You have read the case studies. You have sat through the demo. The agency’s slides are polished, the orchestrator diagram looks impressive, and the word “agentic” appears seventeen times in their proposal. Now you need to decide whether to sign.
The problem is that “agentic AI” has become the most overloaded term in B2B marketing in 2026. Agencies that were running Make.com workflows six months ago have rebranded them as agentic systems. Prompt chains are now “multi-agent architectures.” A single GPT-4 call with a long system prompt is now an “AI agent.” The terminology is being used to justify premium pricing for what is, in practice, deterministic automation.
This guide gives you the 12 questions that cut through the positioning. They are designed to be asked in a discovery call or proposal review. The answers will tell you, with high confidence, whether you are looking at a genuine agentic AI methodology or a sophisticated-sounding automation workflow. Use them before you sign anything.
Why the Briefing Stage Is the Most Important Moment
The briefing stage is where the gap between genuine agentic AI and automation is most visible — if you know what to look for. A true agentic system requires a fundamentally different briefing process than an automation workflow. It needs sector context, scenario mapping, ICP definition, guardrail design, and an understanding of your human-in-the-loop requirements. An automation workflow needs a process map and a trigger list.
If an agency’s briefing questionnaire looks like a project management intake form — timelines, deliverables, channel preferences, budget — you are looking at an automation vendor. If it asks about your sector’s edge cases, your compliance constraints, your ICP’s decision-making hierarchy, and the scenarios where you want human oversight, you are looking at something closer to genuine agentic methodology.
According to Gartner’s 2025 AI Hype Cycle, agentic AI is at the Peak of Inflated Expectations — which means the market is full of vendors claiming capabilities they do not yet have. The 12 questions below are your filter.
The 12 Questions — and What the Answers Reveal
Question 1: Can You Show Me Your Orchestrator Architecture?
A genuine agentic system has a documented orchestrator layer — the decision-making agent that sits above all specialist agents, holds strategic context across the engagement, and decides which agent to invoke next. Ask to see a diagram. Ask how it handles a scenario where two agents produce conflicting outputs. Ask what happens when the orchestrator encounters a situation it has not been trained on.
Red flag answer: “Our orchestrator is the LLM itself — it decides what to do next based on the prompt.” This describes a single-model prompt chain, not an orchestrated multi-agent system.
Green flag answer: A documented architecture diagram showing a distinct orchestrator layer, specialist agents with defined domains, handoff triggers, and a conflict resolution mechanism.
Question 2: How Many Specialist Agents Do You Deploy, and What Is Each One Trained On?
True agentic systems deploy specialist agents — each trained, prompted, and constrained for a specific domain. In a B2B marketing context, you should expect separate agents for ICP qualification, content strategy, paid media signal interpretation, AEO gap analysis, and CRM enrichment at minimum. Each agent should have a documented training scope and sector-specific calibration.
Red flag answer: “We have one main agent that handles everything, with different prompts for different tasks.” This is a single-model system with prompt switching — not specialist agents.
Green flag answer: A named list of specialist agents with documented domains, training data sources, and the specific scenarios each agent has been calibrated to handle in your sector.
Question 3: How Do You Train Agents on Our Sector?
Sector-specific training is the single biggest differentiator between a genuine agentic methodology and a generic tool deployment. The agency should be able to describe a specific process for mapping your sector’s scenarios — the objections your buyers raise, the compliance constraints that govern what an agent can say, the content formats that perform in your vertical, the seasonal patterns in your buying cycle.
Red flag answer: “We give the agent your brand guidelines and a few examples of your content.” This is prompt customization, not sector training.
Green flag answer: A documented scenario-mapping process, evidence of prior work in your sector, and a description of how sector knowledge is embedded at the agent level rather than applied as a wrapper at the output level.
Question 4: How Does the System Handle Edge Cases — Scenarios Outside the Defined Workflow?
This is the question that most cleanly separates automation from genuine agency. Automation workflows break on edge cases — they either fail silently, produce generic output, or require manual intervention. A true agentic system has been trained to recognize edge cases and route them appropriately — either to a different specialist agent, to a human-in-the-loop checkpoint, or to a fallback protocol.
