Hire AI Engineers or Train Your Team? I've Advised Both. Here's What Actually Works.
A fintech client called me last spring. "We've been trying to hire an ML engineer for eight months. We've made three offers — all rejected. One candidate accepted, then got a counter-offer for $50K more and left after two weeks. We're bleeding money on recruiter fees and we still can't ship our AI product."
Six months later, I got a different call from a healthcare company. "We spent $120K on an upskilling program. Our developers can now build basic ML models. But our actual business problem needs production-grade NLP with domain expertise in medical terminology. We're back to square one."
Two different companies. Two different strategies. Both failed. And in both cases, the mistake was the same: they picked an approach without understanding what they actually needed.
Related: AI predictions including the skills gap, hidden costs of building AI teams, and production AI deployment.
The Real Salary Landscape (It's Worse Than You Think)
Let me share numbers from roles I've helped fill in the last 18 months:
| Role | US Market Range | What Competitive Offers Look Like |
|---|---|---|
| ML Engineer (mid) | $150K-$220K | $180K + equity + remote |
| Senior ML Engineer | $200K-$300K | $250K + equity + signing bonus |
| ML Ops / Platform | $160K-$250K | $200K + equity |
| Data Scientist (AI-focused) | $130K-$200K | $170K + equity |
| AI Product Manager | $140K-$220K | Hard to find, often promoted internally |
And the hiring funnel reality:
- Average time to fill an ML engineer role: 5-7 months
- Offer acceptance rate: ~40% (compared to 70-80% for standard engineering roles)
- First-year turnover for AI talent: 25-30% (they get poached)
If you're outside a major tech hub (SF, NYC, London), multiply the difficulty by 2x. If you're competing with FAANG compensation, multiply by 3x.
The Three Strategies (And When Each One Works)
Strategy 1: Hire External AI Talent
When it works: You need deep ML expertise right now for a specific, high-value project. You have competitive compensation. You're in a market where AI talent wants to live.
When it fails: You're trying to build an entire AI capability from scratch through hiring. You can't match top-tier compensation. Your AI roadmap is vague ("we need AI people for... AI things").
A startup I advise hired a single senior ML engineer at $260K. Expensive, yes. But she shipped their core AI feature in 4 months — a feature that directly drove $2M in new ARR. That's an incredible ROI.
Another company hired three junior ML engineers for $150K each. Eighteen months later, they'd built several prototypes but nothing in production. The team lacked someone senior enough to make architecture decisions and push through the hard parts of productionization. They eventually hired a senior person anyway.
My rule: If you hire, hire senior. One senior ML engineer who can ship production code is worth three juniors who need guidance you can't provide.
Strategy 2: Upskill Your Existing Team
When it works: You have strong software engineers who want to learn ML. Your AI use cases are well-defined and not bleeding-edge. You have 6-12 months before you need production results.
When it fails: You need results in 3 months. Your AI challenges require deep research-level expertise. Your engineers aren't motivated to learn ML (this is more common than people admit).
The realistic timeline for upskilling a good software engineer to "productive in AI":
| Phase | Duration | Outcome |
|---|---|---|
| Foundations (Python for ML, basic stats) | 2-3 months | Can run notebooks, understands concepts |
| Applied ML (scikit-learn, basic deep learning) | 3-4 months | Can build simple models, evaluate results |
| Production ML (MLOps, deployment, monitoring) | 3-4 months | Can deploy models, handle real data |
| Specialization (NLP, CV, agents, etc.) | 3-6 months | Can tackle domain-specific problems |
Total: 12-18 months before someone is genuinely productive on production AI work. Anyone who tells you it's faster is selling a course.
The cost: ~$5K-15K per person in training materials and time. The real cost is 12-18 months of reduced productivity while they learn. For a 5-person team, that's significant.
But the upside is huge: trained engineers already understand your domain, your codebase, and your data. They don't need 3 months of onboarding that a new hire would. And they're less likely to leave for a $50K raise.
Strategy 3: The Hybrid (What I Usually Recommend)
Most companies should do both — but in a specific sequence.
Phase 1 (Months 0-6): Bootstrap with external expertise. Hire one senior ML engineer or bring in a consultant (hi). This person defines the architecture, builds the initial system, and establishes patterns. Budget: $150K-250K (hire) or $75K-150K (consultant engagement).
Phase 2 (Months 3-18): Upskill your team in parallel. While the senior person is building, start training 2-4 engineers. They learn by contributing to the actual AI project — not abstract coursework. The senior person mentors them. Budget: $40K-80K in training + reduced productivity.
Phase 3 (Months 12-24): Selective hiring for specialization. By now you know what specific AI skills you actually need (not what you think you need). Hire for those gaps. Maybe it's an NLP specialist, or an ML Ops person, or a second senior engineer. Budget: varies.
Why this works: You get production results in 3-6 months (from the senior hire), build internal capability for the long term (from upskilling), and make targeted hires based on actual needs instead of speculation.
The ROI Math (Honest Version)
For a 5-person AI-capable team over 3 years:
| Approach | Total Cost | Time to First Production System | Risk |
|---|---|---|---|
| All new hires | $2.5M-4.5M | 6-9 months (hiring + onboarding) | High turnover, culture fit |
| All upskilling | $800K-1.2M | 12-18 months | Slow, may not reach required depth |
| Hybrid | $1.5M-2.5M | 3-6 months | Balanced, but requires coordination |
The hybrid approach costs more than pure upskilling but delivers results 6-12 months sooner. At most companies, that time-to-value difference is worth far more than the cost difference.
The Questions I Ask Before Recommending
When a client asks "should we hire or train?" I ask:
- When do you need results? If <6 months → you need external talent. If 12+ months → upskilling is viable.
- How specific is your AI challenge? Generic (chatbot, classification) → upskilling works. Specialized (medical NLP, custom CV) → hire an expert.
- Do your engineers actually want to learn ML? Not everyone does. Forcing it creates resentment and poor outcomes.
- What's your retention risk? If you're in a market where AI talent gets poached every 6 months, upskilling your loyal engineers is more sustainable than hiring.
- Can you define the role? If you can't write a specific job description with concrete deliverables, you're not ready to hire. Train instead — it'll help you figure out what you actually need.
The Mistake Everyone Makes
The biggest mistake I see: treating AI talent strategy as a one-time decision. "We decided to upskill" or "we decided to hire" and then never revisiting.
Your AI needs will evolve. The team that was right for your pilot project won't be right for your production system. The person you hired for model development might not be the person you need for MLOps.
Revisit your talent strategy every 6 months. What skills do you need now that you didn't need before? What skills did you think you'd need but don't? Who on the team has grown into new capabilities?
The companies that do AI talent well aren't the ones who made the perfect initial decision. They're the ones who adjust continuously.
Building your AI team and unsure where to start? I help companies design talent strategies that balance speed, cost, and long-term capability. Let's talk.