.png)
Salesforce recently published an article titled "In 2025, AI Grew Up and Learned to Play by the Rules." The piece makes a critical admission. Large language models alone cannot close the last mile of enterprise AI. They are probabilistic by design. They generate likely outcomes, not guaranteed ones. And in core business operations where policies, regulations, and customer trust are non-negotiable, that is a liability.
Salesforce puts it plainly. To cross the last mile of enterprise AI, agents must be able to reason through complexity, operate on trusted business data, and execute critical actions deterministically. Remove any one of these, and the system breaks down.
The details behind this admission are instructive. In a recent article in The Information, Salesforce's CTO for Agentforce, Muralidhar Krishnaprasad, explained that LLMs start dropping instructions after about eight steps. Sanjna Parulekar, senior VP of product marketing, said "We all had more trust in the LLM a year ago." Another Salesforce executive, Phil Mui, wrote in a blog post that customers struggle with agents "losing track of their primary objectives" when users ask tangential questions. One customer cited in the article, a home security company with 2.5 million customers, could not get Agentforce to reliably send satisfaction surveys until they built a deterministic trigger around it. The AI needed human-designed guardrails to perform a simple, repeatable task.
The technology is not the problem. Salesforce built a foundation that is unmatched in the enterprise market. They are playing the long game, combining deterministic execution with LLM reasoning in a way that can actually generate ROI for businesses. The gap is in how companies implement it.
2025 was not the year headlines promised. But it was not a failure either. It was the year that proved AI agents need the same things human workers need: defined roles, clear instructions, performance management, and oversight. The companies that treated Agentforce as an IT project got drift and dropped instructions. The companies that treated it as workforce orchestration got results.
The difference is not the technology. It is the approach. There is a lot of agent washing happening right now. Everyone wants to slap an agent on something and call it transformation. That is not how this works. Digital workers need to be onboarded like employees. Job descriptions. Success criteria. Coaching. You would not hire a college graduate and say "figure out your job and you're off to the races." Same applies here.
This is something we have been saying at Asymbl since the beginning of 2025. It is also why we chose different language to describe what we do.
We call our AI capabilities digital labor. We call the discipline of coordinating human and digital workers workforce orchestration. And the distinction matters.
We believe the future of work is not human or digital. It is human and digital, orchestrated as one.
We have deployed more than 95 digital workers at Asymbl. We have had failures along the way. Understanding how deterministic workflows impact success is one example of what we learned last year from doing this in our own company. When to use LLM reasoning and when to enforce fixed logic is not something you can learn from a whitepaper. You have to run it, watch it fail, coach it, and fix it. Everything we learn gets distilled back.
The Problem with "Agent" as a Catch-All
The term "agent" has become the default label for any AI that can take action. But it obscures more than it clarifies. When everything is an agent, nothing is clearly defined.
The word collapses three fundamentally different types of work into one:
- Deterministic automation that follows fixed rules every time.
- Agentic AI that uses generative models to interpret, reason, and take action.
- The coordination layer that ensures both work together with human oversight.
This conflation creates real problems. Leaders hear "agent" and assume they are getting intelligence, autonomy, and reliability in one package. They are not. As Mui noted, even Salesforce's "most sophisticated customers" struggle with "goal drifts," where agents lose track of their primary objectives when users ask tangential questions. An agent guiding users through filling out a form can "lose focus" when asked something unrelated. This is not a technology defect. This is what happens when you deploy AI without the workforce management discipline that keeps human employees on task.
Salesforce understood this and built accordingly. Their Agentforce 360 platform combines LLM reasoning with deterministic frameworks. They developed Agentforce Script to define when tasks should use LLM reasoning and when they should follow deterministic logic. Salesforce built the foundation. Workforce orchestration is how you make it deliver ROI.
Why We Say Digital Labor
We use the term digital labor because it reflects how we think about AI in the enterprise. A digital worker is not just an agent. It is a defined role with specific responsibilities, measurable outcomes, and a manager who provides oversight and coaching. It operates within a system alongside human workers.
