There is a word doing a lot of heavy lifting in enterprise AI right now. That word is agentic. It appears in pitch decks, RFPs, agency credentials, and product launch announcements at a rate that has long since outpaced its actual meaning. Gartner estimates that only around 130 of the thousands of vendors currently claiming to offer agentic AI are delivering genuine agentic capabilities. The rest are selling you a workflow in a trench coat.
This is not a semantic complaint. If you are a CMO, AI Director, or transformation leader trying to deploy AI that actually changes how your organisation operates, the distinction between true non-deterministic agentic AI and a deterministic pipeline with an LLM bolted on is the difference between a system that can handle novel situations and one that breaks the moment reality diverges from the script. It is also the difference between a board-level AI transformation story and a very expensive automation project that saves two hours a week.
What Deterministic Workflows Actually Are
A deterministic workflow is a system where the path from input to output is fully specified in advance. Given the same input, the system always follows the same steps and produces the same output. This is not a criticism — deterministic systems are reliable, auditable, and appropriate for a large class of enterprise problems. The issue is not that they exist. The issue is that they are being sold as something they are not.
The tell is in the architecture. A deterministic pipeline typically looks like this: a trigger fires, a sequence of steps executes in order, each step calls a tool or an LLM with a fixed prompt, and the output is assembled from the results. The LLM in this system is not reasoning — it is formatting. It is a very capable text transformer operating inside a rigid container. If the input deviates from what the prompt was designed for, the system either produces garbage or fails. There is no re-planning. There is no negotiation. There is no agent.
Common examples include Zapier chains with GPT-4 nodes, n8n workflows with LLM steps, Make.com automations with AI actions, and the majority of what is currently sold as "AI-powered" marketing automation. These are useful tools. They are not agentic AI.
What True Non-Deterministic Agentic AI Actually Does
A genuine agentic AI system operates on goals, not steps. The agent is given an objective — not a script — and must reason about how to achieve it. This means the system can decompose a novel task into sub-tasks it has never seen before, select tools from its available toolkit based on what the current sub-task requires, adapt its plan when it encounters unexpected information mid-execution, and in multi-agent architectures, negotiate with other agents about what is needed and who should do what.
The output path is not fully predictable in advance. Two runs of the same agent on the same goal may take different routes depending on what the agent discovers along the way. This is not a bug. It is the core capability that makes agentic AI valuable for complex, open-ended tasks. It is also what makes it genuinely difficult to build, test, and govern — which is precisely why most vendors are not actually doing it.
MIT Sloan's definition is useful here: agentic AI systems "incorporate multiple, different agents that are orchestrating a task together." The key word is orchestrating. Not executing a fixed sequence. Orchestrating — which implies dynamic coordination, negotiation, and adaptation among agents with different capabilities and contexts.
The 4-Level Autonomy Framework: Where Most Vendors Actually Sit
Writer.com's production AI team, which has deployed agentic systems at scale with enterprises including Uber and Franklin Templeton, uses a four-level framework that is the most practically useful classification I have seen for enterprise contexts.
Level 1 — Assistive agents automate simple, single-step tasks based on fixed prompts. Input in, output out. No external data, no tool use, no adaptation. Think automated FAQ generation or content summarisation. These are mostly deterministic with minimal probabilistic components. Most "AI copilots" and "AI assistants" sold to enterprises sit here.
Level 2 — Knowledge agents integrate enterprise knowledge through retrieval-augmented generation (RAG). They pull from internal documents and databases to provide context-rich outputs. The retrieval step is deterministic; the synthesis is probabilistic. This is genuinely useful for document-heavy workflows and is where most well-implemented enterprise RAG systems operate.
Level 3 — Action agents connect to external tools and APIs. They can send emails, update CRM records, trigger workflows, and publish content. They have tool-calling capabilities that extend beyond their built-in knowledge. This is where genuine agentic behaviour begins to emerge — the agent is making decisions about which tool to call and when. Most marketing automation use cases that genuinely benefit from AI belong here.
