1. Home
  2. Our Thinking
  3. AI & Automation

AI Automation in 2026: What the Data Really Says (and What It Means for Indian Businesses)

Almost every company says it uses AI, yet few can show profit from it. We went through the best available studies, two real-world cases and the Indian data to separate what works from what is hype — with every number sourced.

AI & AutomationBy 91Technologies··14 min read
AI Automation in 2026: What the Data Really Says (and What It Means for Indian Businesses)

The short version: AI automation works — but only when it is applied to a specific, measurable process, with good data and a human in the loop. Adoption is now almost universal; profitable adoption is not. The gap between those two is where most of the money, and most of the disappointment, is.

This article pulls together the most credible studies, surveys and real-world cases we could verify. It shows what each one proves (and what it does not), then ends with a practical framework, a 90-day plan and a checklist you can use before spending a rupee. We name the source of every number, and we say clearly where a figure is only reported second-hand or where sources disagree.

Contents at a glance

  1. The headline: adoption is high, impact is not
  2. Why this is cheaper to try than ever
  3. Where AI automation measurably works
  4. Two real cases that teach more than any survey
  5. Where it fails: the pilot graveyard
  6. The India picture, including small businesses
  7. What the profitable few do differently
  8. Where to start, by industry
  9. Risks you must manage (legal, privacy, quality)
  10. A simple ROI test, then a 90-day plan

1. The headline: adoption is high, impact is not

The most-cited benchmark is McKinsey's global survey on the state of AI. The 2025 edition (1,993 participants, fielded 25 June – 29 July 2025) found that 88% of organisations use AI in at least one business function, up ten percentage points from the year before. Stanford's 2025 AI Index tells a consistent story for the earlier year: 78% of organisations reported using AI in 2024, up from 55% in 2023 (the attribution of that figure varies between write-ups, but the direction is not in doubt).

Yet only about one-third of McKinsey's respondents said they were scaling AI across the enterprise, and coverage of the report puts the share reporting any enterprise-level earnings impact at roughly 39%. Around 62% said they were at least experimenting with AI agents.

At the top of the curve sits a small group McKinsey calls "AI high performers": roughly 6% of respondents, who attribute 5% or more of their earnings (EBIT) to AI. Read the whole picture together: most companies have started; a third have scaled; a minority can show profit; and about one in sixteen is a genuine winner. That is the real story of 2025–26.

2. Why this is cheaper to try than ever

One reason the experiments keep multiplying: the running cost has collapsed. Stanford's 2025 AI Index reports that the cost of running a system at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024 — from about $20 to about $0.07 per million tokens, driven by smaller and more capable models.

For a business this means two things. First, the model is no longer the expensive part: the real costs are integrating it with your data and systems, designing the workflow, testing, and training people. Second, cheap tools make it very easy to start a pilot with no clear goal — which is exactly how the "pilot graveyard" in section 5 gets filled.

3. Where AI automation measurably works

Strong, independent-ish evidence exists in two areas: customer support and software development.

Customer support

A widely cited study by Brynjolfsson, Li and Raymond (Stanford and MIT, published through the NBER) followed customer-support agents at a large software company after an AI assistant was introduced. It found:

  • 13.8% more issues resolved per hour on average;
  • gains of around 35% for the least experienced agents, with little to no benefit for the most experienced;
  • no significant change in customer satisfaction, and lower agent attrition (reported as 8.6% lower).

The lesson is not "AI replaces agents". It is that AI spreads the habits of your best people to everyone else — especially valuable when you hire and train continuously.

Software development

In a controlled experiment run with GitHub, developers using Copilot completed a programming task 56% faster than those without it. A later field study across several large employers found developers with early access completed about 26% more tasks. A caveat worth knowing: much of this research was funded or run by the vendors who sell the tools, so treat the exact percentages as indicative, not guaranteed.

What the two have in common

Both are high-volume, repeatable, text-heavy tasks with a clear measure of "done". That is the profile of work AI handles well today. The further a task drifts from that profile — vague goals, rare edge cases, high stakes — the weaker the evidence gets.

