Agentic AI, Explained: The 22% Category Built to Act, Not Just Answer

By Diego Navia · BizBlocz · May 2026

Part of the AI Explained series. Start with the overview →

Agentic AI is the category every enterprise vendor is selling and most enterprises are focused on shipping in 2026. Gartner projects that by 2028, roughly a third of enterprise software will embed agentic AI, up from under 1% in 2024. We are one cycle into a thirty-fold move, and almost nobody agrees on what an agent actually is.

Here is the practitioner definition. An agent is a piece of software that sets a goal, plans steps, calls tools, observes what comes back, and decides what to do next, without a human pressing enter for each step. Machine learning predicts. Generative AI drafts. Agentic AI acts.

What agentic AI actually does

The honest frame before any taxonomy: most of what ships in 2026 as agentic sits much closer to deterministic automation (RPA) than the marketing suggests. The two ends of the axis are worth naming in plain English.

Deterministic automation is software that follows a decision graph a developer hand-coded up front. Every branch, every tool call, every escalation is written in advance. An RPA bot clicking through SAP screens in the same order every night is deterministic. Predictable. Brittle.

True agentic autonomy is the other end. The model reads a goal, decides what to do next at runtime, picks its own tools, evaluates what came back, and adjusts without a pre-written script. Flexible. Fragile.

The practitioner question is where on that axis a given system actually sits, not what the slide deck calls it. A rough spectrum is useful.

Level 1: LLM-wrapped workflow. The decision graph is hand-coded. The model phrases the output and occasionally picks from a small menu of templates. The bulk of live Salesforce Agentforce, ServiceNow Now Assist, and SAP Joule deployments sit here. That includes Heathrow's Agentforce-built WhatsApp agent "Hallie," which Salesforce reports resolves roughly 90% of traveler queries without a human handoff.
Level 2: Tool-calling chatbot. The model picks one of a small pre-defined set of tools, executes one step, returns. Microsoft 365 Copilot across Office, Teams, and calendars is the archetype. Accenture has disclosed 100,000 employees on M365 Copilot with a stated commitment to scale to roughly 200,000. Shopify, the first external GitHub Copilot customer in 2022, reports ~80% engineer adoption and, per VP of Engineering Farhan Thawar, roughly 20% team productivity gain; they run Copilot alongside Cursor, Claude Code, and OpenAI Codex behind a unified proxy.
Level 3: Bounded-horizon planner. The model decomposes the goal, calls a fixed tool set, iterates across several steps, stops at a pre-declared rule. GitHub Copilot Workspace (still in technical preview), Cursor agent mode, and most of the engineering-agent category live here. This is where the three classic agent ingredients (planning, tool use, memory across steps) start doing real work. Publicly sourced Fortune 500 production deployments at this level are still thinner than the marketing suggests.
Level 4: Open-horizon autonomous agent. The model plans, calls tools it sometimes discovers, self-corrects across many steps toward a goal stated in plain English. Cognition's Devin, Anthropic's Computer Use, OpenAI's Operator and Deep Research live here. Real, narrow-scope, and they still fail often.

The honest read of production in 2026 is that the overwhelming majority of deployed enterprise agents sit at Level 1 or Level 2. The branding says agentic. The decision graph was pre-written. That is not a criticism. LLM-wrapped workflows deliver real value. It is a naming problem procurement teams should catch before they sign.

Research supports the caution. METR, the independent AI evaluation lab, tracks how long an autonomous task an agent can actually finish; the horizon is doubling roughly every seven months but still caps at a few hours of human work, not weeks. Gartner now distinguishes agentic AI from AI agents for exactly this reason.

What agentic AI does not do is worth naming while we are here, because the market keeps mixing it up. It does not predict which customer is about to churn. That is machine learning. It does not draft the 40-page memo from scratch. That is generative AI. It does not read the handwritten invoice. That is document AI. Agentic calls those services in sequence to get a job done.

AI agent types

Four shapes recur in enterprise deployments, and each one maps onto the spectrum above.

Single-task agents: one job, one tool set. Mostly Level 1. An inbox triage agent that reads tickets, classifies, and routes.
Conversational agents: a chat surface with tools behind it. Level 1 to Level 2 in practice. Salesforce Agentforce is the largest deployed example.
Coding and knowledge-work agents: agents that act inside a developer's or analyst's environment. Level 3. GitHub Copilot Workspace, Cursor, JPMorgan's internal LLM Suite.
Orchestration agents: agents that coordinate other agents. The one shape aiming at Level 4. Microsoft Copilot Studio, UiPath Autopilot, Automation Anywhere AI Agent Studio.

