Once a company decides to use AI, the next question is how much to build. The options form a ladder from least to most effort.

1. Buy an off-the-shelf assistant

Subscribe to a ready-made assistant or a feature inside software you already use. Fastest to deploy, minimal engineering, but limited customization and dependence on the vendor’s roadmap and data terms.

2. Integrate a model through an API

Call a hosted model from your own applications. You control the user experience and workflow; the provider runs the model. Costs scale with usage, so estimate token volumes early.

3. Add retrieval (RAG)

Retrieval-augmented generation, described in a 2020 paper by Lewis and colleagues, lets a model answer using your documents by retrieving relevant passages at query time. It is the most common way to make a general model useful on company knowledge, and updating knowledge is as simple as updating the documents.

4. Fine-tune

Fine-tuning adjusts a model’s weights on your examples. It helps with style, format and specialized behavior, less so with facts that change often. Techniques such as LoRA (low-rank adaptation) make fine-tuning far cheaper than full retraining.

5. Train or heavily adapt your own model

Only a small number of organizations have the data, budget and expertise to justify this. It offers maximum control but requires ongoing investment in evaluation, safety and infrastructure.

How to decide

Need Usually best fit
General productivity Buy
Custom workflow in your app API
Answers grounded in your documents RAG
Consistent tone or output format Fine-tuning
Data cannot leave your premises Open-weight model, self-hosted

Most mature deployments combine approaches — for example an API model with retrieval, plus a small fine-tuned model for a high-volume task.

A worked example: an internal policy assistant

A company wants employees to get answers about HR and IT policies.

  • Buy: an enterprise AI assistant with document upload might be enough for a small company, deployed in days.
  • Integrate via API: a mid-sized company could build a simple chat interface into its intranet, calling a model API.
  • Add retrieval: for hundreds of policy documents that change often, RAG lets the assistant answer from current documents with citations — usually the right choice here.
  • Fine-tune: only if the assistant must adopt very specific formats or terminology that prompting and retrieval can’t achieve.
  • Train your own: almost never justified for this use case.

The pattern generalizes: most enterprise use cases are best served by buying or by an API plus retrieval, with fine-tuning reserved for specific needs.

Total cost of ownership

Licence or API fees are only part of the cost. Include integration and engineering time, data preparation, security and compliance reviews, evaluation and monitoring, training for staff, and ongoing maintenance when models or documents change. A cheaper model that needs extensive engineering can cost more overall than a more expensive turnkey product.

Risk considerations

  • Data protection: where data is processed and whether it’s used for training.
  • Vendor lock-in: how easily you could switch models or providers.
  • Accuracy and liability: who reviews outputs before they affect customers or decisions.
  • Regulation: sector rules and laws such as the EU AI Act may impose requirements depending on the use.

See how RAG works and LoRA fine-tuning.

Bottom line

Start simple, measure, and add complexity only when results demand it.

What to watch next

  • Enterprise assistants with built-in retrieval over company data.
  • Falling costs of fine-tuning and hosting open models.
  • Agent platforms that connect models to business systems.
  • Regulatory requirements for high-risk enterprise AI uses.

Key takeaways

  • Options range from buying an assistant to training your own model.
  • API integration plus retrieval covers most enterprise needs.
  • Fine-tune for specific formats, styles or narrow tasks, not for adding knowledge.
  • Consider total cost of ownership, not just licence fees.
  • Data protection, lock-in, accuracy and regulation shape the decision.

Frequently asked questions

Should we fine-tune to teach the model our company knowledge?

Usually not. Retrieval is better for knowledge that changes; fine-tuning is better for behavior and format.

When does training our own model make sense?

Rarely — mainly for organizations with unique data at massive scale and specific needs that existing models cannot meet.

How do we start?

Pick one use case, try the simplest option first, measure results and move to more complex approaches only if needed.