How to Structure Your AI Team: The Roles You Actually Need at Each Stage

The right AI team structure depends on your stage, and most companies over-hire specialists early and under-hire evaluation and data roles late. Vintti AI places dedicated AI engineers into US companies at every point on that curve, from the founder hiring one person to the Series B team splitting into squads. This guide lays out the structure that works at each stage: which roles, in what order, who reports to whom, and what the team costs per month depending on where you hire.

Start From the Work, Then Name the Roles

AI team design starts with three questions about the work, and the roles fall out of the answers.

1. Is AI the product or a feature of the product? A company whose moat is a model needs research-flavored engineers early. A company adding an LLM feature to a SaaS product needs applied engineers and almost no research.
2. Do you build on foundation models or train your own? Building on OpenAI, Anthropic or open-weight models makes the LLM integration developer the central profile. Training and fine-tuning your own pushes toward the AI/ML engineer.
3. Where does your data come from? If model quality depends on labeled or human-judged data, a data annotation specialist enters the plan far earlier than most teams expect.

Write down the answers before the org chart. Teams that start with titles ("we need an AI team: a data scientist, an ML engineer and an MLOps engineer") end up with three people and no shipped feature.

Scope note: this guide covers product and engineering teams of roughly 5 to 200 people adding or scaling AI capabilities. AI research labs and large enterprises with platform teams have different structures.

The Seven Roles, in Plain Terms

Seven roles cover almost every applied AI team, and most companies need only three or four of them at any given time.

The explainers What Does an AI/ML Engineer Do? and What Does a Prompt Engineer Do? go deeper on the two most misunderstood titles.

‍

Stage 1: The First AI Hire (1-2 People)

Your first AI hire should be a builder who can ship a feature end to end, and for most companies today that is an LLM integration developer or an applied AI/ML engineer, never a researcher.

Structure: one senior builder reporting to the head of engineering or CTO, paired with an existing backend engineer for integration work. If the product is prompt-heavy, a semi-senior prompt engineer can join as the second person, but only after the first feature is live.

What this stage should produce: one AI feature in production, with basic logging of inputs and outputs, inside 90 days. Anything else (a platform, a data strategy, a model registry) is premature.

The most common mistake at Stage 1 is hiring a data scientist to "explore opportunities". Exploration without a shipping engineer produces decks.

‍

Stage 2: The Production AI Team (3-5 People)

Once a feature is live and users depend on it, the team's center of gravity shifts from building to measuring, and the second wave of hires reflects that.

Typical structure:

• 1-2 builders (AI/ML engineer and/or LLM integration developer)
• 1 evals engineer, the hire most teams delay and regret delaying. Every prompt change and model update is a regression risk once customers are watching.
• 1 prompt engineer if prompts are a first-class asset of the product
• 1 data annotation specialist if quality depends on labeled or preference data, which for most LLM products it does by this stage

Reporting: the group reports to an engineering manager or a technical lead who owns the AI roadmap. Product management stays where it is; AI teams that report into product lose engineering rigor, and AI teams with no product counterpart build things nobody asked for.

Ratios that work in practice across Vintti AI placements: roughly one evaluation or data role for every two builders. Teams below that ratio ship fast for a quarter and then spend the next one firefighting quality.

‍

Stage 3: The Data-Centric and Specialized Team (6-12 People)

Past five people, AI teams split by surface or by capability, and the data roles multiply faster than the modeling roles.

Two structures work:

• By product surface: each squad (search, assistant, document processing) has its own builder plus shared evals and data roles. Best when surfaces have different users.
• By capability: a modeling group, an evaluation group and a data group serve the whole product. Best when the same model powers many surfaces.

New profiles at this stage: an NLP engineer when language problems (extraction, classification at scale, multilingual search) outgrow general LLM calls; two or more data annotation specialists, including domain experts if the data is clinical, legal or financial; a senior AI/ML engineer acting as tech lead.

Ratios shift toward data: mature applied teams commonly run one annotation or evals professional per builder.

