Abstract—Surveys describe strong enthusiasm for generative AI among Indian small businesses, yet deployment appears much shallower than intent. This paper synthesises published survey data and controlled studies, adds current Indian subscription prices, and builds a simple payback model to estimate where generative AI earns back its cost for small and medium enterprises (SMEs). We find that evidence of gains is strongest for customer support and writing-type tasks, and is largest for less experienced staff. Evidence is mixed for coding by experts on familiar code, where one randomised trial measured a 19% slowdown. At Indian list prices, a ₹399 per month plan breaks even at under three saved hours per month for a ₹150/hour employee, while plans near ₹2,000 need 13–16 hours. We propose a two-factor framework (cost of checking output, cost of an undetected error) to classify use cases, outline a 90-day pilot plan, and discuss data-protection obligations under India’s DPDP Act. Several inputs are secondary sources or author assumptions, and are flagged in the text.

Index Terms—generative AI, SMEs, MSMEs, India, productivity, return on investment, DPDP Act

I. INTRODUCTION

India is among the most active markets for generative AI. In one survey, 65% of Indian respondents had used generative AI, compared with a global average of 31% [14]. For an owner-managed firm, however, the relevant question is not whether staff have tried a chatbot but whether a paid tool returns more than it costs once review time and training are counted.

This paper addresses three research questions:

RQ1. How large is the gap between stated AI intent and actual deployment among Indian SMEs?

RQ2. What does controlled evidence say about productivity gains, and for whom?

RQ3. At Indian prices, how many hours of saved time make a subscription worthwhile, and which use cases clear that bar?

Our contributions are (i) a consolidated view of adoption signals (Fig. 1) and experimental effects (Fig. 2), (ii) a transparent break-even model using current Indian prices (Fig. 3), (iii) a classification framework for use cases (Fig. 4), and (iv) a staged pilot plan with compliance guidance (Fig. 5).

II. RELATED WORK

A. Adoption surveys

The Vi Business MSME Growth Insights Study 2026 reports that 57% of surveyed MSMEs view AI as a core tool, while coverage of the same study reports that about 25% already deploy AI tools [1]. SAP’s mid-market research found that 96% of Indian firms with 250 to 1,500 employees treat generative AI as a priority, and 49% use AI for forecasting and budgeting to a strong degree [3]. An OECD survey of more than 5,000 SMEs in seven countries found 30.7% GenAI adoption, of which 28.7% of users applied it to core revenue-producing activities [2]. EY reports that 62% of employees in India use AI regularly at work, yet only about 10% of surveyed large enterprises have enterprise-wide deployment [4].

B. Controlled studies of productivity

Brynjolfsson, Li and Raymond studied 5,179 customer-support agents and found a 14% average increase in issues resolved per hour, 34% for novices and minimal for experts [5]. Noy and Zhang report about 40% less time on professional writing tasks with higher rated quality, with weaker writers gaining most [6]. Dell’Acqua et al. found consultants using GPT-4 completed 12% more tasks, 25% faster, at 40% higher quality on tasks inside the model’s capability [7]. A survey table reports roughly 55% faster completion in a coding experiment by Peng et al. [8]. In contrast, METR’s randomised trial with 16 experienced open-source developers working on their own large repositories found that AI access made them 19% slower, even though they believed it made them 20% faster [9].

III. METHODOLOGY

A. Evidence collection

We performed a desk review in October 2026 using web searches for Indian adoption surveys, peer-reviewed or working-paper productivity studies, and current Indian price lists. We prefer primary sources (papers, trial reports, press releases) and use secondary summaries where the primary text was not accessed. Such cases are marked in the reference list. Survey results from different populations are shown together to illustrate a pattern and are not statistically comparable.

B. Payback model

For one subscription seat with monthly price P and an employee loaded cost of w rupees per hour, the break-even number of hours saved per month is

B = P / w                                                    (1)

We use w = ₹150 for clerical or support staff (about ₹25,000 per month over roughly 170 hours) and w = ₹400 for professionals. Both wage values are author assumptions and should be replaced with a firm’s own payroll data. The model ignores training time, review time and tool-switching costs, so it gives a lower bound on required savings.

