At Relevance AI, we're building an agentic automation platform. We often get compared to AI workflow solutions like n8n or Zapier. It's a natural comparison—we both help teams automate work and both integrate deeply with LLMs.
But this comparison misses something fundamental about what each tool is designed to do.
The likes of Relevance and n8n solve different problems; understanding the distinction is key to building an effective automation strategy.
This post is quite theoretical, but the goal is to create a really clear foundation as to what type of work makes most sense for AI Agents.
Firstly, it’s not about LLMs
The most common misconception about agentic automation is that it’s just traditional automation with the introduction of LLMs.
Browse any workflow automation platform—n8n, Zapier, Make—and you'll find LLM integrations everywhere. GPT-4 for text generation, Claude for summarization, even “agentic modules”.
Workflow automations have gotten way better thanks to LLMs. Agentic automation has been made possible thanks to LLMs.
The presence of language models tells you nothing about which type of automation you're looking at. There’s a more fundamental distinction to be found.
How they make decisions: conditions vs. predictions
All automation systems must make decisions to complete tasks. The core technical distinction between agentic and workflow automation is simple:
- Workflow automations decide based on predefined conditions, powered by code.
- Agentic automations decide based on real time predictions, powered by a model.
It’s important to understand that these two forms of automation are both valuable and can be applied to different types of work.
Workflow automation solve tasks where every possible step and the connections between them can be defined. This is, essentially, the foundation of all software.
Software and automation has fundamentally changed our world. It's eliminated a lot of work that used to be manual and tedious. But far from all of it!
The rest of this work, the tasks where it's not feasible to define every condition, has required humans to solve. This is because we're able to take our skills and apply them in a generalised way to a larger range of scenarios.
Consider the calculator as a proxy for software. Zoomed in, the calculator is far more proficient at each individual maths task than the human. Zoomed out, a human is capable of leveraging a less precise but more general understanding of maths to run an accountancy or to discover new forms of physics.
Agentic automation is a new way to solve a large amount of these "undefinable" tasks. Those where a model is able to predict what steps to take and how to connect them in a generalised but highly scalable way.
Let’s explore this further through the lens of an unlikely example: hospital work.
Automation in nursing
Nurses handle an enormous range of responsibilities during every shift.
They administer medications, monitor patients, document care, communicate with medical teams and families, respond to emergencies, and make critical clinical decisions.
Workflow automation: dispensing medication
The medication preparation is particularly demanding - repetitive yet requiring perfect accuracy. Imagine: measuring liquid medications, preparing injections, counting tablets, double-checking dosages against patient weight, cross-referencing drug interactions. One miscalculation could be fatal.
New automated systems have helped. When hospital staff enter a prescription, the system follows exact protocols:
- Drug: Amoxicillin, Dose: 500mg, Frequency: 3x daily → Dispense 21 tablets for 7 days
- If patient_age < 12, calculate pediatric dose
- if patient_weight > 100kg & patient_height < 180cm, print blood test referral
- If drug_interaction_detected = true, alert pharmacist
The system executes these rules with perfect precision. No interpretation, no variation, no judgment calls. It just works, within these defined conditions. This is the model for workflow automation, at its best.
We can add an LLM step to this workflow. Let's say, the details on the patient originate from a note, written quickly by the nurse while triaging.
"patient is eighteen, 6 foot high and 80 kgs"
The automated system might first run these unstructured notes through an LLM to convert them into the expected structure of "patient_age", "patient_height", "patient_weight".
This leverages AI in a new and powerful way, but its still fundamentally a workflow: the steps are predefined and should work the same every time, within expected schemas.
Agentic automation: everything else
Now think of a nurse doing rounds. The nurse has the same goal—patient care—but approaches it completely differently:
- Notices a patient seems uncomfortable and adjusts their position
- Recognizes subtle signs of distress that aren't triggering monitors
- Prioritizes which patient to check based on a combination of scheduled tasks and real-time observations
- Adapts their approach based on each patient's communication style and needs
The nurse makes continuous judgment calls (or, predictions) based on context and experience. This sort of work was never previously possible to automate. Today, this type of work can be delegated to agentic automation: adaptive, contextual, predictive.
