Before Building AI Agents: Read This Marketing Framework First
Before you build AI agents for marketing.
The full framework I teach at Iqra University. What AI actually is, where it breaks, why entire industries are collapsing in months instead of years, and the psychological trap nobody is warning you about. Full slide deck attached.
Most marketers are being told to “start using AI” before anyone slows down to explain what AI actually is, where it breaks, or how a real AI product gets built. Skip the framework and you will spend six months in tool fatigue without a single workflow that survives a Monday morning.
This guide is the theory layer underneath every practical AI marketing workflow I publish. It is also the written version of the guest lecture I delivered at Iqra University, and the full slide deck is available to download further down the page. If you sat through that session and wanted the longer version, this is it. If you missed it, start here.
I am going to cover the parts most “AI for marketers” content skips. The vocabulary that has to land before anything else. The six limitations that decide whether AI helps or hurts your team. How industries are collapsing on month-long timelines instead of decade-long ones. The psychology of the “do it for me” shift and the handicap it is creating in your customers. And the BUILD framework that explains why one company’s AI feels like magic and the next one’s feels like a chatbot in a trench coat.
01What Marketers Are Telling Me Right Now
The honest version.
Before I lay out the framework, I want to share what motivated building it.
Over the last year I have had conversations with marketers across in house teams, agencies, and freelance practices. The pattern in those conversations is what shaped both the lecture and this guide. Here is what they keep telling me, in their own words.
Leadership keeps saying just use AI without telling me what for. I have eleven tabs open and no workflow.
We tried ChatGPT for product descriptions. It invented a feature we do not have. The legal email landed the next morning.
I cannot tell the difference between an AI tool, an AI feature, and an AI agent. Every vendor is using all three words for the same thing.
I keep getting told agentic AI will change everything. Nobody tells me what it actually does in a marketing department on a Tuesday.
I am watching my SaaS lose traffic to ChatGPT and nobody on my team has a plan for it.
Tool fatigue is real. Every week there is a new must have. Half of them break in a month.
The pattern across every conversation: the problem is rarely the tool. It is the absence of a mental model. Without a framework, every AI decision is a gut call. You pick the tool that trended on LinkedIn that week, you build a workflow around it, and three months later you are rebuilding because the tool changed or the use case never fit in the first place.
02So Wait. What Is AI, Actually?
Ninety seconds of vocabulary that has to land first.
Every conversation about AI strategy fails when four words mean different things to different people in the room. Before any framework lands, the vocabulary has to be settled. This is the version I open every workshop with.
There is also one bigger question that hangs over the whole conversation, and it changes how you should approach the rest of this guide. Is AI an invention or a discovery? An invention is built for a job — like a hammer, you know what it does. A discovery is found and then explored — like electricity, we are still finding uses for it a hundred and fifty years later. AI is closer to the second. Which means your job, even as a marketer, is partly to explore. Not just to deploy.
Model
Math that learned a pattern. A model is the underlying mathematical object that recognises something. Show it ten thousand photos of cats and it learns to recognise a cat in the eleven thousandth. The model itself is not a chatbot, not an app, not a product. It is the engine.
Large Language Model (LLM)
A model that learned words instead of images. ChatGPT, Claude, and Gemini are all powered by LLMs. They predict the next word in a sequence based on patterns from massive amounts of text. That is all they do at the core. Everything else you see is scaffolding around that prediction.
Generative AI
AI that creates rather than just classifies. An older model could look at a photo and tell you it was a cat. A generative model can produce a new photo of a cat that has never existed. Midjourney for images, Runway for video, ElevenLabs for voice. The output is new content, not a label.
Agentic AI
AI that acts, not just answers. An LLM responds to a prompt. An agent pursues a goal. The cleanest definition I have come across is from Dharmesh Shah, Co-Founder and CTO of HubSpot:
Software that uses AI to pursue a goal, autonomously or assisted, by breaking it into tasks, monitoring progress, and using tools and other agents to get it done.
