch, 18 Feb 2026). [Synthesis] The gap between total and per-head use suggests that most potential users, including most smaller firms, have yet to adopt.

Failure rates show why method matters more than model choice. The figures below come from surveys and forecasts, not audits:

• MIT NANDA: 95% of generative-AI pilots showed zero measurable profit-and-loss return. Success meant deployment beyond the pilot with measurable KPIs, with ROI measured six months after the pilot. The report (research period Jan–Jun 2025) reviewed 300+ public initiatives, interviewed 52 organisations and surveyed 153 senior leaders; its authors call the figures "directionally accurate" and interview-based. Projects with external partners reached deployment about 67% of the time, against about 33% for internal builds (52-organisation sample). The document is hosted on mlq.ai, so the MIT attribution rests on the document itself (MIT NANDA report PDF; Fortune on MIT NANDA).
• Gartner: more than 40% of agentic projects will be cancelled by the end of 2027, and only about 130 of the thousands of vendors that call themselves agentic really are. This rests on a webinar poll (Gartner, 25 Jun 2025). Only the release headline was re-read on 12 Oct 2026; the "about 130 vendors" figure and the webinar-poll basis come from the release body, which was not re-read.
• McKinsey: in its 2025 survey, 39% of firms reported any impact on enterprise earnings (EBIT), and for most of them it was under 5% (McKinsey State of AI 2025). Press coverage of the 2026 edition puts the share at 37%, with about 6% of firms as high performers (FM Magazine on McKinsey 2026, secondary).
• Deloitte: only one company in five has a mature way of governing autonomous agents (Deloitte 2026 AI report).
• LangChain: respondents rank quality as the top blocker to putting agents into production; 32% cite it (LangChain State of Agent Engineering).

Where AI works, the task is well scoped and a human is close by:

• Customer support: agents with an AI assistant resolved 14% more issues per hour, and 34% more for novices (Brynjolfsson, Li and Raymond, NBER w31161).
• Consulting: BCG consultants completed 12.2% more tasks, 25.1% faster, on work inside the model's capability frontier. On a task outside it, they were 19% less likely to get the right answer (Dell'Acqua et al., HBS).
• Banking: Bank of America says its EricaAssist tool saves nearly a minute per call for more than 18,000 representatives (vendor-reported) (Bank of America newsroom).

The well-known reversals show the opposite. Klarna started rehiring human agents after its CEO said cost had become "too dominant" and quality fell (Maginative on Klarna). Commonwealth Bank of Australia reversed 45 job cuts after a voice-bot launch, once its union showed that call volumes were rising (Canberra Times). [Synthesis] In both cases, staff cuts came before quality and call volumes had been measured over time.

The roles and skills you need: forward-deployed engineers at the centre, evaluators as the quality gate

Every AI automation programme needs the roles below covered, whether by your own staff, by an implementation partner or by a mix. Role definitions here come from vendor job postings and engineering guidance:

• Anthropic describes an agent as "typically just LLMs using tools based on environmental feedback in a loop", and says tool design deserves as much care as prompts (Anthropic, Building effective agents).
• Anthropic defines context engineering as curating what enters the context window "each time we decide what to pass to the model" (Anthropic, Effective context engineering).
• Anthropic builds evaluation work on tasks, trials, graders and transcripts (Anthropic, Demystifying evals).

The pay columns show what these skills cost to hire, whether you employ them directly or pay for them through a partner's rates. US pay comes from the US Bureau of Labor Statistics (BLS), Stack Overflow's 2025 survey and frontier-lab job boards. Indian pay comes from Stack Overflow, PayScale and blogs. Stack Overflow's figures are medians of total annual compensation (salary, bonuses and perks), as its chart data labels them (re-read 11 Oct 2026). Its Indian role samples are small: 25 AI/ML engineers, 15 data scientists and 74 architects, against 80 US AI/ML engineers.

On Stack Overflow's figures, a US AI/ML engineer earns about 11 times as much as an Indian one (189,500 / 17,436 = 10.9). Indian sources disagree on the level. PayScale's self-reported average base pay of about ₹10.2 lakh sits below Stack Overflow's total-compensation median, and a vendor blog cites a Glassdoor range of ₹6-16 lakh (Tredence). [Synthesis] When budgeting an in-house team in India, treat ₹10-17 lakh as the mid-market band, and expect senior engineers of FDE calibre to cost well above it.

The FDE job description is a template for the work you need done. OpenAI lists 26 FDE-titled roles among its 815 openings (re-read 11 Oct 2026). Palantir says it "pioneered this unique position, embedding talented engineers directly with our customers". The work these posts describe is a written scope, a proof of concept, production on the business's own systems, then reusable patterns. [Synthesis] Whether you hire this profile or buy it from a partner, ask for those four outputs.

The labs' Indian job posts are mostly pre-sales architect roles for "large enterprises across India and the Asia-Pacific region" (Anthropic). OpenAI's Bangalore partner role is about "enabling AWS and partners to independently drive deployments" (OpenAI Partner AI Deployment Engineer). Small and mid-sized businesses should therefore expect to implement with their own team or an implementation partner rather than with a lab directly.