Red flag answer: “We handle edge cases by updating the workflow and adding a new branch.” This describes a brittle deterministic system that requires manual maintenance every time a new scenario emerges.
Green flag answer: A description of the edge-case detection logic, the routing mechanism, and examples of edge cases the system has encountered and handled autonomously in prior engagements.
Question 5: What Is Your Dynamic Handoff Logic, and What Triggers a Handoff?
Dynamic handoff logic is what allows agents to recognize when a scenario exceeds their domain and route to a more appropriate specialist, passing the relevant context. The triggers should be condition-based — signals the agent has been trained to recognize — not fixed workflow transitions. Ask for specific examples of handoff triggers in your sector.
Red flag answer: “Handoffs happen when a task is completed — the output of one step becomes the input of the next.” This describes a linear pipeline, not dynamic handoff logic.
Green flag answer: Specific condition-based triggers — for example, “when the ICP qualification agent identifies enterprise complexity signals above a defined threshold, it routes to the enterprise specialist agent and passes the account context, the qualification score, and the specific signals that triggered the handoff.”
Question 6: How Are Guardrails Designed, and Who Defines Them?
Guardrails are the defined boundaries within which agents can reason freely. Well-designed guardrails distinguish between load-bearing constraints — compliance requirements, brand voice, ICP definition — that become hard limits, and habitual preferences — format defaults, channel preferences — that become soft defaults the agent can override when the scenario warrants. The guardrail design process should involve your team, not just the agency.
Red flag answer: “Guardrails are built into the system prompt — we add rules about what the agent should and shouldn’t do.” This describes a static instruction set, not a dynamic guardrail architecture.
Green flag answer: A documented guardrail design workshop, a distinction between hard and soft constraints, and a process for updating guardrails as your strategy evolves without breaking the system.
Question 7: How Does the System Learn From Outcomes Over Time?
This is the question that distinguishes a genuinely non-deterministic system from a sophisticated automation workflow. Automation is stateless — each execution is independent and the system does not learn from outcomes. A true agentic system maintains memory across engagements and feeds outcome signals back into agent calibration. Ask specifically how the system is measurably different at 90 days than at day one.
Red flag answer: “We review performance monthly and update the prompts based on what’s working.” This describes manual optimization, not autonomous learning.
Green flag answer: A documented feedback loop — how outcome signals are captured, how they are fed back into agent training, and evidence from prior engagements showing measurable improvement in agent performance over time.
Question 8: Where Are the Human-in-the-Loop Checkpoints, and Who Decides?
Genuine agentic AI is not fully autonomous — it is designed to know when to involve a human. The human-in-the-loop design should be intentional, not a fallback for when the system fails. Ask which decisions require human approval, how the system surfaces those decisions, and what the handoff to the human looks like in practice.
Red flag answer: “Everything is reviewed by our team before it goes live.” This describes a human-supervised automation workflow, not an agentic system with intelligent human-in-the-loop design.
Green flag answer: Specific decision categories that trigger human review — for example, budget decisions above a defined threshold, content that touches compliance-sensitive topics, or ICP qualification decisions for named accounts — with a documented escalation protocol.
Question 9: Can You Show Me a Case Study From Our Sector With Specific Outcome Data?
Sector-specific case studies with outcome data are the most reliable proxy for genuine methodology. Ask for a case study from your vertical — not a generic AI marketing case study — with specific metrics: pipeline generated, cost per qualified lead, content performance, paid media efficiency. Ask about the edge cases that occurred and how the system handled them.
Red flag answer: A generic case study with percentage improvements but no absolute numbers, no sector context, and no description of how the system adapted to specific challenges.
Green flag answer: A sector-specific case study with absolute outcome data, a description of the agent architecture deployed, specific examples of dynamic handoffs or edge cases handled, and evidence of system improvement over the engagement period.
Question 10: How Do You Separate Deterministic and Non-Deterministic Elements in Your System?
A mature agentic architecture is explicit about which elements are deterministic — fixed rules, compliance constraints, brand guardrails — and which are non-deterministic — agent reasoning, content generation, ICP qualification decisions. The deterministic elements provide stability and compliance; the non-deterministic elements provide adaptability and intelligence. An agency that cannot articulate this distinction is almost certainly running a fully deterministic system.
