AI Automation Use Cases and ROI Evidence

Key takeaways

Facts in this document are as of 12 October 2026. The main study figures, government evaluations and vendor claims were re-checked against primary sources that day; facts not re-read keep their original read dates.

• The strongest evidence is in customer support and clinical documentation, not autonomous agents. A field study of 5,179 support agents found 14% more issues resolved per hour (34% for novices). In three ambient-scribe trials, burnout or exhaustion scores improved in every arm, but time savings were smaller and uneven.
• Gains are real but narrow. BCG consultants were 25.1% faster on tasks inside AI's capability frontier and 19% less likely to be correct on a task outside it. Experienced open-source developers in one trial took about 19% longer with AI while believing they were 20% faster.
• Most organisations use AI; few have measured a profit. Press coverage of McKinsey's 2026 survey puts the share of respondents reporting any enterprise-level EBIT impact at 37%, about the same as in 2025 (secondary); in Deloitte's 2026 report, 66% report productivity gains but only 20% increased revenue. MIT NANDA's contested headline is that 95% of GenAI pilots showed zero measurable P&L return, with ROI judged six months after the pilot. The authors call their interview-based figures "directionally accurate".
• Most company figures are self-reported. Vendor pages report wins such as drafting cut from 15 hours to 15 minutes with no baseline or sample. Large banks' primary disclosures are mostly adoption counts or a single time saving.
• Reversals happen where cost targets came before quality data. Klarna began rehiring human agents in 2025. Commonwealth Bank of Australia reversed 45 voice-bot job cuts after the union argued that call volumes were rising.
• The first labour effect is fewer entry-level hires. Stanford's August 2026 update puts 22-to-25-year-olds in the most AI-exposed occupations about 19% below expected employment, driven by reduced hiring. India's five largest IT firms shed a net 7,389 staff in FY26, according to secondary reports, with AI one cause among several.
• What works is plain and measurable: high-volume tasks with checkable results, a baseline before the pilot, a quality metric beside every efficiency metric, and a route to a human.

How to read the evidence in this document

Mixing evidence grades is how most ROI claims go wrong: a randomised trial and a vendor case study can both report a percentage, but the numbers carry very different weight. This document labels every claim with its grade and, where research rounds conflict, uses the primary source and says why. Some primary pages (the World Economic Forum's report, OpenAI's website, several Indian investor sites) blocked access, so figures from them rest on labelled secondary reports.

The second distinction is between adoption metrics (users, seats, agents deployed, conversations handled, AI product revenue) and outcome metrics (time per task, cost per contact, error rate, customer satisfaction, revenue or profit). "65,000 staff use our LLM platform" is adoption; "calls are a minute shorter" is an outcome. Most 2026 headline numbers are adoption metrics. Related documents in this set cover adjacent ground: Costs and economics for unit prices, Governance for risk controls, Implementation playbook for delivery method, India for the domestic market and Future outlook for forecasts. Terms are defined in the Glossary.

Customer support and coding lead the function catalogue

Ranked by quality of evidence rather than volume of marketing, customer support is the most mature function in October 2026: it alone has a large field study, large company deployments, published per-outcome prices and two documented reversals, which together show what works and what fails. Software development has the most randomised evidence, but the studies disagree. Clinical documentation has the cleanest independent trials, though mainly for healthcare providers. Every other function rests on vendor-reported cases or none. "Maturity" in the table means maturity of evidence (how sure anyone can be that the use case delivers), not how widely it is sold.

Where independent researchers measured the work, gains fall between single digits and about a third; where vendors measured it, claims run to 60-fold (15 hours to 15 minutes) with no baseline. The best-evidenced functions share three traits: high volume, a checkable outcome (an issue resolved, a pull request merged, a note written) and a human still in the loop. MIT NANDA, as reported by Fortune, found that more than half of GenAI budgets went to sales and marketing tools while the biggest measured ROI came from back-office automation (Fortune). That finding is secondary, but it points to the same functions the trial evidence favours.

