LEADER’S NOTE

Generative AI opened the floodgates to AI innovation, particularly in real, scalable, enterprise applications of AI with AI
agents to, as they say, “make AI do what it was designed to do.”
AI agents are software entities that can leverage all the learning, reasoning, context awareness being immersed into
increasingly sophisticated large language models, multimodal models, and now large action/ reasoning models, along with
real-world interactions, to do what humans do. Today, at a nascent scale, but with the promise of immense autonomy and
agency sometime in the future!
Agentic AI systems – multi-agent systems with human + AI collaboration or fully autonomous systems with minimal human
input – will eventually become mainstream enterprise applications. But at the core, AI agents are built on the capabilities of
Achyuta Ghosh previous generations of AI technologies – both predictive and generative – and in that, these systems also inherit inherent
Senior Director and Head of challenges of both. Adoption, scale, and real-world RoI from agentic AI systems will be determined by an enterprise’s
Research, Nasscom ability to overcome these inherent data, trust, AI risk, and human mindset challenges before its competitors do.
Our study, “Enterprise Experiments with AI Agents – 2025 Global Trends,” the first of a two-part study set, offers a peak
into the evolution journey of current AI agents and potential models of the future. The study offers a comprehensive
analysis of the current state of enterprise interest, adoption, preferences, challenges, and risk perception associated with
deploying AI agents in the near-mid-term, based on a global survey of over 100 large and medium-sized enterprises.
Companies are optimistic about the promise of AI agents. Yet, they prefer piloting internally before taking solutions to
clients. Promise is also reflected in the intent to invest dedicatedly, however, foundational readiness with data governance,
infrastructure, GenAI model usage maturity, and more importantly, risk management suited for AI risks are wanting areas.
As a result, enterprises intend to start small, with task-level automation, and scale-up to process/functional scale. This
early experimentation stage itself is complex, as AI agent architectures will have to integrate with existing tech stack, will
need explicit ethical guardrails, observability systems for transparency, and continuous refinement to build scale.
Early days; but the world is watching as to how the human + AI collaboration will evolve to deliver real, practical, and trust-
worthy agentic AI systems. We will certainly revisit these trends.
We hope you find this study useful in the current context of AI agent developments. We welcome your feedback at
[email protected].

2
CONTENTS STUDY METHODOLOGY

Study Snapshots 4 Study Part 1: Enterprise Experiments with AI

Understanding AI Agents Agents: 2025 Global Trends
9 The current study is first of a two-part set of research
▪ What are AI Agents studies. This report intends to apprise readers about
▪ Evolution of AI Agents the core concepts around AI agents, the tech’s
▪ The Horizons of Agentic AI
evolution, and potential future state. The study also
▪ AI Agent Characteristics
▪ Types of AI Agents offers a comprehensive current-state analysis of end-
▪ Multi-Agent Process Execution user adoption from a survey of 100 global enterprises
across 8-9 major regions and over 10 industries.
2025 Global Enterprise Trends
16
▪ Survey Demographics Study Part 2: “Agentic AI Adoption Models”
▪ The AI Foundation – Global Trends
This research will explore emerging tech ecosystem
▪ Experimenting with AI Agents – Global Trends
supporting the rapid evolution of agentic AI solution
Perspectives space, specifically the emerging provider business
26 models, pricing models, partnerships, and innovative
▪ Making Agentic AI Real positioning of agentic AI solutions within the
▪ Learnings from Enterprise Trends
product/platform stack. The study will feature a select
▪ The Human + AI Future
set of end user and provider case studies covering the
Appendix current scape of agentic AI use cases and offerings.
31

3
STUDY SNAPSHOTS
Agents are… Core characteristics of LLM/LMM-based AI agents

Software programs that can sense,

Aids the go/no-go human
assess, act, learn to pursue
decision about the reasoned
complex goals with autonomy and Assess context or action against defined objectives,
predictability intermediate output,
DECISION- or redefine objectives or action
evaluate validity of
MAKING
information, and
LLM-based agents are recent but identify relevant REASONING
rapidly evolving generation of AI action
technology aimed at delivering
human-like process and functional Initiate or effect the
automation at scale. chosen action or a
chain of multiple
actions using tools to

PERCEPTION
ACTION access external
AI agents can be segmented based Sense, interpret and EXECUTION
on their objectives, learning understand the environments
techniques, or the operating environment through
architecture. data collated from
sensors, API calls, or
direct user input LEARNING & (Self) Learn patterns from
Current AI agents have technology, ADAPTATION outcomes, interactions
learnability, and autonomy (such as, RLHF*) and data
limitations and work best with to finetune performance
human-in-the-loop. over time

4
STUDY SNAPSHOTS
Enterprises are strengthening their tech core – AI spend and talent, GenAI capabilities, data foundations, and process flexibility –
to build and deploy AI agents, as revealed in a global survey

1. Agentic AI is the next evolution, building on predictive and generative AI, with strong enterprise interest.

2. 88% of the enterprises now have dedicated AI budgets and emerging AI expert teams crucial to building scalable AI
applications, as AI spend as a percent of tech budgets inches up with 2-in-3 companies spending over 15% dedicatedly on AI.