Neither approach is inherently better; you want the prescription dispenser for medication—precision matters, variation is dangerous and it always works. You want the nurse for patient care—context matters and rigid rules miss important signals and edge cases.
Likewise, your automation strategy needs both. In fact - they are often used together. Agents can use workflows and workflows can trigger agents.
So when should you reach for the nurse over the dispenser?
The three scenerios where you have to use agentic automation
In a hospital, you can write protocols for dispensing medication—exact drugs, exact doses, exact timing.
But you can't write protocols for every situation a nurse might encounter.
Some scenarios resist standardization because there are too many variables, or because understanding requires interpreting semantics, or because the best approach only becomes clear as you gather information.
Here are the three scenarios where traditional automation hits its limits and agentic automation becomes essential.
1. When decisions grow exponentially
Technically, you can define enough “IF/ELSE” conditions to automate literally anything. But in reality, some processes have so many variables and edge cases that mapping them all becomes practically impossible.
The number of potential decisions grow exponentially with each new factor you consider. At a certain point, defining conditions becomes more complex than the work itself.
The best automation is the one you can actually build. Agentic automations, through their generalised approach, are able to handle an exponentially larger set of conditions - even if less accurate on the molecular level.
Example: Dynamic lead scoring
Imagine a B2B software company trying to score leads. They start with basic rules—company size, industry, engagement level. But real lead quality depends on hundreds of subtle factors: hiring patterns, technology stack, competitive dynamics, market timing, leadership changes, funding status, and countless combinations thereof.
A traditional approach would require mapping every possible combination.
- Is a 50-person company with rapid hiring more valuable than a 200-person company with stagnant growth?
- What if the smaller company just hired their twentieth engineer and their CEO previously sold a company in your space?
- What if the larger company is in a declining industry?
An agentic system doesn't need every permutation mapped. It learns patterns from the training data and prompting of its underlying LLM, and applies that learning to new situations.
It can recognize that rapid engineering hiring plus experienced leadership equals high potential, without anyone explicitly programming this rule.
2. When decisions rely on meaning not structure
Human communication is inherently semantic—we convey meaning through context, implication, and nuance, not just through structured data.
Workflow automation operates on syntax (the structure of information) while many business problems require semantics (the meaning of information).
The model powered decision making engine of an AI Agent is able to interpret semantics through language, and sometimes even images, video and audio. This allows it to tackle some of those business problems that workflows can't.
Example: Multi-channel customer relationship management
Imagine a B2B company where the same customer contacts you through multiple channels over several weeks:
- Email 1: "Interesting product, might be useful for our Q2 initiatives"
- LinkedIn message: "Hey, saw your CEO's post about enterprise features"
- Support ticket: "Getting 'permission denied' errors when trying to access the advanced analytics"
- Sales call note: "Customer mentioned they're evaluating 3 competitors"
- Email 2: "Our procurement team is asking about SOC2 compliance"
Traditional automation, even with LLM steps, would process each interaction in isolation—route the support ticket, log the sales call, respond to the compliance question.
But the real insight requires connecting these dots across time and channels: this customer is deep in an evaluation process, hit a technical roadblock that might derail the deal, and is now doing final vendor diligence.
If your agent has access to all these touchpoints it could recognize the pattern: a high-intent prospect facing a potential deal-breaker. It might:
- Escalate the support ticket as deal-critical (not just technical)
- Alert sales that the customer is in active evaluation
- Proactively share SOC2 documentation and enterprise onboarding materials
- Recognize that "permission denied" for advanced analytics might mean they're trying to test the exact features that differentiate you from competitors
This requires a true semantic understanding of the task—interpreting communication, recognizing patterns across time, and making predictions about what these signals mean together. No feasible amount of LLM-powered rules in a workflow could capture this emergent insight.