The difference is structural. A chatbot waits to be asked. An agent decides what to do next. Most “AI agents” on the market today are LLMs with a few tools bolted on and a loop. That is not a flaw. It is the current state. Knowing the distinction protects you from buying agentic positioning that is really just a chatbot.
The Tool Picker Problem
The other vocabulary issue is tool selection. AI is not one thing. It is a toolbox. Picking the right one matters more than picking the most powerful one. Here are the categories every marketer should have a mental map of.
Claude
Best for long writing, structured reasoning, and tasks where tone and nuance carry weight. The default for ghostwriting, briefs, content strategy, anything where the output will be read closely by a human who will notice if it sounds generic.
ChatGPT
Strong general use. Reasoning, code, analysis, multimodal tasks. The Swiss army knife. Best when you do not know which model is best — you can usually get a workable answer here even if it is not the optimal tool.
Gemini
Tight integration with Google Workspace. Strong image and video understanding. Best when your workflow already lives in Google Docs, Sheets, and Drive, or when you need to reason over visual content at scale.
Perplexity
Built for research with citations. When you need current information and you want to verify sources, this is the default. Use it for competitive research, market scans, and any task where freshness and citations matter more than long-form output.
Midjourney
Editorial-grade image generation. Best for brand visuals, mood boards, social, anything that needs to feel intentional rather than generic. Slower to learn than other image tools, but the ceiling is higher.
Runway
The current leader for turning images into video. For ads, social motion, and short-form content where production cost used to make experimentation impossible.
03The Six Limitations Every Marketer Must Know
Honest framing before you build anything.
AI is powerful. AI is limited. Both are true. The marketers who get hurt are the ones who internalise the first sentence and ignore the second. Here are the six limitations that decide whether AI helps your team or quietly damages your brand.
It knows its training data. Not today.
An LLM’s knowledge is frozen at whatever date its training stopped. If you ask it about your competitor’s launch from last week, it either does not know or, worse, confidently makes something up. Without an explicit fix, every output is at risk of being stale.
It forgets when the page fills.
Models have a context window. Once you exceed it, earlier instructions and details fall off. This is why a long ChatGPT thread eventually starts to drift, contradict itself, or forget your original brief. Memory is not magic. It is a buffer with a hard ceiling.
Weak at multi-step logic.
Models pattern match. They do not reason the way a human reasons. Give it a problem with three logical steps and it will often skip one or substitute a similar-looking answer. The fluency hides the gap. The output sounds correct, which is exactly what makes it dangerous when it is wrong.
It does not get smarter from you.
Every new chat starts from scratch. The model is not learning your brand voice, your customer base, or your past wins. Unless you explicitly persist that context — through memory features, fine-tuning, or a system prompt — every Monday morning is day one.
It has no goals. Only responses.
An LLM does not want anything. It does not know what success looks like. It responds to whatever you put in front of it, optimised for sounding helpful. This is the single biggest reason agentic AI matters: without an explicit goal structure on top, you have a very fast responder, not an operator.
It needs tools to touch the world.
An LLM by itself cannot send an email, post to social, query your CRM, or update a spreadsheet. It can only produce text. To do anything in the real world, it needs tools — APIs, plugins, integrations. Anyone selling you a “fully autonomous AI marketer” without explaining the tool layer is selling positioning, not capability.
04Hallucination. The One That Gets People Fired.
I am pulling hallucination out of the limitation list because it deserves its own section. Every other limitation is annoying. Hallucination is the one that ends careers.
A hallucination is when the model produces something that sounds correct, is delivered with full confidence, and is factually wrong. Not a typo. Not a misunderstanding. A fabricated fact, statistic, source, or feature, presented with the same fluency as the true ones around it.
When Hallucination Is Acceptable
For brainstorming, captions, mood boards, first drafts, and anything where variation is the feature, hallucination is not really a problem. You are using the model as a divergent thinking partner. The fact that some of its output is invented is sometimes the point. You filter it through human judgment before it ships.