This framing changes everything. It shifts the conversation from "what can AI do" to "how does this worker fit into our team." It forces clarity about scope, accountability, and integration. And it acknowledges that AI is not a replacement for human judgment. It is a new category of worker that requires workforce management, not IT implementation.
I wrote about this in June: you are not implementing AI, you are hiring it. When companies treat AI as a technology project, they hand it to IT and hope for efficiency gains. When they treat it as a workforce decision, they define roles, assign managers, set performance criteria, and integrate digital workers into existing team structures. The second approach produces measurable results. The first produces pilots that never scale.
At Asymbl, our digital workers operate across 10 business functions: sales development, recruiting, people operations, marketing, finance, and delivery. Each one has a defined role, success criteria, and a human counterpart who manages its work. We do not deploy AI and walk away. We onboard digital workers the way we onboard human employees, with clear expectations and ongoing oversight.
Without human oversight, digital workers become what we call zombie agents: decayed performers that introduce risk instead of reducing it. They require feedback, refinement, and direction from humans who understand the business context.
What Workforce Orchestration Requires
Workforce orchestration is how an organization designs, manages, and scales a blended workforce of human and digital workers. It requires three connected elements working together.
1. Deterministic automation that executes reliably and predictably.
For example, when a candidate submits an application and their years of experience is below a threshold, the system rejects them the same way every time. No interpretation, no variation. This is auditable, compliant, and necessary for any enterprise workflow that touches legal, finance, or regulated operations.
2. Agentic AI that handles ambiguity.
When a digital worker reviews a resume and assesses fit, it uses pattern matching and contextual understanding to surface insights a rule cannot anticipate. When it drafts a personalized outreach message or summarizes a complex document, it draws on language understanding that no deterministic system can replicate. This is where LLMs add value. But this value only materializes when bounded by clear guardrails.
3. Human coordination.
People provide context, judgment, and domain expertise that digital workers cannot replicate. They manage edge cases, build relationships, and make decisions where trust is required.
Remove any one element and the system breaks down. Deterministic logic without reasoning cannot handle ambiguity. Reasoning without deterministic controls cannot guarantee compliance. Either one without human oversight cannot adapt to changing business needs or catch errors before they compound.
Measuring Digital Labor Like You Measure Talent
When AI is framed as a tool, ROI becomes abstract. When it is framed as a worker, ROI becomes concrete.
Most businesses invest in AI without clear ROI metrics. An MIT study in 2025 found that 95% of businesses were getting zero return from their AI pilots. The primary factor was approach. These businesses adopted generic AI solutions that enhanced individual productivity, where capturing P&L performance is challenging. The 5% who realized positive ROI approached AI differently. They built systems that learn from feedback, retain context, and customize deeply to specific workflows.
We measure digital labor using four core metrics: cost to hire efficiency, time to fill reduction, productivity gain, and quality of outcome. These are workforce metrics, not technology metrics. They reflect how digital workers perform as members of a team, not how software performs in isolation.
We are now driving toward a goal to have 30% of output per person delivered by their digital teammate in 2026. That is not a technology target. That is a workforce productivity target. It changes how we hire, how we structure teams, and how we measure performance. Every role now has a question attached to it: what portion of this work should be handled by a digital worker, and what portion requires human judgment? The answer shapes job descriptions, capacity planning, and investment decisions. That is the multiplier effect of workforce orchestration.
The companies that will lead in this era are the ones that stop treating AI as an IT project and start treating it as a workforce strategy. Salesforce gave you the platform. Workforce orchestration is how you make it work.
That is digital labor. And that is how enterprises will cross the last mile.
Ready to explore what workforce orchestration looks like for your organization? You do not need a massive program to start. You can begin with a single job to be done. Asymbl's Digital Labor Advisory practice can help you move from pilot purgatory to measurable results. Connect with our team to discuss your specific use case.
Related reading:
FAQs
Let's connect
Discover how to build a hybrid workforce that delivers meaningful business impact, faster.
.png)
.png)
