The word "agent" has done a great deal of work in the past eighteen months, most of it in slide decks. Vendors describe anything that calls a model twice as agentic. Boards ask why the company does not yet have agents. Transformation leads are left to answer a narrower and more useful question: of the workflows this enterprise actually runs, which ones can an AI agent take on in 2026 without creating a new class of operational risk?
This piece answers that question for AI agents in enterprise workflows with a deployability matrix rather than a prediction. The evidence base is thin but no longer empty. McKinsey's 2025 global survey found 23% of respondents scaling an agentic AI system somewhere in the organisation and another 39% experimenting, yet in no single business function did more than 10% report scaling agents. Gartner, in June 2025, predicted that over 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Both findings are consistent with what practitioners see: agents are real, narrow and expensive to get wrong.
The matrix below rests on four tests that any candidate workflow must pass before an autonomous or semi- autonomous agent should be allowed to act in it. It is an analytical device, not a survey finding, and the scoring of individual workflows is a judgement that readers should check against their own operations.
The state of AI agents in enterprise functions in 2026, measured rather than announced
Three data points anchor the discussion.
First, McKinsey's The state of AI in 2025, a survey of 1,993 participants in 105 nations fielded between 25 June and 29 July 2025, reports that 88% of organisations use AI regularly in at least one function, up from 78% a year earlier. On agents specifically, 23% are scaling an agentic system somewhere, 39% are experimenting, and in any given function no more than 10% are scaling. Agent use is most common in IT, for example service-desk management, and in knowledge management, for example deep research. High performers, about 6% of respondents who attribute 5% or more of EBIT to AI, are at least three times more likely than peers to be scaling agents in most functions.
Second, Gartner's June 2025 press release estimated that only about 130 of the thousands of vendors describing themselves as agentic AI companies are "real", and coined "agent washing" for the practice of relabelling chatbots or robotic process automation as agents when they lack substantial agentic capabilities. A January 2025 Gartner poll of 3,412 webinar attendees found 19% had made significant agentic AI investments, 42% conservative investments, 8% none, and 31% were waiting or unsure. Gartner's forward view is nonetheless bullish: at least 15% of day-to-day work decisions made autonomously through agentic AI by 2028, up from essentially zero in 2024, and 33% of enterprise software applications including agentic AI by 2028, up from under 1%.
Third, for India, EY's AIdea of India Outlook 2026 survey of more than 200 enterprise leaders, published in November 2025, found 23% at early stage or pilots, 47% with multiple GenAI use cases live in production, and 10% scaling across the business. 91% ranked speed of deployment as the top factor in buy-versus-build decisions. That last figure matters for agents: speed pressure is exactly the condition under which the four tests below get skipped.
Read together, the picture is of an enterprise market that has moved past curiosity but is still concentrated in two or three functions where the work is already digital, bounded and logged. That is not a coincidence. Those are the functions where the deployability tests are easiest to pass.
A working definition, and why the definition changes the risk
For the purpose of this analysis an AI agent is a system that, given a goal, decides which steps to take, calls tools or systems to take them, observes the result and continues until the goal is met or a stopping condition is reached. The distinguishing feature is not the language model. It is that the system takes actions in other systems with less than full human review of each action.
That distinction is what separates an agent from a copilot. A copilot drafts; a human sends. An agent sends. The moment an enterprise moves from the first to the second, three things change. The error cost shifts from "a human caught a bad draft" to "a bad action reached a customer, a ledger or a supplier". The audit question shifts from "who approved this" to "what did the system do, in what order, on what evidence". And the vendor-management question shifts from "is the output accurate" to "what is this system permitted to do, and how do we know it stayed within that".
McKinsey's survey records the consequence: 51% of respondents from AI-using organisations have experienced at least one negative consequence, and nearly one-third report consequences from inaccuracy, the most commonly experienced and the most commonly mitigated risk. Explainability is the second most experienced risk but is not among the most mitigated. For agents, which chain many model outputs together, both gaps compound.
The agent deployability matrix
The matrix applies four tests to a workflow. A workflow that passes all four is a candidate for an agent acting with light human oversight. One that fails any test is a candidate for a copilot, a human-in-the-loop design, or nothing yet.