Red flag answer: “Everything is AI-driven — we don’t use fixed rules.” This is either untrue or describes a system with no guardrails, which is a different kind of problem.
Green flag answer: A clear description of which system components are deterministic (and why), which are non-deterministic (and within what boundaries), and how the two layers interact in practice.
Question 11: What Happens to Our Data, and How Is It Used in Agent Training?
Sector-specific training requires data — your ICP definitions, your content performance history, your CRM signals, your campaign outcomes. Ask explicitly how your data is used in agent training, whether it is kept separate from other clients’ training data, and what happens to it when the engagement ends. This is both a due-diligence question and a data governance question.
Red flag answer: “Your data is used to improve our models for all clients.” This means your sector knowledge and performance data are being used to train agents deployed for your competitors.
Green flag answer: A documented data governance policy, confirmation that your training data is isolated to your agent instances, and a clear data retention and deletion protocol at engagement end.
Question 12: How Do You Measure and Report on Agentic System Performance — Not Just Campaign Performance?
Campaign performance metrics — impressions, clicks, conversions, pipeline — tell you whether the marketing is working. Agentic system performance metrics tell you whether the system is getting smarter. Ask for both. Agentic system metrics include agent accuracy rates over time, handoff trigger frequency and accuracy, edge-case resolution rates, and the delta between day-one performance and current performance on comparable tasks.
Red flag answer: “We report on the standard marketing KPIs — traffic, leads, pipeline.” This tells you nothing about whether the agentic system is functioning as designed.
Green flag answer: A reporting framework that includes both campaign performance metrics and system performance metrics, with a clear explanation of how the two are connected and how system improvements translate into campaign outcomes.
How to Score the Answers
After your discovery call, score each answer on a simple three-point scale: 0 for a red-flag answer, 1 for a partial answer that shows awareness but lacks specificity, and 2 for a green-flag answer with documented evidence. A total score of 20 or above suggests a genuine agentic methodology. A score of 12 to 19 suggests a hybrid approach — some genuine agentic elements alongside automation. A score below 12 suggests you are looking at automation rebranded as agentic AI.
The most important questions are 1 (orchestrator architecture), 3 (sector training), 7 (outcome learning), and 10 (deterministic vs non-deterministic design). These four questions are the hardest to answer convincingly without a genuine methodology, and the easiest to identify as red flags when the answers are vague.
What a Genuine Briefing Process Looks Like
If you are working with an agency that has a genuine agentic methodology, the briefing process will feel different from a standard agency onboarding. It will start with a sector scenario-mapping workshop — a structured session to map the specific scenarios, edge cases, objections, and decision-making patterns that occur in your vertical. It will include an ICP definition workshop that goes deeper than firmographics — into the behavioral signals, the buying triggers, and the compliance constraints that define your ideal customer in your specific sector.
It will include a guardrail design session where you and the agency jointly define the hard limits and soft defaults that govern agent behavior. It will include a human-in-the-loop design session where you define which decisions require your approval and what the escalation protocol looks like. And it will include a baseline measurement session where current system performance is documented so that improvement can be tracked over time.
If the onboarding process skips any of these steps, the system being built is not genuinely agentic — regardless of what the proposal says.
The Bottom Line
The 12 questions in this guide are not designed to be adversarial. They are designed to help you identify the agencies that have done the hard work of building genuine agentic methodology — the orchestration architecture, the sector-specific training, the dynamic handoff logic, the outcome feedback loops — from those that have rebranded automation as intelligence.
The agencies that can answer all 12 questions with specificity and evidence are rare. They are also the ones worth signing with. The investment in a genuine agentic system compounds over time — the system is measurably better at 90 days than at day one, and better still at 180 days. The investment in a sophisticated automation workflow does not compound. It plateaus, and then it breaks when the scenario changes.
Use these questions. Ask for evidence, not explanations. And if the answers are vague, trust your instincts — the vagueness is the answer.
*Modi Elnadi is Founder and Director of Marketing and AI Growth at Integrated.Social, a London-based B2B AI marketing agency specialising in agentic AI systems, AEO, and performance marketing across B2B SaaS, FinTech, Cybersecurity, Enterprise Technology, and professional services. Connect on LinkedIn or explore Integrated.Social’s Agentic AI services.*
You have read the case studies. You have sat through the demo. The agency’s slides are polished, the orchestrator diagram looks impressive, and the word “agentic” appears seventeen times in their proposal. Now you need to decide whether to sign.
