Choosing the right AI model for your app

When most people say “off‑the‑shelf model”, they’re talking about foundation models - huge neural networks trained on massive datasets that can handle lots of tasks from day one. Think big players like OpenAI’s GPT models or Google’s Gemini.

But there’s an important distinction to be made here: not all off‑the‑shelf models are foundation models. There are many other pre‑trained models that haven’t been trained at the same massive scale but are still incredibly useful because they’re specialised for particular tasks.

For example, in image generation, you might choose models optimised for creative outputs or efficiency like Google’s Flash models or experimental community offerings like Nano‑Banana (for fast, lightweight generation). On platforms like Hugging Face you’ll find many of these lighter pre‑trained models ready to drop into your stack for specific uses.

These models haven’t learned everything, but they’ve already been trained on data relevant to your task, which means you don’t have to start from zero.

If we stick with the “suit” analogy: using an off‑the‑shelf model is like renting a well‑made suit for a wedding or event - it isn’t bespoke, but it’s designed to be presentable for most occasions, saves a lot of time and money, and can be suitably fit for purpose with a small adjustment here or there.

Here’s why product teams reach for these models:

• Speed to market: You can be prototyping within a day, and live within weeks.
• Lower upfront cost: You avoid the big investment in compute power, engineering, and data needed to train a model from scratch.
• Built‑in intelligence: Because these models have already been trained on huge datasets, they perform well with very little extra input.

In short, they’re ideal for most general use cases, powering everything from AI‑powered customer service bots to smart search features in apps like Shopify, Slack, and Duolingo.

In fact, OpenAI’s GPT models are now used by over 92% of Fortune 500 companies.

If you like the flexibility of pre‑trained models but want more control over cost, data, or privacy, self‑hosting an open‑source model might be the sweet spot. Models like LLaMA, Falcon, or MPT can be deployed on your own infrastructure, giving you complete ownership over how they run.

This approach gives you:

• Tighter cost control: No per‑call API billing, costs are tied to the hardware and energy you run.
• Data privacy: Sensitive information stays on your servers, never sent to third‑party APIs.
• Customisation flexibility: You can adapt or fine‑tune the model without vendor restrictions.

But a caveat: self‑hosting isn’t “set and forget”. You’ll need to manage servers, scaling, monitoring, and maintenance and if your usage skyrockets, those infrastructure costs can add up. It’s worth modelling anticipated token usage and hosting costs before committing.

Sometimes an off‑the‑shelf or self‑hosted model just isn’t enough. That’s when you go a step further: custom training or fine‑tuning.

This might mean:

• Starting with a smaller baseline model that’s already closer to your domain (e.g., a legal‑specialised language model).
• Then fine‑tuning it on your own proprietary data so it learns your company’s terminology, style, and context.

For example, a model already competent in general English might struggle with technical medical language, but a domain‑adapted model fine‑tuned on clinical datasets will understand that nuance.

• You have niche data that generic models won’t understand.
• Your app requires domain-specific language, like legal, medical, or engineering jargon.
• Privacy is a priority, and you don’t want to send your users' data to third-party APIs.
• You need more control over how the AI behaves or responds.

Custom approaches are increasingly used in healthcare apps, advanced legal research tools, and financial platforms where detail and nuance matter. They can reduce hallucinations, tighten up responses, and give your AI features a tone that matches your brand.

And yes, customisation adds cost.

Training or fine‑tuning models can range from thousands to, in some cases, hundreds of thousands of pounds when you factor in engineering time, infrastructure, and ongoing maintenance. Think of it more like building a tailored suit with custom linings and special fabrics: it’s more expensive, but if it has to fit a specific purpose perfectly, it’s worth the investment.

One of the most common misconceptions is that you need to pick one route and stick to it. But hybrid models, where an app blends pre-trained models with custom logic or domain-specific tuning, are often the smartest route.

You might:

• Use an off‑the‑shelf model for general language tasks, then layer on retrieval‑augmented generation (RAG) to bring in results from your private database.
• Start with a pre‑trained computer vision model (a model trained to analyse images) and fine‑tune the last few layers for something specific, like identifying your own product catalogue.

Continuing our tailoring metaphor: a hybrid approach is like buying a decent off‑the‑rack suit and then refining it with extra pockets, reinforced stitching, or zips for specific needs. You get broad capability and then build the features that matter most to you.

Across many AI services (whether you’re calling APIs from a vendor, self‑hosting, or fine‑tuning) you’ll hear the term “token” a lot.

What’s a token?

A token is roughly a piece of text, roughly a word or part of a word. For example:

“AI models are fun”
- that’s about 5 tokens.

Billing and usage limits with many models are measured in tokens processed (input + output). So if your app sends a long chunk of text and expects a long reply, that’s more tokens and that typically means higher cost or usage impact.

Seeing tokens as the currency of AI usage helps you:

• Predict costs more accurately.
• Model hosting and compute requirements.
• Choose where and how to optimise (e.g., shorter prompts, caching, or selective context).

Off‑the‑shelf and smaller pre‑trained models will get you from idea to MVP fast with minimal upfront investment, but pay‑as‑you‑go pricing can scale up with heavy usage.

Custom and self‑hosted models require more engineering effort and infrastructure, but give you complete control and (often) improved long‑term cost predictability.

Hybrid approaches give you flexibility, the speed of existing models with a layer of logic or tuning where it matters most.

If you’re trying to figure out if AI is worth building into your project, don’t worry, you’re not alone. This is one of the most common questions we get asked. The answer depends entirely on your goals, your budget, and your user base.

• Want a chatbot that can handle 80% of customer queries? An off-the-shelf model is a great starting point.
• Need an AI that understands your brand voice down to the nuance? Fine-tuning might be the better route.
• Need the best of both worlds? That’s where app developers like us can help you blend the two.

We’ve helped businesses across travel, retail, and health sectors figure out the best solutions for their app, not by selling them the most expensive option, but by helping them make smart, scalable choices. Sometimes that means using GPT out of the box. Other times, it could mean building a unique AI engine from the ground up.

The most important thing we tell our partners is this: don’t get distracted by the tech. Start with what you want your app to achieve. Then work backwards.

AI models are tools, not silver bullets. But in the hands of the right team, they can make your app smarter, faster, and more valuable to your users.

And if you’re still staring down the AI decision tree and feeling a little overwhelmed, speak to app developers who understand the balance between innovation and pragmatism. Whether you’re a global brand or a local business, it’s worth finding a team that’ll treat your product like it’s their own.