3. When optimization paths emerge from context (and you care)
Some processes can't be optimized in advance because the optimal approach depends on information you discover during execution. The best path forward emerges from context, not from predefined workflows.
It seems obvious - but for this to be a reason to adopt agentic automation, you also need to care about taking the most effecient path every time. A lot of the time, you are OK to make this trade off for the reliability of a workflow automation.
It usually matters for tasks where there's a large cost or time difference at scale between the optimal paths and the configurable workflow.
Example: High quality research
Imagine an investment firm investigating potential acquisitions. Their research process is extremely important, as lots of money rides on the findings.
The optimal path for each business you research varies dramatically based on what you discover.
- Start researching a company, and realise that they're a software company.
- Start researching what their software does and realise you need to learn about the niche industry they service.
- Discover they have a complex IP licensing arrangement with a competitor. Suddenly, intellectual property analysis becomes critical.
- Find that their founder is testifying before Congress next week? Regulatory research is required.
An agent adapts its approach based on discoveries. It might spend 80% of time on financial analysis for one company and 80% on regulatory review for another.
Traditional automation could still, theoretically, cover its bases by researching every topic (regulatory, IP, financial etc.) for every business. But this would cost a lot of money and waste a lot of time at scale.
So usually, this means companies either change their automation to do lower quality, more generic work - or just do things like this manually.
The adoption gradient: why AI Agents have made people care about automation
At Relevance, we're seeing an interesting phenomenon. Teams are coming to us asking for help automating tasks that could have been automated with traditional tools for years.
"We want an AI agent to move files from Dropbox to Google Drive when they're approved."
"We need an agent to send Slack notifications when deals close in our CRM."
These sort of problems have been possible to automate for a long time. Even workflow automation leveraging LLM steps has been possible for years now.
Yet it’s only with the advent of AI Agents that we’ve seen this incredible hunger to automate processes truly blossom.
The reason is - there’s another superpower of agentic automation that makes it very attractive:
Agentic automation has a significantly lower technical barrier to entry.
Binary vs. gradient configuration
Workflow automation is binary in its configuration—it either works perfectly or it doesn't work at all.
You need to map out every step, define every condition, handle every edge case upfront. If you miss something, the automation breaks. If requirements change, you rebuild.
You can start by creating a very basic workflow, connecting something from A to B - but basic workflows can only solve basic problems.
Agentic automation exists on a gradient. You can start with a barebones agent, so long as it has access to the right integrations ("tools" in Relevance AI).
No complex configuration, no workflow mapping—just "Hey, when someone uploads an approved contract to Dropbox, make sure it ends up in the right Google Drive folder."
The agent might not get it perfect initially. Maybe it puts contracts in slightly wrong folders or doesn't understand your naming convention. But here's the key: you can guide it. "Actually, legal contracts go in the Legal folder, not Contracts." "Use the client name from the filename, not the contract ID." The agent adjusts.
If it tries to use a tool incorrectly it can “self heal”, learning from the error and trying again. It may take a few goes, but often agents can figure out how to fix things by themselves.
Basic agents can solve complex problems - they might just require a lot of feedback and take a very inefficient path.
This gradient approach promotes experimentation. Teams that would never attempt traditional automation because the upfront investment feels too high will try agents. Even if they eventually realize their use case is better suited for traditional workflows, they've now:
- Validated that automation provides value
- Understood their process well enough to define it
- Built organizational momentum around automation
Resiliency to third-party systems
When systems change—and they always do—traditional automations break; an API changes its response format, a webpage restructures, a field gets renamed. Everything stops until someone debugs and fixes the workflow.
Agents adapt.
When LinkedIn changes their page layout, an agent doesn't look for specific HTML elements—it looks for information that resembles work experience. When an API adds new fields, an agent incorporates useful data rather than erroring on schema mismatches.