4. Two real cases that teach more than any survey

Klarna: the promise and the course correction

In February 2024 the payments company Klarna announced that its AI assistant had handled about 2.3 million conversations in its first month — roughly two-thirds of its customer-service chats — doing work it estimated as equivalent to 700 full-time agents, with customer satisfaction on a par with humans. The story then became more nuanced. In May 2025 Klarna began recruiting human support agents again and promised every customer access to a person, and its CEO described the cost-cutting push as having "gone too far". By late 2025 the company said the assistant's work was equivalent to around 853 full-time staff.

How to read it: the "700 agents" number is Klarna's own estimate rather than an independent audit, and sources disagree on whether quality problems or a deliberate shift to a hybrid model drove the change. The durable lesson is the same either way — AI handled volume well, and humans remained essential for complex and sensitive cases.

Air Canada: you own what your chatbot says

In February 2024 a British Columbia tribunal (Moffatt v. Air Canada, 2024 BCCRT 149) held the airline liable after its website chatbot gave a customer wrong information about bereavement fares. Air Canada had argued the chatbot was effectively a separate entity; the tribunal rejected that, saying the chatbot is part of the company's website and the company is responsible for all the information on it. The money involved was small. The principle is not: a customer-facing AI speaks for your business, and you are accountable for what it says.

5. Where it fails: the pilot graveyard

Two widely quoted warnings deserve attention — each with a health warning of its own.

  • Gartner (June 2025) predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls.
  • MIT's NANDA initiative (August 2025), in "The GenAI Divide", reported that roughly 95% of generative-AI pilots had not delivered measurable financial returns. This comes from one research team's pilot-level survey, so read it as "most pilots stall", not "AI fails".

Both point the same way: the technology is rarely the problem. The usual causes are a vague goal, messy data, no owner, no risk controls, and automating a broken process instead of fixing it first.

Common mythWhat the evidence suggests
"We need an AI strategy first."You need one measurable process and an owner. Strategy follows results.
"The best model guarantees the best result."Data quality, workflow design and training usually matter more than the model.
"Agents will run the business on their own."Even leaders keep people in the loop; Gartner expects many autonomous-agent projects to be scrapped.
"AI will replace the team."Evidence so far shows tasks changing and less-experienced staff improving fastest, not whole roles vanishing.
"Savings will show up in profit automatically."Saved minutes only become savings if you redeploy or avoid hiring; otherwise they disappear into the day.

6. The India picture, including small businesses

Indian numbers come from several surveys that define "adoption" differently, so compare them with care:

SourceFinding
BCG (2025) — employee survey92% of Indian employees use GenAI tools vs a global average of 72%; only about 36% feel adequately trained
EY–CII (Nov 2025)47% of Indian enterprises have multiple GenAI use cases live; 23% are in pilot
CIO&Leader — State of Enterprise Technology 2025 (firms above ₹5,000 crore turnover)93% plan to raise AI and analytics investment; only 15.8% reach "strategic maturity"
NASSCOM–EY AI adoption indexMoved only from 2.45 (2022) to 2.47 (2024) — a marginal gain in maturity

The pattern matches the global one: very high enthusiasm and usage, middling maturity, and a training gap.

What about MSMEs?

Small-business numbers are patchier and sit in a wide range, because surveys count different things:

  • Trade coverage of MSME studies puts the share of MSMEs already deploying AI tools at roughly 23–25%.
  • A November 2025 report says about nine in ten Indian SMBs are investing in or planning AI adoption — intent, not current use, so it is not comparable.
  • A 2026 retail-focused index covering 10,000 neighbourhood retailers reports 71% using AI tools in daily operations — a sample skewed toward small retailers, so do not generalise it.

Several of these are vendor-sponsored or second-hand, and we could not verify their methods. The honest conclusion: roughly a quarter of small businesses are actually using AI, most of the rest plan to, and cost and skills are the main brakes. That is good news for a disciplined small business — most of your competitors are still at the starting line.

7. What the profitable few do differently

Across the studies, the same practices separate the profitable minority from everyone else. McKinsey finds its high performers are nearly three times as likely to have fundamentally redesigned their workflows when deploying AI, rather than just adding a tool to the old process.

