Back to Blog

Build vs Buy AI Agents: The Team You Need If You Decide to Build

AIEngineeringLeadership

Last updated: November 19, 2026

Buy an agent product when the workflow is generic and the data may leave your walls. Build when the workflow is proprietary, the data is regulated, or the integrations run deep. Of the five seats a build needs, four can be contract for two quarters, at about $510,300 at national medians (KORE1 Contract Rate Index 2026-Q4).

The demo cleared a loan condition in forty seconds, and everyone in the room nodded.

By the end of the second quarter it was still a demo. Nobody had written a test set, so nobody could say whether the agent was right 90 times in 100 or 99, and the two engineers who built the prototype had gone to the next funded project when the budget year turned. Three processors kept clearing conditions by hand, and now they were testing the agent on the side too, which no plan had counted as work. Nobody owned it. (This is a composite of builds I have watched, details changed.)

Build or buy is the easy half of that decision. A decent test settles it in an afternoon. The hard half shows up once you say build, because you are about to open reqs for a team that does not exist, against a budget line somebody already promised the board. So I will give you my test for the call, then the five seats a build needs, what they cost, and the ones that should never be contractors.

I write in partnership with KORE1, a staffing firm that places contract AI, data, and software engineers through its AI and machine learning staff augmentation practice, so it benefits when a build needs extra seats. Weigh my view with that in mind. The cost figures come from the KORE1 index and not from me, and you can check every one.

Engineering executive placing the last of five blank sticky notes, four orange and one gray, in a row on a glass wall to represent the five seats of an AI agent build team

Should You Build or Buy an AI Agent?

An AI agent is software that uses a language model to pick the next tool to call, calls it, reads the result, and repeats until the task is done or it hands the case to a person. The tool might be an API, a database query, or a click inside a system of record.

My test for agents is narrower than the general argument, which I made in why most AI pilots die before production. Read that first if the edge-versus-plumbing question is still open.

This one is about what the agent touches. Picture a workflow that looks about the same at a thousand companies, runs on data a vendor may hold, and only reads from your systems before a person approves what comes out. Buy it. Somebody else can maintain that product while your engineers stay on the roadmap, and the market for things like meeting summaries and first-draft support replies is crowded enough that you were never going to out-build it.

I would build when any one of those three flips. A workflow that is your edge, say the way your underwriters weigh a thin-file borrower, should not go to a vendor who sells the same agent to your competitor next quarter. Regulated data changes the math as well, because the security review for a vendor that will hold your loan files can run longer than the prototype takes to write. Then there is the agent that has to write into the loan origination system, or the ERP, or the claims platform. It spends most of its life in integration work, and a product cannot do that part from outside your walls.

Plenty of cases land in the middle. For those, buy the platform and build the exception queue yourself, a pattern I laid out in buying the platform and building on top of it.

What Does Buying an Agent Product Still Leave You Owning?

A lot, and most of it is not in the sales deck, because the deck stops at the edge of the product. Your company starts there.

Authentication bites first. Somebody has to give the agent a service account, decide what it may write, and rotate those credentials on a schedule, and the Model Context Protocol specification, the way many agents now plug into their tools, says authorization is optional for implementations. So the connector in the demo may have run wide open. Under the hood, I would ask the vendor to prove it did not.

Evals come next. The accuracy figure on the slide came from the vendor’s test set, which is fine for the vendor. You need one built from your own loan files or invoices or support tickets, and it takes longer than anyone budgets. After that, the data access approvals, where security, legal, and the owner of each data set all get a vote and none of them moves at the vendor’s pace.

Then monitoring. The agent starts doing something odd on a Thursday afternoon, say approving a refund it should have escalated. Who switches it off? And what do the three people who used to do that work do until it is back?

Buying takes the construction of the agent off your plate but not the ownership. I once watched a team budget zero engineers for a bought agent, then borrow three from the roadmap to get it connected. Same seats as a build. They just were not on the plan.

If You Build, Which Five Seats Does an Agent Team Need?

Five. With fewer, a seat’s work falls through the gap between two people, and with more you are mostly paying for meetings.

Start with the seat everybody pictures, the agent engineer, who writes the loop itself. Orchestration and tool calling are part of it, but most of the week goes to guardrails, meaning the code that stops the agent from doing something silly with a tool it was handed. The index has no cell for this seat, so I price it at the senior software engineer cell and label that a benchmark, and the AI engineer contract rates page carries the AI-specific numbers.

The integration engineer is the seat I would worry about most. Every system the agent touches has its own API quirks, its own permissions model, and a release calendar nobody mentioned, and the loan system in particular has a rate limit that only shows up when the agent is fast. A mid-level engineer is usually enough, since the work is bounded and the other side is documented. Budget for it.

Then comes the evaluation engineer, the seat that gets cut first and should be cut last. The job is building eval sets from your real files, rerunning them every time a prompt or a model changes, and red-teaming the thing before launch. If you want a data scientist anywhere on this team, it is here, and the NIST AI Risk Management Framework splits AI risk work into Govern, Map, Measure, and Manage, so this seat is the Measure function with a name attached. Cut it and you are guessing.

What does the data engineer do on an agent team? Mostly retrieval. For most agents that means the retrieval-augmented generation pipeline, the data quality checks on what goes into it, and the access controls that keep the agent from reading anything the person who called it could not read, which is where a Snowflake or Postgres permission model earns its keep.

The fifth seat is the product owner, and that one is permanent. This person owns the workflow and the success metric. Not a project manager. The person who can say that 94 percent is good enough for this queue, or that it is not.

Show me the data. Here it is. Four of the five seats come straight from the KORE1 Contract Rate Index 2026-Q4, and the fifth has no bill rate because it should not be a contractor.