Multi agent systems

The Level 4 topology. A planner agent decomposes, worker agents execute, a judge agent reviews. The theory is the most interesting part of the field. The production failure modes are the most expensive. Communication overhead, circular delegation, and silent pass-throughs compound faster than a single-agent loop, and the audit trail is harder to reconstruct after the fact. This is why multi-agent systems remain rare in production outside well-bounded engineering and research-assistant scenarios.

The practitioner test for any of these shapes: strip the hand-written prompt template and the fixed tool list. Does the remaining system still have a sensible action space? If no, it is a workflow with LLM polish. Useful, often valuable. But not the autonomy the slide deck promised.

Where agentic AI lands in the enterprise

Across 127 enterprise subprocesses we mapped, agentic AI is the leading technology in roughly 22% of them. Four patterns are already in production.

Agentic AI use cases in customer service

Heathrow Airport runs a Salesforce Agentforce-built agent called "Hallie" on WhatsApp. Gate locations, security wait times, in-terminal wayfinding, amenity details, and FAQ handling. Salesforce's own reporting puts query resolution at roughly 90% without a human handoff, with case-summary accuracy targeted at 95% and live-chat times cut by about 30 seconds. The shape is the Level 1 end of the spectrum: a hand-coded flow over an LLM that phrases the answer. That is genuinely valuable for a passenger experience team. It is not, by any honest reading, the agent planning and reorchestrating a trip.

AI agents enterprise — knowledge work

JPMorgan's LLM Suite, an in-house portal that taps OpenAI and Anthropic models, has been rolled out to roughly 250,000 of the bank's employees (excluding branch and call-center staff), with about half using it daily per CNBC's September 2025 reporting. The on-the-record headline: Chief Analytics Officer Derek Waldron has said the system produced a credible five-page investment-banking pitch deck for a meeting with Nvidia's leaders in about 30 seconds. Important distinction: LLM Suite is a bespoke internal build, not a commercial product any other bank can buy. That is a pattern worth naming. Tier-one financial and pharmaceutical firms are assembling custom agent stacks on top of frontier model APIs because the shelf products do not yet match their data-segregation, audit, and compliance bar. Thomson Reuters, Bloomberg, and every major consultancy has shipped their own version. The productivity headline is noisy. The shift in where senior-analyst time actually goes is not.

Engineering agents

Shopify was the first external enterprise customer on GitHub Copilot, back in 2022 before ChatGPT. VP of Engineering Farhan Thawar has said publicly that adoption sits around 80% of engineers and estimates the team is roughly 20% more productive as a result; engineers run Copilot alongside Cursor, Claude Code, and OpenAI Codex behind a unified LLM proxy. The public disclosures are about adoption and productivity, not a Copilot Workspace production rollout. Copilot Workspace, Cursor agent mode, Anthropic's coding agents, and the newer CI-integrated engineering agents are where the bounded-planner shape (plan, write, run tests, open the pull request) actually lives. This is the fastest-moving agentic surface in the market right now, and the one where production failure is cheapest to catch: the tests either pass or they do not.

AI workflow automation for everyday work

Accenture has publicly disclosed a deployment of Microsoft 365 Copilot to 100,000 of its employees, with a stated commitment to scale to roughly 200,000 users, through a joint announcement with Microsoft and Avanade in November 2024. It is the version of agentic AI most people will actually touch. A tool-calling layer working across the Office, Teams, and calendar surfaces the user already lives in every day, priced at a per-seat subscription instead of a platform contract, which is why this category is about to move from the boardroom into the inbox. The specific internal use cases Accenture assigns to its Copilot seats are not something the company has broken out publicly, and we will not invent them.

The RPA vendors didn't die. They pivoted.

The most important repositioning in enterprise automation happened quietly in 2024–2025. The RPA category (UiPath, Automation Anywhere, Blue Prism, Microsoft Power Automate) was supposed to be disrupted by LLMs. Instead, the vendors rebuilt.

UiPath rebranded the platform agentic automation. Autopilot plus Agent Builder. The old bots are now robotic agents coordinated by LLM planners.
Automation Anywhere launched the Automation Success Platform with AI Agent Studio. The pitch: generative plus agentic plus RPA as one substrate, not three.
Blue Prism, acquired by SS&C, repositioned as intelligent automation inside the SS&C Blue Prism Next Generation stack.
Microsoft Power Automate added Copilot-native agents in the same canvas that ran RPA. The largest installed base of the four.