3. Overall awareness and usage of GenAI is strong, with over 60% respondents focusing on GenAI tools, infrastructure readiness
and platform-based solutions.

4. However, direct usage of GenAI models is still low, as at best, 51% of companies are actively finetuning foundation models or
LLMs.

5. Data strategies continue to be top priority with 68% companies focusing on streamlining data management and governance
and 62% focused on building data systems for structured and unstructured data stream integration.

6. Encouragingly, 71% of the enterprises are experimenting with building reinforcement learning models for better output
accuracy, while a good majority is working on RAG-based real-time data ingestion.

7. 58% companies are also experimenting with reengineering existing process workflows to deploy agentic AI systems, although
majority of these are hi-tech companies and IT departments within them.

5
STUDY SNAPSHOTS
Experiments with AI agents are advancing, but more internally, at a task automation level with limited process-level deployments,
as client demand is still shaping and organizations realize the need for intense human involvement to train these systems

8. 62% of the global enterprises are experimenting with AI agents, from PoCs to production to some scale.

9. 88% expect that they will allocate dedicated spending to agentic AI solutions in 2025, up to 20% of AI spend.

10. Yet, adoption is nascent and 77% indicate doing task-level agents or process agents with human-in-loop. 46% do indicate
trying autonomous agents, however, these organizations are primarily hi-tech and advanced manufacturing.

11. "Client zero" (internal deployments) is a key initial focus for 76% of the companies, before offering external client solutions,
with 81% indicating that it is their IT operations teams piloting or using the AI agents. Only 31% claim use of AI agents for
customer services and support.

12. On average, for any enterprise, over 50% of its clients are experimenting with or actively seeking AI agents as part of the
offerings, with manufacturing clients taking a lead over services.

13. Over 55% of enterprises believe in organizational-scale impact of agentic AI – in greater capability to respond to new business
opportunities and rapidly transforming information to intelligence for real-time decision making.

14. Contrary to the hype, only 39% believe that using agentic AI systems will free up workforce time for high-order work, crucially
suggesting that AI agents will need sustained human coordination and oversight to deliver the desired outcomes.

6
STUDY SNAPSHOTS
With deep wariness about core risks due to AI, the ability to build trust-worthy agentic AI systems will determine the speed of
adoption despite early optimism, and here, enterprises aiming for human + AI over autonomous, are likely to succeed

15. Majority companies, over 50%, continue to cite data privacy and security (60%), unknown risks of self-learning systems
(56%), and lack of regulatory protection as barriers to scaling GenAI and agentic AI adoption.

16. Surprisingly, only 44% of the companies appeared less concerned about the mindset shifts needed to build human + AI
systems and the limited RoI from such deployments. Only 27% feel that they are limited by lack of AI talent!

17. When addressing AI risks, majority companies still follow the standard enterprise risk frameworks, retrofitted to suit
immediate AI reporting, auditing needs.

18. Just about 43% companies are planning focused AI risk management – observability, traceable systems, continuous review
of responsible AI guardrails, hardware-level audits

19. Fully cognizant of the challenges and risks, enterprises do have a cautious view on adopting agentic AI in terms of
timelines. 47% of the companies expect longer adoption timelines of 18 months or longer, driven by PoC success or
competitive disruption.

20. However, the remaining 53% expects adoption within 12 months, and a deeper look reveals that most of these are
companies with a combination of revenue scale, tech readiness, and AI budgets and teams.

7
STUDY SNAPSHOTS
Inherent preference for human + AI synergistic agentic systems is expected to emerge as AI agents demonstrate autonomy and
trust-worthy agency with human-assisted reinforcement learning, context awareness, real-world reasoning, and ethical alignment

Increasing autonomy and greater trust with real-world AI agency

Phase 1: <12 months Phase 2: 12-24 months Phase 3: 36 months+

Human Agentic Human Agentic Human Agentic

Intelligence Intelligence Intelligence Intelligence Intelligence Intelligence

Humans control all operational, strategic Humans control shifts towards strategic Humans control long-term business sustenance
workflows, ethical checkpoints, and decisions foresight and managing ethics and creativity planning and post-crisis revival planning

AI agents are best at task execution guided by AI agents get better at process execution with AI agents gain significant autonomy in
definitive rules minimal human involvement compliant execution during dynamic situations

Human + AI synergies involve agents Human + AI synergies happen for validation of Human + AI synergies occur at strategic
interpreting natural language commands, and AI decisions and actions during exceptions frontiers with precise coordination
seeking stepwise human validation

8
UNDERSTANDING AI AGENTS
The Definitions – AI agents are defined in multiple ways, but they are essentially software entities that can
sense, reason, learn, act, and adapt

Evolution of Classical AI to Agentic AI – Agents are an evolution from both predictive AI technologies, and
more so, from generative AI models, along with their ability to use external environment, data and tools to
accomplish tasks and goals

Core Characteristics – Most agents typically exhibit perception of the environment, reasoning, decision-
making, action execution, and learning and adaptation through reinforcement learning techniques

