Agentic AI in Business Processes: What Changes for Companies After the May 2026 Announcements

🗓️ What happened in May 2026 and why it marks a turning point

On 29 April 2026 Salesforce made Agentforce Operations available — an orchestration engine that transforms unstructured process documents into digital blueprints executed by specialised AI agents. The vendor claims up to 70% reduction in cycle time and up to 80% of manual tasks eliminated on workflows such as invoice auditing, onboarding, and purchase order rescheduling.

On 5 May, at Think 2026, IBM presented the new generation of watsonx Orchestrate: no longer a handful of agents in production, but thousands of agents built by different teams, on different platforms, governed by a single control plane with near-real-time audit trail. The same day IBM proposed a new operating model for AI in enterprise built on four pillars: coordinated agents, real-time data, end-to-end automation, and hybrid deployment for sovereignty and governance.

On 6 May ServiceNow and Accenture announced a joint forward deployed engineering programme to bring agentic AI on the ServiceNow AI Platform directly into clients' environments. On 20 May Camunda opened the closed beta of ProcessOS, an intelligence layer that discovers, redesigns, and continuously optimises processes as agentic workflows.

Four announcements in three weeks, all converging on one idea: LLMs alone are not enough — a level of orchestration is needed that puts agents inside measurable, governed, and auditable processes.

🧩 From LLMs to agents: what actually changes for business processes

Until late 2024, AI in enterprise meant a chatbot in front of a knowledge base and some summarisation scripts. Agentic AI is a different paradigm: an agent is an LLM with access to real tools (CRM, ERP, email, files, internal APIs) capable of executing a sequence of actions to close a task end-to-end, deciding whether it needs another agent, and stopping when it must hand off to a human.

The leap is not only technological — it is organisational. Many companies are discovering that their processes were designed around human workarounds — "if the PDF looks odd, Marco will catch it" — and break when executed literally by an agent. Adopting agentic AI means first rebuilding processes in an executable form, then delegating them.

• An agent is not a chatbot: it has tools, memory, goals, and can close tasks autonomously.
• The agent works inside a control plane that specifies what it can do and keeps track of everything.
• The value lies not in a single agent but in the orchestration of multiple specialised agents.
• Humans remain in the control room: approvals, exceptions, non-delegable decisions.

🏗️ The three layers of a production agentic AI architecture

The agentic architectures emerging across Salesforce, IBM, ServiceNow, and Camunda share the same three-layer structure. Understanding it helps to separate marketing from substance and to read vendor offerings accurately.

💡 Concrete use cases: where agents generate value immediately

A recurring objection is: "sounds great, but it's not relevant for us". In practice the map of high-intensity use cases is fairly stable and covers the vast majority of SMEs. The ideal candidate is almost always a repetitive, low-exception process that today consumes hours of human work in sorting, reconciliation, and data entry.

🏢 What companies need before starting: data, processes, governance

The number-one risk with agentic AI is the same fallacy as RPA ten years ago: buy the platform, automate a broken process and get a broken process that runs faster. Avoiding this requires three pre-conditions that every serious project addresses before touching an agent.

⚠️ Common mistakes and how to avoid them

When I work with companies on introducing AI agents into their processes, I see the same mistakes repeatedly. The first is starting from the vendor instead of the process: a licence for Agentforce or watsonx Orchestrate is purchased without having selected the use case, and after six months the platform sits idle. The second is confusing automation with agentic: automating a step with an LLM is not agentic AI — it is just one more API call inside a process that remains human. The third is underestimating governance: an agent with access to the CRM and email is a digital employee, and needs permissions, audit logs, and a revocation procedure.

• Start from the process, not the vendor: map first, choose platform second.
• Measure the baseline first: without before/after KPIs you cannot demonstrate value.
• Limit the blast radius: every agent has minimal permissions, never administrative ones.
• Log everything: prompts, tool calls, outputs. That is your audit trail.
• Define from day one who is the human owner of the process.

🗺️ Practical roadmap for companies

For a company that wants to bring agentic AI into its processes today without chasing the latest vendor announcement, the roadmap I recommend has five stages. Start small, on a single measurable workflow, and expand only after proving the model holds. Typically the first three stages fit within a quarter.

More articles from the blog