Every industry has pilots, but few have audited numbers

Industry evidence is thinner than function evidence. Most named, quantified examples come from vendor pages (chiefly Google Cloud's customer stories and Cloud Next '26 summary, plus Anthropic's customer pages), which report outcomes on the vendor's definitions with no baseline or sample. The exceptions are healthcare, which has independent trials, and banking, where some companies published primary figures. Where an industry has no verified quantified case, the table says so. Indian cases are marked [India].

Two cautions apply when reading these cases. Retail and search cases report the most precise figures (no-result searches down from 2% to 0.1%), but they measure conversion and revenue, not labour automation. And the agent counts (800 at GE Appliances, 300 at Tata Steel) and employee reach (75,000 at Merck) are adoption metrics: they show that deployment is real, not that any agent pays for itself (Google Cloud Next '26). Vendor background is in the Gemini, Claude and Companies documents.

Controlled studies show real gains inside a narrow frontier

Only rigorous studies separate the effect of AI from everything else that changed at the same time. Three findings recur: gains are largest for less experienced workers; they vanish or turn negative outside what the model handles well; and people's own estimates overstate what was measured. Where research rounds conflict, the later primary-source reading is used.

The chart leaves out one result because it points the other way: METR's experienced developers took about 19% longer with AI, in a setting (experts in large, mature codebases they knew well) about as close to the edge of the frontier as software work gets. Read together, the studies show single-digit to low-thirties percentage gains on well-scoped tasks, concentrated among less experienced staff, and slower or less accurate work outside the frontier. Brynjolfsson's result carries a staffing lesson most business cases miss: the newest agents gained most and experienced agents little, which favours AI that supports junior staff over AI that removes them.

The perception gap matters as much as the effect sizes. METR's developers believed they were 20% faster while measured slower (METR). The same gap separates the evidence grades. GitHub's lab study says developers "code up to 55% faster" (vendor lab result) (GitHub); Accenture staff report routine tasks "up to 15 times faster" (self-reported) (SiliconANGLE); yet the Accenture arm of the randomised Cui et al. study measured +7.5% to +8.7% more pull requests. Anthropic's usage analysis estimates that Indian users finish in 14.8 minutes tasks that would take 3.8 hours without AI, a 15-fold speed-up against 12-fold globally (Anthropic India brief). That is a model-based estimate from conversation logs, not a measure of changed output, and should never be quoted as a productivity result.

The healthcare trials add two lessons. Uptake was low even inside the trials: clinicians used the scribe in about a third of eligible visits (29.5% to 38% across the two NEJM AI trials), so any per-user gain must be multiplied by actual use. And the well-being gain was more reliable than the time gain: burnout or exhaustion improved in every arm of all three trials, while only one of the two scribes in the UCLA trial cut documentation time significantly (PubMed 41497288), the Wisconsin trial cut note time by 0.36 hours a day but its drop in after-hours work did not survive removing extreme values (PubMed 41625485), and the JAMIA trial found after-hours time largely unaffected (PubMed 41729180). A tool that makes a hard job more bearable can be worth buying, but the business case should say that honestly. GDPval and OSWorld 2.0 measure capability, not business gain, and both show a ceiling: one leading model matched or beat experts less than half the time on GDPval, and in the June 2026 OSWorld 2.0 paper the best agent (Claude Opus 4.8, max thinking) completed 20.6% of long tasks under strict scoring. The best later leaderboard row (Claude Opus 5, Max, 31.4%; Epoch's copy, secondary) is still under a third. Model benchmarks are covered in Top models and Agent frameworks.

Surveys agree: adoption is broad and measured profit is thin

Surveys measure two things that must be read apart: adoption, which is high and consistent, and outcomes, which are low and self-reported. The gap between them is the most consistent finding in the 2025–2026 survey record.

