← Back to journal

Your AI Knows Everything About Your Customers. That's Why It Can't Sell to Them.

Dream Outcome · JournalFig. YOUR-A

Your AI Knows Everything About Your Customers. That's Why It Can't Sell to Them.

A luxury handbag store on 63rd Street in Manhattan has a sales method that makes most marketing technology look like an expensive joke.

When you visit Goyard, a sales rep enters your details into a handheld device and lets you hold whichever bag catches your eye. You leave without buying. Then, at 8pm on Friday, a text arrives with a photo of you holding that bag. You are on your second cocktail with friends. The bag suddenly feels affordable. You buy it Saturday morning.

The rep knew one thing about you: you would be out drinking cocktails on Friday at 8pm. Not your browsing history. Not your lifetime value score. Not your purchase propensity across 47 behavioural segments. One piece of context, delivered at the exact right moment.

Meanwhile, Australian businesses are spending an average of $150,000 per year on marketing automation tools to collect more customer data than ever before. The Australia CRM market is growing at 13% annually. And the segment growing fastest? Personalisation services, at 18.5% CAGR.

The industry's answer to every marketing challenge is the same: collect more data, feed it to AI, personalise harder. The evidence says this approach is breaking.

The Data Graveyard You're Paying to Maintain

Here is the uncomfortable truth about all that customer data you have been collecting.

60 to 73% of enterprise data is never analysed, never acted upon, and never monetised. Industry researchers call it "dark data." For most businesses, the majority of their CRM sits there doing nothing except decaying. B2B contact data loses 22.5% of its accuracy every year. Nearly one quarter of your database went stale while you were reading this paragraph.

The cost is not abstract. Gartner estimates poor data quality costs the average organisation $12.9 million per year. Sales teams waste over 500 hours annually pursuing invalid or outdated leads. And 44% of companies report annual revenue losses exceeding 10% tied directly to CRM data issues.

The response from every MarTech vendor? Buy more tools to collect more data to feed more AI.

That is like responding to a flooded basement by turning on more taps.

What businesses are doingWhat actually happens
Collecting more customer data points60-73% of data is never used
Investing in hyper-personalisation53% of customers report negative experiences
Expanding CRM capabilities76% of users say less than half their CRM data is accurate
Adding AI to process more dataAI accuracy drops from 94% to 61% as context increases

Even AI Gets Worse When You Give It More

If you thought AI would solve the data overload problem by simply processing more of it, the engineering research has bad news.

A 2025 study published at EMNLP found that AI accuracy drops between 13.9% and 85% as input length increases, even when the model can perfectly retrieve all relevant information. The sheer volume of context degrades performance. Not because the AI cannot find the right answer. Because the noise drowns out the signal.

This is not a minor technical quirk. Eleven frontier AI models dropped below 50% of their short-context baseline by just 32,000 tokens. Models advertising 200,000 token context windows become unreliable around 130,000 in practice. And pure noise, context with no useful information at all, performed worse than giving the AI nothing: 8.0% accuracy versus 23.6%.

Read that again. Giving AI bad context is worse than giving it no context at all.

The smartest engineers in AI have arrived at the same conclusion the Goyard sales rep already knew. Andrej Karpathy, formerly of OpenAI and Tesla, coined the term that now defines the field: "Context engineering is the delicate art and science of filling the context window with just the right information for the next step." Shopify's CEO Tobi Lutke agreed. Gartner declared in mid-2025: "Context engineering is in. Prompt engineering is out."

The core skill is not feeding AI more data. It is curating what goes in and what stays out. Anthropic, the company behind Claude, published research showing that a well-engineered context window can raise AI agent task completion from roughly 30% to 90% on the same underlying model. Same AI. Same capability. The only difference: less noise, more signal.

We have written before about why giving AI your business context matters more than better prompts. This takes that argument further. It is not just about giving AI the right instructions. It is about giving it the right information and nothing else.

Marketing Science Already Knew This

The AI engineering world is rediscovering a principle that marketing scientists documented decades ago. Byron Sharp and the Ehrenberg-Bass Institute have spent 40 years studying how brands actually grow. Their research on Category Entry Points tells us something crucial about what customers actually need to hear.

Category Entry Points are the specific needs, occasions, or situations that cause a buyer to think about a product category. "I need a plumber" is a CEP. "My hot water stopped working on a Sunday morning" is a more specific one. Sharp's research across 17 product categories found that brands grow by being linked to more of these entry points, not by knowing more about each individual customer.

The research identifies roughly 20 important CEPs for most categories. But here is the key finding: the first brand a buyer thinks of is the one they buy 70% of the time in B2B decisions, according to Professor John Dawes at Ehrenberg-Bass. The battle is not won by knowing everything about the customer. It is won by being present in the right moment with the right association.

This is exactly what the Goyard rep did. She identified the one Category Entry Point that mattered: the moment when luxury feels justified. Friday cocktails. Second drink. Friends around. That is the entry point. Everything else is noise.

Robert Cialdini's work on Pre-Suasion formalises this into a principle. In his research, Cialdini demonstrated that the moment before a message is delivered determines whether the message works at all. He calls these "privileged moments." The same advertisement, the same offer, the same product recommendation can fail completely or succeed brilliantly depending on what preceded it. Not depending on how much data you have about the customer. Depending on when and how you deliver the message.

