AI made marketing output cheap. Customer understanding is still expensive.

We automated the visible work first.

A marketer can now turn one brief into thirty headlines, six landing-page variants, a nurture sequence and enough LinkedIn posts to make everyone involved reconsider the internet. The output arrives quickly. It is tidy. It is often perfectly acceptable.

That is precisely the problem. Acceptable marketing is becoming abundant. Every team can make more of it, faster. The advantage does not move to the team with the most output. It moves upstream, to the team that understands something true about the customer that competitors have missed.

The visible half of marketing got cheap: reading, summarising, drafting and adapting. The less visible half did not: noticing what a buyer is protecting, hearing the hesitation behind an objection, distinguishing a stated preference from the reason a deal actually stalled.

When output becomes abundant, context becomes the constraint.

Customer data is not customer understanding.

A CRM record can tell you that the buyer is a VP of Marketing at a 500-person company. It cannot tell you that she inherited three disconnected tools, promised the board an efficiency programme and is quietly worried that another platform purchase will make her look careless.

The first description is useful for targeting. The second is useful for persuasion. AI systems are usually fed the first and expected to invent the second. They do. That invention is then called personalisation.

A persona is not automatically better. Most personas are polished containers for old assumptions. They describe a fictional average buyer while the real customer keeps changing. A revenue system needs evidence that is recent, attributable and connected to actual behaviour.

• What customers say when nobody is pitching to them.
• What they tried before they contacted you.
• Which risks slow the decision down.
• Which words appear in won, lost and retained accounts.
• What changed between first interest and final approval.

Build a customer truth stack, not a bigger prompt.

The answer is not to paste more documents into a model and hope that volume becomes insight. Context needs a supply chain. It should move from raw evidence to an explicit interpretation, into the system that uses it, and back again when the market proves the interpretation wrong.

We call this a customer truth stack. Each layer has a different job. Mixing them creates confident mush: a transcript becomes a conclusion, a conclusion becomes a fact, and six months later the system is still writing for a buyer who no longer exists.

Your teams share a customer. Their evidence rarely meets.

Marketing sees campaign response. Sales hears the objection. Customer service sees the consequence after the promise meets reality. Each team owns a different fragment of the same customer, stored in a different tool and interpreted for a different meeting.

This is why adding AI to each department separately can make the organisation less coherent. Marketing's system learns from engagement. Sales' system learns from pipeline. Service's system learns from resolution. All three become more efficient at acting on partial truth.

The commercial opportunity is not merely to give each team an assistant. It is to create a shared context layer that lets customer evidence travel. A lost-deal pattern should change the next campaign brief. A recurring support issue should change the promise sales makes. A high-performing message should be tested against retention, not only clicks.

Audit the truth before you automate the voice.

Choose one recurring piece of revenue work: a campaign brief, an account plan, a renewal summary. Open the context currently used to produce it. For every important claim about the customer, ask where it came from, when it was last observed and what would cause you to change your mind.

If the answer is a slide deck, an intuition or ‘we have always known this’, do not add another prompt. Repair the context loop first. The quality ceiling of an AI system is set long before the model starts writing.

The teams that win with AI will not be the ones that prompt more. They will be the ones that know more, and know why they know it.