The chart combines surveys with different samples, so the bars are not strictly comparable, but every survey shows use running well ahead of reported financial results. MIT NANDA's 95% figure is the most quoted number in this debate and the weakest: it rests on interviews with 52 organisations and a survey of 153 leaders, which the authors themselves call "directionally accurate" rather than company-reported figures. Success required measurable KPIs, with ROI measured six months after the pilot, a window that may be too short for sectors such as financial services (MIT NANDA report; Let's Data Science). Its diagnosis is better supported than its headline. It blames implementation, workflow integration and learning rather than model quality, which matches the reasons press coverage gives for Gartner's agent-cancellation forecast (rising costs, unclear value, weak risk controls). McKinsey gives the more useful planning number: 39% see some EBIT impact, and most of those attribute less than 5% of EBIT to AI (McKinsey). The 2026 edition, as reported, finds that share essentially unchanged at 37%, with about 6% of respondents attributing 5% or more of EBIT to AI (FM Magazine). On these figures, the share reporting any profit impact did not rise in a year of wider adoption.

Gartner's 40% cancellation forecast comes from a webinar poll, so it is a warning, not a measured failure rate; its more practical point is that many use cases sold as "agentic" do not need agents (Gartner). Bick, Blandin and Deming give the most sobering macro figure: across the US workforce, self-reported time savings come to 1.4% of all work hours (NBER), consistent with large gains on a few tasks spread thinly across all work. In India, the PayNearby index pairs very wide informal use (71% of retailers use AI tools daily) with uncertainty about value (26% unsure which features would help), which signals demand for advice, not just software (Newswav). Broader adoption data is in India and Asia-Pacific.

Named company cases: what was claimed and who measured it

Named cases fall into three groups: primary corporate disclosures (banks' releases and annual reports), which are careful, narrow and mostly about adoption; executive statements (interviews and memos), which are bold and unaudited and often contain no outcome; and vendor customer stories, which report large outcomes on the vendor's definitions. None was independently audited.

The most credible outcome figure is also the most modest: EricaAssist cuts nearly one minute per call across 18,000-plus representatives, an agent-assist design that leaves the human on the call, as in the Brynjolfsson study. JPMorgan's annual report shows how a large firm discloses AI when it chooses its words carefully: "value", a 2030 efficiency target and adoption rates, but no measured hours or costs (JPMorgan). Jamie Dimon's shareholder letter adds that "AI will definitely eliminate some jobs, while it enhances others", with a commitment to "support and redeploy" affected staff (CEO letter). Barclays' vendor-published announcement gives scale but no savings.

Salesforce is the hardest case to judge. The fall from about 9,000 to about 5,000 support staff is a headcount fact, not a productivity measure; Salesforce's framing, as reported, stresses fewer support cases, no backfilling and the redeployment of "hundreds" of staff (TechRadar). Agentforce revenue roughly doubled in 2026, but the definition widened in Q2 FY27, so like-for-like growth cannot be calculated. Fin's 76% average resolution rate, the highest automation share in this document, is the acquirer's description of what it bought, on Fin's own definition of an "outcome", with no volumes published. Fin's homepage repeats it across "12,000+ businesses", a lower count than Salesforce's 30,000+ (fin.ai). Per-outcome pricing (USD 0.99 at Fin; at Zendesk €1.50 committed or €2.00 pay-as-you-go per automated resolution on its German page, with USD 1.50 / 2.00 for the US from secondary sources) at least gives buyers a unit to audit (Zendesk). The vendor landscape is in Companies.

Reversals and failures: when the cost case outran the quality data

The documented reversals share a cause: the business case rested on assumed savings, and the quality and volume measurements that would have tested it were missing or ignored. None failed because the model was weak.

CBA is the most useful case for any business planning an automation because its failure was one of measurement, not technology. The business case credited the bot with taking 2,000 calls a week off the queue while volumes were rising, and the bank, whose contact centre had about 2,400 staff when the bot was tested, had to offer overtime (Canberra Times). Any pilot that reports "deflection" without total contact volume, repeat contacts and backlog risks the same error. Klarna's case is weaker evidence: the quality decline rests on the CEO's account, no independent post-mortem measured it, and no verified 2026 service metric was found. 2026 coverage (secondary) says the CEO now stresses investment in people (Realtid, 28 Jul 2026). It is still the precedent most often cited.

