Muddu Sudhakar is the SVP & GM, Agentforce IT Service & HR Service, Salesforce.
While agentic AI promises major improvements in productivity, growth and efficiency, the results have been underwhelming. A recent Deloitte survey found that only a quarter of their respondents moved 40% or more of their AI pilots into production.
This means that achieving sufficient ROI can be challenging. This is worsened by the escalating token costs.
Enterprises try to deal with the problems by taking a buy-versus-build approach. This strategy can be effective with traditional enterprise software, but it often falls short with complex agentic AI workflows.
What to do? A better approach is to use a blend of both: buy and build.
Buy Phase
Generally, the preferred method of enterprise deployment is to buy a system. Some of the benefits include faster time-to-value, built-in security and compliance, ongoing support and upgrades.
But buying an agentic AI platform is different than selecting a traditional enterprise application, say for ERP, CRM or HR. These systems are based on strict rules and clear-cut processes.
An agentic system, on the other hand, has complex large language models (LLMs) as their core intelligence. This means that the responses are based on advanced probabilities.
This is also why the technology is so powerful. Agentic AI can interpret context, engage in reasoning, adapt to change and make grounded decisions.
Consider IT services. A traditional system will route tickets based on keywords and predefined rules. In contrast, an agentic AI system will go through a comprehensive set of workflows: interpret the employee’s issue, review prior incidents, check knowledge articles, evaluate permissions and identify root causes for the issues. Throughout this approach, the agent can take action to update configuration, reset access, trigger a workflow or escalate to a human agent.
Selecting The Right Agentic AI Platform
The selection process for an agentic AI platform is not easy. The technology can be complicated and changes quickly. Still, there are some key elements to look for:
• Orchestration: Complex workflows require multistep processes. This is why there must be a sophisticated orchestration system.
• Out-Of-The-Box Agents: See if there are already agents available for your workflows. They should also be easily integrated with your business data.
• Observability: Agents can easily go down rabbit holes, misfire or go into endless loops. To minimize the issues, an agentic AI platform needs monitoring and tracking capabilities, along with the ability to evaluate results. This should be unified in a mission control system.
• System Of Work: You want the agentic AI platform to seamlessly integrate data and context for where employees and users do their work, such as with collaboration tools, email or chat.
• Headless: This allows for customizing the system and involves using APIs, connectors and model context protocol (MCP), which integrates with agents. The customizations can be done using common tools like command-line interface (CLI).
• Security: Because agents can take mission-critical actions, there needs to be permissions, guardrails, policies and controls.
The Build Phase
Once you have selected and implemented an agentic AI platform, the next step is the build phase. It’s important to focus on one or two use cases. Otherwise, the process can get too complicated and bogged down. This is one of the reasons that many agentic AI projects fail to achieve sufficient ROI levels.
First, target a particular category such as IT, HR, DevOps or finance. All of these can greatly benefit from improvements in workflow automation.
If your agentic AI platform has headless capabilities, the customization process will be much smoother than traditional development approaches. Domain experts do not have to be coders. Rather, they can use vibe coding to put together the customizations.
However, there still needs to be extensive testing. Agentic AI workflows can sometimes be unpredictable. This means there should also be humans in the loop.
Yes, this can be time consuming, but it is well worth the effort. It means that an enterprise can realize the true value of agentic AI.
Conclusion
With agentic AI, a buy and build approach means lower costs and faster deployment. But the key is to stay focused on those use cases that matter and build on them. It also involves strong testing and guardrails. Done right, this approach can help enterprises move beyond pilots and finally achieve meaningful ROI from agentic AI.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?