When AI work needs to happen and the team is full, three options get quoted: an agency, a contractor, or an embedded engineer. Most buyers compare them on day rate, which tells you almost nothing about how the engagement will actually go. The question that predicts the outcome is different: after this work ships, does your team need to be able to change it?
Key Takeaways- Agencies and contractors sell a deliverable at a fixed or capped price. An embedded engineer sells a monthly retainer that your own team directs.- The three models split on one axis: whether the reasoning behind the work stays with your people or leaves with the vendor.- Scope moving mid-project is a change order with an agency, a new invoice with a contractor, and a re-ordered backlog item with an embedded engineer.- The real comparison is not the three of these against each other. It's any of them against a full-time hire: a senior AI engineering search takes 45-75 days to close, and a mid-to-senior US hire runs $290,000-$480,000 fully loaded in year one once compute, tooling, and recruiting fees are counted (KORE1, 2026).- Embedding is the wrong buy in three specific cases, covered below. Say so before the call, not after month two.
Key Takeaways
- Agencies and contractors sell a deliverable at a fixed or capped price. An embedded engineer sells a monthly retainer that your own team directs.
- The three models split on one axis: whether the reasoning behind the work stays with your people or leaves with the vendor.
- Scope moving mid-project is a change order with an agency, a new invoice with a contractor, and a re-ordered backlog item with an embedded engineer.
- The real comparison is not the three of these against each other. It's any of them against a full-time hire: a senior AI engineering search takes 45-75 days to close, and a mid-to-senior US hire runs $290,000-$480,000 fully loaded in year one once compute, tooling, and recruiting fees are counted (KORE1, 2026).
- Embedding is the wrong buy in three specific cases, covered below. Say so before the call, not after month two.
You describe an outcome, a team you mostly do not meet builds it, and a delivery arrives. There is an account manager between you and whoever is writing the code.
This works when the scope is fixed and separable: a marketing site, a defined integration, something with a clean edge around it. It has a real advantage too: a fixed bid gives you a cost ceiling that no other model does.
It goes wrong when the scope needs to move, which for AI work it usually does. An agency has to price the work before anyone has watched a model behave on your data, so the contract protects the estimate. Anything outside the scope document becomes a change order, and every technical decision after signature becomes a commercial decision too. The lag between "we found a problem" and "we agreed on the fix" is the product.
You hire an individual, they build the thing, they leave. Faster to start than an agency and usually cheaper than both alternatives.
The structural problem is the exit. Most contracts are written around handing over code, and code is the least valuable thing being handed over. The reasoning behind it does not transfer. Six months later the retrieval pipeline is a black box that works, and nobody wants to touch it.
Contractors are a good fit when the work is finite and you accept you are buying an artifact rather than a capability.
Someone joins your existing team: your channel, your spec, your repo, your review process. You direct the work and you review what ships.
The knowledge stays because your people were in the loop the whole time. They approved the plan and reviewed the code, so the reasoning lives in the building of the thing, not in a leaving contractor's head.
The cost is that it demands something from you. Somebody internal has to be able to direct and review. If nobody can, embedding does not fix that. It exposes it.
They split on who sets the price, who absorbs a scope change, and what you have to supply on your side. This is the row buyers skip and the one that actually predicts how the engagement goes.
Most buyers compare the first two rows and sign against the ceiling row. The row that actually predicts how the engagement goes is what you have to supply. An agency asks you for a scope. Embedding asks you for attention. Plenty of teams can produce one of those and not the other, and it is worth knowing which kind of team you are before you pick a model that needs the other.
Answer one question honestly: after this work is done, does your team need to be able to change it?
If no, and the scope really is fixed, an agency or a contractor is a reasonable buy, and you should not overpay for embedding you will not use.
If yes, the only model that leaves you able to change it is the one where your people were involved in building it. That is not a sales position. It is just how knowledge works. Nobody understands a system they only watched arrive.
The useful anchor is not the three options against each other. It is any of them against the cost of a local hire, including the months the role sits open, the recruiting spend, and the ramp before that person is productive. A senior AI engineering search takes 45-75 days to close, and a mid-to-senior US hire runs $290,000-$480,000 fully loaded in year one once compute, tooling, and recruiting fees are counted (KORE1, 2026). Measured against that, the day-rate conversation usually turns out to have been the wrong conversation. See how much it costs to hire an AI engineer in 2026 for the full breakdown, or what is a forward deployed engineer for the specific role that popularized the embedded model.
Five questions work on all three models, and the answers separate them faster than a proposal does.
Three cases, worth naming before you get quoted.
If you have no technical person able to direct and review the work, embedding puts the judgment burden back on someone who cannot carry it. That is worse than an agency, because an agency at least expects to lead. If your scope has a clean edge and no maintenance future, you are paying for continuity you will never use. And if the budget only supports a few weeks, do not embed. The model needs enough time for the knowledge transfer to be the point, and a short engagement collapses back into contracting with extra ceremony.
Insight: The failure mode with embedding is rarely the model itself. It's signing without naming, in writing, who on your side is supposed to review the work. That's a five-second question before a contract, and a much harder one to answer six months into it.
We sell one thing: an AI-native engineer embedded in your team, directed and code-reviewed by your own people.
There is no rate card on this site, and that is a real cost to you: you cannot drop us into a comparison spreadsheet without talking to us first. What we do instead is a written intake, then a proposal and a profile of the proposed engineer inside one business day. You interview that engineer before anything is signed. How it works covers the spec and the gates; pricing covers how the retainer is structured.
We are not SOC 2 certified and not ISO 27001 certified. We say so here rather than let it come up on a call. References are available on request rather than published on the site.
Not necessarily, and the comparison is usually the wrong one. An agency's bid caps your spend but caps what you get. A retainer has no ceiling, but you decide what's worth building each month. Compare both against the cost of a local hire before comparing them to each other: a mid-to-senior US hire runs $290,000-$480,000 fully loaded in year one, against a 45-75 day search close (KORE1, 2026).
Yes, and it's a common path. The usual trigger is scope that keeps moving past what a fixed bid anticipated. The switch works cleanly if you can name someone on your side to direct and review the work going forward. Without that, switching models doesn't fix the underlying problem.
You keep it, along with the data and the reasoning behind it. Clean exit, no lock-in, no handover fee is a standing term, not a negotiated concession.
Yes, for the embedded lane specifically. Someone has to approve the plan and review the code, or the model exposes rather than solves your problem. If nobody on your side can do that, the managed lane or an agency will serve you better.
Yes. If the work is a one-off with no maintenance future and you don't plan to touch it again, a contractor costs less and gets you the same artifact. Embedding is for work that stays part of your product.
We do the embedded version, and we are deliberately a bad fit for some of the cases above. If you want a handover and no ongoing relationship, a contractor will serve you better and cost less. We say which teams we work well with for that reason. See the AI team for what an embedded engineer builds toward, or start with a written intake if the work needs to outlive the engagement.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.