Agentic AI adoption is accelerating fast. According to Salesforce’s 11th Annual Connectivity Benchmark Report, organizations currently use an average of 12 agents, with that number projected to climb 67% within two years. Yet nearly half of those agents still operate in isolated silos, disconnected from the systems and people they’re meant to support. Agents monitor systems, generate outputs, trigger workflows, and take action across the enterprise. As autonomy increases, the question shifts from what agents can do to how we design systems people can trust.
Most conversations today focus on agent capability. Far fewer focus on how humans and agents should work together. At Salesforce, the Office of Ethical and Humane Use is centered on the human-AI partnership — grounded in the belief that the most innovative AI is also the most responsible AI. In practice, that means designing not just what AI can do, but how humans and agents work together.
A Day Empowered by Agents introduces a practical framework inspired by how we use AI at Salesforce to help both our teams and our customers think through how to design the human-AI partnership, by making trust something that is designed into every interaction.
Trust doesn’t come from a single product or feature. It’s a shared responsibility between those designing and building AI and those using it.
The role of the builder is changing
As organizations deploy autonomous workflows, the role of the builder is changing.
Developers, architects, admins, and business leaders aren’t just configuring tools anymore. They’re shaping how decisions get made, how work gets handed off, and where accountability sits. Every role — from developer to CEO — plays a part in defining how agents show up in their work.
To make this more concrete, the framework introduces a simple model:
Human task + Agent task + Trust guardrails
It’s straightforward, but it forces clarity on ownership, responsibility, and how work gets done:
- What task does the human own?
- What part of that task does the agent take on?
- What makes that interaction safe, visible, interruptible and controllable?
Trust isn’t something you layer on later. It’s what makes the handoff between human and AI trusted — and in many cases, possible.
How this shows up in real workflows
Take a developer reviewing and deploying code. Their responsibility is making sure what ships is secure and reliable, whether that code is written by a human or generated by an agent. An agent can help by drafting code, analyzing pull requests, surfacing vulnerabilities, and summarizing logs. But the developer can still trace how issues are flagged, and the system pauses at defined boundaries instead of pushing forward on its own. Approval gates stay in place. The agent supports the workflow, but accountability stays with the human.
Or a data analyst. They’re responsible for turning data into decisions. An agent can generate dashboards, surface trends, and flag anomalies, but it stays within defined metrics, runs harmful content checks, and logs validation steps. That means the analyst can move faster without second-guessing the integrity of the output.
For a software architect, the challenge is balancing trade-offs — scalability, cost, and security. An agent can generate options and run compliance checks, but those recommendations are grounded in trusted data, fully traceable, and constrained within defined boundaries. The architect still owns the decision.
Across all of these, the pattern is the same: the agent extends capability, but the human remains accountable. You can explore this across 16 different roles in the experience.
Best practices for designing responsible autonomy
Builders decide how systems are actually used — where agents act, what decisions they influence, and how humans stay involved. Trust lives at that intersection. We’ve outlined a set of best practices for designing trusted human–agent interactions:
1. Measure outcome over output
Agents’ work is evaluated based on how accurate, goal-aligned, and high-quality the results are, rather than how fast or how much they produce. While saving time can be helpful, improving the quality of outcomes is the key to long-term efficiency.
2. Lead with human strengths
When agents assist with rote tasks such as drafts, summaries, or suggestions, they free up people to focus on creativity, empathy, and high-impact, high-stakes work. This approach allows AI to support human strengths and make work more meaningful and effective.
3. Instill clear handoffs
Clear handoffs make it easy to understand which tasks belong to agents and which stay human-led. This balance lets agents take on repetitive, structured work while people focus on judgment-heavy, interpersonal, or strategic work that benefits from human insight. Handoffs should be triggered at defined checkpoints, when work falls outside scope, or when human review or intervention is required.
4. Retain transparency & auditability
Every agent action leaves a clear record so it’s possible to see how and why each action occurs. Including sources increases transparency and helps build trust and shared understanding across teams.
5. Maintain an open feedback loop
AI systems thrive on ongoing evaluation and refinement. Human edits are not just the final step; they are the input that continues to keep the system safe, innovative, and aligned with changing business needs.
6. Guardrails by design
Building concrete guardrails at the start gives teams confidence to work effectively. Boundaries are clearly defined and visible from inception, ensuring trust is core to the process rather than added at the end.
Where to start
Before deploying an agent, start with a few simple questions:
- What outcome is this supporting?
- Where does accountability sit?
- What guardrails are in place?
- Can decisions be reviewed or reversed?
- What remains human, no matter what?
To explore further, the A Day Empowered by Agents Quick Start Guide provides a practical set of prompts and exercises teams can use to evaluate these decisions before deployment — helping translate these principles into real workflows.
As AI becomes more embedded in how work gets done, the differentiator won’t just be who can build agents. It will be who can design how humans and agents work together.
Every workflow encodes decisions about ownership, visibility, and control. Those decisions shape whether people trust the system or avoid it. In the agentic enterprise, trust isn’t a feature you add at the end. It’s a design discipline from the start.
Explore A Day Empowered by Agents to see how trusted guardrails, clear ownership, and human–AI collaboration come together in practice — and download the Quick Start Guide to bring these principles into your own workflows.