Level 4 — Multi-agent systems involve networks of agents collaborating to achieve complex goals. Multiple agents communicate, pass tasks, and execute in coordination. One agent might handle research, another synthesis, another quality assurance, and a fourth execution — with each agent able to request help from the others when it encounters something outside its competence. This is true non-deterministic agentic AI. It is appropriate for complex, open-ended tasks where the path to the goal cannot be fully specified in advance.
The problem Writer.com identifies — and which I see consistently in enterprise AI deployments — is that most mismatched expectations happen when vendors market Level 1 tools as if they were Level 3 or Level 4 systems. The demo looks impressive because it uses a controlled input. The production deployment breaks because the real world does not use controlled inputs.
The Agent Washing Red Flags: What to Look For Before You Sign
The FTC launched Operation AI Comply in 2024 specifically to address deceptive AI marketing claims. The SEC charged Presto Automation in 2025 for misleading investors with inflated AI claims — their "agentic" system required human assistance for over 70% of orders. These are not edge cases. They are symptoms of a market where the incentive to claim agentic capability far exceeds the cost of not delivering it.
The red flags are consistent across vendors. Watch for any system that promises full autonomy while requiring human assistance for routine tasks. Watch for vague claims of "AI-powered" without any explanation of the actual autonomy level. Watch for demos that always use the same controlled input. Watch for an inability to explain the hybrid architecture — specifically, how deterministic and probabilistic components work together and where the boundaries are. And watch for any vendor who cannot describe their failure modes. Every system on the autonomy spectrum has failure modes. A vendor who cannot describe theirs has not thought seriously about production deployment.
The Tests I Use Before Signing Off on Any Agentic AI Deployment
When I evaluate an agentic AI system — whether built by an agency, a vendor, or an internal team — I use four tests that cut through the marketing language quickly.
The first is the novel input test. Give the system an input it was not designed for. Not a wildly out-of-scope input, but a plausible variant that deviates from the training examples. A deterministic pipeline will either produce garbage or fail. A genuine agentic system will attempt to reason about the novel input and either handle it gracefully or fail in a way that explains why.
The second is the mid-task disruption test. Interrupt the system mid-execution with new information that changes the context. Tell it the CRM it was about to update is unavailable. Tell it the research it just completed has been superseded by a new report. A deterministic pipeline cannot adapt. A genuine agentic system will re-plan.
The third is the agent negotiation test. In a multi-agent system, ask to observe the inter-agent communication logs. What are the agents saying to each other? Are they passing fixed data structures along a pre-defined protocol, or are they making requests, receiving responses, and adapting their behaviour based on what other agents tell them? The former is a distributed deterministic pipeline. The latter is genuine multi-agent coordination.
The fourth is the failure mode test. Ask the vendor to show you a case where the system failed and explain why. How did it fail? What was the recovery path? What human oversight was triggered? A vendor who cannot answer this has not deployed their system in production at meaningful scale.
The Right Architecture for Enterprise Agentic AI
The most important insight from working with both deterministic and genuinely agentic systems is that the goal is not maximum autonomy. The goal is right-sized autonomy for the task at hand, with deterministic guardrails where reliability and auditability matter most.
The jazz band analogy from Writer.com's production team is the most useful frame I have found: you need solid musical structure so the musicians can improvise brilliantly. Deterministic infrastructure — schemas, protocols, safety rails, human oversight triggers — provides the structure. Probabilistic intelligence — reasoning, creativity, adaptation, negotiation — provides the improvisation. The best enterprise agentic AI systems combine both deliberately, not by accident.
For most B2B marketing and growth use cases, the practical architecture is a Level 3 action agent layer (tool calling, CRM integration, campaign execution) built on a Level 2 knowledge foundation (RAG over your content, product, and customer data), with Level 4 multi-agent coordination reserved for the genuinely complex, open-ended tasks where the path to the goal cannot be scripted — autonomous competitive intelligence, multi-persona content generation at scale, or cross-system pipeline orchestration.
If you are working with a tech or product team to deploy this, the conversation that matters is not "is this agentic?" It is: what decisions do we want the system to make autonomously, what decisions require human confirmation, and where are the guardrails that prevent the system from taking actions outside its authorised scope? That conversation is more productive than any vendor demo.
