AI software development: choose, integrate or build

When is standard AI software enough, and when does custom development pay off?

Standard AI software is enough for anything a competitor could also buy from a vendor; custom development only pays off for the layer that holds your own knowledge, such as rules, prompts, test sets and integrations. In many cases, the outcome is a combination: buying the model and platform, having the knowledge layer built to your specification, and placing the implementation with a partner. Anyone who doesn't want to run the selection themselves can have it guided with the Vendor Review, which takes 3-6 weeks.

What determines the real cost of AI software?

Integration, data and exit determine the real cost of AI software, not the licence. The Cost criterion in the workable set of selection criteria covers the licence, usage at scale, integration, support and exit over three years.

When should the selection criteria for AI software be set?

Before vendors come into view: the advice is to set the selection criteria for AI software and their weighting at that point, because criteria written after the first demo describe that demo. After that, the advice is to give every vendor the same scenario on their own data.

Does the consultant receive a payment from vendors for a Vendor Review on AI software?

No. For a Vendor Review on AI software, the consultant receives no payment from vendors, so the advice can also be to buy nothing.

Source: article by Dave Willemstein. Updated: October 2026.

Types of AI software

AI software is software in which a model does part of the work: understanding or generating text, recognising images, making predictions, or carrying out tasks. In practice, you’ll run into four forms, from light to heavy:

1. AI features inside software you already have. Your office suite, CRM or service desk gains an assistant or a smart feature. Little integration work, but also little influence over how it works.
2. Standalone AI tools. Chat assistants, writing and analysis tools, transcription. Quick to adopt, usually licensed per user. A separate article on this site covers common categories of AI tools for business.
3. Specialised AI platforms. Software for a single domain, such as document processing, demand forecasting or customer service, that you configure and connect to your own systems.
4. Custom AI software development. An application built for your specific process, usually around a bought-in language model, with your own rules, integrations and evaluation. Examples of what that can deliver are covered on the page about custom AI solutions.

The further down this list you go, the more influence you get over the outcome, and the more you have to manage yourself.

Standard or custom AI software development?

The core question isn’t whether you can have something built, but whether you’ll still want to maintain it in a few years. Only build what sets you apart in the market. If your competitor can buy the same result from a vendor tomorrow, it’s a commodity, and you buy it too.

1. Buy the building blocksLanguage models, infrastructure, standard components: vendors have the scale advantage here
2. Configure the platformSet thresholds, rules and workflows within standard software
3. Have the knowledge layer builtYour rules, prompts, test sets and integrations: small in size, large in value
4. Keep the evaluation yourselfThe criteria you judge the outcome by always stay yours

In many cases, the answer is a combination. You buy the model and the platform, have the layer with your own knowledge built, and place the implementation with a partner. Choosing between standard and custom also calls for an honest cost comparison between the two.

Integrating AI software

The software itself is often the smallest part of the work. Integrating AI software with your existing systems takes the most time and produces the most surprises. Five points to settle upfront:

• Data sources and quality. Which systems need to be opened up, how current and complete is the data, and who manages those sources? If your data foundation is shaky, the AI software is the first thing to show it.
• Permissions and access. The AI shouldn’t see more than the user calling it can see. Check how your source systems’ permissions model is passed through.
• Privacy and data protection. Where is the data processed and stored, is it used for training, and is there a data processing agreement? With personal data, that’s not a detail.
• Logging and oversight. Record which input led to which outcome, and who confirmed it. Without an audit trail, you can’t trace errors back.
• Support after go-live. Who measures quality, who keeps test sets current, and what happens if the vendor replaces the underlying model?

The building and integrating is done by your own IT team or an implementation partner you contract directly. I don’t build alongside them, but I help draw up the requirements and test whether the delivery meets them. How that fits into the whole development journey is covered on the AI development page.

Selection criteria

Set your criteria and their weighting before you see vendors. Criteria written after the first demo describe that demo. A workable set:

Criterion	Key question
Functional fit	Does it cover your scenarios, including the awkward ones?
Performance on your own data	How well does it perform on your material, not the demo material?
Integration	How much work is connecting it to your systems, and who does that?
Data and privacy	Where does the data sit, what happens to it, and is it used for training?
Manageability	Can you adjust thresholds, rules and exceptions yourself?
Cost	Licence, usage at scale, integration, support and exit over three years
Vendor	Continuity, roadmap and support in your language and time zone

Give every vendor the same scenario on your own data, have assessors score it individually first, and calculate the total cost. A full eight-step approach to this process is covered elsewhere on this site, in an insights piece on AI vendor selection.

Don’t want to run that process yourself? I can guide it with a Vendor Review: criteria upfront, scorecards, validation on real data and a TCO comparison including exit costs. The case on vendor selection for an enterprise retailer shows how that plays out in practice. I don’t receive any payment from vendors, so the advice can also be that you buy nothing.

Looking instead for someone who assesses or builds the technical side? A separate article on this site sets out the different AI-development profiles. And if you’re unsure which route fits, book a strategy call.