Your board deck says “AI roadmap” and your budget says two hires. What it does not say is how many AI engineers you need to actually ship the thing, or which roles have to exist around them before a single model call reaches a customer.
So most teams guess. They hire two strong generalists, point them at an LLM, and six months later the demo still works and production still does not. The headcount was not wrong because it was too small. It was wrong because it was the wrong shape.
This is the ratio math, by stage, with the cost numbers attached.
Why the Headcount Question Got Harder in 2026
Three years ago, “we need AI” meant hiring a machine learning engineer to train a model. The team shape was obvious: one ML engineer, one data engineer, done.
That shape is now mostly obsolete for product companies. Foundation models removed the training work and replaced it with integration, evaluation and reliability work. The job moved from “build the model” to “make someone else’s model behave in your product, on your data, under your latency budget, without embarrassing you.”
The labor market reflects the shift. The US Bureau of Labor Statistics projects data scientist employment to grow 33.5 percent between 2024 and 2034, adding 82,500 jobs, while software developers grow 15.8 percent and add 267,700. The raw volume is still in software engineering. The scarcity premium is in the AI-adjacent slice, and that gap is exactly why teams overpay for the wrong seat.
The talent guide from A.Team puts the practical version plainly: most product companies now need AI engineers before they need ML engineers or data engineers. Hiring in the old order is the single most common and most expensive mistake we see.
The Four Roles Behind Every Shipped AI Feature
Before you can answer the headcount question you need the role list. Almost every AI feature that reaches production is carried by some combination of four jobs. Sometimes one person holds two of them. Nobody credible holds all four.
AI product engineer.
Builds the feature: retrieval, prompting, tool calls, agent orchestration, the fallback path when the model returns nonsense. This is the role that ships. If you hire one person, hire this one.
Data or retrieval engineer.
Owns ingestion, chunking, embeddings, freshness and the pipeline that keeps your context correct. Skip this and your retrieval quality quietly degrades until support tickets find it for you.
Evaluation and reliability engineer.
Builds the test harness that tells you whether last night’s prompt change helped or hurt. Almost nobody hires this early and almost everybody wishes they had. We wrote a full breakdown of what an AI evaluation engineer does and how to hire one, because the role is new enough that most job descriptions for it are wrong.
Platform or LLMOps engineer.
Deployment, observability, cost control, model routing, rate limits, incident response. This becomes urgent the month your inference bill stops being a rounding error. The LLMOps engineer role is usually the third or fourth hire, not the first.
Notice what is missing: a research scientist. Unless you are training or heavily fine tuning your own models, you do not need one, and hiring one to build product features is a fast way to burn 250,000 dollars on someone who is bored by your roadmap.
How Many AI Engineers You Need, by Company Stage
Here is the shape that holds up across the Series A to Series B companies we staff. Treat the numbers as the floor for shipping one meaningful AI surface, not the ceiling for ambition.
| Stage | Eng team size | AI headcount | Role split | What you can realistically ship |
|---|---|---|---|---|
| Pre-product AI, exploring | 5 to 15 | 1 | 1 AI product engineer | One feature, one model, manual evals, no agents |
| Series A, AI is a feature | 15 to 40 | 2 to 3 | 1.5 product, 0.5 data, 1 eval or platform shared | Two to three surfaces, basic RAG, an eval suite, cost dashboard |
| Series B, AI is the product | 40 to 100 | 5 to 8 | 3 product, 1 to 2 data, 1 eval, 1 platform | Multi step agents, tool use, SLA grade reliability, on call |
| Series B plus, regulated or multi tenant | 100 to 150 | 8 to 12 | Add 1 security, 1 dedicated retrieval | Audit trails, guardrails, tenant isolation, red teaming |
Ratios That Actually Hold Up
Headcount tables are useful. Ratios are more useful, because they survive your next funding round.
One AI product engineer per shipped AI surface.
A surface is a distinct user facing capability: a copilot, a document summarizer, a support deflection agent. Two surfaces, two engineers. Teams that try to run three surfaces on one engineer ship all three at 60 percent quality.