What This Means for Your AI Transformation Roadmap
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. The organisations that avoid this outcome are not the ones that deployed the most autonomous systems. They are the ones that were precise about what they needed, honest about what they were buying, and deliberate about where human oversight was non-negotiable.
The CMOs and AI Directors I work with who are making genuine progress on AI transformation share a common characteristic: they are not impressed by demos. They ask hard questions about production performance, failure modes, and cost at scale. They insist on seeing the architecture, not just the interface. They know the difference between a workflow and an agent. And they have stopped letting vendors define the vocabulary.
The next time an agency or vendor tells you their system is agentic, ask them which level. Ask them to show you the inter-agent communication logs. Ask them what happens when the input is unexpected. The answer will tell you everything you need to know about whether you are looking at a genuine AI transformation capability or a workflow in a trench coat.
If you want to assess where your current AI stack sits on the autonomy spectrum — and identify the gaps between what you have been sold and what you actually have — our free AI Growth Audit includes an agentic AI readiness assessment covering your current tools, vendor claims, and deployment architecture.
Free Download
The Enterprise Agentic AI Architecture Guide
The companion working document to this article. Covers the 4-Level Autonomy Spectrum, the Four Readiness Questions, the Hybrid Architecture Model, the Decision Rights Matrix, and an 8-point Readiness Assessment with a go/no-go scoring guide. Built for CMOs and AI Directors who are past the evaluation stage.
↓ Download the Guide (PDF, Free)About the Author
Modi Elnadi is Founder and Director of Marketing and AI Growth at Integrated.Social, a London-based B2B AI growth marketing agency. Modi has deployed agentic AI systems across B2B technology, financial services, and professional services organisations, working directly with tech and product teams to bridge the gap between vendor claims and production reality. His work spans the full agentic AI stack — from multi-agent architecture design and prompt engineering to governance frameworks and board-level AI risk communication. He writes for CMOs and AI Directors who are navigating the difference between genuine AI transformation and expensive automation theatre.
There is a word doing a lot of heavy lifting in enterprise AI right now. That word is agentic. It appears in pitch decks, RFPs, agency credentials, and product launch announcements at a rate that has long since outpaced its actual meaning. Gartner estimates that only around 130 of the thousands of vendors currently claiming to offer agentic AI are delivering genuine agentic capabilities. The rest are selling you a workflow in a trench coat.
This is not a semantic complaint. If you are a CMO, AI Director, or transformation leader trying to deploy AI that actually changes how your organisation operates, the distinction between true non-deterministic agentic AI and a deterministic pipeline with an LLM bolted on is the difference between a system that can handle novel situations and one that breaks the moment reality diverges from the script. It is also the difference between a board-level AI transformation story and a very expensive automation project that saves two hours a week.
What Deterministic Workflows Actually Are
A deterministic workflow is a system where the path from input to output is fully specified in advance. Given the same input, the system always follows the same steps and produces the same output. This is not a criticism — deterministic systems are reliable, auditable, and appropriate for a large class of enterprise problems. The issue is not that they exist. The issue is that they are being sold as something they are not.
The tell is in the architecture. A deterministic pipeline typically looks like this: a trigger fires, a sequence of steps executes in order, each step calls a tool or an LLM with a fixed prompt, and the output is assembled from the results. The LLM in this system is not reasoning — it is formatting. It is a very capable text transformer operating inside a rigid container. If the input deviates from what the prompt was designed for, the system either produces garbage or fails. There is no re-planning. There is no negotiation. There is no agent.
Common examples include Zapier chains with GPT-4 nodes, n8n workflows with LLM steps, Make.com automations with AI actions, and the majority of what is currently sold as "AI-powered" marketing automation. These are useful tools. They are not agentic AI.
What True Non-Deterministic Agentic AI Actually Does
A genuine agentic AI system operates on goals, not steps. The agent is given an objective — not a script — and must reason about how to achieve it. This means the system can decompose a novel task into sub-tasks it has never seen before, select tools from its available toolkit based on what the current sub-task requires, adapt its plan when it encounters unexpected information mid-execution, and in multi-agent architectures, negotiate with other agents about what is needed and who should do what.
