AI team structure and cost: people, tools, tokens, review

The cost structure of an AI-augmented engineering team is the sum of four lines: the people, the AI tools they work with, the tokens those tools consume, and the review time that AI-written code demands. A fixed in-house team budgets almost entirely on the first line. An AI-augmented team has to budget all four, and it expects the first line to change shape over time.

That difference is why budgets built for one kind of team break when applied to the other. Finance knows how to price a headcount; it has a harder time pricing a usage meter or the senior hours that disappear into reviewing generated code. This article walks through each line of the cost of an AI-augmented engineering team, compares the structure with a fixed team, and closes with how to build the comparison for your own case. The argument: AI does not remove cost from an engineering team so much as move it. Teams that see where it moves can decide whether the trade works for them.

People: salaries and the full cost of a position

People are still the largest line in any engineering team, with or without AI. The cost of a position is not the salary: it is the salary plus benefits, payroll taxes, recruiting, the weeks before the person is productive, and the time the seat sat empty before they arrived.

Start with the salary itself. In the United States, the median annual wage for software developers was $135,980 in May 2025, according to the Bureau of Labor Statistics. Senior profiles in AI engineering, security and data are, in DevLatam's words, “scarcer, more senior, and far more expensive to leave open.”

Then add what sits on top of the salary. The same agency's employer compensation data shows that, for private industry in June 2026, wages and salaries accounted for 70.0% of employer compensation costs and benefits for the remaining 30.0%. In other words, a salary figure alone understates the cost of an employee by a wide margin, before a single dollar goes to recruiting or onboarding.

Three costs never appear on a payslip but belong in the people line:

• Recruiting and time to fill. Every week a role stays open, someone else covers it, or the work waits.
• Ramp-up. A new engineer spends the first weeks learning the systems, the domain and the unwritten rules of the team.
• Replacement. When someone leaves, the context they built leaves with them, and part of the ramp-up starts again.

The people line is the one AI changes most, but not by removing it. In an AI-augmented team, people are the ones who direct the tools, decide what gets built and own what the tools produce.

Tools and tokens: the lines that didn't exist before

AI coding tools add two lines a traditional engineering budget never had: seats or subscriptions for the tools, and the tokens those tools consume as they read code and write answers. Seats are predictable. Tokens grow with use, and that makes them the line most often underestimated.

The seat line works like any software license: a monthly price per person, sometimes with an included allowance of usage. The token line works like a utility meter. Every request an assistant or an agent makes is billed by the volume of text it reads and writes, and agents that plan, edit, test and retry can make dozens of requests for a single task.

How those meters work, what makes them spike and how to estimate them before the invoice arrives is a topic of its own, covered in detail in the article on AI token costs in your engineering budget. For the cost structure, the point is simpler: tools and tokens are two separate lines, one fixed and one variable, and both belong in the business case from the start rather than in a surprise invoice three months later. A sensible budget treats the token line like the cloud bill: forecast from real usage, reviewed monthly, capped where the tools allow it.

Review time: the cost AI moves rather than removes

AI coding tools make writing code faster, which means there is more code to review. Review is senior time, and senior time is the most expensive time on the team. When AI writes part of the code, a share of the effort moves from writing to checking, and the budget has to follow it.

Google's DORA research program tracks how AI adoption affects software delivery. In its 2025 report on AI-assisted software development, DORA found a positive relationship between AI adoption and software delivery throughput, but a continued negative relationship with delivery stability. Its explanation: AI accelerates development, and without strong automated testing, mature version control and fast feedback loops, more change leads to more instability.

For the cost structure, that finding has a practical meaning. The review step is the control that keeps faster code from becoming more incidents, and it costs hours of the people best qualified to judge the code. A team that adopts AI tools without budgeting for review either slows down at the pull request or ships more problems. Neither shows up in the price of the tool, and both show up in the cost of the team.

What review time includes in an AI-augmented team:

• Reading generated code with the same care as code a person wrote, including the dependencies it pulls in.
• Testing what the tools produced, since a fast first draft still has to prove it works.
• Maintaining the shared assets that make the tools useful: the prompts, evaluations and rules the team relies on. These decay like any other code if nobody owns them.

Where the engineer is employed changes the cost

The country where an engineer is employed changes the people line: salary levels, payroll taxes, benefits and the cost of employing someone all vary by market. That variation is a fact to price in, not a promise of savings. A cheaper market with a scarce profile can still cost more than an expensive one if the role stays open for months.

The employment model matters as much as the market. Hiring someone in another country means someone has to be their legal employer there. DevLatam's model puts that line on its side, as its home page states: “We are their legal employer. Payroll, compliance and employment obligations sit with us, in the country where they live.” The details of what that covers are on the page about who legally employs them.

For the three profiles that are hardest to find right now, AI engineering, security and data, the deciding variable is usually reach rather than salary. Those are the profiles DevLatam describes on its page on AI, security and data engineering, and the ones where a long search costs the most. If you're specifically hiring AI developers in the region, the guide on how to hire AI developers in Latin America covers the search itself.

