Agentic AI: 7 Critical Lessons From Salesforce’s Agentic AI in the Enterprise Report
Agentic AI is forcing organizations to confront how well they understand their own work. Salesforce’s State of Agentic AI in the Enterprise offers useful evidence for a position I have developed through my Serious Insights analysis and reporting on agents: an implementation’s quality depends on the organizational decisions surrounding the technology. Those decisions include what constitutes acceptable work, which knowledge informs it, who can authorize an action, and who remains responsible when the result arrives late, incomplete, or wrong.
I appreciate the report’s attention to methodology. I also see reasons to push its implications further. A successful agent deployment should leave an organization better able to explain and manage its work. An agent that produces a measurable benefit while obscuring dependencies, consuming uncounted human attention, or weakening institutional knowledge creates a more complicated balance sheet than a headline return suggests.
Key Takeaways
- Agent adoption numbers don’t tell the whole story.
- Real ROI includes the work automation leaves behind.
- Knowledge management makes or breaks agent performance.
- Human oversight can conceal costly recovery work.
- Faster returns don’t justify weaker governance.
- Automation can erode the expertise it depends on.
- More agents don’t necessarily mean more value.

1. Read the agentic AI findings through the sample
Salesforce surveyed 2,025 purchasing decision-makers in May 2026. All represented organizations deploying, piloting, or considering agents; 30% had fully deployed them. Only 10% of respondents represented companies with more than 4,500 employees. Outcomes are self-reported, and the study measures association, not causation.
That makes the report useful evidence about committed early participants, but it limits how far its findings can be extrapolated. The deployment percentage cannot describe adoption across all businesses. An organization still deciding whether agents fit its needs is also asking a different question from one comparing implementation approaches.
I would apply the same care to comparisons between industries and governance approaches. Different workflows carry different costs, risks, and opportunities. A faster return may reflect the work selected, existing infrastructure, or management competence. Research that identifies an association helps establish what to investigate. It does not, by itself, establish which intervention will reproduce the result elsewhere.
2. Treat ROI as the beginning of the investigation
Deployers report reaching meaningful ROI in roughly eight months, alongside a 29% lift in customer satisfaction.
Those results caught my attention, but they also raise questions about what goes into the calculation. I always want to see the baseline, the measurement period, and the costs assigned to data preparation, integration, evaluation, review, correction, and ongoing maintenance. Time saved becomes financial value only through some identifiable change in spending, capacity, revenue, quality, or risk.
Consider a hypothetical service agent that reduces handling time while sending the difficult cases to experienced employees. The organization may improve its overall economics. It may also concentrate emotionally demanding work and complex judgment in a smaller group. Measuring the agent’s throughput without measuring that group’s workload would produce an incomplete account of the change.
My preferred unit of analysis is the completed business outcome. For a service workflow, that means a correctly resolved issue, within an acceptable period, at a known total cost. Reopened cases, abandoned interactions, corrections, and escalations should be part of the assessment. Agentic AI should earn its place in operations when those measures remain favorable after the initial implementation team moves on.
3. Data readiness needs knowledge management
Only 31% of deployers fully unified their data before launch. Salesforce’s practical recommendation is to prepare the information required for the specific job.
That is a sensible boundary for an implementation. I would extend it beyond data access into knowledge management. An agent may retrieve a current record and still misunderstand its significance. A customer’s contractual entitlement, an approved exception, a regional practice, and an unresolved disagreement between departments can all change what the organization should do next.
In my interview with NewRocket’s Sean Iannuzzi, we explored how operational context and clear ownership shape dependable agent workflows. The next question is how organizations sustain that context. Who decides which explanation is authoritative? Who retires obsolete guidance? How does an exception become reusable knowledge without becoming an unintended general rule?
A narrowly scoped deployment should produce a maintainable body of knowledge about the work it performs. That includes definitions, sources, owners, limitations, and the circumstances that require judgment. Connecting repositories can make information available. Establishing what that information means must remain organizational work.
4. Increased human involvement deserves its own measures
Eighty-one percent of deployers say human reentry increased as deployments expanded. Salesforce attributes the increase to a mixture of designed checkpoints and agents falling short.
