--- title: When is an AI API wrapper enough, and when do you need more? author: Giacomo Balli published: 2026-09-16 url: https://BigBalli.com/blog/build-vs-buy-ai-api-wrapper tags: ai, build vs buy, api wrapper, generative ai, ai strategy --- # When is an AI API wrapper enough, and when do you need more? *By Giacomo Balli* | *Published: September 16, 2026* **Short answer:** An API wrapper is enough when the AI does a generic task, a person checks the output, and the data is not regulated. Build more, meaning retrieval over your documents, testing, guardrails, and logging, when the AI answers customers, touches regulated data, runs at high volume, or is what you sell. Most owners who ask whether to build AI should buy it first, and most who already paid for a wrapper need to know where it stops being enough. A wrapper is a thin layer of your own software that sends text to a model from OpenAI, Anthropic, or Google and shows the reply. It is quick and cheap to build. It also inherits every weakness of the model underneath, and whatever it tells a customer, your company said. 1. **1. Buy a seat**ChatGPT, Claude, or Microsoft Copilot for generic work 2. **2. API wrapper**Your prompt and screen around a hosted model 3. **3. Wrapper plus**Your documents, a test set, guardrails, logging 4. **4. Custom model**Fine-tuning, only after level 3 is measured and falls short Start at the top. Move down only when the level above fails on measured, real work. ## What exactly is an AI API wrapper? An AI API wrapper is software that sends a prompt to a hosted model, such as Anthropic's Claude or OpenAI's GPT models, and returns the answer inside your own screen or workflow. The wrapper owns the prompt, the interface, and the billing. The model provider owns the model, its accuracy, and every change made to it. Wrappers got a bad name because thousands of startups shipped one and called it a company. For an owner, that criticism is beside the point. You do not need a moat to save your office manager four hours a week on drafting quotes. You need the cheapest thing that works reliably, and for an internal tool a well-written wrapper is often exactly that. It can be running in days. The cost is dependence. Anthropic's [pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing) already lists models such as Claude Opus 4 as retired except on specific cloud platforms. Every wrapper is built on a model with a shelf life, so budget for someone to re-test your prompts each time the provider retires the one you chose. ## When is buying an AI tool better than building? Buying an off-the-shelf AI tool beats building when the job is common across companies: drafting email, summarizing meetings, answering questions about a stack of documents, or writing code. ChatGPT, Claude, and Microsoft Copilot already do those jobs well, and a monthly per-seat subscription costs less than one week of a developer's time. Buying also works more often. MIT's Project NANDA reviewed more than 300 publicly disclosed AI initiatives for its 2025 report [The GenAI Divide](https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf). Tools bought from or built with external partners reached deployment about 67% of the time. Tools built internally got there about 33% of the time. The same report found 95% of organizations getting zero return on generative AI. My test before any build is simple. Could a $20-a-month seat do most of this job? If the answer is yes, run the seat for a month on real work and write down every miss. That log is a better requirements document than anything a vendor will draft for you. ## When is a simple API wrapper enough? A simple API wrapper is enough when four things are true: the task is generic, a person reviews every output before it matters, the data sent is not regulated, and volume is modest. Internal drafting, summarizing call notes, and first-pass sorting of inbound email usually pass all four tests without extra engineering. 1. **Generic task?**Drafting, summarizing, sorting 2. **A person reviews it?**Before a customer or a decision sees it 3. **Unregulated data?**No health, financial, or children's records 4. **Modest volume?**Hundreds or low thousands of calls a month Four yes answers: a plain wrapper will do. Any no: plan the extra layers before you build. Model cost is rarely the obstacle at that scale. Anthropic's pricing documentation works through a support example: about 3,700 tokens per conversation on Claude Haiku 4.5, at $1 per million input tokens and $5 per million output tokens, comes to about $37 for 10,000 conversations. Its Batch API halves token prices for work that can wait, such as overnight classification. At those prices the model bill is small next to the developer who builds and maintains the wrapper. Spend the budget on the prompt, a set of 50 to 100 real examples with known right answers, and a log of every request. Do not spend it choosing the cheapest model. ## When is an API wrapper not enough anymore? An API wrapper stops being enough when the AI speaks for your company, answers from your private data, handles regulated information, or is the thing customers pay for. Each of those situations needs work a plain wrapper leaves out: retrieval over approved documents, automated testing, guardrails, audit logs, and a clear path to a human. SituationWrapper enough?What to add Internal drafting, a person reviews itYesPrompt, test examples, request log Answers customers on your websiteNoAnswers only from approved sources, refusal rules, hand-off to staff Health, financial, or children's dataNoSigned data agreement, data residency, access controls, audit log Tens of thousands of calls a monthSometimesCaching, batching, cheaper model for easy requests, cost alerts The AI is what you sellNoEvaluation suite, feedback loop, ability to switch model providers The legal exposure is already on record. In [Moffatt v. Air Canada](https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html) (2024 BCCRT 149), British Columbia's Civil Resolution Tribunal held Air Canada responsible for a bereavement-fare promise its website chatbot made up. The airline argued the chatbot was responsible for its own statements. The tribunal rejected that and ordered Air Canada to pay the fare difference. Regulated data changes the design before it changes the price. Anthropic, for example, offers US-only processing for its newer models at a 1.1x price multiplier. A wrapper that sends patient or client records to a model needs that kind of control, a signed data processing agreement, and a record of who sent which record and when. ## What does building more than a wrapper involve? Building more than a wrapper means adding four layers around the model: retrieval that feeds it your approved documents, an evaluation set that tests answers before every release, guardrails that block unsafe or off-policy replies, and monitoring that shows