IV. RESULTS

A. Intent versus deployment (RQ1)

Fig. 1 shows a consistent pattern: intent indicators are high (57% to 96%), usage indicators are lower (25% to 62%), and depth indicators are lowest (about 9% to 10%). The 8.8% figure is our own calculation (30.7% × 28.7%) from [2] and approximates the share of all surveyed SMEs using GenAI for revenue-producing work.

Fig. 1. Adoption signals from different surveys, grouped by intent, use and depth. Populations and definitions differ, so compare the pattern and not individual bars. *Author’s calculation from [2]. Sources: [1]–[4].

B. Productivity evidence (RQ2)

TABLE I. CONTROLLED STUDIES OF GENERATIVE AI PRODUCTIVITY

Fig. 2. Reported effects by study. Metrics differ (output per hour, time saved, speed), so interpret direction and approximate size only. Sources: [5]–[9].

Three patterns emerge. First, gains are largest for novices, consistent with the view that models spread the practices of top performers [5], [6]. Second, experts working on familiar, complex material may lose time reviewing near-miss output [9]; the METR authors caution that this result applies to experienced developers on familiar projects and may not extend to juniors or smaller projects. Third, self-perception is unreliable: participants in [9] misjudged their own speed, which argues for measuring real hours and error rates.

C. Indian pricing and break-even (RQ3)

TABLE II. INDIAN LIST PRICES FOR AI SUBSCRIPTIONS (SECONDARY SOURCES)

Prices vary by source, tax treatment and promotion, and several free promotions were reported; they should be verified on vendor pages before purchase. Applying Eq. (1) gives the break-even hours in Fig. 3. A support employee who gains the study-average 14% [5] frees about 24 hours in a 170-hour month, which exceeds every break-even value shown, provided the freed time is redeployed and outputs are reviewed.

Fig. 3. Hours saved per month needed to break even on one subscription seat, from Eq. (1), using list prices in Table II and the wage assumptions stated in Section III-B.

D. Classifying use cases

We classify use cases on two axes: the cost of checking an AI output relative to doing the task directly, and the cost of an undetected error (Fig. 4). Tasks with cheap checking and low error cost, such as FAQ drafts, marketing copy and translation, are candidates to scale first. Customer-facing claims and contract wording have cheap checks but costly errors, so they need mandatory review. Expert work on familiar material has expensive checking and is best adopted selectively. Hiring screens, personal-data decisions and finance calculations combine costly errors with costly checks, or legal and bias risk, and need strict controls or should be avoided. The figure is a conceptual framework proposed by the authors, not an empirically fitted model.

Fig. 4. Proposed framework for classifying generative AI use cases (conceptual; placement of examples is the authors’ judgement).

V. DISCUSSION

A. Implications for SMEs

The evidence supports starting with narrow, reviewable tasks and measuring baseline hours, error rates and response times before and after. The novice effect suggests the largest gains where firms hire freshers or have high staff turnover. The adoption gap in Fig. 1 suggests that many firms buy access without redesigning workflows, which may explain thin deployment.

B. Risks and compliance

Accuracy. 70% of Indian respondents in a Microsoft survey were concerned about AI hallucinations [14]; customer-facing, financial and legal outputs therefore need human review. Data protection. The Digital Personal Data Protection Rules were notified in November 2025 and phase in over about 18 months, with the main duties (security safeguards, notices, breach reporting, erasure) applying from mid-May 2027 [15], [16]. Pasting customer or student personal data into a public AI tool sends it to a third-party processor, so firms should prefer business plans with data controls, adopt a written policy on what may be entered, and use data-processing terms where available. Skills. Mid-sized Indian firms name finding and retaining AI talent as their top risk (39%) [3], which favours training an internal champion per team.

C. Recommended pilot

Fig. 5. Ninety-day pilot plan. Measure hours per task, error or rework rate and response time before and after, and scale only tasks that show a gain.

VI. LIMITATIONS AND THREATS TO VALIDITY

This is a desk study and not a field experiment. Most controlled studies come from the United States and use English-language tasks, so effects on Indian-language work may differ. Survey figures use different populations, years and definitions. Pricing comes from news and comparison sites and not vendor pages. The wage inputs and the use-case placement in Fig. 4 are author assumptions. The Peng et al. result is cited via a survey table [8], and the EY result via a summary article [4]. The payback model excludes training and review time and therefore understates required savings.