Test 1: Is the task bounded? The agent's goal can be stated in advance, the set of permissible actions is finite and enumerable, and completion is unambiguous. Password resets, invoice matching and ticket triage are bounded. "Improve supplier relationships" is not.
Test 2: Are the actions reversible? If the agent acts wrongly, can the action be undone at low cost and within a short time? Creating a draft purchase order is reversible. Releasing a payment is reversible only with effort. Sending a regulatory filing is, practically, not.
Test 3: Is the trail auditable? Every action the agent takes is logged with its inputs, the tools it called, the outputs it received and the decision it made, in a form a human or an auditor can reconstruct later. This is a property of the surrounding system as much as of the agent.
Test 4: Is the data available and clean enough? The agent can reach, with the right permissions, all the data it needs to make the decision, and that data is of a quality the enterprise would accept for a human making the same decision. A separate PanScience Research piece covers data readiness in detail; here it is simply the fourth gate.
The table scores twelve common enterprise workflows against the four tests. Scores are the authors' judgement for a typical large Indian enterprise with mature core systems; a given company may differ.
Two patterns stand out. The deployable rows cluster in exactly the functions McKinsey found agents most common in, IT and knowledge management, and in finance and service operations where the "act" step can be stopped short of the irreversible action. The not-yet rows fail on reversibility or on data, rarely on boundedness. In other words, the obstacle to wider agent deployment in 2026 is not that enterprises lack well-defined tasks. It is that the consequences of error are asymmetric and the underlying data is not yet trusted.
Reversibility as the design variable
If one lesson generalises from the matrix it is that reversibility is the variable an enterprise can engineer. The others are mostly inherited. A task is bounded or not. Data is clean or not, at least in the short run. But the point at which an agent stops and a human takes over is a design decision.
The practical pattern in 2026 is to let the agent do everything up to the irreversible action and then hand over. In finance, the agent proposes journals and the controller posts. In procurement, the agent shortlists and the category manager awards. In compliance, the agent assembles the submission and a named officer files. This design captures most of the labour saving, because the preparation is where the hours were, while keeping accountability where regulators and auditors expect it.
It also happens to align with how Indian regulators are thinking. The RBI's FREE-AI committee report, released in August 2025, lists as high-risk "autonomous AI systems that handle customer interactions, make financial decisions, or move customer funds", recommends that regulated entities adopt a board-approved AI policy covering governance structure, accountability, risk appetite, auditability and a model lifecycle framework, and proposes that the Risk Management Committee or a similar body own AI risk. These are recommendations, not yet binding regulation, but a bank that lets an agent move funds without a human gate in 2026 is working against the direction of supervisory thinking. A separate PanScience Research piece covers the FREE-AI framework for BFSI compliance in depth.
Auditability is a platform property, not a model property
The third test, auditability, is where most agent pilots that pass the first two tests still fail in production. A demonstration agent runs in a notebook; nobody asks where the logs go. A production agent runs across six systems, and the enterprise discovers that it cannot answer a simple question: what did the agent do last Tuesday, on whose data, and why?
The answer in practice is a control layer between agents and the systems they act on. It records every call, enforces which tools each agent may use, applies data access policy, and provides a kill switch. Several PanScience ventures work at this layer and serve as examples of the pattern rather than of any specific outcome. Proteus is described by the company as an LLM gateway that routes, secures and monitors model traffic, which is the enforcement point for logging and policy on every model call an agent makes. Own360 is described as a control plane for AI-native enterprises, the layer from which an organisation sees and governs the agents running across its estate. PanScience reports these as its own products; this piece makes no claim about their deployment scale.
Whatever the tooling, the test is simple: can a non-technical auditor, given a transaction ID, reconstruct the agent's reasoning and actions from logs alone? If not, the workflow fails Test 3 regardless of how well the model performs.
Why cost, not capability, will cancel most agent projects
Gartner's cancellation forecast names escalating costs first. This deserves more attention than it gets. An agent that reasons over a long context, calls tools, re-reads the results and iterates can consume ten to fifty times the tokens of a single model call for the same nominal task. At pilot scale, nobody notices. At the scale of a service desk handling tens of thousands of tickets a month, the inference bill becomes a line item, and the business case written for a copilot no longer holds for an agent.