cost and failure rates in production. Once those exist, the model call is the smallest part of the system. MIT NANDA's report names what most failed pilots lacked: the systems did not retain feedback, adapt to context, or improve over time. That gap is the difference between a demo and a tool staff keep using. It is also where most budgets break, because none of that work shows up in a sales demo. [Gartner predicted in July 2024](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025) that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. It named four causes: poor data quality, inadequate risk controls, escalating costs, and unclear business value. A better model fixes none of the four. Fine-tuning or training your own model comes last. It makes sense when you hold thousands of labeled examples of one narrow task and a wrapper with retrieval has been tested and measured as falling short. Few owner-operated businesses meet both conditions. ## How should an owner decide between build and buy? Decide by running the cheapest option on real work first. Give a few staff a bought AI tool for 30 days and log every miss. Generic misses mean keep buying. Misses caused by missing company data or company rules mean scope a wrapper with retrieval. Build further only when measured errors cost more than the build. 1. **Misses are generic**Keep the bought tool 2. **Misses need company data or rules**Wrapper with retrieval and a test set 3. **Measured errors cost more than a build**Custom system, scoped against the log The 30-day log of misses decides the next step, not the vendor demo. 1. Name the task and the number it should move: hours saved, response time, or error rate. 2. Run a bought tool on that task for 30 days with three to five staff. 3. Sort every miss into model error, missing company data, or missing company rules. 4. If most misses are data or rules, scope a wrapper with retrieval and a test set. 5. Before signing any build, ask the vendor for its evaluation set and its plan for model retirement. ## Related guides - [Where should AI actually fit in my product?](https://BigBalli.com/where-does-ai-fit-in-my-product) - [Build custom software or buy off the shelf?](https://BigBalli.com/blog/build-vs-buy-business-software) - [What are the red flags in a software development proposal?](https://BigBalli.com/blog/software-proposal-red-flags) - [The Second Opinion: an independent review of an AI initiative](https://BigBalli.com/second-opinion) - [Moffatt v. Air Canada, 2024 BCCRT 149 (full decision)](https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html) - [Anthropic: Claude API pricing and worked cost examples](https://platform.claude.com/docs/en/about-claude/pricing) ## Key takeaways - Try a bought AI tool on real work for 30 days before paying anyone to build. - A plain API wrapper is enough for generic, reviewed, unregulated, modest-volume tasks. - Customer-facing AI makes your company liable for its answers, as Moffatt v. Air Canada showed in 2024. - Past a wrapper, the budget goes to retrieval, testing, guardrails, and monitoring, not to the model. - Fine-tune only with thousands of labeled examples and a measured shortfall from retrieval. ## Frequently asked questions ### How much does an AI API wrapper cost to run? Model usage is usually the small part. Anthropic's documentation prices 10,000 support conversations on Claude Haiku 4.5 at about $37 in tokens. The larger cost is developer time for prompts, a test set, logging, and the screen staff use, and that grows with each system the wrapper has to connect to. ### Is it safe to send company data to ChatGPT or Claude through an API? Read the provider's commercial API terms before sending anything sensitive, because business API terms differ from consumer app terms, including how long inputs are kept. For regulated data, also require a signed data processing agreement, data residency where the law requires it, and logs showing who sent which records and when. ### Should we fine-tune our own AI model? Usually not at the start. Fine-tuning helps when a company holds thousands of labeled examples of one narrow task and a wrapper with retrieval has already been tested and measured as falling short. Most business accuracy problems come from missing company data or unwritten rules, which retrieval and clearer prompts fix for less. ### Who is liable when an AI chatbot tells a customer something wrong? The company that deploys the chatbot. In Moffatt v. Air Canada, decided in February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay the fare difference its website chatbot had wrongly promised, and rejected the airline's argument that the chatbot was responsible for its own statements. ## About the author Giacomo Balli is an independent technology advisor in San Francisco. He has built software and mobile apps since 2010, runs a portfolio of more than forty live apps of his own, and reviews software, AI, and vendor decisions for owners before they commit the money. ## Disclosure Giacomo Balli sells fixed-fee independent reviews of technology decisions, including AI initiatives. He does not build or resell software and takes no referral fees. No company named on this page paid to be mentioned. ## Questions this answers **How much does an AI API wrapper cost to run?** Model usage is usually the small part. Anthropic's documentation prices 10,000 support conversations on Claude Haiku 4.5 at about $37 in tokens. The larger cost is developer time for prompts, a test set, logging, and the screen staff use, and that grows with each system the wrapper has to connect to. **Is it safe to send company data to ChatGPT or Claude through an API?** Read the provider's commercial API terms before sending anything sensitive, because business API terms differ from consumer app terms, including how long inputs are kept. For regulated data, also require a signed data processing agreement, data residency where the law requires it, and logs showing who sent which records and when. **Should we fine-tune our own AI model?** Usually not at the start. Fine-tuning helps when a company holds thousands of labeled examples of one narrow task and a wrapper with retrieval has already been tested and measured as falling short. Most business accuracy problems come from missing company data or unwritten rules, which retrieval and clearer prompts fix for less. **Who is liable when an AI chatbot tells a customer something wrong?** The company that deploys the chatbot. In Moffatt v. Air Canada, decided in February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay the fare difference its website chatbot had wrongly promised, and rejected the airline's argument that the chatbot was responsible for its own statements. --- **Tags:** ai, build vs buy, api wrapper, generative ai, ai strategy --- *View the [HTML version](https://BigBalli.com/blog/build-vs-buy-ai-api-wrapper) of this post.*