Read the full story
“We trained our AI.”
No. No, you didn’t.
I see this line everywhere right now. And I get it – it sounds impressive. Like you’ve built a bespoke model that’s learning and evolving inside a company.
But in most cases? What’s actually happened is someone’s plugged a retrieval system (RAG) into an LLM like GPT or Claude.
That’s a solid move. Useful. Valuable. Just… not training.
- You can’t train GPT-4.
- You can fine-tune GPT-3.5 - but only via OpenAI’s API, with limited parameters and strict safeguards.
- You can’t even fine-tune Claude.
Why’s that?
OpenAI and Anthropic really don’t want you messing with their models. They spend hundreds of millions on each model they train.
Imagine the fallout if someone took one of their models and fine-tuned it to do something genuinely harmful. That’s not just a PR nightmare - that’s regulatory heat, legal up their ass, and trust out the window.
And if you really did want to train a model from scratch - say LLaMA 7B - you’d need:
- 20+ days of round-the-clock GPU time
- Hundreds of A100s
- £500K–£1M just for compute (not including data access/cleaning/structuring, infrastructure, or team)
So yeah - almost no one’s training their own foundation model in-house.
You could try fine-tuning, but that still comes with limitations
What you can do, though, is fine-tune a smaller model like GPT-3.5.
And that’s a decent alternative, especially for specific use cases like marketing.
But fine-tuning still has its limits. It works well when you want to:
- Mimic your brand voice
- Follow consistent templates or patterns
- Nail specific product messaging, persona targeting, or objection handling
- Generate structured, repeated content (like repackaged assets or summaries)
But it won’t:
- Develop real-world awareness
- Make strategic decisions
- Replace actual marketing or sales thinking
- Deliver true ABM - there's just too much nuance in these kind of conversations to ever train and AI on them
You still need human input - insights, context, and direction - ideally structured in a way the model can use.
And building that dataset isn’t a quick job. You’ll need to create a large, structured JSON file with 50,000+ clean examples - written, reviewed, and formatted consistently.
That takes time. Investment. Strategy. Alignment across teams. You can't skip the thinking.
It’ll take months, and who knows what tech will be available by then… So, go this route only after trying RAG, prompt engineering, and embeddings
Keep it simple, it's still valuable and you aren't underselling yourself
Building a smart wrapper around someone else’s model is totally fine - this is what I would advocate for. Done well, it slaps and saves a shit load of time.
But we’ve got to stop telling the story that these models are soaking up our company data and magically becoming in-house experts.
- LLMs don’t “learn your dataset”.
- They don't work like people.
- They don’t absorb your nuance.
- They pattern-match and generate based on probabilities.
- They operate best when built from the ground up with isolated datasets, inside specific workflows, bespoke prompts, solving clearly defined problems.
And when we overstate what’s happening, we give people a completely skewed idea of what’s possible.
Right now, loads of businesses think they can build a single AI system that plugs into every data source, understands their entire operation, and acts like a digital COO.
You just can’t - not today.
So let’s be honest.
We’re at a point where the value of AI is obvious - we don’t need to pretend it’s sentient. The real opportunity is in building the right tools, with the right guardrails, for the right use cases.
But if we keep overselling the dream, it becomes harder to have the pragmatic conversations we need to have about adoption, enablement, and impact.
Truth builds trust.
And trust is what actually gets AI working inside a business.