This resilience makes teams more willing to automate edge cases and temporary processes. Why build a brittle workflow for something that might change next month when an agent can handle the variations?
Opening automation up to domain experts
Agents can be instructed with natural language, rather than schemas and conditions. This, crucially, allows them to be configured by domain experts - not just technical resources.
This democratizes automation. The people who understand the work best can directly improve how it's automated. The feedback loop is immediate and intuitive.
This means not only can teams build more automation, but the automation will be more effective as it has been directly guided by subject matter experts.
The compound effect: a reason to invest heavily in agents, that no one is speaking about
The same characteristics that make individual agents easy to configure—natural language interaction, contextual understanding, adaptive behavior—also make them remarkably good at working together.
This creates compound effects that go beyond what you explicitly design.
Workflow integration: Planned connections
Traditional automations integrate through precise, pre-planned connections. System A outputs data in format X, System B accepts format X. Every interaction is designed, documented, and rigid.
With N systems, you need N(N-1)/2 explicit integrations. The complexity grows quadratically. Worse, each integration is brittle—change one side, break the connection.
Agentic collaboration: Emergent cooperation
Agents collaborate using the same natural language interface they use with humans. A Research Agent doesn't need to know the Writing Agent's data schema—it just needs to communicate its needs clearly.
Straight away, this makes it easier for your team to piece separate agentic automations together to solve new problems.
Here's where it gets interesting: because agents can interpret and adapt, it also means they can solve problems you didn't explicitly plan for.
Imagine, you build three agents:
- Research Agent: Gathers market intelligence
- Analysis Agent: Identifies patterns and insights
- Reporting Agent: Creates summaries for stakeholders
You designed a simple pipeline: Research feeds Analysis, Analysis feeds Reporting. Done.
But then something interesting happens. Your Analysis Agent hits a gap in the data and decides to ask follow-up questions: "Research Agent, can you find three more examples of SaaS companies using usage-based pricing?"
Meanwhile, your Reporting Agent realizes the executive team needs different formats: "Analysis Agent, can you break down those pricing insights by company size and send the enterprise segment data directly to our Notion database?"
You never programmed these interactions. The agents figured out they needed to collaborate differently.
This emergent behavior means your automation system can become more capable without additional configuration. New agents can join and immediately start collaborating.
The network becomes smarter than the sum of its parts and your automation surface area grows exponentially.
Delivering on the promise of automation
Automation technology such as software and machinery has improved rapidly since the industrial revolution. Yet somehow, Keynes' prediction of a 15-hour work week never materialized.
If anything, knowledge workers are drowning in more administrative tasks than ever.
Why? Because until now, there was a massive gap in what we could automate.
Most knowledge work isn't structured or predictable. It requires understanding context, making judgment calls, handling ambiguity—all things that only humans could do.
So we automated the easy parts and kept piling the complex, judgment-based work onto humans. Every efficiency gain just created more work that required human interpretation.
Email made communication faster, so we communicated more.
Digital tools made data collection easier, so we collected more data that needed human analysis.
The result? A paradox where technology made individual tasks easier but overall work more overwhelming. We optimized the prescription dispensing but, as any hospital worker could tell you, multiplied the nursing duties.
Agentic automation represents the first technology capable of handling this judgment-based work at scale.
We're not witnessing a revolution where new automation replaces old. We're seeing an expansion of what can be automated.
For the first time, we can automate not just the structured processes but the messy, context-dependent tasks that multiply faster than humans can handle them.
Technology has made the world more complex. Agents can make it simpler.
At Relevance, this is what excites us the most: the potential for a future where technology has finally reduced the amount of repetitive, draining work that dominates the work week.
Where work itself is more inherently joyful, with AI Agents taking care of the parts of your job that make you dread Monday mornings.
This isn't about replacing humans—it's about finally delivering on technology's original promise.
