How to Build AI Models?

Should You Build at All? The Taker, Shaper, Maker Decision

You want a custom AI model — but before you write a line of training code, the more valuable question is whether you should train one at all. That decision matters more than any architecture choice you'll make later.

The pressure to adopt is real: McKinsey's global survey found 65% of organizations now regularly use generative AI, nearly double the figure from just ten months prior. Yet the economics of building from scratch have moved sharply the other way — Epoch AI's trend data shows frontier training costs rising roughly 3.5× per year since 2020, with gigawatt-scale data centers costing around $38 billion in upfront capex each. And even if you have the budget, Deloitte's research found 37% of organizations cite lack of talent as the top barrier to scaling AI.

This is where McKinsey's Taker, Shaper, Maker framework earns its keep. Instead of asking "how do I build a model?", it forces a sharper question: which posture fits your data, budget, and timeline?

• Taker — use off-the-shelf models through APIs or embedded tools. Fastest to production, no code required, but no differentiation.
• Shaper — fine-tune a pre-trained model with proprietary data. Custom behavior without frontier-scale costs, though full fine-tuning still demands multiple powerful GPUs running in tandem.
• Maker — build from scratch. Total independence and IP control, but prohibitive costs for all but a handful of organizations.

For most teams, the answer isn't one posture but a hybrid: buy for common tasks, build where proprietary data matters most. McKinsey's research found companies using pre-built AI tools cut time-to-market by up to 50% versus custom builds — a strong argument for taking wherever the task is generic. Meanwhile, Gartner projects that by 2026 more than 80% of enterprises will use custom models fine-tuned for industry-specific language, so shaping is becoming the default middle path.

There's also a newer rung on this ladder that costs almost nothing: agent skills. These SKILL.md instruction packages turn a general-purpose agent into a specialist without any training, loading context progressively so reference material only consumes tokens when used. Anthropic's own guidance is to treat skill installation like installing software and use skills only from trusted sources — which is why SkillzBench is building a free, verified directory with security scanning and sandboxed benchmarking, so you can adopt skills without auditing each one yourself.

Decide your posture first. The rest of the build process only makes sense once you know which race you're actually running.

The Customization Ladder: Prompting, RAG, and Fine-Tuning

Here's the paradox of custom AI: most "custom" needs never require training a model at all. According to IBM's framework, three optimization methods form a ladder of increasing resource intensity — and they work best when combined, not chosen between.

Rung one is prompt engineering. It is the least time-consuming method and can be done manually with no extra compute. IBM is blunt about its importance: "Further training and greater data access cannot compensate for poor prompting." Master this first, because every rung above it depends on getting prompts right.

Rung two is retrieval-augmented generation (RAG). This rung demands data science expertise to organize datasets and build data pipelines that feed current, accurate information to the model at query time. RAG is the right choice when up-to-date facts matter more than deep specialization.

Rung three is fine-tuning. IBM calls it "arguably the most demanding" method, requiring compute-intensive data preparation and training. Full fine-tuning requires multiple powerful GPUs running in tandem, putting it out of reach for smaller teams without cloud budgets.

The ladder in brief:

• Prompt engineering — cheapest, foundational, no extra compute
• RAG — needs data pipelines and expertise to keep knowledge current
• Fine-tuning — most demanding, multiple GPUs, best for doing one thing very well
• Combination — IBM notes the methods "are not mutually exclusive and are often combined for optimal outcomes"

Here is the encouraging part: the ladder is getting cheaper to climb. IBM's trend analysis shows that a 2-billion-parameter model (IBM Granite 3.3 2B Instruct) now scores 80.5% on HumanEval — beating the original GPT-4's 67% from just two years earlier while being 900× smaller. Inference costs have dropped dozens of times over in under two years, with algorithmic improvement running at roughly 400% per year.

That efficiency trend opens a fourth, even cheaper rung: packaging expertise as agent skills. These SKILL.md instruction packages load progressively — metadata costs about 100 tokens at startup, and bundled resources cost zero tokens until accessed — turning a general-purpose agent into a specialist with no training at all. Anthropic's guidance is to create one "when you keep pasting the same instructions, checklist, or multi-step procedure into chat," but also to treat installing third-party skills like installing software, since a malicious skill can leak data. SkillzBench is being built to solve exactly that: a free directory and MCP connector with security scanning and sandboxed benchmarking for every skill, so you can climb this rung without the audit burden.

The Build Lifecycle: From Data to Deployment in Six Steps

Building a custom AI model sounds like one giant leap, but in practice it's a sequence of six well-defined steps — and skipping any of them is where most projects quietly fail.

Step 1: Data collection and cleaning. Every model starts with raw data, and its quality determines everything downstream. As Domyn's technical team puts it, "the quality of the training data directly affects the model's accuracy" — a point that matters especially in regulated industries like banking and insurance.

Step 2: Feature engineering. Clean data becomes useful data here, as raw inputs are transformed into signals the model can actually learn from. This is where domain expertise pays off more than raw compute.

Step 3: Architecture selection. You choose the model family and size that fits your compute budget and task. Scale dominates this decision: Epoch AI's research shows frontier training compute has grown roughly 5× per year since 2020, which is why most teams adapt existing architectures rather than inventing new ones.

Step 4: Training and validation. The model learns patterns, then gets tested against data it has never seen. Validation is what tells you whether the model generalizes or has simply memorized its homework.

Step 5: Tuning. Hyperparameters, thresholds, and training data get adjusted to squeeze out accuracy. IBM's guidance is blunt: "further training and greater data access cannot compensate for poor prompting" — the same discipline applies to tuning.