Indian supply is tight. NASSCOM and Deloitte expect India's AI talent pool to grow from about 6 to 6.5 lakh people to more than 1.25 million by 2027. A Deloitte quote in the same article puts skills growth at about 13% a year against 25% to 35% growth in demand (Entrepreneur India); another summary gives a 15% rate, so treat the growth figures as rough. An IIM Ahmedabad and BCG study says India's top 500 firms need 25,000 to 30,000 advanced AI/ML practitioners within three to five years (Careers360 on IIMA study). Global capability centres (GCCs), the in-house offshore centres of foreign companies, compete for the same people. India had more than 1,700 GCCs employing 1.9 million people in FY2024 (NASSCOM GCC report). In July 2026, Naukri's JobSpeak index showed AI/ML hiring up 33% year on year, and GCC hiring up 5% nationally, with Chennai up 34% and Hyderabad up 22% (ANI wire report on Devdiscourse, 3 Aug 2026, secondary; Naukri's own page could not be read).

[Synthesis] A pilot team. This uses the roles above, staffed in-house, by a partner or both. For comparison, one implementation vendor says it staffs teams of 3 to 8 engineers and deploys in 2 to 3 weeks (Ideas2IT, vendor claim).

[Synthesis] Building the in-house team. If you build your own capability, hire FDE-style generalists first, because they can run discovery, build and hand over. Build evaluation skills second; the research found no posted pay for evaluator roles anywhere, yet this is the role that turns a pilot into evidence. Then train juniors as evaluators and approvers. If a partner does the building, keep at least the process owner, the evaluator and the release approver on your side, so that someone in the business can judge the evidence. AI helps novices most (+34% in the support study). In ADP payroll data, workers aged 22 to 25 in the occupations most exposed to AI are about 19% below the employment trend of less-exposed peers, mainly through reduced hiring; the authors call the finding descriptive, not causal (Stanford Digital Economy Lab). A business that keeps training juniors builds a pipeline others are cutting.

Certifications cost US$200 or less each, and three vendor tracks have changed

Certifications matter twice: when you train your own staff and when you judge a partner's claims. Earlier research listed Microsoft's AI-102 exam as current. Its page now says it is retired, so this table follows the newer pages. Anthropic launched its certifications with the Claude Partner Network in March 2026, and AWS replaced the English version of its ML Engineer exam with a beta in September 2026. Fees are list prices before tax and partner discounts. The Anthropic, AI-103, AI-500, Google Professional ML Engineer, Databricks, NVIDIA and AWS rows were re-read on 11 Oct 2026.

Discounts. Staff of Anthropic's Select, Preferred and Global Premier partner firms get 50% off exams, and Global Premier partners pay nothing until 31 Dec 2026. Microsoft's Exam Replay offer bundles one exam voucher with one retake, both usable within 12 months (Microsoft offers).

[Synthesis] Certifications are a signal, not a legal requirement. The European Commission says the EU AI Act's AI-literacy duty requires no certificate; an internal training record is enough (European Commission AI literacy Q&A). For your own staff, spend accordingly:

• One vendor certificate per senior engineer, matched to the cloud platforms your business uses.
• AB-620 or AI-103 if you build on Microsoft.
• The free courses for everyone, with completions kept as the training record.

At list prices of US$125 to US$200 an exam, one exam each for four engineers costs US$500 to US$800 (₹48,000 to ₹76,800).

When you evaluate a partner, check that its certificates are current and match your platforms. Claude certificates last 12 months (renewal is free if done on time) and are open only to staff of Claude Partner Network firms; only the Developer and Architect exams count toward partner eligibility. A certificate shows that someone passed an exam, not that the partner has shipped a system like yours.

How to run and gate a project: a nine-stage method grounded in vendor guidance

No single vendor publishes the whole method. Together, Microsoft, AWS, Google, OpenAI and Anthropic cover every stage. They agree on the order: rule out agents where code will do, score and rank, pilot the hardest step, let business metrics decide, invest in tests and tools, then run with telemetry. Use the stages to run your own projects, and as the yardstick for any partner's proposal: each exit gate is a point where you can stop paying.

Score opportunities on CAF's three axes plus three tie-breakers

CAF's business impact covers strategic fit, value and change timeframe. Technical feasibility covers risks, safeguards and technology fit. User desirability covers the main users, their value proposition and their resistance. [Synthesis] For SMEs, add three tie-breakers: volume times error cost, data readiness, and the agent privilege required (lower is better). The privilege test follows OWASP and the Five Eyes guidance, which both say to start narrow and low-risk. MIT NANDA found that more than half of generative-AI budgets went to sales and marketing tools, while the biggest measured returns came from back-office automation (Fortune, above). So score document processing, reconciliation and support triage before marketing content.

Pilot KPIs that act as go or stop gates

Measure quality and throughput together. Klarna and Commonwealth Bank both judged on cost alone.

[Synthesis] Gate rules.

• Go if four things hold: the pass rate meets the threshold agreed before the build (pass^k for customer-facing flows), quality is no worse than the baseline, cost per task is within budget, and no high-severity security finding is open.
• Change direction if easy cases pass but the hardest step fails. Narrow the scope or use a simpler design.
• Stop if safeguards remain unclear, if the data or the consent to use it cannot be secured, or if the hardest step still fails after two rounds of fixes.

Evaluations, hardening and monitoring in practice

Start the tests small. Anthropic says "20-50 simple tasks drawn from real failures is a great start", and regression tests "should have a nearly 100% pass rate". A task that fails all of 100 attempts usually means the task is broken, not the agent. Anthropic's research team began with about 20 test questions, and one prompt change moved success from 30% to 80%. Hamel Husain's three levels set the rhythm: automated checks on every change, regular review of logged runs, and A/B tests after major changes.

