AI Marketing Insights
AI is good at language.It is bad at being responsible.
That single distinction explains almost every AI marketing project that works and almost every one that embarrasses the business that built it. Here is where we use it, where we refuse to, and how to tell the difference before you have spent the money.
The useful frame
Ask what the task actually requires
Most AI decisions get made by asking whether a tool is impressive. The better question is what the task genuinely demands, because that determines whether a language model is the right instrument at all.
-
Reading unstructured language
An enquiry, a voicemail transcript, a chat thread, a review. Extracting meaning from text a human wrote in their own words is exactly what these models are built for, and they are very good at it. This is where the return is.
-
Moving structured data
Copying a phone number from a form into a CRM field does not require intelligence. It requires reliability. Plain automation does this correctly every time for a fraction of the cost, and putting a model in the middle only introduces a new way for it to go wrong.
-
Being accountable for a decision
Quoting a price, making a promise, handling a complaint, interpreting a contract. A model cannot be held responsible for these, cannot be disciplined for getting them wrong, and cannot read the room. The liability stays with your business either way.
Where it earns its place
Five jobs AI does better than the current process
These have one thing in common. Each involves reading language, and each has a human either supervising the output or picking it up immediately afterwards.
- Routing — reading what someone actually wrote and sending it to the right person by service, urgency and territory
- Qualification — scoring an enquiry on the content of the message rather than only the form fields
- Follow-up drafting — writing the next message from the real context of the thread, for a human to approve and send
- Summarisation — turning a call, a transcript or a long thread into a CRM note nobody had to type
- First-line support — answering the genuinely repetitive questions, at two in the morning, from your own documented answers
Where it does not belong
Five places we will tell you not to use it
Every one of these fails in the same way: the model produces something fluent and wrong, and by the time anyone notices, a customer has already acted on it.
- Quoting prices or committing to timelines — a confident wrong number becomes a dispute you have to honour
- Handling complaints or anything emotionally loaded — fluency reads as indifference exactly when it should not
- Anything requiring professional judgement — legal, medical, financial or safety-related answers need a qualified human
- Publishing content unreviewed — unchecked output at volume damages search visibility and credibility together
- Replacing the relationship — in most service businesses the reason people buy is that a person took them seriously
Doing it properly
Four rules that keep an AI deployment safe
The difference between a system you trust and one you quietly switch off six weeks later is almost never the model. It is the design around it.
-
Define the escalation before the automation
Decide in advance what the system does when it is unsure — who it hands to, how quickly, and what the customer sees in the meantime. A model that never escalates will eventually improvise, and improvisation is the failure.
-
Keep a human on anything outbound that commits you
Drafting is safe. Sending is a decision. Approval on anything that quotes, promises or apologises costs seconds and removes most of the risk in the entire system.
-
Constrain it to your own material
Answers drawn from your documented services, pricing rules and policies are checkable. Answers drawn from the model’s general knowledge about your industry are guesses in your brand voice.
-
Evaluate it against real cases, continuously
Collect actual enquiries, check the classifications and drafts against what a good human would have done, and review that sample regularly. Without measurement you are not running an AI system, you are hoping.
Next step
Bring us the process you are thinking of automating.
Book a strategy call and describe it. We will tell you honestly whether AI belongs in it, whether plain automation would do the job better, or whether it should stay with a person.