Types of AI Agents – Agents today can be categorized by their objective setting (task or goal), learning
technique with human-in-the-loop or without, and the operational setup of single versus multi-agent systems

Process Automation with Agentic AI – A typical human + AI agentic system for a simple and unobvious
process, such as URL parsing, can involve many complexities, emphasizing the need for careful design
WHAT ARE AI AGENTS: Programs that sense, assess, act, and learn, like humans

Gartner defines intelligent (agentic)

According to MIT, Agentic AI AI as “goal-driven software entities
OpenAI calls agents “systems that
systems are “designed to pursue that take instructions, create plans,
intelligently accomplish tasks,”
complex goals with autonomy and use tools, and carry out tasks
capable of executing complex
predictability.” autonomously—reaching goals
workflows to meet objectives.”
without needing step-by-step
human prompts.”

McKinsey describes an AI agent as

IEEE defines an intelligent agent as
a “software component that acts
a “software that acts
on behalf of a user or system to
autonomously—perceiving its
perform tasks, orchestrate
environment, making decisions,
workflows, coordinate multiple
taking actions, and learning over
agents, apply logic to problems,
time.”
and evaluate answers.”

Sources: MIT, IEEE, McKinsey, Gartner, OpenAI. Nasscom analysis. 10

EVOLUTION OF AI AGENTS: Predictive AI to generative AI to AI agents that “act”
1950s 2025
Predictive AI: 1956 onwards
Generative AI: 2022 onwards
Agentic AI: 2024+

Predictive AI Generative AI Agentic AI

Concept Deterministic machine learning (ML) algorithms Probabilistic ML models that "learn" the underlying AI systems equipped with goals, perception, reasoning,
that “learn” patterns in existing structured data patterns and structures of existing data (text, planning, actioning, and interaction capabilities given
to categorize/predict new input pattern images, audio, etc.) to generate novel and diverse a context or environment, to achieve specific
outputs that somewhat resembles training data objectives, using predictive and GenAI and other tools

AI ▪ Supervised learning – pre-labeled data for ▪ Natural language processing and generation ▪ Perception (pre-trained datasets, RAG datasets)
classification (NLP/ NLG) ▪ Reasoning (context awareness, CoT, etc.)
Capabilities ▪ Unsupervised learning – machine-determined ▪ Pattern recognition from vast datasets using ▪ Decision-making
/ Training labels for clustering GPT, GAN, VAE, and RAG models ▪ Action execution (access to external tools)
Techniques ▪ Image or speech pattern recognition ▪ Reinforcement learning (RL) ▪ Learning and adaptation (RL-self or human)
▪ Chain of thought (CoT) reasoning ▪ Workflow orchestration
▪ Mixture of Experts (MoEs)

Application ▪ Recommendation systems ▪ Content understanding, summarization and ▪ Task-based agents for independent tasks
▪ Speech recognition generation ▪ Goal-based workflow agents
Scenarios ▪ Expert systems ▪ Novel multimodal (text, video, image, audio, ▪ Complex process optimization agents
▪ Computer vision code) content creation ▪ Personal assistants
▪ Prediction engines ▪ Personalization engines ▪ Self-learning or adaptive agents tending to semi- or
▪ Natural language processing (NLP) ▪ Synthetic data generation fully autonomous execution

Limitations ▪ Unable to process unstructured data ▪ Lacks reasoning behind patters ▪ Based on GenAI, hence behavior unpredictability
▪ No ability to generate content ▪ Limited self-awareness, context knowledge ▪ Enterprise security, privacy risks
▪ Cannot use real-time context ▪ Requires frequent pre-training rounds with new ▪ Human oversight overload due to current tech
▪ Requires AI specialist talent data limitations
▪ Available to all humans, but poses security, ▪ Available to all, of particular interest to enterprises
privacy, accuracy, bias, ethics, cost risks but high enterprise TCO

Sources: Nasscom analysis. For key AI definitions, refer to Glossary on page 32. 11
THE HORIZONS OF AGENTIC AI: Collaborative vs. autonomous systems

AI systems that deliver on shared goals and AI models and algorithms that make decisions
Human + AI aligned intent with active human-in-the-loop with little to no human oversight in a dynamic Autonomous
Systems emphasizing collaboration and clear division of world, demonstrating not only greater autonomy Systems
labor but led by human expertise and control. but also increasingly human-like agency.