When an AI model grades outputs, check it against people. Eugene Yan finds side-by-side comparison "more stable" than scoring outputs one at a time. OpenAI's GDPval study shows why the check matters: its automated grader agreed with human experts 66% of the time, while human experts agreed with each other 71% of the time (GDPval, arXiv).

Fix the tools before the prompt. Building its coding agent, Anthropic "spent more time optimizing our tools than the overall prompt"; requiring absolute file paths removed a whole class of errors. An agent that rewrote tool descriptions cut task time by 40%. OpenAI notes that some agents handle more than 15 distinct tools while others struggle with fewer than 10 overlapping ones. Every loop needs exits: a final-output tool, a reply with no tool call, an error, or a turn limit. Claude Managed Agents, still in beta, added per-session budgets with a hard spending cap in August 2026 (Claude Managed Agents).

Train approvers, and plan for model changes. The capability-frontier result shows approvers must learn where AI fails. Expect frequent model migrations: Anthropic launched new models on seven dates between 9 June and 7 October 2026, and its release notes list breaking API changes for Opus 5.5 and Sonnet 5.5 (Claude release notes).

Six reference architectures, ordered by how much autonomy they grant

Anthropic separates two kinds of system:

• Workflows, where "LLMs and tools are orchestrated through predefined code paths".
• Agents, where "LLMs dynamically direct their own processes and tool usage".

It says to add complexity "only when it demonstrably improves outcomes". Google's agents white paper describes an agent as a model, tools and an orchestration layer (Google Agents whitepaper). AWS maps reasoning, retrieval, coding, voice, workflow and multi-agent patterns to its services (AWS agentic AI patterns).

The costs in the table below are model charges per task at list prices. The low figure assumes a small model (Haiku 5.5 or GPT-6 Luna) and the high figure a mid-tier one (Sonnet 5.5 or GPT-6.1 Sol). The token counts per task are assumptions (Claude pricing docs; OpenAI pricing; details in Costs and economics). Haiku 5.5 bills prompts above 100K tokens at $0.50 / $2.50 per million instead of $0.10 / $0.50, so the low figure is a lower bound for long agent loops on Haiku. Reasoning (thinking) tokens bill as output tokens at Anthropic and OpenAI, even when the thinking text is not returned (Claude extended thinking; OpenAI reasoning guide). OpenAI also reprices the whole request above 272K input tokens (GPT-6.1 Sol $4 / $15; GPT-6 Luna $0.20 / $0.75) (OpenAI pricing). Long or reasoning-heavy loops therefore cost more than these figures.

Workflows and RAG should be the SME default. CAF says standard RAG follows "a deterministic retrieval process", whereas agents add "nondeterministic behavior that requires robust testing and governance". Start with pgvector inside an existing Postgres database (pgvector repo) and the MIT-licensed Docling parser (Docling repo). Add a dedicated vector database only when scale demands it. OCR for Indian scripts and Hindi search quality were not tested for any tool, so pilot them on your own documents first (options in Indian languages).

The single tool-using agent does most agent work. Agents use about 4 times the tokens of chat, so cap the turns. Browser-controlling agents are a fragile special case. The best model in the OSWorld 2.0 paper (June 2026), Claude Opus 4.8 at maximum thinking, completed only about 20.6% of 108 long-horizon tasks under strict scoring. Later leaderboard rows put Claude Opus 5 at 31.4% (Epoch's copy, secondary) (OSWorld 2.0 paper). Anthropic warns that its model "will follow commands found in content even when they conflict with your instructions" (Anthropic computer-use docs).

Use multi-agent designs only where the value covers the tokens. Anthropic's research system beat a single agent by 90.2% on an internal benchmark, and running subagents in parallel "cut research time by up to 90%". Token usage alone, though, explained 80% of the performance variance on the BrowseComp test, so much of the gain is bought with tokens.

Human approval is a design element, not an add-on. OWASP says to require human approval for high-impact actions and to enforce permissions in the systems the agent calls, not in the model. GDPR Article 22(3) gives people the right to human intervention in automated decisions (GDPR Art. 22). Colorado's 2026 law lets consumers request "meaningful human review and reconsideration" (Colorado SB26-189).

Cost voice agents per minute, all in. Retell's example adds up to $0.11 a minute before phone charges: $0.04 for the LLM, $0.055 for voice infrastructure and $0.015 for text-to-speech. LiveKit charges $0.01 a minute per agent session plus model costs (LiveKit pricing), and Bland $0.12-0.14 a minute (Bland pricing). For Indian languages, Sarvam charges ₹3 per 1,000 characters of speech output and ₹30 per hour of transcription (Sarvam API pricing); Indian languages compares Indian-language speech, models and OCR. In India, budget for telephony too (Plivo lists ₹0.38 and ₹0.25 a minute, vendor-reported) (Plivo India voice pricing) and check TRAI's rules. In TRAI's 2026 draft amendment, a fully automated voice agent appears to fall within the A2P call category (TRAI-hosted comments on the draft); the final text was reported as notified on 18 Sep 2026 (secondary; not read), and no explicit AI-voice rule is confirmed. Meesho says its voice bot cut some support-call costs by 75%, a company claim with no stated baseline (TechCrunch on Meesho).