  1. Pick one workflow, not "AI strategy". A measurable process with volume — support triage, invoice entry, lead qualification, report drafting.
  2. Redesign the workflow, don't just add a chatbot. Decide what the AI does, what a person checks, and what happens on exceptions.
  3. Keep a human in the loop for anything customer-facing, financial or legal — at least until accuracy is proven on your own data.
  4. Measure before and after. Time per task, error rate, cost per ticket, response time. No baseline means no proof of ROI.
  5. Train people. The BCG training gap is real: tools that staff do not trust or understand sit unused.
  6. Own the data. Clean, current, well-organised data (price lists, policies, FAQs, product details) decides whether answers are right.
  7. Assign an owner. One named person is accountable for results, not a committee.

8. Where to start, by industry

These are practical starting points drawn from the evidence above and from the kinds of systems we build — suggestions, not statistics:

IndustryGood first automationsKeep a human on
Retail & e-commerceOrder-status and returns replies, product-description drafts, WhatsApp enquiry triageRefunds, complaints, pricing exceptions
Healthcare & clinicsAppointment reminders, intake forms, FAQ answers, report summaries for staffAnything clinical or diagnostic
Finance & accountingInvoice and receipt data entry, reconciliation hints, document checksApprovals, credit and compliance decisions
Manufacturing & distributionEnquiry-to-quotation drafting, purchase-order matching, stock and dispatch reportsPricing, supplier commitments
Real estate & servicesLead qualification, site-visit scheduling, follow-up remindersNegotiation and contracts
Education & trainingAdmissions enquiries, content drafting, attendance and fee remindersAssessment and counselling
Agencies & IT servicesProposal drafts, ticket triage, code assistance, weekly client reportsScope, estimates, client relationship

Notice that none of these needs a futuristic autonomous agent. Most of the proven value today sits in unglamorous, repeatable tasks inside a CRM, helpdesk or accounts system you may already use.

9. Risks you must manage

  • Accountability. As the Air Canada ruling showed, you are responsible for what your AI tells customers. Limit what it may promise, link to your official policies, and review conversations.
  • Hallucinations. AI can state wrong things confidently. Ground answers in your own documents, show sources where possible, and escalate when unsure.
  • Data privacy. India's Digital Personal Data Protection Act, 2023 applies when you process customers' personal data. Know what you send to which vendor, and where it is stored.
  • Security. Give automations the least access they need, log what they do, and keep an off-switch.
  • Vendor lock-in and cost drift. Prefer designs where you can change the underlying model, and track usage cost monthly.
  • Change management. Tell staff what is changing and why; involve the people who do the work.

10. A simple ROI test, then a 90-day plan

Before commissioning anything, answer these questions for one candidate process. (The numbers are an illustrative example, not a statistic.)

QuestionExample
How many times a month does the task happen?1,500 support emails
How long does a person take per item?6 minutes
What share could AI draft or resolve with human approval?Assume 50%
Time saved per month1,500 × 50% × 6 min = 75 hours
Value of that time vs. cost of build and runningIf a staff hour costs ₹300, that is ₹22,500/month — compare with set-up and monthly costs, and decide whether you will redeploy the time or avoid a hire

If the saving is thin, skip it. If it is clear, follow a short plan:

DaysWhat happensOutput
1–15Pick one process; record today's time, cost and error rate; name an owner; gather the documents and data it needsBaseline and one-page scope
16–45Build the smallest version with human approval on every output; test on real past cases; fix mistakesWorking pilot, accuracy measured
46–75Run it live with a small group; log every correction; train the teamReal-world results vs baseline
76–90Decide with numbers: expand, adjust or stop; relax approvals only where accuracy is provenGo / no-go decision and a next-step plan

That discipline alone puts you ahead of most of the "95%".

A one-page checklist before you start

  • Is there one named process with clear volume and a measurable result?
  • Do we have a baseline (time, cost, errors) from before the change?
  • Is the data the AI needs clean, current and allowed to be used this way?
  • Who owns the result, and who reviews the AI's work?
  • What can the AI never say or do on its own?
  • How will staff be trained, and what happens to the time we save?
  • What is our off-switch, and how do we log what it did?

How 91Technologies can help

We build custom software, CRMs and automation for Indian businesses. If you want to know whether AI automation makes sense for one of your processes, we can run a short assessment, build a fixed-scope pilot and measure the result before you commit further. Talk to us or see what we have delivered.