The Human + AI Lens: Collaborative Agentic System The AI-Dominant Lens: Autonomous Agentic System

Human Human

Human-defined prompts

A Permeable Distinction?
Reasoning and
decisioning
2a. Human Tasks
Goal Goal 1. GOALS Goal
2. ORCHESTRATE
proposed AI clarification SET proposed AI
2b. AI Tasks Goal
Reasoning and clarification 1. GOALS
decisioning SET
Prompt APIs APIs APIs
analysis
4. ORCHESTRATE Feedback
Tools Tools Tools
into next
action
5. Output 3. Prompt
Presentation Analysis
APIs APIs APIs 3. Output
Presentation
Tools Tools Tools

Sources: Nasscom analysis. *Autonomous agentic AI systems may or may not fully align with the levels of autonomy defined around autonomous driving or autonomous vehicles. 12
AI AGENT CHARACTERISTICS: Decoding the core of an AI agent

Core characteristics of LLM*/LMM*-based AI agents Additional Features

Multimodal input and output:

Make a go/no-go decision about
Process diverse inputs (text, voice,
the reasoned action against
Assess context or video) and/or collaborate with
defined objectives, or redefine
intermediate output, other agents to effect multimodal
DECISION- objectives or action plan
evaluate validity of input and output.
MAKING
information, and
identify relevant REASONING Multi-agent communications and
action systems: Communicate with other
agents aligned to the same
Initiate or effect the outcome, or different agents from
chosen action or a external environments to execute
chain of multiple complex workflows.
actions using tools to
PERCEPTION

ACTION access external

Sense, interpret and EXECUTION Workflow optimization: Optimize
understand the environments resource allocation and workflow
environment through efficiency through enhanced
data collated from memory and reasoning through
sensors, API calls, or reinforcement learning.
direct user input LEARNING & (Self) Learn patterns from
ADAPTATION outcomes, interactions Autonomy: Operate independently
(such as, RLHF*) and data with no extraneous control of the
to finetune performance system, in environments that offer
over time such independence.

Sources: Nasscom analysis. For key AI definitions, refer to Glossary on page 32. 13
 Model Based ▪ Model-based – Constructs perception-based internal
Agent model of the physical world, regularly updated and
aligned with present context

▪ Goal-based – Assesses probability of achieving preset

Objective-Based Goal Based
goals based on each next move, aimed at minimizing
Agents Agent
goal distance
Agents based on defined outcomes
▪ Utility-based – Evaluate multiple actions by balancing
Utility Based trade offs between different outcomes to maximize the
TYPES OF AI Agent defined utility functions

AGENTS: Complete Human ▪ Human supervision – Human input injunction and

Supervision control through the action pathway to outcome
Learning Agents
Based on Self-Learning ▪ Self learning – Learn through techniques such as
Agents based on learning technique System reinforcement learning without human supervision
decisions,
learnability, ▪ Single agent – Single agent performs the defined
Single Agent System
and system actions independently, typically task-based agents

complexity Sequential Agent

▪ Sequential agents – Single or multiple agents perform a
set of steps within a task sequentially, action and
System output at each step determining the next
Systems-Based Agents ▪ Multi-agent – Multiple agents collaborating in a shared
Multi Agent environment (memory, context, tools) to perform a
Agents based on operational setup System certain task, achieve a goal, or deliver utility

▪ Hierarchical agents – Layers of sequential or multi-

Hierarchical Agent agent systems with higher-level agents orchestrating
System tasks across lower-level agents
Sources: Nasscom analysis. For key AI definitions, refer to Glossary on page 32. 14
MULTI-AGENT PROCESS EXECUTION: Complex workings of human + AI agentic systems
For Illustration Only
A SIMPLE For Human Operator: Parse a URL and scrap (or copy) relevant data from the site.
TASK For Agentic System: Continuous, self-learning, automated parsing pipeline

Requirement and Context Parsing & Data Processing Quality Control & Continuous Updates
Output Definition Extraction Validation
TYPICAL
▪ Parse a given URL ▪ Check URL ▪ Perform data ▪ Clean data per usage ▪ Set update frequency
PROCESS ▪ Scrap data ▪ Parse scrapable data cleaning guidelines ▪ Create a calendar
STEPS ▪ Input data into pre- types ▪ Load data in ▪ Maintain data log for trigger
defined format ▪ Scrap data readable format reporting ▪ Reinitiate process

Agent confirms frequency of update with user,

Output for user check after orchestrator go-ahead
Human can (re)define process, frequency
Human can redefine output format
1. Human user
HUMAN
provides URL Prompt, URL not understood Share audit log with user
Data format access failure User can refresh ethical AI rules
Validate data
Auto process restart
formats;
Final output
Agent 1: Orchestrator Agent 2: Agent 3: Data Agent 4: Quality Agent 5: Continuous
Agent Parsing Agent Processing Agent Assurance Agent Monitoring Agent
2. Understand prompt 9. Open a browser 14. Invoke ETL Tools 19. Access ethical AI 24. Record completion
3. Identify requirement (Internet API) 15. Segregate rules 25. Check frequency
SYSTEM OF 4. Establish intent 10. Read the URL structured data 20. Validate output 26. Validate update
5. Check URL access 11. Scrap data (tool) 16. Normalize data against rules cycle with human
AI AGENTS
6. Plan agent sequence 12. Access a data store 17. Categorize data 21. Assess accuracy 27. Set trigger for
7. Identify tools (DB API) 18. Load data in 22. Give confidence Parsing Agent
Share data
8. Initiate audit log 13. Initiate data dump ETL failure readable format scores 28. Update
formats 23. Create audit log Orchestrator Agent
URL scraping failure, Compliance success,
final output to user KEY
data access failure Planned agentic flow
Report success of data load Ethics compliance failure, Failure Pathway
Orchestrator instructs user sharing Failure in creating audit log Success Reporting
Sources: Nasscom analysis. 15
GLOBAL ENTERPRISE TRENDS
Survey Demographics – a global survey of 100 large-medium sized enterprises across major regions and
industries.