What the Team Costs Per Month, by Stage

Team structure and hiring geography together decide the monthly bill, and the nearshore version of each stage costs well under half of the US version. Monthly compensation for semi-senior profiles, from the Vintti AI Roles Benchmark (US base salary versus LATAM total compensation for dedicated nearshore hires):

US totals grow another 20-30% with benefits and payroll taxes. Midpoints are the average of each range; the full dataset by level is at vintti.com/ai/salary-benchmark. Many Vintti AI clients run mixed teams: US-based technical leadership with dedicated nearshore builders, evals and data roles, which keeps direction in-house and puts the headcount where the differential is largest.

The Roles You Do Not Need Yet

Several fashionable titles belong to a later stage than the one most companies hire them at.

• Head of AI before there are three AI engineers to lead. Hire the builder first; promote or hire the lead when there is a team.
• MLOps or AI platform engineer before you have more than two models in production. Until then, the builders own their deployment.
• Research scientist unless the model is the moat. For products built on foundation models, research hires produce papers and prototypes and rarely features.
• Data scientist as first AI hire for an LLM product. Analytics is a different job from shipping AI features.
• AI ethics or governance lead as a full-time role before Stage 3. Governance matters from day one, and until then it is a responsibility of the tech lead and legal, with written policies.

Three Anti-Patterns Vintti AI Sees Most Often

The same three structural mistakes account for most AI teams that stall.

1. All builders, no measurers. Four engineers shipping prompt changes with no evaluation suite. Quality degrades invisibly until a customer notices.
2. The AI team as a separate island. An AI group that does not sit inside product engineering ships demos rather than features. Embed the roles in squads, or keep a strong product counterpart.
3. Renting the brain. Outsourcing the entire build to an agency and keeping nobody who understands the system. Agencies have their place for scoped 0-to-1 builds; the ongoing capability should live with people on your team, whether hired locally or as dedicated nearshore placements.

‍

The Team Design Checklist

• Answered the three questions (AI as product or feature; build on or train; where data comes from)
• First hire is a builder who can ship end to end
• One AI feature in production within 90 days of the first hire
• Evals role planned for the quarter after the first feature ships
• Data annotation planned as soon as quality depends on labels or preferences
• Ratio of at least one evaluation or data role per two builders at Stage 2
• AI roles report into engineering with a named product counterpart
• No Head of AI, platform or research hires before the stage that justifies them
• Hiring geography decided per role, with nearshore for roles where the differential is largest

(Note: Vintti AI is a nearshore staffing and recruiting firm. We are not an Employer of Record (EOR), a managed service provider, or a BPO. We source and place dedicated AI professionals across all seven roles above, as long-term embedded roles or project-based placements, and your company employs or contracts them directly. While Latin America is our core talent pool, we also recruit US-based and global candidates when the role calls for it.)

Frequently Asked Questions

Who should be the first AI hire at a startup?

A senior builder who can ship a feature end to end: an LLM integration developer if you build on foundation models, an applied AI/ML engineer if you train or fine-tune your own. A researcher or data scientist as the first hire is the most common and most expensive mistake.

‍

How many people does an AI team need?

One to two to ship the first feature, three to five to run it well in production, six to twelve once AI powers several surfaces. Headcount beyond that usually means splitting into squads.

‍

Do I need an evals engineer, or can builders handle evaluation?

Builders can handle it for the first feature. Once customers depend on the feature and prompts or models change weekly, a dedicated evals engineer pays for itself in avoided regressions. It is the hire teams most often delay.

‍

Should the AI team report to product or to engineering?

Engineering, with a named product counterpart. AI teams reporting into product tend to lose rigor; AI teams with no product partner build things nobody asked for.

‍

Can Vintti AI staff an entire AI team, or only individual roles?

Both. Vintti AI places individual dedicated hires and builds out full teams across the seven roles, as long-term embedded roles or project-based placements. Typical pace per role: a shortlist in about 7 days, first interviews within 1-2 weeks, and a completed hire in 2-4 weeks.

‍

What does a nearshore AI team cost compared with a US team?

Well under half at the semi-senior level, per the Vintti AI Roles Benchmark: a four-person Stage 2 team runs about $38,900 per month in US base salaries versus about $16,900 in LATAM total compensation, before US benefits and payroll taxes.

‍

Structuring an AI team well means hiring in the order the work demands: a builder first, measurement and data as soon as customers depend on the feature, specialists only when the problems outgrow generalists. Get the sequence right and the org chart takes care of itself.

‍

Build your AI team with dedicated nearshore hires - start with a conversation