The output path is not fully predictable in advance. Two runs of the same agent on the same goal may take different routes depending on what the agent discovers along the way. This is not a bug. It is the core capability that makes agentic AI valuable for complex, open-ended tasks. It is also what makes it genuinely difficult to build, test, and govern — which is precisely why most vendors are not actually doing it.
MIT Sloan's definition is useful here: agentic AI systems "incorporate multiple, different agents that are orchestrating a task together." The key word is orchestrating. Not executing a fixed sequence. Orchestrating — which implies dynamic coordination, negotiation, and adaptation among agents with different capabilities and contexts.
The 4-Level Autonomy Framework: Where Most Vendors Actually Sit
Writer.com's production AI team, which has deployed agentic systems at scale with enterprises including Uber and Franklin Templeton, uses a four-level framework that is the most practically useful classification I have seen for enterprise contexts.
Level 1 — Assistive agents automate simple, single-step tasks based on fixed prompts. Input in, output out. No external data, no tool use, no adaptation. Think automated FAQ generation or content summarisation. These are mostly deterministic with minimal probabilistic components. Most "AI copilots" and "AI assistants" sold to enterprises sit here.
Level 2 — Knowledge agents integrate enterprise knowledge through retrieval-augmented generation (RAG). They pull from internal documents and databases to provide context-rich outputs. The retrieval step is deterministic; the synthesis is probabilistic. This is genuinely useful for document-heavy workflows and is where most well-implemented enterprise RAG systems operate.
Level 3 — Action agents connect to external tools and APIs. They can send emails, update CRM records, trigger workflows, and publish content. They have tool-calling capabilities that extend beyond their built-in knowledge. This is where genuine agentic behaviour begins to emerge — the agent is making decisions about which tool to call and when. Most marketing automation use cases that genuinely benefit from AI belong here.
Level 4 — Multi-agent systems involve networks of agents collaborating to achieve complex goals. Multiple agents communicate, pass tasks, and execute in coordination. One agent might handle research, another synthesis, another quality assurance, and a fourth execution — with each agent able to request help from the others when it encounters something outside its competence. This is true non-deterministic agentic AI. It is appropriate for complex, open-ended tasks where the path to the goal cannot be fully specified in advance.
The problem Writer.com identifies — and which I see consistently in enterprise AI deployments — is that most mismatched expectations happen when vendors market Level 1 tools as if they were Level 3 or Level 4 systems. The demo looks impressive because it uses a controlled input. The production deployment breaks because the real world does not use controlled inputs.
The Agent Washing Red Flags: What to Look For Before You Sign
The FTC launched Operation AI Comply in 2024 specifically to address deceptive AI marketing claims. The SEC charged Presto Automation in 2025 for misleading investors with inflated AI claims — their "agentic" system required human assistance for over 70% of orders. These are not edge cases. They are symptoms of a market where the incentive to claim agentic capability far exceeds the cost of not delivering it.
The red flags are consistent across vendors. Watch for any system that promises full autonomy while requiring human assistance for routine tasks. Watch for vague claims of "AI-powered" without any explanation of the actual autonomy level. Watch for demos that always use the same controlled input. Watch for an inability to explain the hybrid architecture — specifically, how deterministic and probabilistic components work together and where the boundaries are. And watch for any vendor who cannot describe their failure modes. Every system on the autonomy spectrum has failure modes. A vendor who cannot describe theirs has not thought seriously about production deployment.
The Tests I Use Before Signing Off on Any Agentic AI Deployment
When I evaluate an agentic AI system — whether built by an agency, a vendor, or an internal team — I use four tests that cut through the marketing language quickly.
The first is the novel input test. Give the system an input it was not designed for. Not a wildly out-of-scope input, but a plausible variant that deviates from the training examples. A deterministic pipeline will either produce garbage or fail. A genuine agentic system will attempt to reason about the novel input and either handle it gracefully or fail in a way that explains why.