AI and GenAI Strategy – Enterprises are maturing in their AI strategy with dedicated budgets and teams;
GenAI maturity, particularly enterprise use of LLMs/SLMs, is emerging.

Enterprise Tech Readiness – Data strategies are also maturing as focus, while on data governance and
unstructured data management, is shifting to building data feedback loops to train AI agents.

Agentic AI Strategy – Organizational intent is strong in investing in AI agents with focus on task-based and
process agents for internal usage before client offerings.

Challenges and Benefits of AI Agents – Benefits expected are transformational at an enterprise level,
however, most enterprises struggle with data privacy and security, unexpected outcomes from AI, and the
currently prohibitive costs of scaling AI agents.
SURVEY DEMOGRAPHICS: Global awareness and adoption survey of 100 enterprises

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 17
THE AI FOUNDATION: AI spend is strategic now; GenAI readiness is emerging

Nature of enterprise AI strategy, 2025 AI spend as percent of tech budget, 2025

With dedicated budgets 27%
and AI-skilled talent 26%
24%
becoming central to an 68%
enterprise’ AI strategy…
14%
11%
…AI budgets as part
of tech budgets are 20%
inching upwards, 7% 6%
from sub-10% before
2023 to 16-20% on Dedicated Dedicated Formalizing Within 6-12 6-10% 11-15% 16-20% 21-25% Greater
average in 2025 teams and budgets, teams budget and months than 25%
budgets being planned team

Enterprise GenAI adoption strategies, 2025

The second important
GenAI productivity tools (co-pilots, assistants) 73% foundation for AI agents,
AI-ready infrastructure 66% that of GenAI models
or LLM variants, needs
Internal/ external GenAI platforms 62% better adoption…
Finetuned commercial LLMs 51%
…than the current 51%
MLOps/ LLMOps framework 47% finetuning LLMs and
Build own SLMs 40% 40% building use
case specific SLMs
Make own LLMs 23%

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 18
THE AI FOUNDATION: AI-ready data pipelining strategies are high on priority

Prioritized data strategies to fast-track adoption of Reinforcement learning is a technique through which models learn
AI agents, 2025 optimal behavior and reasoning through trial and error loops in
dynamic environments. It helps models learn continuously without
episodic re-training needs. RL or RLHF (with human feedback)
Creating data feedback loops for continual work by building data feedback loops for continual agent learning
agent learning 71% through automated data labeling, user feedback integration, and
outcome-tracking pipelines.

This technique is core to making AI agents learn from their

Streamlining of data management and
governance 68% environments and respond without coded logic or human input at
every step. Over 70% of the companies globally are investing in
these data feedback techniques, indicating growing interest in
deploying AI agents.
Data systems for structured and
unstructured data stream integration 62% Priorities are divergent at present, based on existing enterprise
data readiness and tech capabilities. While a significant percent of
hi-tech companies are building RAG-based real-time data
ingestion techniques to minimize hallucinations with better
Reengineered relevant processes around context, companies more mature in their AI spending are actively
agentic AI flows 58%
working on reengineering processes around agentic AI workflows.

Still, majority of global enterprises continue to revamp existing

data systems to handle unstructured data and also streamline
RAG-based real-time data ingestion 55% governance, two major aspects around data readiness for AI that
will only amplify with AI scale.

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 19
EXPERIMENTING WITH AI AGENTS: Encouraging signs of early adoption at 62% firms

▪ 62% of global enterprises are actively working on AI agents: AI Enterprise AI agent adoption strategy, 2025
agents make AI implementation practical and results-oriented.
From PoCs to production and scale-up, significant enterprise
action is shaping up to deploy AI agents, compared to the
Identifying use cases,
relatively constrained adoption of GenAI when the technology
PoCs soon
was introduced.

▪ Hi-tech companies and those with strong AI budgets are 16%

21% PoCs ongoing
moving fast: Over 25% of enterprises in either subset have
reported PoC-to-production deployments, while over 35% have
indicate successful controlled production and potential for
scaling agents.
PoC-to-production in
▪ Companies in the wait-n-watch mode: Interestingly, large 8% process
companies with revenues above $5 Bn are adopting a more
cautious approach as 25% of them indicate planning PoCs in
next 6-12 months, whereas, only 11% of the more mature AI 33% Successful controlled
companies with dedicated budgets and teams are yet to take
production, scale-up in 6
any steps towards experimenting with AI agents 21% months
▪ Manufacturing ahead of Services: 62% of manufacturing
companies, across process and discrete, have initiated PoCs, Planning PoCs in next 6-
moved to production, or started scaling up, compared to 12 months
around 57% of services companies.