One data or retrieval engineer per three AI product engineers.
Below that ratio the product engineers write their own pipelines, which works until it does not, and then nobody owns the mess.
One evaluation engineer per five AI engineers, starting at five total.
Before five people, evaluation is a shared discipline with a rotating owner. After five, it needs a name attached to it or it stops happening.
One platform or LLMOps engineer once monthly inference spend passes roughly 15,000 dollars.
That threshold is a rough rule from what we see in client accounts, not a law. The point is that the trigger is spend and incident volume, not headcount.
One AI security engineer once you handle customer data in agent context or expose tool calling to users.
Prompt injection is not theoretical once an agent can act. We covered the scope of the AI security engineer role separately, including what to test for.
If your agents are stalling before production despite having the headcount, the problem is usually role shape rather than role count, and we broke that failure pattern down in why agentic AI projects get killed.
The Cost Math Nobody Runs Before Hiring
Here is where the ratio conversation gets uncomfortable. A Series B shape of six AI engineers, hired full time in the US, is a serious annual commitment.
The BLS puts the median annual wage for software developers at 135,980 dollars as of May 2025. AI specialists in US metros routinely clear that by a wide margin, and base salary is roughly 70 percent of the real number once you add benefits, payroll tax, equity, recruiting fees and ramp time. We ran that full calculation in the true cost of hiring a US AI engineer, and in our breakdown of AI engineer salary by level, city and country.
| Approach | Effective cost per engineer per year | Time to first day | 6 person team, year one |
|---|---|---|---|
| US full time hire, AI specialist | 240,000 to 320,000 dollars fully loaded | 60 to 90 days | 1.4M to 1.9M dollars |
| US agency or consultancy | 150 to 250 dollars per hour | 2 to 6 weeks | 1.8M to 3M dollars |
| Staff augmentation, LATAM and APAC | 30 to 35 dollars per hour all in | About 48 hours to candidates | Roughly 375K to 440K dollars |
Competitor and agency rates above are publicly reported ranges and vary by role, seniority and contract length. Confirm current terms directly with any vendor before you build a budget on them.
Divogue places pre-vetted, AI-fluent engineers from LATAM and APAC at 30 to 35 dollars per hour all in, with candidates in about 48 hours, full US timezone overlap, a two week risk free trial and no upfront cost. The reason the ratio conversation matters more than the sourcing conversation: getting the shape right at four people beats getting the shape wrong at eight, whatever you pay per hour.
How to Phase the Hires Over Two Quarters
You do not hire the Series B shape in one quarter. You sequence it, and each hire is gated on evidence rather than on the roadmap slide.
Hire 1, immediately: AI product engineer.
Gate: you have one clearly defined surface and a user who has asked for it. Not a strategy, a surface.
Hire 2, weeks 4 to 8: second AI product engineer or data engineer.
Gate: the first engineer is spending more than 30 percent of their week on retrieval plumbing. That is your signal the data role is real.
Hire 3, month 3: evaluation engineer.
Gate: you have shipped a prompt change you could not confidently say was an improvement. Everyone hits this. Most teams ignore it for two more quarters.
Hire 4, month 4 to 6: platform or LLMOps.
Gate: inference spend, latency complaints, or your first production incident caused by a model provider.
Hires 5 and 6: scale product engineers against surfaces.
Gate: a second and third surface with committed customers.
Contractors and augmented engineers are genuinely better than full time hires for hires 2 through 4, because those are exactly the roles whose necessity you are still proving. If you are new to the model, our guide to staff augmentation covers how the contracts and the IP assignment work.
Where This Advice Breaks Down, and Where In-House Wins
Time to be honest about the limits, including our own.
If you are training or fine tuning your own foundation models, throw this table out.
That is a research organization, not a product team. The ratios invert, you need research scientists and serious infrastructure people, and staff augmentation is the wrong instrument entirely. Hire full time, pay top of market, and expect a long ramp.
If your AI feature is your entire competitive moat, build it in-house.