A sound business case for an enterprise AI agent therefore needs a unit-cost line that includes the full action loop, not the single prompt. It also needs an assumption about containment rate, the share of cases the agent completes without escalation, because every escalation costs both the agent's tokens and the human's time. A separate PanScience Research piece builds a unit-cost model for one workflow; the point here is that agent economics are a different animal from copilot economics and should be modelled as such before procurement, not after.
What to do in the next twelve months
For an Indian enterprise that has not yet deployed an agent, the matrix suggests a sequence.
Start in IT and internal knowledge work. These rows pass all four tests in most large organisations, the data is already in digital systems, the users are technically literate and tolerant of early rough edges, and the actions are low-stakes and reversible. McKinsey's finding that this is where agents are most common globally is a reason for confidence, not a reason to look elsewhere for differentiation.
Build the control layer before the second agent, not after the fifth. Logging, tool permissions, data policy and a kill switch cost little to put in for one agent and a great deal to retrofit across many. The enterprises that will be able to scale agents in 2027 are the ones whose audit trail already exists.
Design the hand-over point explicitly. For every workflow in the matrix, write down the action the agent may not take. Make that a configuration, not a convention.
Measure containment and unit cost from day one. An agent that resolves 40% of cases autonomously at a cost below the human equivalent is a success. One that resolves 90% at three times the cost is a demonstration.
Treat vendor claims as claims. Gartner's figure of roughly 130 real agentic vendors among thousands is an industry estimate, not a list, but it is a reminder to ask any vendor to show the action log of a live deployment, not a demo.
Conclusion
The honest state of AI agents in enterprise workflows in 2026 is that they are deployable, with light oversight, in a specific and recognisable set of bounded, reversible, auditable, data-rich tasks, and that this set is smaller than the marketing suggests and larger than the sceptics allow. McKinsey's survey puts scaled agent deployment at under 10% in any single function; Gartner expects a large minority of projects to be cancelled on cost and value grounds within two years. Neither figure argues against agents. Both argue for choosing the workflow with the four tests in hand, engineering the hand-over point, and building the audit layer first. The enterprises that do that will find the matrix fills in from the top down over the next three years. The ones that do not will supply Gartner's cancellation statistic.
Methodology and sources
The deployability matrix and its four tests are an original analytical framework developed for this article. The scoring of the twelve workflows is the authors' judgement for a typical large Indian enterprise with mature core systems and is illustrative, not the result of a survey. Observed figures come from the following sources, each opened and read for this piece. McKinsey, The state of AI in 2025, survey of 1,993 participants in 105 nations fielded 25 June to 29 July 2025 (https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-state-of-ai-2025): 88% regular AI use, 23% scaling agents, 39% experimenting, no more than 10% scaling in any function, agent use most common in IT and knowledge management, high performers about 6%, 51% experienced a negative consequence, inaccuracy the most common. Gartner press release, 25 June 2025, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over- 40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027): cancellation forecast and causes, about 130 real vendors, agent washing, January 2025 poll of 3,412 attendees, 15% of decisions and 33% of applications by 2028. EY India, The AIdea of India: Outlook 2026, Chapter 2 survey findings, published 16 November 2025, 200+ Indian enterprise leaders (https://www.ey.com/en_in/insights/ai/agentic-ai-india/aidea-survey-findings): 23%, 47%, 10% and 91% figures, as verified by the editor. RBI, Report of the Committee on Framework for Responsible and Ethical Enablement of AI (FREE-AI), released 13 August 2025 (https://rbi.org.in/scripts/PublicationReportDetails.aspx? ID=1306): high-risk use cases at paragraph 4.4.44, Risk Management Committee role at 4.4.45, Recommendation 14 on a board-approved AI policy; these are committee recommendations and not binding regulation. PanScience Innovations facts, including the descriptions of Proteus and Own360, are company-reported from panscience.xyz, read 24 October 2026, and no deployment claims are made for them.