When Hallucination Is Fatal
Pricing. Product specs. Compliance claims. Health information. Legal language. Financial guidance. Anything attributed to a real person or source. In any of these categories, a confidently wrong output is not an inconvenience. It is a lawsuit, a regulatory action, or a public retraction. The model does not know the difference between an okay-to-be-wrong context and a never-be-wrong context. You do. That is the job.
05The Fix Layer. Every Limitation Has a Workaround.
This is the section most “limitations of AI” articles never get to. Every limitation above has an engineering workaround, and understanding them is what separates a marketer who uses AI from a marketer who builds with it. These are the building blocks of any serious AI product.
Retrieval Augmented Generation (RAG) and web search.
If the model does not know about your data or about today, give it access. RAG lets the model search a custom knowledge base before answering. Web search lets it pull live information. Perplexity is a consumer example. Inside a marketing workflow, this means the model can reference your brand guidelines, customer reviews, or your latest product launch instead of guessing.
Short-term and long-term memory layers.
Modern AI products separate session memory (this conversation) from persistent memory (everything I know about this user or brand). Persistent memory survives across sessions. This is how ChatGPT now remembers your preferences and how custom assistants stay in character across weeks.
Chain-of-thought prompting and reasoning models.
Asking the model to think step by step before answering measurably improves multi-step accuracy. A newer category of reasoning models is built around this internally — they take longer but they break the problem down before responding. For any marketing task involving comparison, scoring, or decisioning, this is the gear shift that matters.
Feedback loops and fine-tuning.
If you want the model to learn your brand voice over time, you have two paths. Lightweight: build feedback into your workflow so the model sees what worked and what did not. Heavyweight: fine-tune a model on your own writing or examples. Both teach the system about you — they just sit at different cost levels.
Agents and explicit goal structures.
This is what people actually mean when they say agentic AI. You give the system a goal, a set of tools, and a loop. The model breaks the goal into tasks, calls tools to make progress, and monitors whether it is getting closer to done. Purpose is no longer missing — it is encoded in the agent’s structure.
APIs, plugins, and the internet.
To touch the world, the model needs hands. APIs to your CRM, your email tool, your CMS, your analytics. Plugins to read documents, browse the web, or generate images. The interesting modern AI products are mostly thin LLM layers wrapped around rich tool integrations. The intelligence is the LLM. The leverage is the tools.
06The PMF Treadmill. Why This Matters More Than Your Tool Choice.
Product market fit used to be a milestone. Now it is a moving target.
The framework is one thing. The market context is another. The reason every marketer needs at least a working grasp of AI is not because every team needs an agent next quarter. It is because product market fit has stopped behaving the way it used to.
Product market fit used to be a milestone. You found it, you defended it, you scaled it. Each major technology shift compressed the window. The internet shift moved that milestone over years. Mobile compressed it. AI is compressing it again into months. The visual below makes the scale clear.
Brands That Got Caught Out
This is not theoretical. Two of the most visible casualties of the AI shift dropped harder, faster, than anyone modelling traditional disruption would have predicted. The pattern is the same in both cases: a category leader, a strong defensible business, and a twelve-month window from “fine” to “structurally damaged.”
What used to feel like a slow erosion is now a cliff. There is no twelve-month grace period for an incumbent caught flat-footed. The window slams shut. And the visible damage is the part you see — the invisible part is the customers who quietly left and never came back.
It Is Already Happening In Fashion
If you think your industry is too creative, too tactile, or too brand-driven to be exposed, fashion was telling itself the same story two years ago. Here is where the most aesthetic, taste-driven industry in the world is right now.
The point is not that AI is doing fashion now. The point is that if AI is doing fashion now, your category is not protected by being creative either. The question is not whether AI reshapes your space. The question is whether you reshape with it, or get reshaped by it.