The second is the mid-task disruption test. Interrupt the system mid-execution with new information that changes the context. Tell it the CRM it was about to update is unavailable. Tell it the research it just completed has been superseded by a new report. A deterministic pipeline cannot adapt. A genuine agentic system will re-plan.
The third is the agent negotiation test. In a multi-agent system, ask to observe the inter-agent communication logs. What are the agents saying to each other? Are they passing fixed data structures along a pre-defined protocol, or are they making requests, receiving responses, and adapting their behaviour based on what other agents tell them? The former is a distributed deterministic pipeline. The latter is genuine multi-agent coordination.
The fourth is the failure mode test. Ask the vendor to show you a case where the system failed and explain why. How did it fail? What was the recovery path? What human oversight was triggered? A vendor who cannot answer this has not deployed their system in production at meaningful scale.
The Right Architecture for Enterprise Agentic AI
The most important insight from working with both deterministic and genuinely agentic systems is that the goal is not maximum autonomy. The goal is right-sized autonomy for the task at hand, with deterministic guardrails where reliability and auditability matter most.
The jazz band analogy from Writer.com's production team is the most useful frame I have found: you need solid musical structure so the musicians can improvise brilliantly. Deterministic infrastructure — schemas, protocols, safety rails, human oversight triggers — provides the structure. Probabilistic intelligence — reasoning, creativity, adaptation, negotiation — provides the improvisation. The best enterprise agentic AI systems combine both deliberately, not by accident.
For most B2B marketing and growth use cases, the practical architecture is a Level 3 action agent layer (tool calling, CRM integration, campaign execution) built on a Level 2 knowledge foundation (RAG over your content, product, and customer data), with Level 4 multi-agent coordination reserved for the genuinely complex, open-ended tasks where the path to the goal cannot be scripted — autonomous competitive intelligence, multi-persona content generation at scale, or cross-system pipeline orchestration.
If you are working with a tech or product team to deploy this, the conversation that matters is not "is this agentic?" It is: what decisions do we want the system to make autonomously, what decisions require human confirmation, and where are the guardrails that prevent the system from taking actions outside its authorised scope? That conversation is more productive than any vendor demo.
What This Means for Your AI Transformation Roadmap
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. The organisations that avoid this outcome are not the ones that deployed the most autonomous systems. They are the ones that were precise about what they needed, honest about what they were buying, and deliberate about where human oversight was non-negotiable.
The CMOs and AI Directors I work with who are making genuine progress on AI transformation share a common characteristic: they are not impressed by demos. They ask hard questions about production performance, failure modes, and cost at scale. They insist on seeing the architecture, not just the interface. They know the difference between a workflow and an agent. And they have stopped letting vendors define the vocabulary.
The next time an agency or vendor tells you their system is agentic, ask them which level. Ask them to show you the inter-agent communication logs. Ask them what happens when the input is unexpected. The answer will tell you everything you need to know about whether you are looking at a genuine AI transformation capability or a workflow in a trench coat.
If you want to assess where your current AI stack sits on the autonomy spectrum — and identify the gaps between what you have been sold and what you actually have — our free AI Growth Audit includes an agentic AI readiness assessment covering your current tools, vendor claims, and deployment architecture.
Free Download
The Enterprise Agentic AI Architecture Guide
The companion working document to this article. Covers the 4-Level Autonomy Spectrum, the Four Readiness Questions, the Hybrid Architecture Model, the Decision Rights Matrix, and an 8-point Readiness Assessment with a go/no-go scoring guide. Built for CMOs and AI Directors who are past the evaluation stage.
↓ Download the Guide (PDF, Free)About the Author
Modi Elnadi is Founder and Director of Marketing and AI Growth at Integrated.Social, a London-based B2B AI growth marketing agency. Modi has deployed agentic AI systems across B2B technology, financial services, and professional services organisations, working directly with tech and product teams to bridge the gap between vendor claims and production reality. His work spans the full agentic AI stack — from multi-agent architecture design and prompt engineering to governance frameworks and board-level AI risk communication. He writes for CMOs and AI Directors who are navigating the difference between genuine AI transformation and expensive automation theatre.
