Question one: is the job common or specific to you?

Start by describing the job in a single sentence. Then ask how many other businesses would write roughly the same sentence.

If the answer is thousands, you are almost certainly looking at a buy. Transcribing meetings, drafting marketing copy, summarising documents, answering routine customer questions and extracting data from standard invoices are common jobs. Vendors have already met the edge cases and built the integrations, so you would be paying to rediscover what they know.

If the sentence only makes sense inside your company, the balance shifts. Pricing that depends on your particular mix of contracts, forecasting that relies on data only you hold, or a workflow that runs across three of your own systems in a specific order are all signs that an off-the-shelf tool will force you to bend your process around its assumptions.

Question two: does it give you an edge?

The second question is strategic. If this AI capability works brilliantly, does it make you better than your competitors, or does it simply bring you up to the same standard as everyone else?

For anything that keeps you level with the market, buy. There is no prize for building your own version of something your competitors can purchase for a monthly subscription. Save your effort for the areas where your knowledge, data or way of working is what customers choose you for. That is where a tailored system compounds over time.

A useful test is to imagine your main competitor buying the same tool tomorrow. If that would not worry you, the capability is not a differentiator, and buying is the sensible route.

Question three: where does your data live, and how messy is it?

AI is only as good as the data it can reach. Many tools look excellent on their own sample data and then struggle when they meet a real business, with its duplicate records, inconsistent naming and information split across systems that were never designed to talk to each other.

Before choosing either path, map where the data for this job actually lives and what condition it is in. If the tool can connect cleanly to the systems you already use and your data is in reasonable shape, buying becomes much easier. If the data is scattered and needs significant work to bring together, be honest that this work is needed whichever route you take.

Question four: what does it really cost over three years?

The sticker price is the least useful number in the comparison. Buying looks cheap at first, but costs can rise with usage, with the number of seats, or when you need a higher tier for the features or volumes you actually require. A good habit is to model a realistic month of your own work against the pricing, rather than trusting the entry plan on the landing page.

Building looks expensive at first, and it is. The bigger cost, though, is what comes after launch: monitoring, maintenance, retraining as your data changes, and keeping the people who understand the system. Engineers at Google made the same point in their widely cited paper on hidden technical debt in machine learning systems, which found that real-world machine learning systems commonly carry heavy ongoing maintenance costs. A custom AI system is not a project with an end date. It is a capability you agree to look after.

Lay both options out over three years, including the people needed to run them.

Question five: how much control do you need?

Some jobs need more control than a vendor can offer. If the AI makes decisions that affect customers, money or regulated information, you may need to explain exactly how each output was reached, keep the data within particular boundaries, or change the logic quickly when rules change.

Ask any vendor how they handle these needs. Can you see why the tool produced a given answer? Where is your data processed and stored? Can you set limits on what it is allowed to do on its own? The UK Information Commissioner’s Office and The Alan Turing Institute have published practical guidance on explaining decisions made with AI, and it makes a good checklist for what a vendor should be able to show you. If the answers are vague and the stakes are high, that is a strong signal in favour of building, or at least of keeping a human firmly in the loop.

Question six: can you live with the vendor’s roadmap?

When you buy, you also buy into someone else’s priorities. Features you rely on can change, prices can move and products can be discontinued. So check how easily you could leave. Can you export your data and any configuration you have built up? Would switching to an alternative take a week or a year?

The third option: blend

In practice, the best answer is often neither pure build nor pure buy. Many businesses use proven tools for the standard parts of a process and build a thin, tailored layer where their needs are different. A company might buy a strong document-reading tool and then build its own logic to apply company-specific rules to what that tool extracts. That approach gets you the speed of buying and the fit of building, without taking on more than you need to own.

At Critical Future, the most productive build or buy conversations I have start with exactly this kind of mapping: separating the parts of a problem that are common from the parts that are specific, before anyone discusses technology.

Putting the framework to work

If you want a quick way to apply all of this, score your project against six questions:

1. Is the job common across many businesses, or specific to yours?
2. Would doing it brilliantly give you a real competitive edge?
3. Is your data accessible and in reasonable shape?
4. Which option is cheaper over three years once running costs are included?
5. Do you need control or explainability a vendor cannot offer?
6. How easily could you switch if the vendor changed direction?

Mostly common jobs, no competitive edge, accessible data and modest control needs point firmly to buying. Specific jobs, a real edge, heavy control needs and data only you hold point towards building. Anything in between is usually a candidate for a blend.

The takeaway

Build or buy is not a question about whether your team can build AI. Plenty of teams can. It is a question about where your effort creates the most value. Buy what is common, build what makes you different, and be honest about the data work and running costs either way. Make the decision on those terms and you will spend your AI budget where it actually moves the business forward.