Core differentiating IP developed by people whose equity depends on it compounds in ways contract work does not. We place engineers into a lot of teams, and we still tell founders this: the one or two people who own your most defensible surface should be employees.
If you are pre-seed with under 500,000 dollars in the bank, hire zero AI engineers.
Use the off the shelf APIs, have your best existing engineer spend two days a week on it, and revisit at 20 customers. Adding a specialist headcount to a five person team to chase an unvalidated feature is how runway disappears.
If you need same room, same hours, physical presence work, distributed staffing loses.
Full US timezone overlap is not the same thing as sitting in your office. Some teams genuinely work better in person, and we would rather say so than place someone into a setup that will fail in month two.
Where we do win: proving out roles 2 through 4 before you commit a salary line, absorbing surge work against a deadline, and filling evaluation, retrieval and LLMOps seats that are hard to hire for domestically at any speed. If you want the screening rubric we use, it is in how to vet an AI engineer.
Key Takeaways
- Headcount is the wrong first question. Team shape is the right one: one AI product engineer per shipped surface, then data, evaluation and platform around it.
- Series A teams need two to three AI engineers. Series B teams need five to eight. Under 15 engineers total, one is usually correct.
- Hire the AI product engineer first, not the ML researcher. Foundation models moved the work from training to integration and reliability.
- Gate each hire on evidence: retrieval plumbing eating a week, a prompt change you cannot evaluate, an inference bill that hurts.
- A six person US AI team costs 1.4M to 1.9M dollars fully loaded in year one. The same shape through staff augmentation at 30 to 35 dollars per hour lands near 375K to 440K dollars.
- Build your moat surface in-house. Prove everything else with augmented capacity first.
Frequently Asked Questions
How many AI engineers does a Series A startup need?
Two to three for most Series A companies with 15 to 40 engineers total. That typically breaks down as one and a half AI product engineers, half a data or retrieval engineer, and one shared evaluation or platform engineer. Below two people you can ship one AI surface well, not three.
Should I hire an ML engineer or an AI engineer first?
An AI engineer, in almost every product company. ML engineers build and serve models. AI engineers integrate existing models into product features through retrieval, prompting, tool calling and agent design. Unless you are training your own models, the integration work is your bottleneck.
When do I need a dedicated evaluation engineer?
Once you have roughly five AI engineers, or earlier if you have shipped a prompt or model change you could not confidently say was an improvement. Before that, evaluation can be a rotating responsibility. After that, it needs an owner or it silently stops happening.
Can contractors fill AI engineering roles, or do they need to be employees?
Both, split by purpose. Roles whose necessity you are still proving, plus surge capacity and hard to source specialties like retrieval, evaluation and LLMOps, work well as augmented staff. The one or two people who own your core differentiating surface should be employees with equity.
What does a six person AI team actually cost per year?
Roughly 1.4 to 1.9 million dollars fully loaded for US full time hires, once you add benefits, payroll tax, equity, recruiting fees and ramp time on top of base salary. Through staff augmentation at 30 to 35 dollars per hour all in, the same six person shape lands closer to 375,000 to 440,000 dollars.
The Bottom Line
The question “how many AI engineers do we need” has a defensible answer, and it is smaller than most founders fear and differently shaped than most job descriptions assume. One product engineer per surface. One data engineer per three. Evaluation at five. Platform when the bill hurts.
Get that shape right and a four person team out-ships a badly composed eight person one. Get it wrong and no amount of headcount saves the roadmap, because the gap was never capacity. It was ownership.
Pick the first role. Gate the next one on evidence. Prove the middle of the team with augmented capacity before you commit salary lines to roles you are still defining.
Staff Your AI Team Without Guessing at the Shape
If you want a second opinion on the shape before you open a req, we will walk through your roadmap and tell you which roles you actually need and in what order, including the ones we are not the right fit for. Candidates in about 48 hours, 30 to 35 dollars per hour all in, full US timezone overlap, a two week risk free trial and no upfront cost.
Book a 15 minute call to map your team shape, or reach us through the Divogue contact page if you would rather start over email.