The Build vs Buy vs Partner Decision for ML

The Three Options

Every company at some point faces this decision: we need ML capabilities. Do we build an in-house team, buy a SaaS solution, or partner with a consulting firm?

The answer depends on three things: how core ML is to your competitive advantage, your budget and timeline, and the complexity of your problem. Here's the framework we use with our clients.

Option 1: Build In-House

What it looks like: Hire ML engineers, data engineers, and MLOps engineers. Build your own infrastructure. Own everything.

The Real Cost

Most companies dramatically underestimate the cost of building in-house ML capabilities.

Year 1 costs (US market):

• Senior ML Engineer: $180-250K (salary + benefits)
• ML Engineer: $140-190K
• Data Engineer: $150-200K
• MLOps Engineer: $160-210K
• Infrastructure (GPU compute, storage, tools): $50-150K
• Recruiting costs: $60-100K (20-25% of first-year salary per hire)

Minimum viable ML team: 3-4 engineers = $500-800K/year before they ship anything.

Time to first production model: 6-12 months (including hiring, onboarding, infrastructure setup, data preparation, model development, deployment).

When to Build

Build in-house when:

• ML is your core product or primary competitive advantage
• You have >$1M annual budget for ML
• Your timeline is 12+ months
• You need continuous model iteration and domain expertise accumulation
• You're willing to invest in infrastructure and tooling

When Not to Build

Don't build when:

• You need results in less than 6 months
• The ML component is supporting, not core
• You can't attract and retain ML talent in your location/compensation range
• You don't have the data infrastructure to support ML

Option 2: Buy SaaS

What it looks like: Subscribe to an ML-powered SaaS product. Someone else handles the models, infrastructure, and maintenance.

The Real Cost

SaaS ML products typically charge per prediction, per user, or per data volume:

• Document processing: $0.01-0.10 per page
• Image classification: $0.001-0.01 per image
• NLP APIs: $0.01-0.05 per 1K tokens
• Predictive analytics platforms: $2-10K/month

Year 1 cost: $24K-120K depending on volume and complexity.

When to Buy

Buy SaaS when:

• The problem is well-commoditized (OCR, basic sentiment, standard object detection)
• Speed to market is critical (need results in weeks, not months)
• You don't need to own or customize the model
• The data can leave your infrastructure (no regulatory constraints)
• Volume economics make sense at your scale

When Not to Buy

Don't buy when:

• Your problem requires domain-specific training data or custom models
• You need to run inference on-premise or on edge hardware
• Vendor lock-in is unacceptable
• Your data can't leave your infrastructure (defense, healthcare, financial services)
• You need sub-10ms latency that cloud APIs can't provide

Option 3: Partner (Consulting)

What it looks like: Engage an ML consulting firm for scoped engagements. They bring the expertise, you retain the IP.

The Real Cost

ML consulting ranges widely:

• Freelancers: $100-200/hr ($16-32K for a 4-week project)
• Boutique firms: $150-400/hr ($24-64K for a 4-week project)
• Big 4 / enterprise consultancies: $300-600/hr ($48-96K for a 4-week project)
• Value-based pricing: $5-50K per project (our model)

Year 1 cost for ongoing engagement: $60-300K depending on scope and model.

When to Partner

Partner when:

• You need production ML but can't justify a full-time team
• The project requires specialized expertise you don't have
• You need results in 4-12 weeks
• You want to retain ownership of models and IP
• You're evaluating whether ML is worth deeper investment

When Not to Partner

Don't partner when:

• You need long-term, continuous model development (a retainer can work, but in-house is eventually better)
• The consulting firm wants to own the IP
• The engagement scope is unclear
• The consulting firm has no production deployment experience

The Decision Framework

Answer these five questions:

1. Is ML core to your competitive advantage?

• Yes → Lean toward Build (but maybe Partner first to accelerate)
• No → Buy or Partner

2. What's your timeline?

• < 3 months → Buy or Partner
• 3-12 months → Partner or Build
• 12+ months → Build

3. What's your annual ML budget?

• < $50K → Buy
• $50-500K → Partner
• $500K+ → Build or Partner + Build

4. Can your data leave your infrastructure?

• Yes → All options viable
• No → Build or Partner (no SaaS)

5. How custom is your problem?

• Standard (OCR, sentiment, basic classification) → Buy
• Domain-specific but well-defined → Partner
• Novel and continuously evolving → Build

The Hybrid Approach

In practice, the best approach is often a combination:

Phase 1: Partner — Engage a consulting firm for a scoped sprint ($5-20K, 4-8 weeks). Get a production model deployed, learn what works, validate the business case.

Phase 2: Evaluate — With a working system, you now know the real requirements, challenges, and value. This informs the build vs buy decision with actual data, not assumptions.

Phase 3: Scale — Either grow the partnership into a retainer, transition to in-house by hiring (using the consulting firm's architecture and code as a foundation), or switch to a SaaS solution if one fits.

This is exactly how we structure our engagements at Opulion. Start with a low-risk sprint, prove value, then scale in the direction that makes sense.

The Hidden Costs to Watch

Regardless of which option you choose, watch for these hidden costs:

• Data preparation: 60-80% of ML project effort, regardless of who's doing the work
• Integration: Connecting the ML system to your existing infrastructure
• Maintenance: Models degrade over time and need retraining
• Monitoring: You need to know when the model stops working
• Opportunity cost: Time spent on ML is time not spent on other priorities

The cheapest option isn't always the best option. The best option is the one that gets a working, monitored, maintainable ML system into production fastest — because that's when you start learning whether ML actually moves the needle for your business.