Large software vendors have moved beyond chatbots to offer AI agents that can navigate multiple corporate systems and carry out multistep work on behalf of staff. Microsoft, Google, Salesforce, ServiceNow, Oracle and SAP have each introduced products that place an AI intermediary between employees and the applications they use, promising to complete routine tasks and surface finished work for human review.
What the new agents do
These agents are intended to do more than answer questions. They can connect to different business systems, automate sequences of actions and sometimes adopt a persistent identity that lets them participate in team workflows. The approaches differ by vendor, maturity and the types of tasks targeted.
| Vendor | Agent / offering |
|---|---|
| Salesforce | Redesigned Slackbot, integrated into Slack |
| Microsoft | Copilot Cowork for Microsoft 365 |
| Gemini agent for work and coworker agents with identities | |
| ServiceNow | Personal agent paired with NVIDIA's OpenShell |
| Oracle | Agentic applications for finance and supply chain |
| SAP | Agents across finance, HR and supply chain |
Why leadership decisions matter
Putting an AI agent between staff and the systems they rely on creates strategic dependencies. The first agent a workforce adopts can shape how information flows, who gains visibility of corporate data and which vendor’s roadmap dictates future capabilities. That makes procurement a governance decision as much as a technology one.
- Control — Which agent gets access to multiple systems and data sets?
- Process — Are you automating a sound workflow, or embedding existing errors?
- Vendor influence — Which supplier will steer your future working practices?
Executives should test whether existing processes need fixing before automating them. Automating a flawed workflow can make bad outcomes faster and harder to spot. For example, the vendors describe scenarios where an agent gathers missing documentation or reconciles records, handing back corrected work for review. That can free staff to focus on exceptions and judgement, but only if the underlying process has already been rationalised.
Security, policy and integration limits
Vendors are building safeguards so agents respect company policy, but some controls still rely on how systems are integrated. ServiceNow’s agent, for instance, is presented alongside NVIDIA’s OpenShell to limit what the agent can see and do — though the effectiveness of such constraints depends on implementation. Where agents bridge multiple applications, security boundaries, permissioning and audit trails must be carefully designed and tested.
There is also a procurement risk: whoever deploys the first agent can influence which other tools staff accept and which data flows become standard. That may accelerate vendor lock-in or concentrate sensitive access with a single provider.
Practical steps for organisations
Companies considering agent technology should align adoption to business strategy rather than vendor roadmaps. Practical steps include:
- Audit existing processes to identify whether automation will fix root causes or merely speed flawed workflows.
- Map data access needs and design least-privilege controls to limit what agents can read or change.
- Evaluate how an agent’s identity and integrations will affect procurement, visibility and future platform choices.
Organisations should also pilot agents in limited, high-value areas and measure error rates, exception handling and operational risk before wider roll-out. Thoughtful governance and careful integration — not just the shiny features vendors demonstrate — will determine whether agents genuinely reduce workload or merely shift problems around faster.
The new generation of AI agents promises to change how routine work is done. That potential makes it imperative for corporate leaders to decide deliberately what tasks to hand over to software, which vendor controls critical touchpoints and how to protect data and decision-making along the way.