The agent failures carry a different warning. The Replit incident is one user's account, but each failure mode (an ignored instruction, a destructive action, invented data, false status reports) maps onto a control any deployment can include: separate environments, approval gates, and checking outcomes against the real system rather than the agent's report (The Register). Capability evidence agrees: in the OSWorld 2.0 paper (June 2026) the best agent, Claude Opus 4.8 at maximum thinking, completed 20.6% of long tasks under strict scoring, and the best later leaderboard row, Claude Opus 5 (Max), reaches 31.4% (Epoch's copy of the board, secondary) (arXiv), so long, unsupervised automation of screen interfaces is not yet a safe default for business operations. Agent security controls are covered in Governance.

Labour impact: fewer junior hires, not mass displacement (yet)

Labour evidence comes in three kinds of measure, and confusing them produces most alarming headlines. Exposure indices (IMF, ILO) measure task overlap with AI, not job loss. Employer attributions (Challenger) count how firms describe announced cuts. Payroll and occupation data (Stanford, Yale) measure what happened to employment. Only the last shows outcomes, and so far it shows a hiring effect concentrated among young workers.

The best-supported finding is that AI is so far reducing entry-level hiring more than it is causing layoffs, and mainly where it automates rather than augments (Stanford Digital Economy Lab; Brian Heger), though the authors stress that the evidence is descriptive. That matches the firm-level evidence: the largest gains go to novices (Brynjolfsson), and company policy has moved to hiring gates (Shopify, Duolingo) rather than mass dismissals. It also reconciles Challenger with Yale: firms increasingly attribute cuts to AI while the occupational structure barely moves. Even the attributions have eased: AI was the top cited reason every month from March to July 2026 (July: 10,970 cuts, 33%), then fell to fourth in August (3,462) and fifth in September (3,961, about 9%). Forecasts beyond 2026 are in Future outlook.

India's IT sector: headcount fell in FY26, and AI is one cause among several

India's large IT services firms posted their first net headcount decline in about two decades: the top five (TCS, Infosys, Wipro, HCLTech and Tech Mahindra) shed a net 7,389 employees in FY26, after adding 12,718 the year before, according to secondary reports not checked against filings (Forkast). IANS reported a fall of more than 7,000 (IANS). In July 2025 TCS planned to cut about 2% of its global workforce, roughly 12,000 people (DT Next; Gulf News). Its headcount fell by 19,755 in Q2 FY26 and by 11,151 in Q3, to 582,163 (The Quint; PeopleMatters). Another report gives a full-year net reduction of more than 23,000, to 584,519, followed by a net addition of 9,279 in Q1 FY27 (Entrepreneur India). TCS's Q2 FY27 results (reported 8 Oct 2026; secondary, several outlets agree) give a headcount of 598,056, up 4,258 on the quarter (Outlook Business). That equals 584,519 + 9,279 + 4,258 (arithmetic check, consistent with the year-end series), so headcount has risen for two quarters after the FY26 fall. The TCS series has not been reconciled across sources and should be checked against company filings before it is cited.

The companies themselves describe AI as a partial cause. TCS's CEO reportedly said the cuts were not driven solely by AI and that redeployment failed for roles such as manual testing (DT Next), while its Chief HR Officer said TCS had doubled its intake of fresh graduates with advanced AI skills (The Quint). Secondary coverage reports that some clients want 20–30% price cuts on new contracts, and gives a single-outlet figure of an 8,440 fall in Infosys's Q4 headcount (Let's Data Science). AI is also becoming a revenue line, on secondary figures for Q1 FY27: TCS's annualised AI revenue at about USD 2.6 billion (The Tribune), which TCS put above USD 3 billion in Q2 FY27 (reported 8 Oct 2026; about USD 3.1 billion in one outlet, more than 10% of revenue; secondary) (Outlook Business); Infosys, which "trims guidance as AI revenue climbs" in a TechMarketView headline (article body behind a login; no figure confirmed) (TechMarketView); and HCLTech's AI revenue up 62% year on year (DQ India).

