Blog

Build, buy or get help? How to choose the right AI solution

Breytt 6 min read
A team in discussion around a table with laptops

Once a company has found a task that AI can handle, the next question arrives quickly: should we build the solution ourselves, buy something off the shelf, or bring in help? Some companies spend months on this decision. Others make it in a single meeting, based on a demo someone happened to see. Neither is a good idea.

Fortunately, the decision gets much simpler when you ask the right questions. Few of them are technical. They're about cost, risk, time and the one thing that's usually forgotten: who will look after the solution six months from now.

Three routes, three different commitments

Buy. A ready-made solution, usually on a subscription. You get something that works straight away, and someone else handles updates, operations and security. In return, you adapt to the product rather than the other way round, and you depend on the vendor's pricing and roadmap.

Build. A solution you own and control. It can fit your processes and data perfectly. But building it is only the beginning. Someone has to maintain it, keep up with changes to the models it relies on and fix it when something breaks.

Get help. An outside partner helps you choose, adapt or build, and hands over both a solution and the know-how behind it. This route costs more up front but often saves expensive mistakes. The key is that the knowledge stays with you when the project ends.

None of these routes is right or wrong in itself. Each suits particular circumstances, and the same company may well choose differently for different tasks.

Five questions that settle it

Work through these five questions for each project before you sign a contract or start building.

  1. Is the task specific to you? If thousands of companies face the same problem, such as summarising meetings or answering common questions, there's a good chance a solid product already exists. If the task depends on processes or knowledge unique to your business, the case for building gets stronger.
  2. What data does it need? A solution that needs access to sensitive data, such as contracts, customer records or financials, calls for a careful look at where that data is stored and processed. That applies whether you buy or build.
  3. Who will maintain it? Ask who looks after the solution in six months, when the underlying model has been updated and the person who built it has moved on to another project. If nobody can answer, building is probably not the right route, at least not yet.
  4. How quickly do you need results? A ready-made product can be in use within days. A custom build takes weeks or months. Sometimes the sensible move is to start with an off-the-shelf tool, learn from it and build later, once your needs are clearer.
  5. What's the total cost? Compare costs over two to three years, not just the initial outlay. Include licences, usage fees, set-up and integration work, training and maintenance. The cheapest option on day one is rarely the cheapest over time.

The traps on each route

Each route has its typical pitfalls, and they're worth knowing before you set off.

When buying, companies often forget to start small. Licences are bought for the whole workforce before anyone knows who will use the product, or for what. Start with a defined group and expand once real usage appears. Read the data terms carefully, too. The free version and the business version of the same service are often governed by different rules.

When building, it's easy to forget how far a prototype that works in a demo is from a solution that works in daily operations. Error handling, access control, monitoring and documentation often take more time than the AI itself. Plan for that work from the start.

When getting help, the risk is becoming dependent on the consultant. Insist that documentation, access and know-how stay with you, and that someone in-house takes an active part in the project.

A good consultant makes themselves unnecessary.

Test before you commit

Whichever route you choose, it pays to test before making a big commitment. Many off-the-shelf products offer a trial or a small plan you can use to see how they perform on your real data, not just on the vendor's showcase examples.

If you're leaning towards a custom build, start with a simple prototype that solves a narrow slice of the problem. It doesn't need to be pretty. It needs to tell you whether the technology can handle the task and whether the people who will use it actually want to.

Decide in advance what the test should prove. What quality is good enough? How much time does the solution need to save to be worth it? Without those criteria, tests tend to end in a vague feeling that it's "quite clever", which is no basis for either carrying on or stopping.

The answer is usually a mix

In practice, few companies need to pick a single route for everything. A sensible split is often to buy general-purpose tools for everyday work, build a thin layer of integration where they connect to your own systems, and bring in help for anything that needs specialist expertise, such as security, architecture and vendor selection.

The key is to make the decision deliberately for each project, not once for the whole company. What works for meeting summaries won't necessarily work for a knowledge base built on confidential documents.

Write the decision down, along with the reasoning behind it. A year from now, when the market has moved on, you'll want to see why you chose that route and whether the assumptions still hold.

Next steps

This is exactly what we cover in our talk, Build, Buy, or Partner?, designed for leaders preparing to invest in AI. It walks through the decision process, the real costs, security considerations, vendor selection and the questions boards need answered.

If the decision is already made and you want to put a solution to work, we can help with Applied AI, from choosing the solution to everyday use.