Use an off-the-shelf AI tool when the task is generic, drafting, summarising, transcription, translation, and the tool can see the data it needs. Build something custom when the value depends on your own content, your own workflow, or data that cannot leave your systems. The decision is rarely about model quality, because most custom work uses the same models the tools do.
This guide covers the three real options, what each costs to build and to run, how to evaluate whether either is working, and the data questions that decide the answer more often than features do.
Key takeaways
- Custom rarely means a custom model: It usually means your data and workflow wrapped around a hosted model.
- Usage cost is the line people miss: AI has a build cost and a per-transaction cost that scales with volume.
- Data residency decides many cases outright: If content cannot leave your systems, the tool list shortens fast.
- Evaluate before you commit: Run both against fifty real examples; opinions about AI quality are unusually unreliable.
The three options, not two
| Option | What it is | Typical cost | Best for |
|---|---|---|---|
| Off-the-shelf tool | A product with AI inside it | Per-seat subscription | Generic tasks; fast start; no engineering |
| Custom application on a hosted model | Your workflow and data, someone else's model | $5,000–$30,000 build, plus usage | Your content, your process, standard infrastructure |
| Self-hosted or fine-tuned model | You run the model | $30,000+ and dedicated capacity | Data residency requirements or very high sustained volume |
The middle option is where most business value sits and the one buyers most often skip, because the framing "build or buy" hides it. You are not choosing between a subscription and training a model; you are usually choosing between a generic product and a thin, specific application built on the same underlying model.
Buy when the task is generic
Off-the-shelf wins when the work is not specific to your business: drafting and editing text, summarising meetings, transcription, translation, image generation, coding assistance.
You get immediate availability, no engineering cost, continuous improvement as vendors update, and a support contract. The trade is that the tool cannot see your systems, cannot follow your process, and every competitor has the same capability.
Buy also when volume is low. If a task happens fifty times a month, a subscription almost certainly beats anything you would build.
Build when your data or workflow is the value
Custom work is justified when the answer depends on content only you hold, your policies, your product catalogue, your case history, or when the AI step sits inside a workflow that has to trigger something else in your systems.
The reliable pattern is narrow: retrieval over your approved content, with sources shown, scope limited to topics you have covered, and an explicit way to say it does not know and hand to a person. An assistant with no exit produces its best guess, which is the failure mode that destroys trust fastest.
The tasks that work share a shape: high volume, unstructured input, and a human who can quickly tell whether the output is right. Anything that must be exactly right without review is where AI is weakest, regardless of build or buy.
Understand the two-part cost
AI is priced unlike other software. There is a build cost and a usage cost that grows with traffic.
| Approach | Setup | Ongoing |
|---|---|---|
| Hosted model API | $5,000–$20,000 | Per-token usage, scales directly with volume |
| Retrieval over your content | $10,000–$30,000 | Usage plus storage and re-indexing |
| Self-hosted or fine-tuned | $30,000+ | Dedicated capacity billed whether used or not |
Dedicated capacity is the trap. It bills continuously while most business traffic is intermittent, which is why self-hosting is justified by regulation or sustained scale rather than by preference.
Model the running cost at realistic volume before committing. A per-request cost that looks trivial in testing becomes a meaningful line item in production, and it is the number most projects fail to forecast.
Let the data boundary decide
Answer these before comparing features, because they eliminate options faster than any capability comparison:
- Can this content leave your infrastructure at all?
- Does your jurisdiction or your customer contracts constrain where processing happens?
- Does the vendor train on your inputs by default, and can that be switched off in writing?
- What retention applies to prompts and outputs?
- If the content includes personal data, what is your lawful basis under GDPR or Turkish data protection law?
For health, financial or legal content these frequently rule out consumer-grade tools regardless of how well they perform.
Evaluate both against real examples
Opinions about AI output quality are unusually unreliable, because a fluent wrong answer reads better than an accurate terse one.
Assemble fifty real examples with known correct outcomes. Run each candidate against them, score the results, and count the corrections a reviewer would have to make. That evaluation set is also what tells you later whether a model change improved things or quietly made them worse, the artefact teams most often wish they had built first.
Measure the reviewer's edit rate over time. It should fall. If it does not, the system is not learning from corrections and you are paying for a suggestion engine.
Keep a human on consequential decisions
Automate low-consequence, reversible decisions with monitoring. Keep a person deciding anything that affects someone's money, employment, health or legal standing, in the EU this is a regulatory expectation, not only good practice.
Design the review loop before launch: who checks, how fast, what they do when it is wrong, and how the correction improves the system rather than being lost.
The pragmatic sequence
Start with an off-the-shelf tool for a month, even if you expect to build. It costs almost nothing, tells you whether the task is genuinely suited to AI, and produces the evaluation examples you will need.
If it works but cannot reach your data or fit your workflow, that is your specification for a custom build, and you will have written it from evidence rather than from a demo.
Related guides
- For what this looks like in practice, see Digital Transformation Roadmap: A 7-Step Business Guide.
- A second perspective worth having is Admin Dashboard Development: Features, Architecture, and Cost Drivers.
Should Custom AI vs Off-the-Shelf Tools: How to Decide turn into models and pipelines someone has to maintain, Discuss Your Data Product.
Ali Boran Gazel