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 20
EXPERIMENTING WITH AI AGENTS: 88% expect dedicated spend on AI agents in FY26

Potential spend allocation for AI agents, 2025 Over 3-in-4 enterprises are experimenting with rule-based
conversational AI assistants as a quick win solution to test AI agents,
along with task-based and process agents with human-in-loop.
Ad-hoc tech spend 2%
Part of AI While the aspiration could be autonomous agents across the board,
10% spend
spend, no Dedicated major interest and piloting has emerged from large companies above
Up to 10% of AI $5 Bn revenue, not surprisingly, given the complexity and cost of
allocation yet within AI Spend
spend for agents creating self-learning AI models.
33%

Types of AI agents in enterprise focus, 2025

55%
Up to 20% of AI
spend for agents
Process agents with human-in-loop 77%

Task-based AI agents 77%

With nearly 90% of global enterprises expecting some form of
definitive spend for AI agents, likely from within dedicated AI
budgets,… Rule-based AI assistants 75%
…adoption of AI agents is expected to get a strong boost in
FY26. The trend is uniform across all kinds of industries, revenue Self-learning autonomous agents 46%
segments, or regions.

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 21
EXPERIMENTING WITH AI AGENTS: Structured use cases maturing; personalization still
evolving

Enterprise tasks that AI agents are performing well today, 2025

Majority enterprises are experimenting with AI agents to optimize
existing processes: Automation of repeatable tasks that require
structured human input and exception-based human oversight, such
as data entry, L1/L2 customer service, hospital registrations, etc. Process automation with regular
human interaction 65%

AI-based agents in industrial robotics or collaborative robots Tasks at the interface of physical or
(cobots) with well-established human-machine interface (HMI) is a industrial AI and humans 60%
fast-emerging frontier: AI-powered quality control in manufacturing,
predictive maintenance, or AI-assisted robotics in industrial or
surgical scenarios, are rapidly shaping use cases for AI agents. Standardized content search,
summarization, and generation 58%

Standardized natural language-based tasks or content generation

draw on the strengths of natural language processing (NLP) and Goal-oriented personalized assistance 49%
generative AI in structured data environments.

Orchestration across multiple AI agents

Agents with greater agency are still evolving. These are key to or rule-based bots 41%
building agentic systems that deal with deeper personalization,
context-shifting, goal-(re) orientation, and high-degree of external
Empathetic, intent-based tasks that
interaction. Particularly, tasks that require empathetic reasoning 27%
mimic human behavior
prove limiting for today’s AI agents to interpret and execute.

0% 10% 20% 30% 40% 50% 60% 70%

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 22
EXPERIMENTING WITH AI AGENTS: “Client zero” and client-focused, both are active
scope

Current organizational scope of AI agents, 2025 With GenAI, and now AI agents, a shift in the implementation
approach has emerged, with focus on “client zero,” or internal
successful deployments, before taking solutions to the clients.

Internal functional automation (HR, F&A, 3-in 4 global enterprises are looking at functional scope for AI
employee services, etc.) 76% agents: Functional AI agents are multiple agents in a multi-agent
system (MAS) coordinating across small, interconnected tasks in a
process, and through multiple processes across a function.

Building agentic AI solutions for clients 67% Inherently, such agentic systems work on persona-based context,
dynamic reasoning, frequent data refreshes, and networked
configurations. Such agentic systems also depend heavily on
continuous reinforcement learning, with human feedback in the
Internal standardized task automation (test beginning. Highly standardized functional automation serves as a
script generation, code validation, etc.) 67% strong use case for autonomous agents over time.

Over 80% of hi-tech companies are focused on building AI agents

for their clients. These companies are also prioritizing internal
Value chain integration amongst key supply functional and cross-functional automation with AI agents.
chain partners 61%
With GenAI, value chain optimization has emerged as an
important use case as models learn from the value chain
Internal cross-functional process interactions and data flows to optimize resource, asset, and time
automation (lead-to-cash, etc.) 57% efficiencies. Over 60% of enterprises are focused on piloting multi-
agent systems for autonomous value chain operations.

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 23
EXPERIMENTING WITH AI AGENTS: ITOps, manufacturing clients lead early adoption

Enterprise functions with ongoing AI agent experiments, Client sectors with ongoing AI agent experiments, 2025
2025

Public services 43%

IT operations 81%
Transport & logistics 52%
IT Help desk 49% Education 52%
CPG & retail 56%
Customer Services 31%

Services
Healthcare lifesciences 57%
Production/ Operations 27% Telecom 59%
Media & entertainment 60%
Supply Chain and Logistics 25% Energy & utilities 62%
Sales & Marketing 18% Banking & FS 62%
Insurance 63%
Finance & Accounting 18% Travel & hospitality 69%
HR (includes L&D) 15%

Manufacturing
Industrial mfg. 71%
R&D 13%
Corporate Communications 13% Hi-tech manufacturing 69%

Legal 8% Pharmaceuticals 50%

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 24
EXPERIMENTING WITH AI AGENTS: Enterprise-scale benefits anticipated, however
scaling AI agents has pertinent challenges

Perceived benefits from agentic AI systems, 2025 Current challenges with scaling task-based AI agents, 2025

