Agentic AI in the Enterprise: Agents That Do the Work

Agentic AI in the enterprise has moved past the demo stage in many organizations, but the results are uneven. McKinsey's 2026 survey finds 40% of respondents at large organizations (revenue above $1 billion) report scaling AI agents, up from 27% a year earlier. The share at smaller organizations stayed flat at 22% (McKinsey). Adoption is real. So is the gap between adopting agents and getting value from them.

That gap is the procurement problem of the moment. Buyers are being sold "agents that do the work," and some of them genuinely do. Deloitte's August 2026 survey of more than 500 U.S. business and IT leaders found that only 5% say their business processes are highly prepared for AI agents. Just 15% have scaled orchestrated, cross-functional multi-agent adoption (Deloitte). The technology is not usually the bottleneck. The workflow around it is.

What makes an agent different from a chatbot

A chatbot answers a question. An agent pursues a goal. MIT Sloan describes agentic AI as systems that are semi- or fully autonomous and able to perceive, reason and act. They integrate with other software to complete tasks with minimal supervision (MIT Sloan). That shift from answering to acting is what changes the risk profile.

IBM breaks the mechanics into planning, tool use and reasoning with self-correction. The agent decomposes a goal into subtasks. It calls on external data, APIs or other agents to fill knowledge gaps. Then it revises its plan as results come back (IBM). IBM's legal research example shows the pattern: a low-cost classifier routes queries, and only complex cases reach a more capable research agent. Contract review time fell from 90 minutes to 45 (IBM).

The practical consequence for a buyer is that an agent touches live systems, not just a chat window. It writes to the CRM, raises a ticket, routes an approval. That is why the vendor question is not "do you have agents?" It is "which systems does the agent write to, under whose approval, and what happens when it is wrong?"

For a structured starting point on agent types and controls, the guide to AI agents in the knowledge hub sets out the basics before a vendor conversation.

Where agents are earning their keep

Coding is the clearest early win. About two in ten respondents in McKinsey's 2026 survey are scaling software coding agents, rising to 31% at larger enterprises. And 32% say their organizations decided against buying one or more software products because they could be built internally with agentic coding tools (McKinsey). That last figure deserves attention. Agents are starting to reshape software budgets, not just productivity.

Operational use cases are spreading too. Capgemini's research lists security operations, supply chain and similar multistep workflows among the areas where agents orchestrate entire processes rather than single steps (Capgemini). TechTarget's roundup describes security operations centers where agents scan for emerging threats, investigate anomalies and take corrective action without human intervention (TechTarget).

Vendors are packaging these patterns into repeatable products. OpenAI's agent offerings, for example, cover qualifying and routing inbound leads, reviewing IT tool requests and drafting marketing content at scale, with approvals built in (OpenAI). The pattern to notice is that every credible example names a specific workflow with a defined start and end.

When you shortlist suppliers, ask each one to name the workflow, the systems it touches and the approval step. Vendors that work with OpenAI models are listed in the OpenAI technology directory. Teams that need a partner to build and run these workflows can compare generative AI service providers.

The scaling gap is a process gap

Capgemini's 2025 survey of 1,500 senior executives found only 2% of organizations had deployed agents at scale. Another 12% were at partial scale, 23% were running pilots and 61% were still exploring (Capgemini). Microsoft's survey of 500 enterprise decision-makers across 13 countries reached a similar conclusion. What separates the fast from the slow is not budget or technical firepower. It is preparation (Microsoft).

Microsoft's data adds a useful detail. Nearly 80% of organizations say they cannot share data across teams in ways that make agentic work possible (Microsoft). An agent that cannot see the customer record, the ticket history and the contract cannot finish the job. Data access is a design decision, not an IT afterthought.

Deloitte puts the cost of skipping that preparation plainly. Many organizations layer agents on top of existing processes for quick returns. Yet only one in five leaders say they are prepared to redesign processes for autonomous operation (Deloitte).

A practical way to close the gap is to run a one-page process review before any pilot. Draw the current workflow, list every handoff and name the owner of each step. Mark where a human makes a judgement call today. Those marks become the approval points in the agent design. If the page cannot be drawn, the process is not ready for an agent.