The most defensible reading is that pricing pressure and a change in skill mix matter at least as much as direct automation: routine delivery work such as manual testing is being removed, and the work being added needs AI skills. Indian AI use is unusually concentrated on software: software-related tasks made up 45.2% of Indian Claude.ai use in Anthropic's February 2026 India brief, the highest share of any country (Anthropic India brief). In the June 2026 release (April and May 2026 data; consumer plans only), India had 6.92% and 7.12% of all Claude.ai usage and a usage-per-capita index of 0.31 and 0.30, where 1.0 is proportional to working-age population (Anthropic Economic Index data). The tasks IT services firms are paid for are exactly those being assisted, priced down and restructured. India's global capability centres employed more than 1.9 million people across more than 1,700 centres in FY2024, according to NASSCOM-Zinnov (NASSCOM), but the sources do not split out their AI share. The wider picture is in India.

How to pick and measure automations

The evidence turns into a four-part method: choose tasks where AI is known to work, define success as pairs of efficiency and quality metrics, measure a baseline before anything changes, and know the unit economics. This is a synthesis of the sources cited, not a single published framework; delivery method is covered in the Implementation playbook.

Selection criteria

KPIs: pair every efficiency metric with a quality metric

Every reversal in this document came from tracking efficiency without quality, or deflection without volume. Report each efficiency KPI beside a quality guardrail and a unit cost, and keep adoption metrics separate.

Baseline before pilot

The most common gap in the evidence is the missing baseline: most vendor cases give no "before" figure, and MIT's 95% is a finding about measurement as much as about AI. The protocol borrows what made the rigorous studies credible: randomised or staggered rollout, outcomes measured in the real system, and repeated trials.

Unit economics per task

Model cost rarely decides whether an automation pays. The cost model behind Costs and economics priced six typical tasks at October 2026 list prices from vendor pricing pages (Claude pricing; OpenAI pricing). Its token profiles are assumptions, not measurements; it excludes retries, evaluation runs, monitoring, human review and integration; and its 1.2 to 2 times production multiplier is also an assumption. Model names are price points, not recommendations. The premium column uses GPT-5.5 at USD 5 input, 0.50 cached and 30 output per million tokens (under 272K context), re-read on OpenAI's pricing page on 11 October 2026; it is an older, higher-priced model, not OpenAI's current flagship.

For comparison, vendors charge USD 0.99 per Fin outcome and USD 1.50 (committed) or USD 2.00 (pay-as-you-go) per Zendesk automated resolution (US figures from secondary sources; Zendesk's German page, read 12 Oct 2026, shows €1.50 committed and €2.00 pay-as-you-go), and voice agents cost USD 0.07 to 0.31 a minute on Retell's price list (Intercom pricing; Zendesk pricing; Retell AI pricing). Against the mid-tier five-turn resolution (USD 0.052), Fin's fee is about 19 times the model cost and Zendesk's about 29 to 38 times (on the secondary USD figures); on the low-cost tier the multiple runs into the hundreds. Most of what a buyer pays for, and most of the risk, therefore sits in integration, evaluation, exception handling and support, not in tokens, and model choice moves cost materially only for long agent loops such as code changes. Even there the low-cost tier splits: the code-change profile runs to 120K tokens of context, and Haiku 5.5 charges five times its base rate on prompts above 100K tokens, so its last turns are billed at USD 0.50 / 2.50 and the run costs about USD 0.28 rather than the USD 0.136 at base rates. Any base-rate Haiku estimate for a long agent loop is therefore a lower bound. Detailed pricing is in Costs and economics and Tool stack.

What this means for your business

Start with assistance, not replacement. The best-evidenced cases (Brynjolfsson's support agents, Bank of America's EricaAssist) keep a human in the loop and make that person faster; the reversals (Klarna, CBA) came from replacement designs with assumed savings. An agent-assist design with a human escalation path is easier to measure and far less likely to end in a public reversal or the kind of backlash Duolingo and CBA faced.

