Salesforce urges enterprises to stop treating AI agents as one uniform capability

Salesforce is making a case that the biggest mistake companies can make with “agentic AI” is assuming there is one strategy that works everywhere inside an organization. In a new perspective published by Salesforce News, the company argues that AI capabilities vary sharply depending on the kind of task involved, including how reliably an answer can be judged and how much context the system receives.

The article traces why that misunderstanding happens. It points to demonstrations where swarms of AI agents can write code, test and fix errors, and improve through feedback loops, making it easy to conclude that the same approach will generalize across departments. Salesforce cautions that this assumption breaks down because AI can excel in some areas and fail in others, at times in ways that are hard to catch before decisions are made.

To structure that unevenness, Salesforce says enterprise teams should ask two questions about any proposed AI agent workflow. First, whether the task is something the agent can score against clear criteria (objective verification). Second, whether the agent has access to the right context to act meaningfully (including information that may live in people’s experience rather than in formal process documents). The answers are not binary; Salesforce frames them as a spectrum that influences how work should be divided between humans and AI.

Salesforce describes four broad categories that emerge from those two questions. In tasks where the problem is well understood and results are verifiable, agents can often take on significant execution. Coding is the example used in the perspective, along with scientific and engineering work where outputs can be tested against requirements. Salesforce argues that in this category, AI gains can compound because feedback and tests confirm whether the work is correct.

But the same reliability does not automatically transfer to tasks where verification is possible yet the agent lacks the right training or feedback specific to that kind of work. Salesforce points to operational scenarios such as ensuring timely customer delivery. The article says these situations can be verifiable at a deadline level, yet expert judgment depends on tacit knowledge, like which suppliers routinely miss turnaround times or what exceptions senior employees make in practice. Salesforce’s suggested human role here is to “stitch together” missing context and make it available to AI.

The perspective also highlights a third category where the agent has relevant context but there is no objective way to verify the answer in time to inform action. Salesforce places many marketing, positioning, and strategy tasks into this bucket. The problem, in the company’s telling, is not that AI will always produce incorrect output, but that it may produce plausible work that “feels” new while actually converging with what other firms generate. Salesforce warns that multiple competitors could ask similar questions and receive confident, polished answers that later prove too similar.

In that setting, Salesforce says humans must curate. The article emphasizes human taste as the ability to judge whether output is distinctive, surprising, and aligned with the organization’s values rather than merely competent. Where eventual outcomes can be measured, Salesforce suggests companies can use those results to sharpen what they ask for next time, potentially moving some tasks toward a more verifiable category.

Finally, Salesforce addresses tasks where there is little objective testability and the situation is genuinely novel, such as investment decisions in markets a company has never entered. In these cases, the article argues that even detailed analysis from AI (for example, reviewing financials and comparable deals) may not indicate whether the decision will strengthen the business or create costly risk years later. Salesforce’s position is that AI can still help by surfacing assumptions, proposing alternatives, and arguing against a favored option, but it should not be treated as a decision-settling mechanism.

The closing takeaway is less about deploying AI more broadly and more about designing governance around task type. Salesforce says leaders should ask, for each workflow, what kind of work it is and what humans should do in that specific loop. The company frames verification and context as matters of degree, and suggests that tasks can move between categories over time as companies build better rubrics, capture feedback from outcomes, and develop processes for checking AI outputs.

Salesforce does not provide product-specific announcements in the piece, nor does it cite performance metrics or case studies with quantified results. Instead, the argument is primarily conceptual, grounded in the company’s view of where agentic systems are strongest and where they may fail quietly. That means the actionable guidance is governance and workflow design rather than a concrete implementation roadmap or benchmark of what enterprises should buy or build. Still, for organizations wrestling with AI adoption, the framework offers a common language for deciding when to automate, when to augment, and when to keep humans accountable for the final call. Next, executives may look for internal pilots that classify workflows along Salesforce’s two-question model and then test whether human curation, context capture, and verification practices actually reduce errors and decision regret over time.