AI team structure: an AI-augmented team vs. a fixed team

A fixed in-house team has a stable structure: a set headcount, costs that are mostly salaries, and a budget that changes when someone is hired or leaves. An AI-augmented team is structured to change: it starts with more people while it learns the business, and is designed to need fewer as that knowledge turns into assets the team keeps.

DevLatam describes this model on its page on AI-augmented teams: a squad that starts at around six people and is designed to get smaller as the context it builds becomes prompts, evaluations, pipelines and playbooks your team keeps. The page is just as clear about the limits: “We don't promise six-to-two, and we don't promise a date.”

That changes the shape of the budget in three ways:

• The people line is front-loaded. The first months carry the most headcount, because learning the business is the expensive part and can't be skipped.
• The token line grows as the people line shrinks. As the team gets smaller, tokens become a larger share of the cost of the work. The page puts it directly: the token cost is “smaller than the headcount it replaces, but a new line in your budget.”
• Ownership becomes a cost. Someone on your side has to own the assets the team leaves behind. Without an owner, they decay and the team quietly grows back.

The model also depends on the work itself. The same page lists the conditions: the work has to compound, the context has to leave people's heads, and someone has to own the artifacts. Repetitive work with a stable definition of done shrinks the team; one-off exploratory work doesn't, and a fixed team, or a single engineer joining through staff augmentation, is the better structure for it.

When a fixed team is the better structure

A fixed team is the better structure when the work doesn't repeat, when the knowledge can't be written down, or when nobody on your side can own what an AI-augmented team leaves behind. In those cases the token, tool and review lines add cost without the people line ever coming down.

Exploratory work is the clearest case. A team building something new, where every task is its own problem, doesn't accumulate the kind of reusable context that lets a squad get smaller. DevLatam's own page says so: if the work turns out not to compound, the honest move is “to convert the engagement to traditional staffing, not to keep going.”

Three other situations point the same way:

• The real criteria live in one person's judgment. If the rules that matter can't be written down, there's nothing for the assets to hold, and the team can't shrink around them.
• The budget has to be forecast to the dollar. Salaries are predictable; tokens are not. Some organizations would rather carry a slightly higher, stable cost than a lower, variable one.
• The team isn't ready to change how it works. An AI-augmented structure asks your side to own prompts, evaluations and pipelines. If nobody has time for that, the structure won't hold.

None of these is a failure. They are the cases where adding a named engineer to a stable team, paid for the time worked and directed by your leads, does the job better than a structure designed to shrink.

How to build the cost comparison for your own team

A useful comparison puts both structures side by side over the same period, usually twelve months, with every line priced, including the ones a quick estimate leaves out. The goal is not to prove one option is cheaper but to see where the cost sits in each.

For each option, write down:

1. People: headcount by month, with the full cost of each position, not just salary.
2. Time to fill: how long until each seat is productive, and what the wait costs.
3. Tools: seats and subscriptions, priced from the vendor's current plan page.
4. Tokens: an estimate from a pilot week on real work, not from seat prices.
5. Review: the senior hours per week that go into reviewing generated code.
6. Ownership: who maintains the shared assets, and how much of their time that takes.

If the comparison shows that the work doesn't repeat enough for the team to shrink, the fixed structure wins, and that is a valid answer. If it shows the opposite, the AI-augmented structure earns its token line.

Frequently asked questions

How do you calculate the fully loaded cost of an employee?

The fully loaded cost of an employee is the salary plus benefits, payroll taxes, recruiting, equipment, software and ramp-up time. For a quick estimate in the United States, BLS data show wages at 70.0% of private-industry compensation costs in June 2026, so dividing a salary by 0.70 approximates pay plus benefits.

What is the 10-20-70 rule for AI?

The 10-20-70 rule is a BCG guideline that splits the work of an AI transformation into 10% algorithms, 20% technology backbone and 70% people and processes. Read against an engineering budget, it suggests the tool and token lines are the smaller share of the effort; review, ownership and changes in how the team works carry most of it.

How expensive are LLMs to run?

Running an LLM costs what the tokens it reads and writes cost, plus the review time its output needs. Through an API, prices are set per million tokens and vary by model; an agent working on a large codebase can spend far more than a chat assistant, so a pilot week on real tasks gives the reliable figure.

Can a team host its own LLM?

Yes, a team can host an open-weight model on its own infrastructure, which replaces the per-token bill with the cost of servers, GPUs and the people who run them. Self-hosting gives more control over data, but it adds an operations line to the budget, and it makes sense mainly when usage is steady enough to keep the hardware busy.

Seeing where the cost moves

The cost of an AI-augmented engineering team is not lower by definition. It is distributed differently: less in people over time, more in tools, tokens and review. Budgets that only count heads miss the new lines; budgets that only count tools miss that people still decide what gets built and own what the tools produce.

The next step is to price your own case with all four lines on the table. The three ways to fill the role lay out the structures side by side, and the page on AI-augmented teams explains what has to be true for a team designed to get smaller to work.