Favour back-office work that can be checked. Extraction, triage, routing and template drafting share the traits of the best-evidenced functions: high volume, verifiable output and a model cost per item of fractions of a cent. MIT's secondary finding that back-office automation produced the largest measured ROI points the same way. When a vendor proposes an agent, ask why a fixed workflow or rules will not do, as Gartner's "agent washing" warning suggests.

Measure a baseline before you buy. Most buyers have no baseline, which helps explain why AI use runs so far ahead of reported EBIT impact. A fixed-price discovery phase that ends in a written baseline, an evaluation set and agreed stopping rules is useful in itself and makes any later ROI claim credible in a way a vendor case study cannot. For budgeting, published US boutique price lists follow a sequence of fixed-fee assessment, fixed-scope build and monthly retainer; one lists an assessment at USD 3,500 to 8,500 and a build sprint at USD 18,000 to 60,000 (vendor-published) (Main & Machine).

Compare quotes with outcome benchmarks, and check the evidence grade of every claim. Per-outcome prices of USD 0.99 to 2.00 per resolution are public reference points for any quote, and model cost is a small part of them, so ask what the rest pays for: exception handling, monitoring and support. Set expectations from controlled studies (gains of roughly 7.5% to 34% on well-scoped tasks, possibly negative outside the frontier), not from vendor pages.

Indian SMEs need advice and a policy before agents. Informal use is widespread (71% of surveyed retailers use AI tools daily) but 26% are unsure what would help (Newswav), and governance lags: a quarter of organisations in ISACA's poll have no AI policy, and two-thirds of office workers in a vendor survey have used unauthorised AI tools. A written AI-use policy and a short list of workflows that would actually benefit are sensible first steps. The DPDP deadlines of 13 November 2026 and 13 May 2027 mean consent, retention and breach handling belong in any automation touching Indian customer data, and fintechs and other RBI-regulated firms should track RBI's FREE-AI recommendations on AI disclosure and failure reporting (a committee report, not yet a direction); see Governance.

Plan for the labour shift. The FY26 headcount fall at large Indian IT firms, and TCS's statement that redeployment failed for roles such as manual testing, signal pricing pressure on routine outsourced delivery work and, by inference, more experienced engineers in the hiring market. The skills the evidence keeps pointing to are evaluation design, integration, exception handling and change management. The Stanford finding argues for framing automation to your staff around capacity and service quality rather than job cuts: AI's largest measured gains go to junior staff, so removing them gives up the gain.

Questions to ask any vendor or partner: How is "resolved" or "done" defined, in writing? What baseline will be measured before the pilot? Which quality metric sits beside each efficiency metric? Is each figure quoted from a trial, a survey or the vendor? Where does a human approve external messages, payments or record changes?