The problem is that “agentic AI” has become the most overloaded term in B2B marketing in 2026. Agencies that were running Make.com workflows six months ago have rebranded them as agentic systems. Prompt chains are now “multi-agent architectures.” A single GPT-4 call with a long system prompt is now an “AI agent.” The terminology is being used to justify premium pricing for what is, in practice, deterministic automation.
This guide gives you the 12 questions that cut through the positioning. They are designed to be asked in a discovery call or proposal review. The answers will tell you, with high confidence, whether you are looking at a genuine agentic AI methodology or a sophisticated-sounding automation workflow. Use them before you sign anything.
Why the Briefing Stage Is the Most Important Moment
The briefing stage is where the gap between genuine agentic AI and automation is most visible — if you know what to look for. A true agentic system requires a fundamentally different briefing process than an automation workflow. It needs sector context, scenario mapping, ICP definition, guardrail design, and an understanding of your human-in-the-loop requirements. An automation workflow needs a process map and a trigger list.
If an agency’s briefing questionnaire looks like a project management intake form — timelines, deliverables, channel preferences, budget — you are looking at an automation vendor. If it asks about your sector’s edge cases, your compliance constraints, your ICP’s decision-making hierarchy, and the scenarios where you want human oversight, you are looking at something closer to genuine agentic methodology.
According to Gartner’s 2025 AI Hype Cycle, agentic AI is at the Peak of Inflated Expectations — which means the market is full of vendors claiming capabilities they do not yet have. The 12 questions below are your filter.
The 12 Questions — and What the Answers Reveal
Question 1: Can You Show Me Your Orchestrator Architecture?
A genuine agentic system has a documented orchestrator layer — the decision-making agent that sits above all specialist agents, holds strategic context across the engagement, and decides which agent to invoke next. Ask to see a diagram. Ask how it handles a scenario where two agents produce conflicting outputs. Ask what happens when the orchestrator encounters a situation it has not been trained on.
Red flag answer: “Our orchestrator is the LLM itself — it decides what to do next based on the prompt.” This describes a single-model prompt chain, not an orchestrated multi-agent system.
Green flag answer: A documented architecture diagram showing a distinct orchestrator layer, specialist agents with defined domains, handoff triggers, and a conflict resolution mechanism.
Question 2: How Many Specialist Agents Do You Deploy, and What Is Each One Trained On?
True agentic systems deploy specialist agents — each trained, prompted, and constrained for a specific domain. In a B2B marketing context, you should expect separate agents for ICP qualification, content strategy, paid media signal interpretation, AEO gap analysis, and CRM enrichment at minimum. Each agent should have a documented training scope and sector-specific calibration.
Red flag answer: “We have one main agent that handles everything, with different prompts for different tasks.” This is a single-model system with prompt switching — not specialist agents.
Green flag answer: A named list of specialist agents with documented domains, training data sources, and the specific scenarios each agent has been calibrated to handle in your sector.
Question 3: How Do You Train Agents on Our Sector?
Sector-specific training is the single biggest differentiator between a genuine agentic methodology and a generic tool deployment. The agency should be able to describe a specific process for mapping your sector’s scenarios — the objections your buyers raise, the compliance constraints that govern what an agent can say, the content formats that perform in your vertical, the seasonal patterns in your buying cycle.
Red flag answer: “We give the agent your brand guidelines and a few examples of your content.” This is prompt customization, not sector training.
Green flag answer: A documented scenario-mapping process, evidence of prior work in your sector, and a description of how sector knowledge is embedded at the agent level rather than applied as a wrapper at the output level.
Question 4: How Does the System Handle Edge Cases — Scenarios Outside the Defined Workflow?
This is the question that most cleanly separates automation from genuine agency. Automation workflows break on edge cases — they either fail silently, produce generic output, or require manual intervention. A true agentic system has been trained to recognize edge cases and route them appropriately — either to a different specialist agent, to a human-in-the-loop checkpoint, or to a fallback protocol.
