Build or buy AI? How to pick the right approach

What do you need AI for?

Pick the job closest to yours to see where we'd start.

• Help staff write, summarise and brainstorm

  Start withUse a ready-made AI app

  The big chatbots already do this well. Pay for a business plan: your conversations aren't used for training by default and you get admin controls.

  Or: Turn on the AI in software you already use. If staff spend the day in one tool, such as email or documents, its built-in AI may be enough.

• Answer customer questions from our help pages and policies

  Start withConnect it to your documents (RAG)

  Answering from your own documents is exactly what RAG is for, and the source links let customers and staff check the answer.

  Or: Turn on the AI in software you already use. Most help-desk software now sells this ready-made. Build your own only if theirs can't reach your sources or follow your rules.

• Help people find the right document among thousands

  Start withOrdinary software: search, rules or a database lookup

  Often the answer is better search, not a chatbot: people want the document itself. Search that matches by meaning finds it even when the words differ.

  Or: Connect it to your documents (RAG). Add RAG on top if people also want a short answer with the sources.

• Give exact answers from our systems: order status, balances, stock

  Start withOrdinary software: search, rules or a database lookup

  The answer already sits in a database, so look it up with ordinary software. A model must never be the one making up a number.

  Or: Build on a model's API. AI can still understand a messy question and word the reply, while the number comes from your system (function calling).

• Do the same task thousands of times: sort emails, read invoices

  Start withBuild on a model's API

  A careful prompt through an API, checked against a test set of real examples, handles most of these jobs.

  Or: Fine-tune a model. Fine-tune only if prompting can't reach the accuracy or cost you need at that volume.

• Complete multi-step work across systems: update, book, file

  Start withLet it take actions (an agent)

  Start with one narrow workflow, give it the smallest access it needs and have a person approve anything that can't be undone.

  Or: Build on a model's API. If the steps are always the same, a fixed sequence you design is cheaper and more reliable than letting the AI plan.

• Use AI on data that must never leave our own servers

  Start withAn open-weight model on your own machines

  An open-weight model on your own machines keeps every word in-house. You take on the hardware and the people to run it.

  Or: Through your cloud provider. Often enough: your cloud provider can run models inside the account and region your IT team already approved.

• Predict something from our own numbers: demand, fraud, churn

  Start withTrain a model on your own data

  This is predictive AI, and it's the one area where building your own model is normal. Start simple and compare against a basic rule.

The eight options, from least to most work

"Buy" means someone else built and runs it. "Build" means your team or a contractor owns it, even when the model itself comes from an AI company.

1. 1Ordinary software: search, rules or a database lookupNo AI

   Good search, a few clear rules or a simple lookup in your own system. Modern search can match by meaning as well as by keywords.

   Good for

   Answers that must be exact (an order status, a balance) and people who want the document itself, not a summary of it.

   Wrong choice when

   Questions are open-ended, worded in many different ways, or need an answer written from several sources.

   A "Where is my order?" page that reads the delivery status straight from the system.

   First version: days

2. 2Use a ready-made AI appBuy

   Give staff a business plan of a chatbot such as ChatGPT, Claude, Gemini or Microsoft Copilot. Nothing to build.

   Good for

   Writing, summarising, brainstorming, research and analysis: general work that each person checks themselves.

   Wrong choice when

   The AI must follow your process, reach your systems or serve customers directly.

   A sales team drafts proposals and meeting notes with a shared chatbot plan.

   First version: todayBest AI chatbots →

3. 3Turn on the AI in software you already useBuy

   Most help-desk, CRM, office and accounting software now has AI features: reply drafts, summaries, a help-centre bot.

   Good for

   Common jobs inside one tool, where the vendor already has your data and has done the testing.

   Wrong choice when

   You need data from several systems, or the feature can't be tuned to your rules and tone.

   A help-desk tool's bot answers customers from the articles already in its help centre.

   First version: daysAI tools by job →

4. 4Build on a model's APIBuild

   Your own small app sends text to a model from an AI company and uses the reply. Most of the work is a careful prompt and testing.

   Good for

   One repeated task with a clear right answer: sorting emails, pulling fields from invoices, tagging feedback.

   Wrong choice when

   A ready-made tool already does the job, or nobody will test and maintain it after launch.

   Every incoming invoice is read by a model that fills in supplier, date and total for the finance system.

   First version: days to weeksAPI →Prompting basics →

5. 5Connect it to your documents (RAG)Build

   Before answering, the system searches your documents and gives the best passages to the model, so it answers from them and can cite them.

   Good for

   Questions about your own policies, products and manuals, where the answer is written down somewhere.

   Wrong choice when

   The documents are out of date or contradict each other, or the answer is a number that sits in a database.

   Staff ask "What's our refund rule for group bookings?" and get an answer with a link to the policy page.

   First version: weeksRAG explained →Vector database →

6. 6Let it take actions (an agent)Build

   The model doesn't just answer: it uses tools to look things up, update records or send messages, over several steps.

   Good for

   Multi-step work across systems that follows a pattern but still needs judgement, such as handling a refund request end to end.