The warm beverage priming study is one of Cialdini's most cited examples. Participants who were holding a warm drink rated a stranger as significantly warmer and more trustworthy than participants holding a cold drink. Same stranger. Same interaction. Different context. Different outcome.

Your CRM has 200 data points about each customer. Your AI is processing all of them. And the thing that actually determines whether the customer buys is whether you caught them at the right moment with the right frame.

The Personalisation Paradox

If more data led to better marketing, hyper-personalisation would be the answer to everything. The data says otherwise.

Gartner surveyed 1,464 buyers in 2025 and found that 53% of customers reported negative experiences from personalised marketing. Those customers were 3.2 times more likely to regret a purchase. The industry is hitting what researchers call the "personalisation plateau," where additional data and targeting precision produce minimal improvement while complexity and creepiness continue to rise.

Meanwhile, a study of 43,000 participants found that time-personalised communications achieved a 58% higher response rate than non-time-personalised messages. Not content-personalised. Not behaviour-personalised. Time-personalised. Knowing when to send the message, not what to put in it.

This maps directly to what Rory Sutherland has argued for years. His central thesis is that we do not value things, we value their meaning. And meaning is determined entirely by context.

His favourite example: Uber's killer feature is not faster cars. It is the map. The map does not reduce waiting time. It reduces the psychological pain of uncertainty. "Waiting seven minutes for a train with a countdown clock is less frustrating than waiting four minutes thinking 'when is this damn train going to arrive?'" The intervention that made the London Underground's biggest improvement in passenger satisfaction was not faster trains. It was dot-matrix countdown boards. Same wait. Different context. Completely different experience.

Sutherland's principle applied to marketing AI: "If you change the context of something to a tiny degree, you can completely change its meaning, the emotional response it generates, and therefore, the sales it generates."

The business that knows everything about its customer but delivers the message in the wrong moment loses to the business that knows one thing and delivers it perfectly. We explored a related dynamic in how your ads cost more when nobody trusts you yet. Trust is not built by data volume. It is built in specific moments of relevance.

The Curation Framework: What to Keep, What to Cut

So what does context curation look like in practice? The AI engineering world has converged on four strategies that translate directly to marketing.

StrategyAI engineering termMarketing application
WritePre-build essential contextDefine your brand's Category Entry Points before launching campaigns
SelectRetrieve only what is relevant to this taskFor each campaign, identify the 1-2 data points that determine conversion
CompressSummarise, do not dumpReduce customer profiles to decisive signals, not comprehensive histories
IsolateKeep different tasks in separate contextsDo not mix acquisition data with retention data in the same AI workflow

For an SME running Google Ads, this means something specific. Instead of feeding your AI every piece of customer data you have, ask: what is the one thing about this customer's situation that determines whether they convert?

For a plumber, it is not the customer's age, income, or browsing history. It is that their hot water stopped working this morning and they need someone who can come today. That is the CEP. That is the privileged moment. That is the context your ad needs to match.

For a B2B equipment supplier, it is not the prospect's company size or industry vertical. It is that their current supplier just missed a deadline and they are looking for a backup. That is the moment. Everything else is noise.

The Goyard rep did not build a 360-degree customer view. She asked one question: "What are you doing Friday night?" Then she used that single data point to deliver the perfect message at the perfect time. As we have written about before, the businesses that win are not necessarily better. They are available at the moment that matters.

What the C-Suite Already Suspects

Here is perhaps the most telling statistic from all this research. A Korn Ferry survey of 1,600 executives found that 71% of C-suite leaders say having too much data causes them to fall back on gut instinct instead of evidence-based decisions.

The people running businesses already feel this. They have more dashboards than ever and less clarity than ever. And 36% report that increased data availability has made their strategic decision-making slower, not faster.

The irony is sharp. We built all this technology to avoid gut decisions. Then we collected so much data that gut decisions are all we have time for. The data-to-action gap is not a technology problem. It is a curation problem.

What This Means for Your Business

Stop asking "what else can we learn about our customers?" Start asking "what is the one thing we need to know right now?"

For your AI tools: Less input, better output. If you are using AI for ad copy, customer segmentation, or email sequences, test what happens when you give it fewer, more decisive data points instead of the full customer profile. The engineering research is unambiguous: accuracy improves when you cut the noise. For your campaigns: Identify your three to five Category Entry Points. These are the specific situations that trigger someone to need your service. Build campaigns around those moments, not around demographic profiles. A plumber who shows up for "emergency hot water repair Sunday morning" beats a plumber who shows up for "males 35-54 homeowners in Adelaide." For your CRM: Audit what you are actually using. If 60% of your data is dark data, you are paying to store information that is actively degrading your AI's performance. Clean it. Cut it. Keep the signals. Lose the noise.

The luxury handbag store on 63rd Street understood something that the entire MarTech industry is still catching up to. The goal was never to know everything about the customer. The goal was to know the one thing that makes them buy.

Your AI does not need more data. It needs better curation.

Further Reading


Dream Outcome is an Australian digital marketing agency helping SMEs grow through Google Ads, Facebook Ads, and Email Marketing.
Book a free call

Ready to grow profitably?

Get a free digital marketing plan tailored to your business. No lock-in, no long commitments.

Book my free call  →