VII. CONCLUSION

Generative AI pays back for Indian SMEs mainly in support, drafting and translation-type work, particularly for newer staff, and can break even within hours of saved time per month at entry-tier Indian pricing. It does not pay back automatically for expert work on familiar material, high-stakes calculations or personal-data decisions. Future work should run firm-level pilots in Indian languages, measure review time directly, and track compliance costs as DPDP obligations take effect.

REFERENCES

[1]     Vi Business, “MSME Growth Insights Study 2026: 57% MSMEs view AI as a Vital Scale Driver,” press release, Mumbai, Jun. 25, 2026. [Online]. Available: myvi.in; also cxotoday.com/startups-smes/over-57-of-msmes-view-ai-as-a-vital-scale-driver-25-already-deploy-ai-tools/

[2]     CXOToday, “The Algorithm Next Door: How AI Could Power India’s Small-Business Revolution,” 2026 (reports OECD SME survey and Quantive studies; secondary source). [Online]. Available: cxotoday.com/expert-opinion/the-algorithm-next-door-how-ai-could-power-indias-small-business-revolution/

[3]     IndiaAI, “Indian midmarket businesses are already prioritizing Gen AI faster than rest of world: SAP report.” [Online]. Available: indiaai.gov.in/article/indian-midmarket-businesses-are-already-prioritizing-gen-ai-faster-than-rest-of-world-sap-report

[4]     LetsDataScience, “Indian Enterprises Adopt AI, Create New Roles,” summarising EY research of Jan. 19, 2026 (secondary source). [Online]. Available: letsdatascience.com/news/indian-enterprises-adopt-ai-create-new-roles-e4bf1c66

[5]     E. Brynjolfsson, D. Li, and L. R. Raymond, “Generative AI at Work,” NBER Working Paper 31161, 2023; Quarterly Journal of Economics, 2025. [Online]. Available: nber.org/papers/w31161

[6]     S. Noy and W. Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence,” 2023, as summarised in ITIF, “Evidence Shows Productivity Benefits of AI,” Jun. 2024 (secondary source). [Online]. Available: itif.org/publications/2024/06/11/evidence-shows-productivity-benefits-of-ai

[7]     F. Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” 2023, as summarised in [6] (secondary source).

[8]     “An Automated Survey of Generative Artificial Intelligence: Large Language Models, Architectures, Protocols, and Applications,” arXiv:2306.02781 (tabulates Peng et al.; secondary source). [Online]. Available: arxiv.org/abs/2306.02781

[9]     METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” Jul. 10, 2025. [Online]. Available: metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

[10]    TechCrunch, “OpenAI launches a sub-$5 ChatGPT plan in India,” Aug. 18, 2025. [Online]. Available: techcrunch.com/2025/08/18/openai-launches-a-sub-5-chatgpt-plan-in-india

[11]    CometAPI, “True Cost of ChatGPT in India: A Comprehensive Guide for 2026,” Jan. 2, 2026 (secondary source). [Online]. Available: cometapi.com

[12]    TechMitra, “Claude AI Pricing India 2026: Compare vs ChatGPT and Gemini,” Jul. 2026 (secondary source). [Online]. Available: techmitra.in/claude-ai-pricing-india-vs-chatgpt-gemini/

[13]    Cloudfy Systems, “Google Workspace Business Starter India.” [Online]. Available: cloudfysystems.com/blog/google-workspace-business-starter-india

[14]    BoomLive, “India’s Generative AI Adoption Hits 65%, More Than Double Global Average,” reporting Microsoft Global Online Safety Survey. [Online]. Available: boomlive.in/web-stories/news/indias-generative-ai-adoption-hits-65-more-than-double-global-average-2345

[15]    Hogan Lovells, “India’s Digital Personal Data Protection Act 2023 brought into force.” [Online]. Available: hoganlovells.com/en/publications/indias-digital-personal-data-protection-act-2023-brought-into-force-

[16]    AZB & Partners, “India’s DPDP Act: Phased Rollout and Key Compliance Milestones,” Nov. 14, 2025. [Online]. Available: azbpartners.com/bank/indias-digital-personal-data-protection-act-phased-rollout-and-key-compliance-milestones/