Rapid information-to-intelligence Data privacy & security risks 60%

for accelerated decision-making 57%
Unknown risks of self-learning AI 56%
Better ability to respond to new Inconsistent/no regulations 51%
business opportunities 55%
Prohibitive infrastructure cost 48%
Reduction in failure rates from PoC-to-production failure 47%
human-led process optimizations 47%
High AI ownership TCO 45%
Adaptable, expertise-based Mindset resistance to human+AI 44%
organizational structures 46%
Limited or unproven RoI 44%
Freeing-up of workforce for Tech stack integration challenges 38%
higher-order work 39%
AI & sustainability conflict 38%
Rapid R&D along with superior Limited scalable use cases 38%
product-market fit by design 38%
Concerns with data sharing 33%
Seamless, omni-channel customer Limited data for AI models 30%
experience 35%
Limited supply of AI talent 27%

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 25
PERSPECTIVES
Making Agentic AI Real – Managing the risks associated with AI will require enterprises to think beyond
existing enterprise risk frameworks, majority organizations are wanting in this space.

Learnings from Enterprise Indications – The collective learnings from this early enterprise adoption and
inclination survey indicates a strong intent, transformational potential of agentic AI, but equally limiting
challenges and risks starting with organizational data readiness.

The Human + AI Future – Agentic AI evolution in the near-mid term is expected to hinge a lot more on
synergistic operations and feedback loops with human operators, rather than in fully autonomous operations.
MAKING AGENTIC AI REAL: Ability to build trust-worthy AI will determine adoption
timelines for truly enterprise-scale agentic systems, despite enterprise optimism

Risk management actions that enterprises plan to prioritize Expected timelines by when enterprises expect meaningful
to become prepared for agentic AI systems, 2025 agentic AI adoption or competitive disruption, 2025

I. AI audits and risk reporting fully-integrated with enterprise risk Agentic AI will have no
management (ERM) impact on our business
13%
II. AI led IT observability (AIOps) to ensure a complaint tech stack 25% Within 6 months
In 2 years or more
III. Proactive compliance with national or global AI regulations, even 10%
guidelines
IV. AI risks are handled on an as-is basis
V. Procure/build only fully responsible and traceable AI systems
VI. Continuous monitoring and upgrades to enterprise responsible AI 24%
guardrails 28%
Within 18 months Within 12 months
VII. Rigorous review of enterprise-level responsible AI guardrails
VIII. Specific hardware assurance and model observability audits

Less than 50% of the companies have initiated specific risk-mitigating …Which may further elongate the adoption of truly impactful agentic AI
systems at an enterprise-scale, despite 53% of the enterprises sounding
initiatives focused on curbing ethical, accuracy-related, observatory,
optimistic around adoption in less than a year, although driven more by
and trust-worthiness risks emanating from agentic AI usage… potentially looming disruption than demonstrable value or RoI.

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 27
LEARNINGS FROM ENTERPRISE TRENDS: Enterprises are rapidly experimenting with AI
agents, although readiness for multi-agent, collaborative agentic systems will take time
I. Agentic AI is the next evolution, building on predictive and generative AI, with strong enterprise interest as enterprises are rapidly experimenting
with AI agents, with 62% already active in various stages.

II. Dedicated AI budgets are a crucial factor in enabling AI-related business decisions going from paper to execution, and this is likely to reflect into
agentic AI spend as more of AI spend gets steered towards process-automating agents

III. However, while overall awareness with generative AI is strong, maturity with usage of models for enterprise use cases is still emerging as at best,
51% of companies are actively using finetuned foundation models or LLMs

IV. AI-ready data strategies continue to be top priority, with companies focusing more on unstructured data and real-time ingestion. It is encouraging
to see focus on building continual data feedback loops for reinforcement learning, RAGs, and real-time data ingestion as priorities.

V. Agentic AI adoption depends on creating more specific data strategies, such as building agent-ready data architecture, investing in high-quality,
task-specific datasets, and establishing data standards for agentic use.

VI. As a result, the current focus for AI agent implementation is on optimizing existing processes and functional automation, more so starting with
highly-structured tasks and functions, such as IT operations and help desk that are non-risky, internal and within full enterprise control in case of
any business disruption

VII. "Client zero" (internal deployments) is a key initial focus before external client solutions since organizations are cautious of brand and reputational
damage and legal consequences, that too in an unstandardized regulatory environment.

VIII. Therefore, the long-term (5-10 years) future of enterprises mulling agentic AI adoption is to either not act and get fully disrupted by competition in
the near-mid term or it is to move rapidly beyond “agent washing” to build truly collaborative human + AI agentic systems.