Starter stack: plain code, permissive licences and a deprecation calendar [Synthesis]

The choices in this table are recommendations; the licence terms and prices in it are sourced. The principles come from the vendor guidance: call model APIs directly first, adopt a framework only when the team understands what it does underneath, build tests before scaling, and trace every run from day one. Fuller comparisons are in Tool stack and Agent frameworks.

Plan for churn. In 2026, OpenAI shut down its Assistants API (26 Aug) (OpenAI changelog), and Claude Sonnet 4.5 retires on 30 Nov. OpenAI's deprecations page also shuts legacy snapshots such as o4-mini, o3-mini, gpt-4.1-nano and gpt-image-1 on 23 Oct 2026, and gpt-5.3-codex, gpt-5.4-nano and gpt-5.1 on 1 Apr 2027, so do not build new work on those models (OpenAI deprecations, above). Three of the testing and tracing tools above changed owners. Keep the model layer thin, version every prompt and keep a regression test suite; then a migration takes days, not weeks.

AI coding tools for your developers are cheap. Claude Code comes with Claude Pro ($20 a month billed monthly, or $200 a year) and Max ($100 or $200) (Claude pricing). It is proprietary software, not open source: its licence reads "All rights reserved" and puts use under Anthropic's Commercial Terms (Claude Code licence). GitHub Copilot costs $10 a month for Pro, $39 for Pro+ and $100 for Max, with a free tier (GitHub Copilot plans). Cursor costs $20 a month, or $40 per user for teams (Cursor pricing). Devin's Team plan is $80 a month plus $40 per user (Devin pricing). The Gemini CLI free tier allows up to 1,000 requests a day (Gemini CLI quotas). These are US-dollar prices. Anthropic reportedly bills Claude plans in rupees from 13 Jul 2026, with GST included (Pro ₹2,399 a month, or ₹2,000 a month on annual billing; single-source) (Techstory on rupee billing); UPI is not supported. GST on foreign AI services was not checked, so get tax advice.

What implementation work costs and how it is priced: a buyer's view

Implementation partners price their work in four main ways: fixed-fee assessments and build sprints, monthly managed service, fees per outcome, and seat or credit subscriptions. The figures below are published prices. Use them to benchmark quotes and to compare buying with building: in-house pay is in the roles table above, and model cost per task in the architecture table.

Assess, build and run: what published partner prices look like

Main & Machine budgets model usage and third-party subscriptions separately from its fees, and Softomate quotes care, licences, hosting and usage separately. [Synthesis] Check whether any quote includes model usage, licences and hosting or leaves them to you, and who carries the cost when token prices change.

Indian firms rarely publish package prices. Carmatec says costs "vary based on the complexity of processes" (Carmatec). Infyom prices by "workflow complexity and hosting preference" (Infyom). Antino says clients "pay for outcomes shipped rather than hours logged" (Antino). At ₹96 to the dollar, the US assessment price alone is ₹3.4 to 8.2 lakh. [Synthesis] That is a large sum for a diagnosis at many Indian SMEs. Ask for a short, fixed-scope discovery, or for the fee to be credited against the build, as Main & Machine and OptiWork do.

Fixed-scope packages

Some firms sell named packages with a fixed scope, price and timeline, like Softomate's menu, OptiWork's starting prices by type of work, or Main & Machine's fixed sprint quotes. [Synthesis] A fixed price is safest for you when the package maps to one reference architecture, such as a document-extraction workflow, a RAG policy assistant, a support-triage agent with human approval or an Indian-language voice line, and comes with a known test suite, cost per task and set of controls. Ask which architecture a package assumes, and what happens if your workflow does not fit it.

Outcome-based pricing

At list prices, the model cost of one five-turn resolution is $0.0027 on a small model and $0.052 on a mid-tier one. The $0.99 to $2.00 outcome fees are therefore 19 to 38 times the mid-tier cost, and roughly 370 to 750 times the small-model cost (own arithmetic). The gap pays for retries, testing, monitoring, integration upkeep and human review, so compare an outcome fee with your own fully loaded cost per resolution, not with the model cost. AWS advises charging for agents "through usage-based models, internal ROI metrics, and commercial offerings", which covers charging AI costs back to internal teams as well as vendor pricing.

[Synthesis] Accept outcome pricing only when your own system can count the outcome, baseline volumes are known, and the contract defines what counts, the way Fin bills at most one outcome per conversation.

Digital-worker and seat subscriptions

[Synthesis] A "digital worker" subscription, where an AI agent does a defined job for a monthly fee, is worth buying when it bundles a fixed scope (such as invoices processed), a monthly cap, named human supervision and a monthly KPI report. Compare its price with the seat and outcome prices above, and be wary of any vendor that presents the "worker" as a person.

Partner badges and directories: what they tell you

Large IT services firms already deliver AI at scale. TCS reported annualised AI revenue above $3B for Q2 FY27 (about $3.1B in one outlet), more than 10% of revenue and up from $2.6B in Q1 (8 Oct 2026; secondary, several outlets agree) (Outlook Business on TCS Q2). Coverage of Infosys's June 2026 quarter is headlined "Infosys trims guidance as AI revenue climbs" (TechMarketView on Infosys, secondary; body unreadable, so no AI share is quoted). Accenture said on its results call that more than 400 clients began their first advanced-AI work with it in FY26 (Motley Fool transcript of Accenture Q4 FY26 call, secondary). Accenture's Q4 FY26 release (8-K, 1 Oct 2026) gives revenue of $18.7B and new bookings of $22.17B but no generative-AI bookings figure (Accenture 8-K exhibit). [Synthesis] Size and partner tier do not show whether a partner suits your project. Large firms bring scale and top partner tiers; smaller specialists may fit SME-sized projects, Indian languages or WhatsApp better. Judge either kind on the evidence in the next section.