Open questions and unverified items

Sources

1. NBER Working Paper 31161, "Generative AI at Work"
2. Bank of America newsroom, EricaAssist with generative AI (July 2026)
3. Salesforce, definitive agreement to acquire Fin (June 2026) (returned HTTP 404 on 12 Oct 2026; the 76% also appears on fin.ai)
4. Fin (fin.ai) homepage (read 12 Oct 2026)
5. TechCrunch, Meesho cuts customer call costs by 75% (26 Nov 2024)
6. Cui et al., three field experiments with software developers (presentation PDF)
7. Microsoft Research, Cui et al., "The Effects of Generative AI on High-Skilled Work" (publication page and abstract)
8. METR, Measuring the impact of early-2025 AI on experienced open-source developer productivity (10 Jul 2025)
9. Harvard Business School, Dell'Acqua et al., "Navigating the Jagged Technological Frontier"
10. PubMed 41497288, NEJM AI, ambient AI scribes randomised trial
11. claude.com customer story, EvenUp
12. Google Cloud customer story, Fifth Dimension
13. Anthropic, Barclays scales Claude (1 Oct 2026)
14. Google Cloud customer story, Gordon Food Service
15. Fortune, MIT report: 95% of generative AI pilots failing (18 Aug 2025)
16. MIT NANDA, "The GenAI Divide: State of AI in Business 2025" (PDF hosted on mlq.ai)
17. Google Cloud customer story, Hershey Entertainment & Resorts
18. Google Cloud customer story, Toolstation
19. Google Cloud customer story, Kraft Heinz
20. Google Cloud Next '26 wrap-up (25 Apr 2026)
21. Google Cloud customer story, Schrödinger
22. JPMorgan Chase Annual Report 2025 (PDF)
23. PubMed 41625485, NEJM AI, pragmatic randomised trial of ambient AI and practitioner well-being
24. Google Cloud customer story, FairPrice Group
25. Google Cloud customer story, Virgin Media Ireland
26. SiliconANGLE, Accenture deploys Copilot to 700,000+ users (27 Apr 2026)
27. TechRadar, Salesforce cuts 4,000 support jobs
28. claude.com customer story, Zendesk
29. Anthropic, Bengaluru office and partnerships across India (16 Feb 2026)
30. GOV.UK, Microsoft 365 Copilot experiment: cross-government findings report (HTML report; read 12 Oct 2026)
31. GOV.UK, Evaluating the impact of Microsoft Copilot in HMRC (phase 3 trial; read 12 Oct 2026)
32. GOV.UK, An evaluation of DWP's Microsoft Copilot 365 trial
33. METR, We are changing our developer productivity experiment design (24 Feb 2026)
34. PubMed 41729180, JAMIA, randomised crossover trial comparing ambient scribes
35. arXiv 2510.04374, GDPval (abstract)
36. arXiv 2510.04374v1, GDPval (HTML full text)
37. arXiv 2606.29537, OSWorld 2.0 (abstract; submitted 28 Jun 2026, v2 13 Jul 2026)
38. GitHub blog, quantifying GitHub Copilot's impact in the enterprise with Accenture
39. Anthropic Economic Index, India country brief (16 Feb 2026)
40. Let's Data Science, MIT report documents GenAI pilot ROI gap
41. McKinsey, The state of AI in 2025
42. FM Magazine, Companies’ financial value from AI holds firm in 2026 (McKinsey State of AI 2026 coverage)
43. South China Morning Post, over 40% of agentic AI projects forecast to be scrapped
44. Gartner press release, over 40% of agentic AI projects will be cancelled by end of 2027 (25 Jun 2025)
45. Deloitte, The State of AI in the Enterprise (2026 report)
46. NBER Working Paper 32966, "The Rapid Adoption of Generative AI"
47. Anthropic Economic Index, January 2026 report
48. Anthropic Economic Index, June 2026 report
49. Hugging Face, Anthropic/EconomicIndex dataset (release of 26 Jun 2026)
50. Storyboard18, EY: 62% of Indian employees use GenAI regularly
51. Newswav, PayNearby MSME Digital Index 2026
52. ISACA, 2026 AI Pulse Poll
53. US Census Bureau, AI use by businesses (May 2026)
54. Gallup, AI use at work indicator
55. PagerDuty, shadow AI workplace survey 2026
56. KTVU, Salesforce AI layoffs and Marc Benioff
57. Fox 35 Orlando, Salesforce CEO: AI agents handle half of customer work
58. Salesforce Q4 FY26 results (25 Feb 2026)
59. Salesforce Q1 FY27 results
60. Salesforce Q2 FY27 results (26 Aug 2026)
61. Fin pricing
62. Microsoft FY26 Q4 press release (29 Jul 2026)