07The Four Customer Expectation Shifts
What your customers now expect from every brand they touch.
Industry shifts are abstract. Customer expectation shifts are personal. This is the part most marketers are already being judged on, whether they know it or not. Customer expectations have moved across four axes simultaneously. If you are still building for the old version of these expectations, you are building for last year’s customer.
Each of these is a customer telling you something specific. Together they say: I do not want to do the work anymore. I want the result. Every brand that figures out how to deliver the result without making the customer assemble it is winning. Every brand still selling the assembly is being slowly disqualified.
08The AI Handicap. The Psychology Trap Nobody Is Warning You About.
The shift that is reshaping your customer is also reshaping you.
Here is the part of the AI conversation that almost nobody writes about. Every shift above is real and worth building for. But underneath them is a psychological transformation that cuts in two directions, and ignoring half of it will hurt your career.
The customer expectation shift is also a customer capability shift. When you spend two years pressing one button to get a result, you slowly lose the ability — and crucially, the willingness — to assemble the result yourself. Your customers are not just demanding “do it for me.” They are losing the muscle to do it any other way. This is the handicap.
A customer who has used AI for product research for eighteen months has lost the skill of researching products without AI. They cannot compare options unless something compares for them. They cannot summarise a long document unless something summarises it. This is not laziness. It is atrophy. Muscles you do not use, you lose. Cognitive muscles included.
For marketers, this changes the brief. Your audience now expects synthesis, not raw information. They will not read a comparison table. They will read a recommendation. They will not assemble a checklist. They will follow a workflow. The pre-AI customer wanted information. The post-AI customer wants decisions.
Here is the harder one. Marketers are getting handicapped too — and faster than their customers. Every time you let AI write a first draft you did not need help with, you weaken your own writing muscle. Every time you let it research something you could have researched, you lose the rep. Over months, marketers who outsource everything to AI find they cannot strategise without it. They cannot find an angle without it. They cannot write a brief without it.
This is the part nobody flags. The marketers who will be most valuable in three years are the ones who used AI without becoming dependent on it. Who kept their judgment sharp by exercising it. Who used AI as a force multiplier, not as a prosthesis.
The third trap is the one that gets people fired. AI delivers everything with the same confident, fluent voice. A correct fact and a fabricated one read identically. Marketers who lose their critical reading muscle — because the AI’s output sounds so right — start shipping AI confidence as their own judgment. The first time that lands a wrong number in a board deck or a fabricated quote in a press release, the conversation ends quickly.
Build a deliberate practice of reading AI output skeptically. Out loud if you have to. The fluency is the threat, not the asset.
How to Stay on the Right Side of the Handicap
Three habits I see in the marketers who are using AI well without becoming captured by it.
- Draft before you prompt. On anything important, write your own first draft — even three sentences — before opening AI. This keeps your synthesis muscle active. AI improves your draft instead of replacing your thinking.
- Disagree with the model in writing. When AI gives you an answer, write down where you think it is wrong before accepting it. Even when you are right to accept it, the friction protects your judgment. Yes-clicking is how the handicap forms.
- Spend one day a week unplugged from AI. Not as a wellness move. As a calibration move. The day you cannot work without AI is the day you find out how much of “your” work was actually yours.
09The BUILD Framework
How AI products actually get built. And why most teams stop at the bottom.
This is the framework that ties everything above together. BUILD is a five-layer model for how AI products are constructed. It is also a diagnostic — you can use it to figure out which layer your team is currently operating at, and which layer the brands eating your lunch are operating at.
Raw AI capability.
The model itself, used directly. ChatGPT writing a caption. Claude drafting a brief. No knowledge of your brand, no memory, no tools. Most teams stop here and then complain that AI output is generic. Generic in, generic out. The base layer is where everyone starts. It is not where anyone wins.
Add knowledge, tools, and memory.