[Synthesis] Budgeting an implementation

• Budget for three phases, not one. Published US prices follow three steps: a fixed-fee assessment, a fixed-price build sprint and a monthly managed service. Add model usage, licences and hosting, which are often quoted separately, and the time of your own process owners and approvers.
• For Indian SMEs, a shorter sequence may fit better: a short, low-cost discovery, a fixed-price build of a single workflow, then a small monthly fee covering monitoring, test reruns and model migrations.
• Check guarantees. Some firms offer time-bound delivery guarantees and fee credits (see the table above); ask whether yours will.
• Read price-review clauses. If prices can be reviewed when token prices or exchange rates move, the clause should work in both directions.
• Compare like with like. Judge quotes against these benchmarks, against in-house pay and tool costs, and against hours and cost per task measured in your own pilot, not against token estimates.

How to evaluate an implementation partner [Synthesis]

A good partner can help: in MIT NANDA's interview sample, projects with external partners reached deployment about 67% of the time, against about 33% for internal builds. That rests on 52 organisations, so choose a partner on evidence, not on the statistic. The checks below are built from the facts in this playbook.

When your team or partner delivers with AI agents: review and disclosure duties

Using AI coding, testing and drafting agents to build automation lowers cost, whether your own developers use them or a partner does. It also creates three duties: check the output, contain the agents, and be clear about which work is AI and which is human.

The productivity evidence is mixed:

• Field experiments: the published abstract reports three randomised experiments at Microsoft, Accenture and an anonymous Fortune 100 company, with 4,867 developers and 26.08% more completed tasks (SE 10.3%), and larger gains for less experienced developers (Cui et al., Microsoft Research). The authors' slides report 5,114 developers at Dell, Accenture and Microsoft and pooled pull-request gains of about 25% to 31% (column mapping provisional; Dell's estimate not significant) (Cui et al. slides). The two versions are not reconciled.
• Randomised trial: in METR's 2025 trial, 16 experienced developers took about 19% longer with AI tools, yet afterwards believed they had been 20% faster (METR, 2025 study). In a follow-up (24 Feb 2026), raw estimates were 18% less time for 10 returning developers and 4% less for 47 new ones; both confidence intervals include zero, and METR called the signal unreliable (METR, 24 Feb 2026).
• Developer survey: 46% of developers distrust AI accuracy, and 66% name "AI solutions that are almost right, but not quite" as their top frustration (Stack Overflow 2025, AI).
• Google: says about 75% of its new code is AI-generated and approved by engineers (vendor-reported) (ETV Bharat). Anthropic says that for coding agents "human review remains crucial".

The lesson is to measure productivity with timed records, not with what people say they saved. Anthropic's harness for long-running coding agents shows how to structure supervised work (Anthropic, Effective harnesses). A first agent writes a setup script, a progress log and a first git commit. Later agents build one feature at a time from a JSON feature list in which every feature starts marked as failing; JSON was chosen because the model is "less likely to inappropriately change or overwrite" it than Markdown. Every session ends with committed, working code, which leaves the reviewer small units to check.

Review duties [Synthesis]

Contain coding and delivery agents

Automatic approval by a classifier reduces risk but has been bypassed. Claude Code's auto mode became the default for Pro, Max and Team plans on 14 Aug 2026. In Anthropic's own test (vendor-reported), only 13.6% of 1,053 testers refused a dangerous command slipped into their session, while auto mode would have blocked 89%. On 27 Aug 2026, researcher Johann Rehberger reported bypassing auto mode about 80% of the time (Simon Willison, prompt-injection tag). Willison recommends running unattended agents in a container or virtual machine with no access to home directories, SSH keys or cloud credentials.

Three 2025 incidents show the cost of skipping this. Malicious versions of the Nx package used AI command-line tools on developers' machines to steal credentials (Wiz on AI-invoking malware). A compromised Amazon Q extension tried to run its agent with --trust-all-tools --no-interactive (same source). A Replit agent deleted a production database despite a code freeze, then produced false test results (The Register on Replit). The minimum safeguards are separate environments, no blanket tool permissions, outbound allowlists, and human approval for anything destructive.

Disclosure duties

[Synthesis] Being honest about AI use. Label every AI agent as AI wherever its name appears, with no human photos or job titles. State what humans actually approve. Make chat and voice agents on your own site and channels say they are AI. Expect the same of any partner: it should describe its human team accurately, name people only with consent, say which work its AI agents did and who reviewed it, and publish your results only with your written consent and only if true.

Data terms and contract clauses to ask for

Model vendors' terms set the limits on what any implementation partner can promise you:

• Anthropic: the customer owns outputs and Anthropic does not train on customer content. Its IP indemnity excludes customer inputs, modifications and combinations with other technology, and liability is capped at the previous 12 months' fees (Anthropic Commercial Terms). Zero data retention needs approval, and Anthropic still keeps safety-classifier results (Anthropic ZDR article). Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 require 30-day retention, so ZDR is unavailable on them unless Anthropic expressly authorises it. Batch (29-day retention), the Files API, code execution, programmatic tool calling, the MCP connector, Agent Skills and Claude Managed Agents are not ZDR-eligible for any model. Flagged chats may be kept up to 2 years (Anthropic API and data retention).
• OpenAI: keeps abuse-monitoring logs "for up to 30 days" by default, and zero retention requires prior approval. Its managed Agents API, Assistants, threads and vector stores keep data until deleted and are not eligible for zero retention (OpenAI data controls).
• Microsoft: its copyright commitment for Azure OpenAI applies only if the customer uses a required system prompt and keeps a testing report (Microsoft Customer Copyright Commitment).
• Google: its indemnity lists Google's own models, not third-party ones (Google indemnified services).
• AWS: excludes cases where the customer fails to turn on the available safety filters (AWS Service Terms).

