The Default Should Be Buy

Here's the uncomfortable truth: Most companies that build custom AI shouldn't.

Building AI is expensive, slow, and risky. It requires specialized talent that's hard to hire and retain. It demands ongoing maintenance forever. And in most cases, someone else has already built what you need.

The instinct to build is often ego, not strategy.

This chapter helps you make the right call — and avoid the vendors who will take your money without delivering value.

The True Cost of Building

The Obvious Costs

• AI/ML talent (expensive and scarce)
• Computing infrastructure
• Data engineering
• Project management
• Time to develop

The Hidden Costs

Maintenance forever: Models degrade. Data changes. Systems need updates. You're signing up for ongoing costs indefinitely.

Talent retention: AI talent is mobile. When your ML engineer leaves, you have a problem.

Opportunity cost: Resources spent building could be spent on your actual business.

Technical debt: Custom systems become legacy systems. Someone will inherit your choices.

Security and compliance: You own the risk. Vendor shifts some risk to them.

Typical Build Costs

Very rough order of magnitude:

Simple model: $100K-300K, 3-6 months Moderate complexity: $300K-1M, 6-12 months
Complex system: $1M-5M+, 12-24+ months

Plus 20-40% annually for maintenance.

These numbers surprise most leaders. Building is expensive.

When to Buy

Default to buying when:

• Your use case is common (many companies have this problem)
• Vendors exist with proven solutions
• You don't have AI talent in-house
• Time to value matters more than customization
• The problem is not a competitive differentiator

Common buy scenarios:

• Customer service chatbots
• Email/document processing
• Fraud detection (unless you're a bank)
• HR screening
• Basic forecasting
• Content generation
• Standard image recognition

For most companies, 80%+ of AI opportunities should be solved with purchased solutions.

When to Build

Consider building when:

• Your use case is genuinely unique
• No vendor solves your specific problem
• The AI is a competitive differentiator
• You have data no one else has
• You have (and can retain) the talent
• Long-term economics favor building

Common build scenarios:

• Proprietary algorithms that create competitive advantage
• Unusual data types or formats
• Highly specialized domain requirements
• Integration requirements no vendor supports
• Strategic capability you want to own

The bar should be high. "We're unique" is usually wrong.

The Middle Ground: Build on Top of Buy

Often the best approach: Buy foundation, customize on top.

Examples:

• Use OpenAI/Claude APIs, build custom applications
• Use cloud ML platforms, train on your data
• Use no-code AI tools, customize workflows
• Use vendor products, integrate with your systems

This gets you:

• Foundation capabilities without building from scratch
• Customization where you need it
• Faster time to value
• Lower risk

Evaluating Vendors

When you decide to buy, you need to evaluate vendors without getting fooled.

Red Flags

Demo-driven selling: All polish, no substance. Ask to see failures and edge cases.

Vague accuracy claims: "State-of-the-art accuracy" without numbers. Demand specifics.

No reference customers: If they can't connect you with happy customers, why not?

Roadmap selling: Features promised but not delivered. Buy what exists.

Opaque pricing: Unclear costs that balloon after commitment.

Lock-in by design: Proprietary formats, hard to export data, difficult to leave.

Overselling AI: Everything is "AI-powered" but they can't explain what AI actually does.

Questions to Ask Vendors

About the product:

• How specifically does AI work in your product?
• What accuracy can we expect, and how was that measured?
• What are the failure modes and edge cases?
• What data do we need to provide?
• How long to implement and see results?
• What does the pricing model look like at scale?

About customers:

• Who are your comparable customers?
• Can we talk to reference customers?
• What results have they achieved?
• What problems have they encountered?
• Why have customers left?

About the company:

• How long have you been in business?
• What's your funding/revenue situation?
• What's your AI team look like?
• What's your roadmap and how is it set?
• What happens if you're acquired?

Due Diligence Process

Step 1: Market scan Identify 3-5 potential vendors. Research online. Analyst reports. Peer recommendations.

Step 2: Demo and discovery See the product. Ask hard questions. Understand capabilities and limitations.

Step 3: References Talk to actual customers. Ask about implementation, results, and problems.

Step 4: Pilot Test with your data, your users, your processes. Verify claims.

Step 5: Negotiate Contract terms, pricing, SLAs, data ownership, exit provisions.

Pilot Best Practices

Define success upfront: What metrics prove value? Agree before starting.

Use real data: Curated demos don't predict production.

Include skeptics: Enthusiasts will make anything work. Include the doubters.

Set a time limit: Two to four weeks. Longer pilots become commitments.

Document everything: Results, problems, learnings. You'll need this for decision-making.

Evaluate integration: How hard will full deployment be?

Contract Considerations

When you've found your vendor, negotiate carefully:

Data Rights

• Who owns the data you provide?
• Can they use your data to improve their models?
• What happens to your data if you leave?
• How do you export your data?

Service Level Agreements

• What uptime is guaranteed?
• What response time for issues?
• What remedies for failures?

Pricing Protection

• How are costs calculated?
• What happens if usage grows?
• What are the renewal terms?
• How does pricing compare to building after 3 years?

Exit Provisions

• What's the notice period?
• How do you migrate away?
• What data portability exists?
• What formats are available?

Change Control

• How are product changes communicated?
• What happens if features you depend on are removed?
• What's the deprecation policy?

The Build Decision Framework

If you're seriously considering building:

Checkpoint 1: Is this a core differentiator? If this AI capability isn't central to competitive advantage, default to buy.

Checkpoint 2: Do we have the data? Building requires data you control. Is it available and usable?

Checkpoint 3: Do we have the talent? Building requires ML engineers, data engineers, MLOps. Do we have them? Can we keep them?

Checkpoint 4: Do we have the patience? Building takes time. Can we wait 12-24 months? Will the opportunity still exist?

Checkpoint 5: Do we have the budget? Including maintenance forever. Is leadership committed?

If any checkpoint fails, reconsider.

Hybrid Approaches

Often the best strategy isn't pure build or buy:

Start with buy, build later: Validate the use case with a vendor. If it's strategic and working, consider building your own.

Buy foundation, build differentiation: Use APIs and platforms for commoditized capabilities. Build only where you need to be unique.

Build pilot, buy scale: Build a prototype to prove value and learn. Use vendor for production.

Partner: Some vendors will customize extensively or build jointly. Explore partnership models.

Making the Decision

For each opportunity:

1. List vendor options (5-minute web search reveals most)
2. If strong vendors exist, default to buy
3. If building seems necessary, run through the build checkpoints
4. If checkpoints pass, consider hybrid approaches first
5. Build from scratch only if genuinely required

Most roads lead to buying. That's okay. Your competitive advantage is probably not in AI infrastructure — it's in how you serve customers, what you sell, and how you operate.

AI is a tool. Buy the best tool. Focus on using it well.