Red flag answer: “We handle edge cases by updating the workflow and adding a new branch.” This describes a brittle deterministic system that requires manual maintenance every time a new scenario emerges.
Green flag answer: A description of the edge-case detection logic, the routing mechanism, and examples of edge cases the system has encountered and handled autonomously in prior engagements.
Question 5: What Is Your Dynamic Handoff Logic, and What Triggers a Handoff?
Dynamic handoff logic is what allows agents to recognize when a scenario exceeds their domain and route to a more appropriate specialist, passing the relevant context. The triggers should be condition-based — signals the agent has been trained to recognize — not fixed workflow transitions. Ask for specific examples of handoff triggers in your sector.
Red flag answer: “Handoffs happen when a task is completed — the output of one step becomes the input of the next.” This describes a linear pipeline, not dynamic handoff logic.
Green flag answer: Specific condition-based triggers — for example, “when the ICP qualification agent identifies enterprise complexity signals above a defined threshold, it routes to the enterprise specialist agent and passes the account context, the qualification score, and the specific signals that triggered the handoff.”
Question 6: How Are Guardrails Designed, and Who Defines Them?
Guardrails are the defined boundaries within which agents can reason freely. Well-designed guardrails distinguish between load-bearing constraints — compliance requirements, brand voice, ICP definition — that become hard limits, and habitual preferences — format defaults, channel preferences — that become soft defaults the agent can override when the scenario warrants. The guardrail design process should involve your team, not just the agency.
Red flag answer: “Guardrails are built into the system prompt — we add rules about what the agent should and shouldn’t do.” This describes a static instruction set, not a dynamic guardrail architecture.
Green flag answer: A documented guardrail design workshop, a distinction between hard and soft constraints, and a process for updating guardrails as your strategy evolves without breaking the system.
Question 7: How Does the System Learn From Outcomes Over Time?
This is the question that distinguishes a genuinely non-deterministic system from a sophisticated automation workflow. Automation is stateless — each execution is independent and the system does not learn from outcomes. A true agentic system maintains memory across engagements and feeds outcome signals back into agent calibration. Ask specifically how the system is measurably different at 90 days than at day one.
Red flag answer: “We review performance monthly and update the prompts based on what’s working.” This describes manual optimization, not autonomous learning.
Green flag answer: A documented feedback loop — how outcome signals are captured, how they are fed back into agent training, and evidence from prior engagements showing measurable improvement in agent performance over time.
Question 8: Where Are the Human-in-the-Loop Checkpoints, and Who Decides?
Genuine agentic AI is not fully autonomous — it is designed to know when to involve a human. The human-in-the-loop design should be intentional, not a fallback for when the system fails. Ask which decisions require human approval, how the system surfaces those decisions, and what the handoff to the human looks like in practice.
Red flag answer: “Everything is reviewed by our team before it goes live.” This describes a human-supervised automation workflow, not an agentic system with intelligent human-in-the-loop design.
Green flag answer: Specific decision categories that trigger human review — for example, budget decisions above a defined threshold, content that touches compliance-sensitive topics, or ICP qualification decisions for named accounts — with a documented escalation protocol.
Question 9: Can You Show Me a Case Study From Our Sector With Specific Outcome Data?
Sector-specific case studies with outcome data are the most reliable proxy for genuine methodology. Ask for a case study from your vertical — not a generic AI marketing case study — with specific metrics: pipeline generated, cost per qualified lead, content performance, paid media efficiency. Ask about the edge cases that occurred and how the system handled them.
Red flag answer: A generic case study with percentage improvements but no absolute numbers, no sector context, and no description of how the system adapted to specific challenges.
Green flag answer: A sector-specific case study with absolute outcome data, a description of the agent architecture deployed, specific examples of dynamic handoffs or edge cases handled, and evidence of system improvement over the engagement period.
Question 10: How Do You Separate Deterministic and Non-Deterministic Elements in Your System?
A mature agentic architecture is explicit about which elements are deterministic — fixed rules, compliance constraints, brand guardrails — and which are non-deterministic — agent reasoning, content generation, ICP qualification decisions. The deterministic elements provide stability and compliance; the non-deterministic elements provide adaptability and intelligence. An agency that cannot articulate this distinction is almost certainly running a fully deterministic system.