[Synthesis] Clauses to ask for in a partner contract:

• Disclosure of AI use in delivery, and a commitment to human review of every release and of any decision about a person.
• A list of the models and sub-processors, with notice before any change.
• Data purpose, retention and cross-border terms that match DPDP and the vendors' own terms.
• No training on your data without written consent.
• An indemnity you have checked against the model vendors' exclusions (inputs, modifications, combinations, disabled filters), since those cover exactly what a partner builds.
• For outcome-priced work, an exact definition of what counts as an outcome.
• A handover of versioned prompts, test suites, configuration and run logs, so you can change models or partners.
• Confirmation that every component's licence allows the way it is hosted for you.
• Breach-notice timings that let you inform affected people and the Data Protection Board "without delay" (Rule 7(1), 7(2)) and file the detailed report "within seventy-two hours" (Rule 7(2)), from 13 May 2027 (DPDP Rules, G.S.R. 846(E); commencement, G.S.R. 843(E)). For cyber incidents, also allow for CERT-In's 6-hour reporting rule (CERT-In s.70B Directions, 2022).
• If prices can be reviewed when token prices or exchange rates move, a clause that works in both directions.

Risks and pitfalls that sink automation programmes

Four free frameworks cover most of these risks. The OWASP Agentic Top 10 runs from ASI01 Agent Goal Hijack to ASI10 Rogue Agents, built on "Least Agency". MITRE ATLAS lists 120 attack techniques and 73 case studies in its September 2026 release. The CSA AI Controls Matrix has 243 control objectives in 18 domains, mapped to ISO/IEC 42001 (CSA AICM). Google SAIF adds controls for agent observability, permissions and user control (Google SAIF controls). ISO/IEC 42001 certification is not inherited from a cloud provider: "Your organization is not automatically certified by association" (AWS ISO 42001 FAQ). Pursue it only if a customer requires it, and do not accept a cloud provider's certificate as a partner's own. See Governance and security.

90-day plan for your business [Synthesis]

Day counts assume a start in mid-October 2026, so the dated platform items below fall inside the windows shown.

12-month plan [Synthesis]

Months count from the same mid-October 2026 start, so month 3 ends in mid-January 2027 and India's main DPDP duties (13 May 2027) start at the end of month 7.

Platform and compliance calendar. All dates come from the sources cited above.

What this means for your business

[Synthesis] Start with method, not models. Most organisations still fail to turn pilots into measured gains (MIT NANDA's 95%; 37% to 39% reporting any EBIT impact). Rule out agents wherever plain code will do, score use cases on CAF's three axes, pilot the hardest step against go or stop KPIs signed before the build, and accept results only from test suites and timed baselines.

Decide who builds, but keep the judgement in-house. The labs' Indian hiring is mostly pre-sales for large enterprises, so most businesses will work with their own team or an implementation partner. In MIT NANDA's small sample, partner-led projects reached deployment about twice as often as internal builds. Either way, keep the process owner, the evaluator and the release approver on your side. If you hire in India, budget ₹10-17 lakh for mid-market AI engineers and more for FDE calibre, and allow for tight supply (AI/ML hiring up 33% year on year in July 2026).

Budget for three phases plus running costs. Published US prices run from $3,500-8,500 for an assessment and $18,000-60,000 for a build sprint to managed service from $1,500 a month; usage, licences and hosting are often extra. Model cost is about $0.0001 to $0.10 per task for workflows, RAG and single agents at list prices, so judge $0.99 to $2.00 outcome fees against your fully loaded cost.

Deadlines. Migrate off OpenAI Agent Builder, Evals and Claude Sonnet 4.5 by 30 November 2026, move Indian WhatsApp accounts to INR billing by 31 December 2026, and prepare for Colorado's laws on 1 January 2027 and India's main DPDP duties on 13 May 2027.

Risks and tests. The main risks are prompt injection, over-privileged tools, runaway token costs and cutting staff before quality is proven. Test with pass^k, injection test sets and a second model provider, and measure productivity with timed records, not self-reports.

Ask any partner which design it proposes and why not a simpler one; to see its test suite, cost per task and transcripts; which models, regions and sub-processors will handle your data; which work its AI agents do and who approves it; what its quote leaves out; whether it earns referral fees on tools it recommends; and what it hands over if you change partner.

