By Keshavan Thiagaraja
Every impressive AI model you use today started life as a generalist. It became useful the same way people do: through specialised training. That second stage, fine tuning, is where raw capability turns into dependable business behaviour. Most enterprises misunderstand it. Many pay for that misunderstanding twice: once in wasted spend, and again in missed advantage.
I write this as an enterprise architect, not as a vendor. Fine tuning is not a feature you buy. It is an architecture decision you own.
How a model actually improves
A frontier model is built in layers, and each layer answers a different question.
Pretraining answers: what does the model know? The model reads vast amounts of text and code and learns the patterns of language, logic, and software. This stage is enormous, expensive, and done by the model providers, not by you.
Supervised fine tuning answers: how should the model behave? The provider, and later the enterprise, shows the model curated examples of good work: a question paired with a strong answer, a ticket paired with a correct resolution. The model shifts from predicting text to performing tasks.
Preference optimisation answers: which behaviour do humans actually want? Techniques such as reinforcement learning from human feedback, and its simpler successor direct preference optimisation, train the model on comparisons: this answer is better than that one. Human experts create and judge hard examples. The model learns judgement, not just knowledge.
Evaluation answers: is any of this true? Rigorous benchmarks and human review measure whether the model genuinely improved or merely sounds better. No serious lab ships without it. No serious enterprise should either.
Notice the pattern. Models do not get better because they grow. They get better because humans keep teaching them what good looks like, and keep testing whether they learned it. That is a supply chain of expertise, and it is now a permanent feature of the AI economy.
The decision most enterprises get wrong
Here is the systemic critique. Enterprises reach for fine tuning first because it sounds decisive, when the architecture usually says it should come last.
Prompt engineering changes what you ask. Fine tuning changes what the model is. Retrieval augmented generation changes what the model can see. Three levers, three different problems.
Prompting is for behaviour you can describe. If clear instructions and examples fix the problem, you are done, at near zero cost, with instant reversibility.
Retrieval is for knowledge that changes. Product catalogues, policies, case law, prices. RAG fetches current facts at question time, so the answer stays fresh and traceable to a source. Fine tuning bakes knowledge in, where it silently goes stale.
Fine tuning is for behaviour you cannot describe. House style across ten thousand documents. Domain judgement that experts apply but cannot fully write down. Strict output formats at high volume, where shorter prompts and smaller tuned models cut cost per call. It teaches skills and tone, not fresh facts.
The honest sequence is prompting first, retrieval second, fine tuning where evidence demands it, and often all three together. A tuned model that speaks your language, fed by retrieval that knows today's facts, steered by disciplined prompts. Each choice you skip in that sequence is technical debt with a compounding interest rate.
The economics have changed
Five years ago fine tuning meant retraining giant models at giant cost. Parameter efficient methods ended that era. Techniques such as LoRA and QLoRA freeze the original model and train small adapter layers, typically well under one percent of the parameters. Fine tunes that once needed a GPU cluster now run on a single card or a modest cloud job. The cost barrier fell. The competence barrier did not.
Because the real cost of fine tuning was never compute. It is data. Curating a few thousand genuinely excellent examples of your best work is the hard part, and it is exactly the asset your competitors cannot copy. Your data is your moat. The model is increasingly a commodity.
Governance is the silent failure mode
A tuned model is a new asset with a lifecycle. It can drift as the base model updates. It can memorise sensitive training data. It can quietly encode yesterday's policy. Treat every fine tune the way you treat a production system: versioned, evaluated before release, monitored after release, and owned by a named team. If personal data enters the training set, your privacy obligations follow it in. In Australia that means the Privacy Act; in Europe, GDPR and the EU AI Act. Design for audit before the regulator asks, not after.
What CEOs should do
First, stop funding model worship and start funding data curation. Direct your teams to build one clean, high quality dataset from your best human work. That asset appreciates. Model choices depreciate.
Second, demand an evaluation gate. No tuned model reaches production without a benchmark that reflects your real tasks and a human review of its failures. If your team cannot show you a score, they are shipping hope.
Third, ask one question in every AI review: did we try prompting and retrieval first? Make the cheap levers prove insufficient before the expensive one is pulled.
The CTO checklist
1. Map every AI use case to prompting, retrieval, or fine tuning, and record why.
2. Stand up a small evaluation harness before your first fine tune, not after.
3. Prefer parameter efficient tuning, and version every adapter like code.
4. Keep personal and regulated data out of training sets unless legal has signed off.
5. Assign an owner for model drift, with a retraining trigger, not a retraining habit.
6. Report cost per task, not cost per token. That is the number the board can act on.
The strategic point
Models get better through structured human teaching. Enterprises get better AI the same way. The winners will not be the companies with the biggest models. They will be the companies that teach best: clean data, honest evaluation, disciplined governance. That is not a procurement exercise. It is an engineering culture.
Where is your organisation on that curve: still prompting, already retrieving, or ready to teach?
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
#ArtificialIntelligence #FineTuning #EnterpriseArchitecture #LLM #RAG #AIStrategy #MachineLearning #AIGovernance #CTO #CEO #DigitalTransformation
Originally published on LinkedIn by Keshavan Thiagaraja. Read the original on LinkedIn.