This is one of the report’s most useful findings because it challenges simplistic assumptions about the relationship between scale and labor. It also needs careful interpretation. More interventions across a larger deployment do not necessarily mean more interventions per transaction. The operational question is how the rate, duration, difficulty, and purpose of those interventions change.
I would separate planned approvals from recovery work. A person approving a consequential action is performing a designed responsibility. A person repeatedly reconstructing context that an agent lost is absorbing a system defect. Both involve human effort, but they imply different choices about staffing, architecture, and investment.
My work on orchestrating agentic AI systems that humans can trust places human judgment at consequential points in the workflow. That design requires more than adding an approval button. Reviewers need time, relevant evidence, authority to stop execution, and a manageable volume of decisions. An organization cannot claim meaningful oversight while rewarding people primarily for clearing the queue.
5. Faster returns do not settle the governance decision
Organizations with lighter governance reported ROI in 7.2 months versus 9.3 months for those with heavier governance. Stronger governance correlated with earlier problem detection.
The tempting interpretation is that governance imposes a measurable delay that leadership can trade against faster gains. I would resist making that calculation from these averages. The groups may differ in the consequences of failure, the complexity of their workflows, and the requirements they must satisfy. The findings do not establish that removing controls from a particular deployment would accelerate its return, or that it would be safe to do so.
To that point, there is also a mismatch between when benefits become visible and when liabilities emerge. An incorrect commitment may take months to surface. A permissions failure may remain undiscovered. Early financial performance cannot resolve those uncertainties.
Governance should therefore be designed around specific decisions and consequences. A reversible internal draft can support considerable experimentation. An external commitment requires a different level of authority and validation. The organization should define those boundaries before execution, along with the evidence needed to investigate a failure and the means to contain it.
I would also ask whether controls remain effective when the agent, model, connected application, or underlying policy changes. Governance that works only for the configuration tested at launch has a limited useful life.
6. Preserve the expertise that makes oversight possible
Among respondents whose initiatives slowed, stalled, or failed, 39% wished they had invested more in AI fluency and reskilling.
Training deserves that investment, but the workforce question extends into how expertise develops. Organizations often learn through work that automation aims to remove. Reviewing ordinary cases helps people recognize unusual ones. Preparing an analysis teaches the relationships behind its conclusions. Handling exceptions reveals where formal policy fails to describe reality.
If agents absorb much of that experience, leaders need a deliberate replacement for the learning it once provided. Otherwise, they risk creating a workforce expected to supervise decisions it has had fewer opportunities to understand.
I would make expert development part of the implementation budget: case reviews, supervised practice, discussion of failures, and opportunities to challenge the system’s recommendations. Employees should also help define the work before automation begins. Their knowledge includes the accommodations and informal coordination that process diagrams frequently omit. Capturing that knowledge requires participation and trust, especially when the same initiative raises questions about future employment.
7. Require more evidence as agents coordinate other agents
Salesforce closes the report by considering superagents and AI-assisted agent creation.
My concern is the multiplication of dependencies. A coordinating agent can report success while a downstream action remains incomplete. Several agents can reinforce the same incorrect assumption. A local improvement can move cost or risk into another department. Evaluating each component independently will not reveal every failure that emerges from their interaction.
Before increasing autonomy, I would require evidence that the complete workflow handles missing information, unavailable tools, conflicting instructions, and interrupted execution. Responsibilities should remain identifiable through delegation. The organization should be able to determine which action occurred, under whose authority, using which information, and with what verified result.
Scenario thinking offers another useful discipline here. Test the business case against changed conditions: higher operating costs, reduced model quality, a vendor withdrawal, a new constraint on data use, or an unavailable human reviewer. The purpose is to discover dependencies while there is still time to change the design.
Salesforce’s report reinforces my view that agentic AI requires a more explicit account of how an enterprise works. The next stage should be judged by whether organizations can sustain reliable outcomes, develop their people, preserve their knowledge, and adapt when assumptions fail. Deploying more agents is relatively easy to count. Demonstrating that the organization has become more capable requires a better set of questions.
For more serious insights on AI, click here.
Did you find this analysis, Salesforce on Agentic AI: 7 Critical Lessons From Salesforce’s Agentic AI in the Enterprise Report, useful? If so, please like, share, or comment. Thank you!
The cover image is AI-generated from the author’s prompt.

Leave a Reply