The practitioner point: RPA was agentic AI's training wheels. Deterministic workflow orchestration, credentials vaulting, human-in-the-loop approval, audit logging, idempotent retries: the unglamorous plumbing every production agent needs. The LLM planner on top is genuinely new. The fifteen years of enterprise wiring underneath is not. Companies with a mature RPA practice have a head start most AI-native agent startups simply do not have.

The risk is the same risk we flagged for machine learning. A large share of what is being sold as agentic transformation in 2026 is an RPA contract renewal with a chat pane. The seat count renewed. The capability graph barely moved. Both things can be valuable. They are not the same purchase, and the procurement conversation is cleaner when the team names which one it actually bought.

Where another category leads

The procurement clarity gets sharper when agentic is positioned in relation to its five sibling categories.

Agentic is the action layer. It calls everything else. Forecasting the number is machine learning's job. Drafting long-form language or images belongs to generative AI. Reading the unstructured PDF is document AI. Understanding email intent is NLP. Inspecting a physical part is computer vision. In production, an invoice-exception agent calls document AI to extract, machine learning to score risk, generative AI to draft the supplier response, and an RPA bot to post the adjustment to the ERP. The six AI technology categories compose. Agentic is the category that makes the composition possible.

Why agentic AI is 22% of enterprise AI value

The BizBlocz aggregate AI value split we publish puts agentic AI at 22% of aggregate enterprise AI value, second only to machine learning. Two reasons carry the share.

First, agentic is the only category that closes the loop from decision to action. Machine learning tells you the invoice is suspect. Document AI extracts the fields. Generative AI drafts the supplier email. The agent actually sends the email, updates the record, and opens the exception ticket. Without the action layer, the other five categories produce artifacts a human still has to ferry between systems.

Second, the deployment substrate already exists. The RPA installed base, the ITSM workflow engines, the iPaaS platforms: all of it was built for deterministic automation. Agentic AI inherits it. That is a very different starting position than generative AI, which had to invent its own runtime.

None of this makes the share easy to capture. Agents fail in new ways. They loop, they hallucinate tool calls, they over-delegate to each other, they pass silently when they should have stopped. Scoring agentic readiness is structurally different from scoring ML readiness, which is why the AI Value Assessment tool treats them as different problems. The related question, which enterprise platforms actually own the process IP agents need to act on, is the spine of the SaaSpocalypse piece.

The practitioner angle

Herbert Simon, 1971: a wealth of information creates a poverty of attention. Agentic AI is the category sold as the cure for that poverty. The agent absorbs the attention so the human does not have to. Sometimes it does. Sometimes it just moves the attention from reading the ticket to auditing the agent's decision on the ticket.

The discipline is the stopping rule. Before you deploy an agent, name the condition under which it stops and escalates. Refund under $50 auto-approved, over $50 escalate. Migration pull request with zero failing tests auto-merged, otherwise hold. Outbound email drafted but queued for human send until 1,000 clean sends pass. An agent without a stopping rule is not an agent. It is a loop. And it can be expensive.

Agentic AI is the category doing the most changing this year. Which of your subprocesses are genuinely agentic, meaning the deliverable is an action, not an answer, and which ones have been dressed up as agents because agent sold better than workflow in the 2026 budget cycle?

Next in the series: Generative AI, Explained — the category that went from zero to ubiquitous in 36 months, and the one the market keeps confusing with all the others.

Also in the AI-Explained series: Machine Learning, NLP, Document AI, Computer Vision. Start with the overview →

Sources: Gartner Hype Cycle for Artificial Intelligence (2025); Gartner projection on agentic AI share of enterprise software by 2028. Salesforce press release, June 11, 2025 ("Hallie" at Heathrow). Accenture/Microsoft/Avanade joint press release, November 14, 2024 (100,000 M365 Copilot deployment). First Round Review interview with Shopify VP of Engineering Farhan Thawar (~80% Copilot adoption, ~20% productivity estimate). CNBC, September 30, 2025 (JPMorgan LLM Suite, Derek Waldron on the 30-second pitch deck for the Nvidia meeting). METR research on task-horizon doubling. Public reporting on the 2024–2025 platform repositionings from UiPath, Automation Anywhere, Blue Prism (SS&C), and Microsoft Power Automate. Subprocess-level estimates are BizBlocz aggregate research, an analysis of 127 enterprise subprocesses and 245+ data points across 30+ independent research publications. Directional, not decimal-precise.