Open questions and unverified items

• Pay: no role-specific pay was found for evaluators, LLMOps engineers, AI governance leads or context engineers. OpenAI and Anthropic hide pay on Indian postings, and AmbitionBox, Glassdoor, Naukri and Indeed could not be read. Stack Overflow's Indian medians rest on small samples (25 AI/ML engineers). Hourly and monthly rates of Indian firms and directories had no readable primary source and are left out, as are some "AI employee" product prices.
• Certifications: Microsoft's exam prices (its pages say only that price depends on the country), Google's validity periods, NPTEL exam fees and FutureSkills Prime's AWS terms were not found. OpenAI's certification rests on secondary sources. Anthropic's newsroom lists "$100 million to train 10,000 engineers" (2 Oct 2026); only the headline was read (Anthropic newsroom).
• Method: McKinsey's "One year of agentic AI" and Google's November 2025 "Introduction to Agents" could not be read, and no primary source on process mining was found.
• Partners: Claude Partner Network tier criteria, referral fees and OpenAI's own partner terms were not found.
• Compliance: open are Colorado's final rules, and whether the X.AI v. Weiser stay reaches SB 26-189; whether CERT-In's 10 Jun 2026 guidelines bind service firms (their body is unread; the 2022 s.70B Directions do bind, with 6-hour reporting); whether India's labelling rule covers synthetic voices on live calls; TRAI's final TCCCPR Third Amendment text (reported notified 18 Sep 2026, secondary; not read); and the RBI FREE-AI report PDF (the press release is primary: a report, not a direction).
• WhatsApp: two vendor pages describe 1,000 free service messages per number per month from 1 Oct 2026, then the utility rate (₹0.1150 excl. GST), while Meta's page (11 Oct, summarising fetch) still says they are free: vendor-reported, not confirmed. Gallabox's rupee rates are 1.15 times the other vendor's, for reasons it does not state.
• Infrastructure: no vendor's own API processes requests in India (OpenAI's India option covers storage at rest only). AWS Bedrock offers India-only routing for the previous Claude generation, Kimi K3 and GPT-5.6 Terra/Luna, but not for Claude 5.5 or GPT-6. AWS's regional-availability page lists Claude Opus 5 and Sonnet 5 as In-Region in Mumbai on bedrock-mantle, but the model cards omit it. Indian-language OCR and search quality are untested; no independent Indic WER or document test was found.
• Prices and vendor counts: Fin's customer counts conflict (Salesforce: 30,000+ companies; fin.ai: 12,000+ businesses). Zendesk's US committed rate rests on an earlier read and secondary sources (the US page redirects). The Microsoft 365 Copilot Business promotion dates conflict (secondary), and ChatGPT Business and Enterprise seat prices are secondary or unresolved. Rupee invoicing for the Claude API is unresolved.