The upgrade layer is where the fixes from Section 05 get added. The model now has access to your brand voice document, your customer data, your CRM, your CMS. It can remember context across sessions. This is the layer where AI output starts to feel specifically yours, not just AI.
Learn from feedback.
The system now tracks what worked. Click-through rates, conversions, engagement. Outputs get rated. Over time, the system gets sharper because it is seeing what your audience actually responds to, not just what looks good in a draft.
Goals and autonomy.
This is the first agentic layer. The system is given a goal — “launch a campaign for product X to segment Y by date Z” — and plans the steps itself. It decides what content to produce, what channels to use, when to deploy. The human supervises the goal. The system owns the plan.
Multi-agent teams.
The top layer. A team of specialised agents working in coordination. A research agent profiles the audience. A content agent produces creative. An ads agent deploys and optimises. A reporting agent surfaces what is working. They communicate, they hand off, they specialise. The human is now operating at the executive level — setting direction, reviewing outputs, intervening when needed.
10Where Are You? The Maturity Ladder.
The same five-layer logic, applied to your team’s day-to-day state. Be honest with yourself reading this. The point is not to look advanced. It is to see clearly where you actually sit.
Most teams I see are at Level 2. They feel busy with AI but cannot point to a single repeatable workflow. The jump from Level 2 to Level 4 is the highest-leverage move available in marketing right now, and it is almost entirely about framework, not about tools.
11What Is In It For The Marketer Who Adapts
I have spent ten sections being measured. This section is the part where I am willing to be enthusiastic, because the upside is genuinely big for the marketers who do this right.
The marketer who climbs from BASE to UPGRADE becomes ten times more productive on the production side of their job. Captions, briefs, first drafts, research summaries — work that used to take a day takes an hour. That hour-saved-per-task compounds over months into entire days a week.
The marketer who climbs from UPGRADE to IMPROVE becomes the one in the room with actual data on what their audience responds to, week by week, instead of guessing. They stop arguing about taste and start arguing about evidence. That changes their seniority.
The marketer who climbs from IMPROVE to LEAD stops doing production work at all. They are now setting direction for a system that produces, and reviewing the output. Their job description has changed from “do the marketing” to “design the marketing operation.” That is a different career.
The marketer who climbs from LEAD to DELEGATE is running the marketing output of a company several times their team’s size. They are coordinating agents, not channels. They are operating at a level of leverage that did not exist three years ago.
12Get the Full Slide Deck
The lecture this guide is based on was delivered at Iqra University as a guest workshop on AI integration for marketers. The slides cover the same framework with visuals, the workshop pauses I used with the audience, and additional examples that did not make it into the written version. If you want the deck for personal reference, for a team discussion, or to teach the framework yourself, it is available below as a PDF.
No email gate. The framework is more useful in the world than in a spreadsheet of leads. If you adapt the deck for your own teaching, attribution is appreciated but not required for internal team training.
13Honest Limitations of This Framework
Most “frameworks for AI” content skips this section. I think it is the most important one.
It will date faster than my other writing
The vocabulary section is stable. The limitations section is mostly stable. The BUILD framework will hold for a few years. But specific tool recommendations move quickly. Treat the tool picker as a snapshot, not a permanent guide. Re-evaluate every six months.
It does not replace strategic judgment
A framework helps you make better decisions. It does not make decisions for you. Knowing that your team is at BASE does not tell you whether to climb to UPGRADE next quarter, or whether your business even needs to. That call is yours. The framework just makes the call clearer.
Implementation is harder than understanding
Reading this guide is the easy part. Building the UPGRADE layer for your team — actually connecting your brand voice document, customer data, and tooling — is real work. The framework is the map. The work is the terrain. Do not confuse the two.
Some of this will be wrong
I am building this thinking in public while the technology is moving. Specific claims about model capabilities, market shifts, and tool recommendations will age unevenly. The core structure — vocabulary, limitations, fixes, BUILD layers, the handicap psychology — should outlast the details.