Trust and cost are the brakes

Two constraints show up repeatedly. The first is trust. Capgemini found trust in fully autonomous agents fell from 43% to 27% within a year (Capgemini). That does not mean buyers are rejecting agents. It means they want oversight. Deloitte's data agrees: 75% of leaders say human collaboration with agents creates more value than agent-powered automation alone (Deloitte).

The second is money. About 20% of McKinsey respondents say AI-related operating costs, including token costs, have constrained their AI use (McKinsey). Agents that loop, retry and call tools can burn budget quietly. Any business case should include a cost ceiling per workflow, not just a productivity estimate.

Set the ceiling in the pilot contract. Agree a monthly cap, an alert threshold at 80% of the cap and a named person who can pause the agent. Ask the vendor for a sample cost report before you sign. A vendor that cannot show token and tool usage per run is not ready for production.

Worth noting: survey definitions of "adoption" and "at scale" differ. G2's 2025 buyer poll found 57% of companies with agents in production, far above Capgemini's 2% at scale (G2). The gap reflects different questions and samples. Compare carefully before quoting a number in a business case.

A due diligence checklist for agent projects

1. 1Name the workflow end to end, including every system the agent reads from and writes to.
2. 2Define the approval points. Which actions require a human sign-off, and who owns that decision?
3. 3Ask how the agent's plans and tool calls are logged, and whether each decision is auditable.
4. 4Test failure behaviour: what the agent does when a tool times out, returns bad data or loops.
5. 5Agree a cost model, including token and tool usage, with a cap and an alert threshold.
6. 6Set the exit criteria for the pilot before it starts, and how the agent is retired if it fails.
7. 7Check data readiness across the functions involved, not just the pilot team.

Add three checks to the list. First, ask for a rollback plan with a tested time to restore. Second, confirm who on the vendor side is on call when the agent misbehaves at 2am. Third, require a written handover so your team can run the agent without the vendor. These are the terms that separate a pilot from a product.

What this means for buyers

Start with a workflow that has a clear owner and a measurable output, such as code review, lead routing or ticket triage. Write the process map before the statement of work. Make the vendor demonstrate failure handling, logging and approval gates, not just a happy-path demo. Ask for one reference at a comparable scale and call it.

Then prepare the organisation. Deloitte found 43% of leaders expect agentic AI to significantly disrupt their workforce within 12 to 18 months. Yet half say their organizations are not adequately investing in AI-related workforce transformation (Deloitte). Budget for retraining and role redesign alongside the licence. Treat cost control and auditability as contract terms, with caps and reporting obligations written in.

When you are ready to test the market, post your project with the workflow, the systems and the controls you need. Clear briefs attract vendors who can answer the hard questions.

What this means for vendors

Lead with the workflow, the integrations and the controls, not the model. Buyers who have read the readiness data will ask what happens when the agent is wrong. Arrive with an answer: rollback, human escalation, audit trail, cost ceiling. Show a reference deployment at a comparable scale, because the gap between pilot and production is where most deals stall.

Be precise about autonomy. If humans approve high-risk actions, say so. That matches what most buyers actually want. And help the client redesign the process rather than automate the old one. Deloitte's finding that 74% of leaders expect nearly half of business processes to be rebuilt around agents within four years sets the expectation your engagement should meet.

Frequently asked questions

What is agentic AI in the enterprise?

Agentic AI in the enterprise means AI systems that pursue goals and act on live business systems, not just answer questions. They plan tasks, call tools and adapt as results come back, often with human approval at key steps.

How many enterprises have scaled AI agents?

Estimates vary by survey. McKinsey reports 40% of large organizations scaling agents in 2026, while Capgemini found only 2% at full scale in 2025. Compare definitions before quoting a figure.

What should buyers check before signing an agent contract?

Check the workflow map, the systems the agent writes to and the human approval points. Also require logging, failure handling, a cost cap and a tested rollback plan.

Why do agent pilots stall before production?

Most stalls come from process and data gaps, not model quality. If teams cannot share data or nobody owns the workflow, the agent cannot finish the job reliably.

The bottom line for buyers and vendors

Agentic AI in the enterprise rewards preparation over enthusiasm. Buyers should map the workflow, set approval gates and cap the cost before a pilot starts. Vendors should sell the workflow and the controls, and show a reference at the buyer's scale. The organizations that treat agents as a process change, not a software purchase, are the ones getting value.