How conflicting sources were resolved

Sources

1. X-Rates monthly average
2. Gallup AI indicator
3. US Census BTOS
4. Bick, Blandin and Deming, NBER w32966
5. Anthropic India brief
6. TechCrunch, OpenAI and Tata data centre deal, 18 Feb 2026
7. Fortune on MIT NANDA
8. Gartner, 25 Jun 2025
9. McKinsey State of AI 2025
10. FM Magazine on McKinsey State of AI 2026
11. Deloitte 2026 AI report
12. LangChain State of Agent Engineering
13. Brynjolfsson, Li and Raymond, NBER w31161
14. Dell'Acqua et al., HBS
15. Bank of America newsroom
16. Maginative on Klarna
17. Canberra Times, CBA reverses AI job cuts
18. Anthropic, Building effective agents
19. Anthropic, Effective context engineering
20. Anthropic, Demystifying evals
21. Microsoft AB-100
22. Anthropic Applied AI Architect, Bangalore
23. Stack Overflow 2025, Work
24. OpenAI FDSE posting
25. OpenAI job board API
26. Palantir FDSE
27. Anthropic Manager FDE
28. OpenAI Applied AI Engineer, Delhi
29. Sierra posting
30. PayScale ML engineer India
31. Storyboard18, AI and ML engineer pay in India
32. Databricks certification
33. Eugene Yan, LLM evaluators
34. Hamel Husain, evals
35. BLS software developers
36. BLS data scientists
37. PayScale data scientist India
38. OpenAI practical guide
39. Palantir job board API
40. Sierra job board API
41. AWS, Operationalizing agentic AI
42. Microsoft AB-731
43. Tredence, AI engineer salary 2026
44. OpenAI Partner AI Deployment Engineer
45. Entrepreneur India
46. Careers360 on IIMA study
47. NASSCOM GCC report
48. Devdiscourse (ANI), Naukri JobSpeak July 2026
49. Ideas2IT homepage
50. Stanford Digital Economy Lab, Canaries in the coal mine
51. Anthropic certification FAQ
52. Architect – Foundations
53. Architect – Professional
54. Microsoft AI-103
55. Microsoft renewal
56. Microsoft AI-500
57. AB-620 study guide
58. Microsoft AB-730
59. AWS AI Practitioner
60. AWS MLA
61. Google GenAI Leader
62. Google PMLE
63. Databricks GenAI Engineer
64. NVIDIA NCA-GENL
65. Hugging Face agents course
66. Anthropic Academy
67. Claude FDE path
68. NPTEL FAQ
69. FutureSkills Prime
70. Gend.co, OpenAI certifications
71. Microsoft AI-102 page
72. Microsoft offers
73. European Commission AI literacy Q&A
74. Microsoft CAF business plan
75. Microsoft CAF agents
76. Hogan Lovells on DPDP
77. OWASP LLM06
78. OWASP Agentic Top 10 PDF
79. Mayer Brown on Five Eyes guidance
80. Anthropic, Writing tools for agents
81. Copilot Studio guidance
82. Anthropic multi-agent research system
83. Salesforce on Fin (returned 404 on 12 Oct 2026; not re-read)
84. GDPval, arXiv
85. Claude Managed Agents
86. Claude release notes
87. Google Agents whitepaper
88. AWS agentic AI patterns
89. Claude pricing docs
90. OpenAI pricing
91. Retell pricing
92. pgvector repo
93. Docling repo
94. OSWorld 2.0 paper, arXiv 2606.29537
95. Anthropic computer-use docs
96. GDPR Art. 22
97. Colorado SB26-189
98. LiveKit pricing
99. Bland pricing
100. Sarvam API pricing
101. TechCrunch on Meesho
102. Claude data residency
103. LiteLLM pricing
104. OpenRouter pricing
105. OpenAI Agents SDK repo
106. Claude Agent SDK repo and README
107. Google ADK repo
108. LangGraph repo
109. Microsoft Agent Framework releases
110. OpenAI deprecations
111. Google Agent Starter Pack README
112. Temporal pricing
113. Activepieces pricing
114. n8n licence
115. Flowise discussion
116. Dify repo
117. Qwen3-Embedding card
118. BGE-M3 card
119. Pinecone pricing
120. Unstructured pricing
121. Langfuse pricing
122. LangSmith pricing
123. Phoenix repo
124. Helicone joins Mintlify
125. Snyk acquires Invariant Labs
126. Llama Guard 4 card
127. Playwright MCP repo
128. Browser Use repo
129. Skyvern repo
130. vLLM repo
131. TGI repo
132. Lambda pricing
133. Meta WhatsApp pricing
134. OpenAI changelog
135. Claude pricing
136. Claude Code LICENSE.md
137. GitHub Copilot plans
138. Cursor pricing
139. Devin pricing
140. Gemini CLI quotas
141. Main & Machine pricing
142. Main & Machine AI implementation
143. OptiWork.ai homepage
144. Softomate guide
145. Carmatec AI automation page
146. Infyom n8n automation services
147. Antino AI services
148. Fin pricing
149. Zendesk pricing
150. Sacra on Decagon
151. Microsoft 365 Copilot pricing
152. Microsoft Copilot Studio pricing
153. Agentforce pricing
154. Globant AI Pods
155. Anthropic, Claude Partner Network
156. Claude service partners
157. Microsoft Partner Success Core
158. Google Cloud Partner Network
159. n8n Partners
160. Zapier partner directory
161. Wikipedia, OpenAI
162. Outlook Business on TCS Q2 FY27 (secondary)
163. TechMarketView on Infosys, 24 Jul 2026
164. Motley Fool transcript of Accenture Q4 FY26 call
165. Cui et al. slides
166. METR, early-2025 experienced developer study
167. METR, uplift study update (24 Feb 2026)
168. Stack Overflow 2025, AI
169. ETV Bharat, AI-generated code at Google
170. Anthropic, Effective harnesses
171. Simon Willison, prompt-injection tag
172. Wiz on AI-invoking malware
173. The Register on Replit
174. EU AI Act Article 50
175. NatLawReview on the AI Omnibus
176. Addleshaw Goddard on the final AI Omnibus
177. Colorado AG rulemaking
178. NYC AEDT FAQ
179. Sidley on the FTC rule
180. Bar and Bench on FREE-AI
181. Anthropic Commercial Terms
182. Anthropic ZDR article
183. OpenAI data controls
184. Microsoft Customer Copyright Commitment
185. Google indemnified services
186. AWS Service Terms
187. Willison on Atlas
188. Brave on Comet
189. Invariant on GitHub MCP
190. MITRE ATLAS changelog
191. HN listing, SpaceX and Cursor headlines
192. xAI regional endpoints
193. GPT-6 Astra system card
194. Anthropic newsroom
195. Claude Opus 5.5
196. ISACA AI Pulse Poll
197. PagerDuty survey
198. TeamLease RegTech on CERT-In
199. CSA AICM
200. Google SAIF controls
201. AWS ISO 42001 FAQ
202. Gemini API pricing
203. Anthropic Economic Index data, June 2026 release
204. MIT NANDA, State of AI in Business 2025 (PDF hosted on mlq.ai)
205. fin.ai homepage (vendor-reported)
206. Salesforce completes acquisition of Fin, 10 Sep 2026 (vendor)
207. Claude extended thinking
208. OpenAI reasoning guide
209. Plivo India voice pricing (vendor)
210. TRAI-hosted comments on the draft TCCCPR Third Amendment, April 2026
211. Promptfoo blog
212. Llama 4 Community License
213. IndiaAI price calculator
214. IndiaAI eligibility criteria
215. E2E Networks H100 pricing (vendor)
216. Techstory on Claude rupee billing in India (single-source)
217. Zendesk Germany pricing
218. Microsoft Learn, Copilot Studio billing and licensing
219. Accenture Q4 FY26 earnings, 8-K exhibit
220. Cui et al., Microsoft Research publication page
221. X.AI LLC v. Weiser docket (CourtListener)
222. iPleaders on the IT Rules 2026 (secondary)
223. Gala Law on the 2026 IT Rules amendment (secondary)
224. RBI press release on the FREE-AI report, 13 Aug 2025
225. Anthropic API and data retention
226. DPDP Rules, G.S.R. 846(E), Gazette of India
227. DPDP commencement, G.S.R. 843(E), Gazette of India
228. CERT-In s.70B Directions, 28 Apr 2022
229. Bedrock regional availability
230. CERT-In guideline CISG-2026-03 listing
231. AI Act Service Desk, implementation timeline