Build vs Buy: AI Personalization Infrastructure

TL;DR

• Software infrastructure projects routinely exceed initial estimates. A McKinsey and Oxford University study [1] of 5,400 IT projects found they run 45% over budget on average, and that software projects carry the highest risk of overruns
• The real cost is not engineering hours but learning delay. Every month building infrastructure is a month your AI product stays generic and you do not learn what your users actually need
• The decision framework comes down to one question: is personalization infrastructure your competitive advantage, or is what you do with the personalization your competitive advantage

Build vs buy for AI personalization infrastructure comes down to whether personalization is your core product or a capability your product needs to be good. Software infrastructure projects consistently exceed estimates. According to research from McKinsey and the University of Oxford [2], large IT projects run 45% over budget on average, with every additional year spent on a project increasing cost overruns by 15%. This post covers the iceberg problem of hidden infrastructure costs, a decision framework based on competitive advantage, and why the real cost is delayed learning about your users.

The Iceberg Problem

Building a personalization prototype is straightforward. Store some user data, use it to customize responses, ship it. The prototype works within a week.

The iceberg is everything below the waterline. Google researchers described this phenomenon in their influential NeurIPS paper Hidden Technical Debt in Machine Learning Systems [3], finding that only a small fraction of real-world ML systems is composed of the actual ML code. The vast majority is surrounding infrastructure: data pipelines, configuration, monitoring, and serving systems.

Data model evolution. Your initial user model will be wrong. You will discover new dimensions that matter, realize some dimensions are noise, and need to restructure your data model multiple times. Each restructuring risks breaking existing user models.

Consistency at scale. A user model that works for 100 users does not work for 100,000. Query patterns change. Storage costs scale non-linearly. Real-time personalization requires caching strategies, data pipelines, and consistency guarantees.

Edge cases that multiply. Users who delete their account and come back. Users who share devices. Users whose goals change dramatically. Users who actively mislead the system. Each edge case requires engineering time and testing.

The evaluation problem. How do you know your personalization is working? You need an evaluation framework, alignment scores, fit metrics, and A/B testing infrastructure. That is itself a significant engineering project.

Ongoing maintenance. Models drift. User behavior changes. New product features require new understanding dimensions. The personalization layer is never done. According to Stripe’s Developer Coefficient report [4], developers already spend roughly 42% of their time on maintenance tasks like debugging and refactoring. Adding a custom personalization layer increases that burden significantly.

The Decision Framework

The build-vs-buy decision for AI personalization comes down to one question: is the personalization layer your competitive advantage, or is what you do with personalization your competitive advantage?

Geoffrey Moore’s core vs context framework [5] applies directly here. “Core” activities create sustainable differentiation. “Context” activities are necessary but not differentiating. The strategic move is to invest engineering resources in core and outsource context wherever possible.

If you are building a personalization platform, if understanding users IS your product, then build. You need deep control, custom architecture, and the ability to innovate at the infrastructure level.

If personalization is a capability your product needs to be good, if you are building a financial advisor, a health coach, an educational platform, or a productivity tool that needs to know its users, then buy. Your competitive advantage is the financial advice, the health coaching, the curriculum, or the workflow. Not the infrastructure that stores user models.

This is the same logic that applies to every build-vs-buy decision. Thoughtworks’ strategic framework for evaluating third-party solutions [6] frames it well: buying offers “proven capabilities quickly, at the cost of customization and control,” while building provides “exactly what you want, but at greater expense and effort.” The key is knowing which category personalization falls into for your business.

The real cost is opportunity cost. The months and senior engineers spent on infrastructure are not free. What could they have built instead? Better product features? Faster iteration cycles? Competitive differentiation?

When Building Makes Sense

Building your own makes sense in specific situations:

Your personalization is truly novel. If your product requires a fundamentally different approach to user modeling that no existing tool supports, not just different features but a different paradigm, building may be necessary. This is rare.

You need maximum data control. If regulations or contracts prohibit any user data from touching third-party systems, building in-house may be required. Though most modern platforms offer data residency and privacy guarantees that satisfy SOC 2 and GDPR requirements.

Personalization IS your product. Netflix, for example, has invested hundreds of engineers over two decades into its recommendation system. Research published in 2025 [7] found that replacing Netflix’s recommendation system with a simpler algorithm would reduce engagement by 4-12%. When personalization is the product, the investment makes sense.