Step 6: Deployment. The model goes into production — but the work doesn't end. Gartner warns that models can lose up to 20% of their accuracy per year without retraining, so maintenance belongs inside the build, not after it.

For large language models, the middle of this lifecycle expands into three distinct phases:

• Pre-training — the model learns basic language capabilities from trillions of tokens of internet-scale data.
• Supervised fine-tuning — the model is fed curated inputs alongside reference answers to learn task-specific behavior.
• Alignment — techniques like Direct Policy Optimization use pairs of accepted and rejected responses so the model learns human preferences, not just correctness.

Not every "custom AI" project needs all six steps, of course. For teams customizing existing agents rather than training models, agent skills — reusable instruction packages that load progressively and cost almost nothing until used — offer a zero-training path to specialist behavior. Anthropic's own guidance is to treat installing a skill like installing software, using only trusted sources. SkillzBench is building exactly that trust layer: security scanning, version pinning, and sandboxed benchmarking for the skills you equip your agents with, all free.

The No-Training Shortcut: Agent Skills as Packaged Expertise

Here's a customization option that requires no GPUs, no datasets, and no training runs at all. Agent skills — reusable SKILL.md instruction packages — turn a general-purpose agent into a specialist by loading workflows, context, and best practices on demand, according to Anthropic's official documentation. They work across Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents.

The economics are striking when you compare them to model training. Frontier training compute has grown 5× per year since 2020, and training costs have climbed 3.5× per year, per Epoch AI's trends data. Skills sit at the opposite end of that spectrum, using progressive disclosure to keep costs near zero:

• Metadata of roughly 100 tokens loads at startup, telling the agent when the skill is relevant.
• Detailed instructions, under 5,000 tokens, load only when the skill is triggered.
• Bundled files and reference material cost zero tokens until the agent actually accesses them.

Anthropic's advice on when to write one is refreshingly practical: create a skill whenever you find yourself pasting the same instructions, checklist, or multi-step procedure into chat. That recurring prompt becomes packaged expertise you can reuse across every project and share with your team. For business owners facing the talent gap — 37% of organizations cite lack of AI talent as their top barrier to scaling, per Deloitte — skills offer domain expertise without hiring a data science team.

There is one serious catch. Because skills can invoke tools and read files, a malicious skill can do real damage — leaking data or taking harmful actions inside your environment. Anthropic's guidance is blunt: use skills only from trusted sources, and treat installing one like installing software, not like dropping a text file into a folder.

That warning matters because skills are spreading faster than anyone is vetting them. Published research has found that roughly 26% of community skills carry at least one vulnerability, and most users have neither the time nor the security background to audit bundled files line by line. Reading every script in every skill you install is a full-time job nobody signed up for.

This is the gap SkillzBench is being built to close. The free, independent directory continuously crawls and deduplicates skills from GitHub, then verifies each one: pinning every version to a commit SHA, running security scans, checking licenses, and judging quality. Every skill is benchmarked in a sandbox with the same model, with and without the skill, so you can see the measured Lift in percentage points before installing anything. A 0–100 trust score combines security, quality, license, maintenance, and community signals — and money never touches the scores.

The takeaway from this rung of the ladder: if your "custom AI" need is really repeated expertise, you may not need to train anything. Package the knowledge, verify it, and let your existing agents do the rest.

Your Action Plan: Building Custom AI Without Breaking the Budget

You don't need a research lab to build custom AI — you need a disciplined ladder of escalating investment. Frontier training compute has grown roughly 5× per year since 2020, with costs rising 3.5× annually, per Epoch AI's trend data. The good news: most "custom AI" never requires training anything from scratch.

Start with prompting. It's the least resource-intensive optimization method, requiring no extra compute and no data pipeline. IBM's guidance is blunt: further training and greater data access cannot compensate for poor prompting, so master this layer before spending a dollar elsewhere (IBM Think).

Add RAG when accuracy and currency matter. If your AI must answer with current, factual, proprietary information — policies, product specs, regulations — retrieval-augmented generation beats fine-tuning. It requires data science effort to organize datasets and build pipelines, but keeps knowledge updatable without retraining.

Fine-tune only for one narrow, high-value task. Fine-tuning is the most demanding approach: data preparation and training are compute-intensive, and full fine-tuning requires multiple powerful GPUs running in tandem. Reserve it for the single behavior prompting and RAG can't nail.

Package what works as agent skills. When you keep pasting the same instructions or multi-step procedure into chat, that's the signal to codify it (Anthropic's own guidance). Skills use progressive disclosure — metadata costs about 100 tokens at startup, and bundled resources cost zero tokens until accessed (per Anthropic's documentation). Your action plan in one view:

• Prompt-engineer first — free, fast, foundational
• Layer RAG for accurate, current knowledge
• Fine-tune one narrow task only when the ROI justifies GPUs
• Convert repeatable workflows into agent skills

One caution: Anthropic explicitly warns to treat skill installation like installing software, since malicious skills can invoke tools harmfully or leak data. SkillzBench is building an answer to exactly that — a free directory that crawls and dedupes skills, pins every version to a commit SHA, security-scans and license-checks each one, and benchmarks skills in a sandbox with and without the skill to measure Lift. Trust scores and Lift are always kept separate, and the free MCP connector (currently in development) will let you equip any agent with only the verified skills you choose. Browse the directory, join the waitlist, and test skills that automate the rest of your model-building workflow — skills, proven.