Sources

  • McKinsey, The state of AI in 2025: Agents, innovation, and transformation (survey of 1,993 participants, June–July 2025) — mckinsey.com/mgi/our-research/the-state-of-ai-2025. High-performer share (about 6%) and workflow-redesign finding as summarised in coverage of the report.
  • Stanford HAI, AI Index Report 2025 — inference-cost decline (over 280-fold, Nov 2022 – Oct 2024) and organisational adoption (78% in 2024, up from 55%); attribution of the adoption figure varies between summaries.
  • Brynjolfsson, Li & Raymond, Generative AI at Work — NBER digest, "Measuring the Productivity Impact of Generative AI" — nber.org/digest/20236/measuring-productivity-impact-generative-ai
  • Peng et al., GitHub Copilot controlled experiment (2023), and later field studies (Cui et al.) — as summarised by the Information Technology and Innovation Foundation (itif.org, 2024)
  • Klarna announcements (Feb 2024) and later coverage (May and Nov 2025) in Entrepreneur, Finextra, Business Insider India and tech.co — Klarna's figures are company estimates, not independently audited.
  • Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, February 2024) — via American Bar Association and McCarthy Tétrault commentary.
  • Gartner press release on agentic AI projects, June 2025 — as reported by CRN Asia and Outlook Business
  • MIT NANDA, The GenAI Divide: State of AI in Business 2025 — as reported in trade press
  • BCG AI Radar / employee survey 2025 — as reported by Tribune India and The Hans India
  • EY–CII study on GenAI in Indian enterprises (Nov 2025); EY AIdea of India outlook 2026
  • CIO&Leader / CXOToday, State of Enterprise Technology Survey 2025; CXOToday MSME coverage; IANS (Nov 2025) report on SMB AI plans; PayNearby MSME Digital Index 2026 (retail sample)
  • NASSCOM–EY AI adoption index, as reported by W.Media

Figures were checked on 15 November 2026. Several are quoted from secondary coverage of the original reports and are labelled as such; definitions of "adoption" differ between surveys, so they are not directly comparable. The industry table, risk list, ROI example and 90-day plan are our own practical guidance, not research findings.

#AI automation#generative AI#AI agents#business automation#India#ROI
FAQ

Quick answers

Does AI automation actually improve productivity?
Yes, in the best-studied cases. A Stanford/MIT field study of customer-support agents found about 14% more issues resolved per hour, and a GitHub-run experiment found developers finished a coding task 56% faster. The gains are real but uneven: they were largest for less experienced workers and small for the most experienced.
Why do so many AI projects fail?
The research points to the same causes again and again: no clear business case, poor data, unclear ownership, weak risk controls and treating AI as a side experiment instead of redesigning the workflow. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 for these reasons.
Is AI automation only for large companies?
No. The most reliable wins — drafting, summarising, data entry, support triage, lead qualification — need little infrastructure and work at small scale. The surveys we found are mostly about large enterprises, so SMEs should start with one narrow process rather than a big programme.
How much does AI automation cost?
It depends entirely on scope. A single well-defined workflow (for example, auto-classifying and answering common support emails with human approval) is a small project. Enterprise-wide agents are a different order of magnitude. Always start with a fixed-scope pilot and measure before expanding.
Who is responsible if an AI chatbot gives a customer wrong information?
You are. In Moffatt v. Air Canada (British Columbia tribunal, February 2024) the airline was held liable for wrong information given by its website chatbot, and the tribunal rejected the argument that the chatbot was a separate entity. Ground answers in your own policies, limit what the bot may promise and review its conversations.
How long should an AI automation pilot take?
About 90 days for one well-defined process: two weeks to set a baseline and scope, a month to build and test on past cases, a month to run it live with human approval, and a final two weeks to decide with numbers whether to expand, adjust or stop.
Will AI replace my team?
The strongest evidence so far shows AI changing tasks, not replacing whole roles: it speeds up routine work and lifts less experienced staff. The businesses that benefit most use AI to handle volume while people handle judgement, relationships and exceptions.
Get started

Let's build something great together

Tell us what you need — a website, an app, custom software or a marketing push. You will get a clear proposal with scope, timeline and cost.

  • Expert consultation, no obligation
  • Response within one business day
  • Transparent pricing and milestones

Tell us about your project

We usually reply within one business day.

AI Automation Analysis 2026: Real Data, ROI & What Works