   Wrong choice when

   The steps are always the same (ordinary automation is cheaper and more reliable), or mistakes can't be undone.

   An assistant checks a refund request against the policy, prepares the refund and waits for a person to approve it.

   First version: weeks to monthsAI agents →Human in the loop →

7. 7Fine-tune a modelBuild

   Train an existing model further on hundreds or thousands of your own examples, so it picks up a style, format or narrow skill.

   Good for

   A narrow task done at high volume, where a smaller fine-tuned model can match a big one for less, or a very specific format.

   Wrong choice when

   You want it to know your facts: fine-tuning changes behaviour, RAG adds knowledge. Try examples in the prompt first.

   A model fine-tuned on 5,000 past tickets sorts new ones into the company's 40 categories.

   First version: weeks to monthsFine-tuning →LoRA →

8. 8Train a model on your own dataBuild

   Build a prediction model from your own records: classic machine learning, not a chatbot. Training a language model from scratch is almost never worth it.

   Good for

   Predicting from your numbers: demand, fraud, which customers might leave. Chatbots are poor at this; your history is the advantage.

   Wrong choice when

   You don't have years of clean, labelled data, or a simple rule already works almost as well.

   A shop forecasts next month's sales of each product from three years of orders.

   First version: monthsHow a model is built →Predictive AI →

Side by side

A rough guide for a typical small or mid-sized project. Privacy depends less on the approach than on where the model runs.

The second choice: whose model, and where it runs

Options 4 to 7 all need a model you choose yourself. You can use it in three ways, and this is where most privacy questions are settled.

• The AI company's own service

  Use OpenAI, Anthropic, Google or another lab directly, through their business plans or API.

  Good

  The most capable models, nothing to run, pay as you go. Their business plans and APIs don't use your data for training by default.

  Watch out

  Your data leaves your network, so check how long it's kept and where. Models get retired and prices change, so plan to re-test.

  OpenAI's policy: openai.comAnthropic's policy: privacy.claude.com

• Through your cloud provider

  Amazon, Microsoft and Google offer many companies' models inside the cloud account you may already use.

  Good

  Your data stays under the contracts and security your IT team already approved, often with a choice of country or region.

  Watch out

  Not every model, or every new feature, arrives there at the same time, and setup takes more technical work.

• An open-weight model on your own machines

  Download a model such as Llama, Gemma, Qwen, Mistral or gpt-oss and run it on your own servers or laptops.

  Good

  Nothing leaves your building, there are no per-question fees and nobody can retire the model on you.

  Watch out

  You pay for the hardware and the people to run, update and secure it. The best open models are often a step behind the very best closed ones, and licences differ.

  Open-weight models →Run AI on your computer →

Data rules differ for free personal plans, so give staff a business plan rather than letting them use their own accounts for work.

What real decisions teach

• Klarna, 2024 to 2025In 2024 the payments company said its AI assistant was doing the work of 700 customer-service agents. In May 2025 its chief executive said the cost-cutting had led to lower-quality service, and it began hiring people again.Measure quality, not just cost, and keep an easy route to a person.Source: entrepreneur.com
• MIT study, 2025A study of company AI projects found about 95% of pilots showed no measurable return. Projects bought from specialist vendors or built with partners succeeded about 67% of the time, far more often than companies' own builds.Buying is often the safer first step. The study was preliminary and some experts question its method, so treat the numbers as a signal, not a law.Source: fortune.com

More in AI fails: real mistakes and what they teach.

What the price tag leaves out

Whatever you pick, these costs come with it. The more you build, the more of them are yours.

• TestingA set of real questions with the right answers, re-run after every change, and someone who owns it. How evals work →
• Keeping documents currentAn assistant that answers from your documents is only as up to date as they are.
• Watching after launchLogs, user feedback and a person who reads them every week.
• Model changesProviders retire models and change prices. Each switch means re-testing.
• Security and legal checksData agreements, attacks such as prompt injection, and laws like the EU AI Act for uses such as hiring. AI safety and the rules →
• People's timeTraining staff, and handling the cases the AI hands over or gets wrong.
• Usage that growsPay-per-use bills rise with volume. Work out the cost at full scale, not in the pilot, and set spending limits.
• Getting outHow hard is it to switch? Keep your own copy of your data, prompts and tests.

One company, three answers

An illustrative example: a 250-person travel agency with three jobs for AI.

1. Staff write supplier emails and day-by-day trip plans

   2 Bought a business chatbot plan

   General writing that each person checks. Live the same week.

2. Agents look up rules in 400 supplier contracts and visa pages

   5 Built RAG on the contract library

   The answers are written down but hard to find. A contractor built it in six weeks; one person keeps the documents current.

3. Customers ask "Is my booking confirmed?"

   1 Ordinary lookup, with AI to read the question

   The answer is in the booking system. The AI understands the message; the status comes from the database.

Nothing here needed a custom model. The one real build, the contract assistant, was the job where no ready-made tool could reach the documents.

Questions to ask before you decide

Related: the AI tech stack, where each of these pieces sits, and how to tell if an AI system works.