You have deep expertise and time. If you have a team with specific expertise in user modeling and you are not under time pressure to ship, building can produce a better-tailored solution. But the time-pressure caveat is critical. The Standish Group’s CHAOS research [8] consistently finds that large projects succeed less than 10% of the time, while small, focused projects achieve roughly 90% success rates.

When Buying Makes Sense

Buying makes sense in the more common situations:

You need personalization to make your product good. Your product is a financial advisor, not a personalization platform. You need to know users to give good advice, but the advice is your value. Not the knowing infrastructure.

You are under time pressure. Competitors are shipping personalized experiences. Your users are churning because the product feels generic. You need personalization now, not months or years from now.

Your engineering talent is scarce. Senior ML engineers command median salaries above $185,000-$285,000 [9] depending on experience and location. Using them to maintain infrastructure they did not join to build is a retention risk.

You want to learn fast. Every month without personalization is a month without data on what personalization does for your product. Buying lets you start learning immediately and iterate on what matters, the product experience, rather than the plumbing.

1// Start personalizing in days, not months← What buying looks like
2import { Clarity } from '@clarity/sdk';
3
4const clarity = new Clarity({ apiKey: process.env.CLARITY_KEY });
5
6// Learn about the user from each interaction
7await clarity.observe(userId, {
8  context: 'product_usage',
9  observation: userMessage
10});
11
12// Get the user model for personalization
13const selfModel = await clarity.getSelfModel(userId);
14
15// Your team focuses on: what to DO with this knowledge
16// Not on: how to store, query, and maintain it

The Hidden Cost Nobody Calculates

There is a cost that never appears in build-vs-buy spreadsheets: the cost of delayed learning.

When you spend months building personalization infrastructure, you spend those months without data on how personalization impacts your product. You do not know which understanding dimensions matter most. You do not know how users respond to personalized experiences. You do not know what quality level is good enough.

When you buy, you start learning on day one. Within weeks, you know which dimensions matter. Within months, you have data on personalization’s impact on retention. By the time a build-your-own team is finishing v1, you have iterated through multiple versions of your personalization strategy, guided by real user data.

The learning advantage compounds. Every month of real-world data makes your personalization better. The team that starts learning earlier has a significant, durable advantage. As McKinsey’s 2024 Global Survey on AI [10] found, 65% of organizations now use generative AI regularly, nearly double the previous year, but most still struggle to move from pilot to production scale. The teams that reduce time-to-learning close that gap faster.

Trade-offs

Buying means some loss of control. You depend on a vendor’s roadmap, reliability, and pricing decisions. Mitigate this with contracts, SLAs, and data portability guarantees.

Building means deeper understanding of the problem. Teams that build their own develop intuition for user modeling that teams who buy do not. This matters if personalization is core to your long-term strategy.

Hybrid approaches exist. Some teams buy initially to start learning, then build their own once they understand the problem deeply enough. This is often the optimal path: buy for speed, build for control once you know what you need.

What to Do Next

1. Estimate your real build cost honestly. The McKinsey/Oxford study [11] found large IT projects average 45% over budget. Include ongoing maintenance headcount. Compare this to the subscription cost of buying. More importantly, calculate the opportunity cost: what would those engineers build instead?

2. Identify your competitive advantage. Write one sentence: Our competitive advantage is [X]. If X involves user modeling infrastructure, consider building. If X involves the product experience that user modeling enables, consider buying.

3. Start learning today. Whether you build or buy, start collecting user understanding data now. Every day without a user model is a day your AI stays generic. See how Clarity gets you to personalization in days.

You can build your own database. You can also build your own payment processor. The question is the same: is infrastructure your competitive advantage, or is what you build on top of it? Start personalizing today.

References

1. McKinsey and Oxford University study
2. research from McKinsey and the University of Oxford
3. Hidden Technical Debt in Machine Learning Systems
4. Stripe’s Developer Coefficient report
5. core vs context framework
6. Thoughtworks’ strategic framework for evaluating third-party solutions
7. Research published in 2025
8. Standish Group’s CHAOS research
9. $185,000-$285,000
10. McKinsey’s 2024 Global Survey on AI
11. McKinsey/Oxford study

Related

• When to Add Personalization to Your AI Product
• What Segment and Amplitude Miss About Your Users
• The Personalization Stack Is Broken: Here’s the Missing Layer