63. Business Standard, Infosys unveils over 200 AI agents under Topaz (29 May 2025)
64. The Verge, Duolingo goes AI-first (29 Apr 2025)
65. The Verge, Shopify CEO memo on AI and hiring (7 Apr 2025)
66. JPMorgan Chase CEO letter to shareholders 2025 (PDF)
67. Zendesk pricing (redirects by region; no US page was reachable on 12 Oct 2026)
68. Zendesk pricing, German regional page (read 12 Oct 2026)
69. Maginative, Klarna dials back its AI customer service strategy (May 2025)
70. Entrepreneur, Klarna CEO reverses course by hiring more humans
71. Tech.co, Klarna reverses AI overhaul
72. Realtid, Klarna's CEO: "we went too far" (28 Jul 2026, in Swedish)
73. Canberra Times, Commonwealth Bank reverses AI chatbot job cuts, admits error
74. ACS Information Age, CBA reverses AI-driven job cuts
75. Staffing Industry Analysts, Australia's biggest bank reverses plan to replace jobs with AI
76. The Register, Duolingo ditches more contractors in AI-first refocus (29 Apr 2025)
77. Financial Times, Duolingo CEO on going AI-first (headline)
78. The Register, Replit agent deleted production database (Jul 2025)
79. Stanford Digital Economy Lab, "Canaries in the Coal Mine?" (revised 12 Aug 2026)
80. Brian Heger, summary of Stanford "Canaries in the Coal Mine" August 2026 update
81. AI Weekly, Stanford employment effects of AI
82. Let's Data Science, Stanford finds no broad AI job displacement
83. Fortune, Yale Budget Lab on AI and the labour market (2 Feb 2026)
84. Yale Budget Lab report (PDF, node 1495)
85. Yale Budget Lab report (PDF, node 1419)
86. Challenger, Gray & Christmas, May 2026 job cuts report (PDF)
87. Challenger, Gray & Christmas, June 2026 job cuts report (PDF)
88. Challenger, Gray & Christmas, July 2026 job cuts report (PDF)
89. Challenger, Gray & Christmas, August 2026 job cuts report (PDF)
90. Challenger, Gray & Christmas, September 2026 job cuts report (PDF, 1 Oct 2026)
91. PropertyCasualty360, AI-related job cuts continue to climb (10 Jun 2026)
92. Outlook Business, AI overtakes other reasons for US job cuts
93. IMF, Gen-AI: Artificial Intelligence and the Future of Work (SDN/2024/001)
94. IMF SDN/2024/001 full text
95. ILO Working Paper 140, Generative AI and jobs: a refined global index of occupational exposure
96. ITEdgeNews, Future of Jobs Report 2025
97. Malay Mail, WEF report: 78 million jobs to be created by 2030 (8 Jan 2025)
98. Unleash, five things HR needs to know from WEF's 2025 Future of Jobs report
99. Forkast, India's IT giants cut jobs for the first time in 20 years
100. IANS, top IT firms' headcount falls over 7,000 in FY26 (25 Apr 2026)
101. DT Next, TCS likely to lay off 12,000 employees in FY26
102. Gulf News, TCS to cut 12,000 jobs
103. The Quint, TCS workforce shrinks in Q3 FY26
104. PeopleMatters, TCS headcount falls by 11,151 in Q3 FY26
105. Entrepreneur India, headcount declines from TCS to TechM
106. Let's Data Science, Indian IT giants face AI-driven FY26 reset
107. The Tribune, TCS Q1 FY27 results: AI revenue up 13.6% QoQ
108. Outlook Business, TCS Q2 FY27 results: profit, revenue, AI and workforce (8 Oct 2026)
109. TechMarketView, Infosys trims guidance as AI revenue climbs (24 Jul 2026)
110. DQ India, HCLTech Q1 FY27: management on AI growth
111. NASSCOM, India GCC Landscape Report: the 5-year journey
112. Anthropic, Building effective agents (19 Dec 2024)
113. Anthropic, Demystifying evals for AI agents (9 Jan 2026)
114. OpenAI, A practical guide to building agents (PDF)
115. Gazette of India, G.S.R. 843(E), DPDP Act commencement (13 Nov 2025)
116. Gazette of India, G.S.R. 846(E), Digital Personal Data Protection Rules, 2025 (13 Nov 2025)
117. Bar and Bench, RBI's FREE-AI framework (18 Sep 2025)
118. Reserve Bank of India, press release on the FREE-AI committee report (13 Aug 2025)
119. Hamel Husain, Your AI product needs evals (29 Mar 2024)
120. Eugene Yan, Evaluating LLM-evaluators
121. Google AI for Developers, Gemini API changelog
122. Claude Platform documentation, pricing
123. OpenAI API pricing
124. Intercom pricing
125. Retell AI pricing
126. Main & Machine pricing