Red flag answer: “Everything is AI-driven — we don’t use fixed rules.” This is either untrue or describes a system with no guardrails, which is a different kind of problem.
Green flag answer: A clear description of which system components are deterministic (and why), which are non-deterministic (and within what boundaries), and how the two layers interact in practice.
Question 11: What Happens to Our Data, and How Is It Used in Agent Training?
Sector-specific training requires data — your ICP definitions, your content performance history, your CRM signals, your campaign outcomes. Ask explicitly how your data is used in agent training, whether it is kept separate from other clients’ training data, and what happens to it when the engagement ends. This is both a due-diligence question and a data governance question.
Red flag answer: “Your data is used to improve our models for all clients.” This means your sector knowledge and performance data are being used to train agents deployed for your competitors.
Green flag answer: A documented data governance policy, confirmation that your training data is isolated to your agent instances, and a clear data retention and deletion protocol at engagement end.
Question 12: How Do You Measure and Report on Agentic System Performance — Not Just Campaign Performance?
Campaign performance metrics — impressions, clicks, conversions, pipeline — tell you whether the marketing is working. Agentic system performance metrics tell you whether the system is getting smarter. Ask for both. Agentic system metrics include agent accuracy rates over time, handoff trigger frequency and accuracy, edge-case resolution rates, and the delta between day-one performance and current performance on comparable tasks.
Red flag answer: “We report on the standard marketing KPIs — traffic, leads, pipeline.” This tells you nothing about whether the agentic system is functioning as designed.
Green flag answer: A reporting framework that includes both campaign performance metrics and system performance metrics, with a clear explanation of how the two are connected and how system improvements translate into campaign outcomes.
How to Score the Answers
After your discovery call, score each answer on a simple three-point scale: 0 for a red-flag answer, 1 for a partial answer that shows awareness but lacks specificity, and 2 for a green-flag answer with documented evidence. A total score of 20 or above suggests a genuine agentic methodology. A score of 12 to 19 suggests a hybrid approach — some genuine agentic elements alongside automation. A score below 12 suggests you are looking at automation rebranded as agentic AI.
The most important questions are 1 (orchestrator architecture), 3 (sector training), 7 (outcome learning), and 10 (deterministic vs non-deterministic design). These four questions are the hardest to answer convincingly without a genuine methodology, and the easiest to identify as red flags when the answers are vague.
What a Genuine Briefing Process Looks Like
If you are working with an agency that has a genuine agentic methodology, the briefing process will feel different from a standard agency onboarding. It will start with a sector scenario-mapping workshop — a structured session to map the specific scenarios, edge cases, objections, and decision-making patterns that occur in your vertical. It will include an ICP definition workshop that goes deeper than firmographics — into the behavioral signals, the buying triggers, and the compliance constraints that define your ideal customer in your specific sector.
It will include a guardrail design session where you and the agency jointly define the hard limits and soft defaults that govern agent behavior. It will include a human-in-the-loop design session where you define which decisions require your approval and what the escalation protocol looks like. And it will include a baseline measurement session where current system performance is documented so that improvement can be tracked over time.
If the onboarding process skips any of these steps, the system being built is not genuinely agentic — regardless of what the proposal says.
The Bottom Line
The 12 questions in this guide are not designed to be adversarial. They are designed to help you identify the agencies that have done the hard work of building genuine agentic methodology — the orchestration architecture, the sector-specific training, the dynamic handoff logic, the outcome feedback loops — from those that have rebranded automation as intelligence.
The agencies that can answer all 12 questions with specificity and evidence are rare. They are also the ones worth signing with. The investment in a genuine agentic system compounds over time — the system is measurably better at 90 days than at day one, and better still at 180 days. The investment in a sophisticated automation workflow does not compound. It plateaus, and then it breaks when the scenario changes.
Use these questions. Ask for evidence, not explanations. And if the answers are vague, trust your instincts — the vagueness is the answer.
*Modi Elnadi is Founder and Director of Marketing and AI Growth at Integrated.Social, a London-based B2B AI marketing agency specialising in agentic AI systems, AEO, and performance marketing across B2B SaaS, FinTech, Cybersecurity, Enterprise Technology, and professional services. Connect on LinkedIn or explore Integrated.Social’s Agentic AI services.*
