How to decide which AI features are worth turning on

Features vs. tiers

AI enablement is no longer a project with an end date; it is an ongoing operating capability. The companies that treat it that way will likely spend the next several years making faster, defensible calls while everyone else relitigates the same questions release after release.

"Enterprise leaders should evaluate AI features based on operational impact, not novelty. The question is not whether the feature sounds innovative. The question is whether it improves business coordination, decision-making and execution across teams," said Walia.

As a result, there's a strong need for a strategic plan to to ensure business value and measurable outcomes are harvested from AI use in operations. It begins with a standing discipline that ties every AI feature enablement decision to a named owner, a measurable outcome and an honest risk assessment. 

"The pressing decision for enterprise leaders is who in the organization is qualified to authorize AI features, on what evidence and with what review cadence. Most enterprises let individual product owners flip toggles without ever defining that boundary," said Brian Fending, CISM and managing director of Ordovera Advisory.

Without clear governance, AI sprawl is common and meaningful investment returns are harder to achieve. But sometimes pinning down what goes where on top of a strong foundation remains a challenge.

"The decision is not simply which features to enable; it is which tier is the right home for each use case," Vargo said. To clarify, the tiers he is referring to are embedded AI and enterprise AI platforms.

The temptation might be to place a use case in the software where it is most often managed and let the embedded AI assist with those workflows. "The principle is straightforward: Use what is built into the solution as far as it can reliably go and route to your enterprise AI platform only when the embedded capability hits its ceiling," Vargo said.

But cost structures and deployment constraints also vary between the two approaches. Embedded AI offers "native data access within that system, no integration overhead and vendor-maintained model updates," said Vargo. But it is also limited, for example, in "constrained customization, narrow data scope, opaque model governance and per-seat add-on pricing that compounds at enterprise scale." Enterprise AI platforms generally offer the opposite.

Also consider that costs for both tiers will change over time as an organization's usage grows.

"Embedded AI features typically carry per-seat pricing that scales with head count; enterprise AI platforms run on token-based consumption costs that have trended downward but are not fixed," Vargo said.

"Neither structure offers a fully locked-in cost trajectory and organizations should pressure-test both against expected usage volume rather than comparing today's rates alone," Vargo added.

Assistive vs. autonomous

The type of AI in question also affects how organizations prioritize feature adoption. Embedded assistive AI tools tend to be ideal for meeting summaries, content drafting, enterprise search and workflow recommendations.

"The vendor's model is sufficient for these tasks, the relevant data lives inside the system, human oversight is maintained and the cost is typically justified by the productivity return. These are the features worth enabling early; this is where embedded AI delivers on its promise," Vargo said.

These are the features worth enabling early; this is where embedded AI delivers on its promise.

Autonomous or semi-autonomous AI capabilities such as automating transaction approvals, modifying records and triggering cross-system workflows without review are much riskier to use and warrant close inspection before enabling them. "Here, the embedded model's limits on data scope, auditability and governance often make the enterprise AI platform the right home, even at a higher integration cost," said Vargo.

When presented with agents as embedded features -- such as SAP Joule, Workday Sana, Microsoft Copilot, ServiceNow agents, Salesforce agents, UC assistants and autonomous endpoint tools -- organizations should evaluate them carefully before activation.

"The features to pilot carefully are the ones that start creating work products for human review, such as email drafts, ticket triage suggestions, code completion and expense categorization," Fending said. The features "worth declining for now, or at least moving more slowly on," he added, are those that "act autonomously across systems your users have permission to change."

Fending cited examples of autonomous activities such as emailing, calendar booking with external parties, data movement and record updates. These tasks, he said, "have a blast radius matching the user's access permissions, not the feature's intended scope." A feature scoped to "help with email," Fending explained, "inherits every system that the user's identity can reach."

The practical framework is simple, said Diptamay Sanyal, principal engineer at cybersecurity company CrowdStrike. Sanyal recommends asking the following questions of each AI feature before enablement:

• What data does it touch?
• Who is accountable when it fails?
• Can you observe what it actually did after the fact?
• Is the blast radius bounded?

 "If those questions don't have clear answers before deployment, the feature is aspirational, not operational. That's the line," Sanyal said.

Pam Baker is a freelance journalist and the author of books including ChatGPT for Dummies and Generative AI for Dummies. Baker is also an instructor on AI topics for LinkedIn Learning and a member of the National Press Club, the Society of Professional Journalists and the Internet Press Guild.