Sources: Nasscom survey of Global 100 enterprises in April 2025. nasscom analysis. 28
THE HUMAN + AI FUTURE: We believe that human + AI will be a more sustainable,
ethical, trust-worthy approach to building agentic AI autonomy and agency
For Illustration Only

Increasing autonomy and greater trust with real-world AI agency

Phase 1: <12 months Phase 2: 12-24 months Phase 3: 36 months+

Human Agentic Human Agentic Human Agentic

Intelligence Intelligence Intelligence Intelligence Intelligence Intelligence

Complete operational, strategic, ethical Strategic intelligence tending to legal, Long-term strategy with ethical, spiritual,
Human intelligence; creative storytelling emotional, environmental foresight; ethics; societal oversight; post-crisis revival planning
Input creative output

AI Agent Task-based intelligence; rule-based instruction Process intelligence established; emerging Self-regulating, autonomous operations;
Execution execution; limited goal orientation functional focus with agent2agent interaction compliance monitoring; alignment with EQ

Human exception handling; agent assistance in Seamless execution collaboration in complex Autonomous execution of human strategy, with
Human+AI natural language processing; outcome processes; AI-led decisions; rising ethical control increasing agentic inputs; precision
Synergy validation coordination

Sources: Nasscom analysis. 29

 “A lot of people in industry and
computer science research are
creating fancy agents, but very few
are looking at the interactions
between humans and these tools,”

MIT Research

30
APPENDIX
Glossary

Acknowledgements

About nasscom

Disclaimer
GLOSSARY

Prompt A short instruction or idea given to an AI or computer to help it understand what to do or create

RAG Retrieval-Augmented Generation : An AI method where it first looks up facts, then uses them to give a better answer

Reinforcement Learning from Human Feedback :A way to train AI using human preferences to make its answers more helpful and
RLHF
safe

NLP Natural Language Processing: A branch of AI that helps computers understand and use human language

Context The background information an AI uses to understand and respond correctly

Human + AI
When humans and AI work together, combining human judgment with AI speed and data skills for better results
Systems

LLM An advanced AI trained on huge amounts of text to understand and generate human-like language

Orchestration Coordinating multiple AI tools or agents to work together smoothly to complete complex tasks

Multi Modal AI that can understand and use different types of input like text, images, audio, or video together

32
ACKNOWLEDGEMENTS
We sincerely thank the survey participants and their leadership for generously sharing their time and valuable insights, which have been vital in shaping this
report. We thank our survey partner, Dynata, for enabling a global survey across major world regions and over 11-12 industries, for curating, hosting the survey
and providing data collation services. We also extend our appreciation to the Nasscom AI team for continuous guidance on the subject and helping refine our
analysis.

Research Authors

Nasscom Insights Team

Janhvi Juyal
Senior Associate, AI

Namita Jain
Director, Digital Transformation

Achyuta Ghosh
Senior Director and Head of Research

33
 ABOUT NASSCOM
Nasscom represents the voice of the $280 billion+ technology industry in India with the vision to establish the nation as the world’s leading technology
ecosystem. Boasting a diverse and influential community of over 3500+ member companies our network spans the entire spectrum of the industry from
DeepTech and AI start-ups to multinationals and from products to services, Global Capability Centres to Engineering firms. Guided by our vision, our
strategic imperatives are to accelerate skilling at scale for future-ready talent, strengthen the innovation quotient across industry verticals, create new
market opportunities - both international and domestic, drive policy advocacy to advance innovation and ease of doing business, and build the industry
narrative with a focus on Trust, and Innovation. And, in everything we do, we will continue to champion the need for diversity and equal opportunity.

nasscom Insights is the in-house research and analytics arm of nasscom generating insights and driving thought leadership for today’s business leaders
and entrepreneurs to strengthen India’s position as a hub for digital technologies and innovation.

Visit our e-community at https://community.nasscom.in

Plot 7 to 10, Sector 126, Noida – 201303, India

Phone: +91-120-4990111
Email: [email protected]

www.nasscom.in
Visit our community at
www.community.nasscom.in

34
 DISCLAIMER
The information contained herein has been obtained from sources believed to be reliable. nasscom and its advisors & service providers disclaim all warranties
as to the accuracy, completeness or adequacy of such information. nasscom and its advisors & service providers shall have no liability for errors, omissions or
inadequacies in the information contained herein, or for interpretations thereof. The material or information is not intended to be relied upon as the sole basis
for any decision which may affect any business. Before making any decision or taking any action that might affect anybody’s personal finances or business,
they should consult a qualified professional adviser.

Use or reference of companies/third parties in the report is merely for the purpose of exemplifying the trends in the industry and that no bias is intended
towards any company. This report does not purport to represent the views of the companies mentioned in the report. Reference herein to any specific
commercial product, process or service by trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement,
recommendation, or favouring by nasscom or any agency thereof or its contractors or subcontractors.

The material in this publication is copyrighted. No part of this report can be reproduced either on paper or electronic media without permission in writing from
nasscom. Request for permission to reproduce any part of the report may be sent to nasscom

Usage of Information

Forwarding/copy/using in publications without approval from nasscom will be considered as infringement of intellectual property rights.

35
 Plot 7 to 10, Sector 126, Noida – 201303, India
[email protected]

Phone: +91-120-4990111 | Email: [email protected]

www.nasscom.in

www.nasscom.in
Visit our community at
www.community.nasscom.in

www.community.nasscom.in
/nasscomOfficial /nasscom /nasscomVideos

/nasscomOfficial /nasscom /nasscomVideos

Copyright © 2025