SeatContract or permanentMonths needed in the first two quartersIndex cell and bill range per hour6-month cost at the median
Agent engineerContract6Senior software engineer, United States national, modeled. $118 to $142, median $131. A benchmark, because the index has no AI agent engineer cell.$136,600
Integration engineerContract6Mid-level software engineer, United States national, modeled. $96 to $122, median $108.$112,200
Evaluation engineerContract6Senior data scientist and ML engineer, United States national, modeled. $107 to $132, median $122.$127,100
Data engineerContract6Senior data engineer, United States national, modeled. $118 to $140, median $129.$134,400
Product ownerPermanent employeeFrom day oneNo bill rateNot priced
Four contract seatsAbout $510,300

Four contract seats for an AI agent build (agent, integration, evaluation and data engineers) cost about $510,300 for two quarters at national index medians (KORE1 Contract Rate Index 2026-Q4).

Ranges run from the 25th to the 75th percentile, bill rates include the 24% employer burden, and six months is 26 weeks at 40 hours. The cells are modeled, not observed. They start from the Bureau of Labor Statistics OEWS May 2025 wages for software developers (SOC 15-1252), data scientists (SOC 15-2051), and database architects (SOC 15-1243). Then comes the burden, and then the bill-to-cost range KORE1 sees on its own placements. The contract rate methodology page has the full method.

Add the four six-month costs and you land on the $510,300. Cut the plan to one quarter and the same four seats cost about $255,200. All of this is one agent on one workflow, so a wider scope adds seats before it adds anything else.

Price your agent build team.

Send the seats, the seniority, the months, and how sensitive the data is through the bill rate quote, and you get hourly bill rates by role to set beside this table.

Two software engineers at a shared desk reviewing blurred green and orange result charts and blank printed grids during an AI agent evaluation review

Which Seats Can Be Contract for the First Two Quarters, and Which Cannot?

Four can. The agent, integration, evaluation, and data seats each have a defined output in the first two quarters, which is a working loop, connected systems, an eval suite, and a pipeline, and defined output with an end date is the plain case for contract staffing.

Two things stay permanent from the first day, and the product owner is one of them. The other is whoever owns the evals once the agent is live. That person can be a contractor you convert, but it has to be somebody, because an agent that is wrong still returns a clean HTTP 200 and nothing pages anyone.

KORE1 IT-role assignments run a median 136 days (n=410; 79% of KORE1 engagements are in California). A six-month build is longer than that, so plan on at least one seat turning over before you finish. Plan for it. The engineer you want to keep converts to your payroll at no fee after 1,040 hours, which is 26 weeks at 40 a week, and across KORE1 contract-to-hire placements 37.5% convert (207 of 552). Decide who you want while the work is still going. The last week is too late.

Every seat bills hourly and you direct the work.

What Breaks If a Funded Agent Stalls in Pilot Before 2027?

Devin Hornick, a partner at KORE1, put one question in his content email that I now ask engineering leaders too. “What technology initiatives are already funded that you can’t push into 2027?” If the agent is on that list, a stall has a price.

What breaks is dull. The budget is spent and nothing is in production, the workflow is still run by hand by the same people who now maintain a pilot on top of it, and the model bill and the vendor contract keep arriving anyway.

The cause is rarely the model. Nobody wrote an eval suite and nobody was named as the owner, so no one could say the agent was ready, and “not yet” became the standing answer through two budget reviews. I made the case for the test set before the features in evals before features. For budget season the short version is that the evaluation engineer and the product owner are the two seats that prevent a stall, and they are two of the five.

Who Should Run the Build, and How Many Managers Is Too Many?

Devin told our content team that with AI he is seeing “so many more managers” on the calls. He was talking about hiring calls, where the hiring manager used to be the only voice and is now one of several. I see the same drift on builds. Every VP wants a say in the agent, and each one carries a slightly different picture of what done means, which is how a six-week build turns into a six-month negotiation.

Two names fix most of it. One product owner with written decision ownership, and one technical lead. The product owner decides what the agent is for and when it is good enough. The technical lead decides how it gets there. Everybody else is consulted. Consulted is a real category, by the way, and it has an end date.

Write down what each of them decides, literally on one page. Which cases the agent may handle alone, which it hands to a person, what accuracy is the release bar, who can switch it off. Skip that and the decisions get made in hallway conversations, and a hallway decision never makes it into the repository. Who owns models across teams once you have a second agent is a separate problem, and I covered it in who owns the model across engineering, data, and product.

What Happens to the Agent When the Contract Seats Roll Off?

It either keeps running or it quietly rots, and which one depends on what you made the contractors leave behind. Before the last month, ask for these.

  • The eval suite, runnable by someone who did not write it.
  • Runbooks for the failures you have actually seen in staging, not the ones anyone predicted.
  • Prompts and tool definitions committed to your own repository, because a prompt that lives in a contractor’s notebook or a vendor console walks out the door with them.
  • A named on-call owner.
  • A decision on which contractor converts to your payroll, made before anyone updates a resume.

Test the handover the way you would test a backup, by making someone new run it. If they cannot, you do not have a handover yet. You have a hope.

Senior engineer sliding a closed gray binder across a conference table to a colleague with an open laptop during an AI agent project handover

From KORE1: what we place and how it works

KORE1 places W2 contract AI, data, and software engineers that you direct, billed to you hourly. Contractors are KORE1 employees, paid every Friday, and they can convert to your payroll at no fee after 1,040 hours. Contract placements carry no replacement guarantee. We have recruited technology talent since 2005, and 92 percent of the people we place are still in the job twelve months later.

Start with contract AI and data engineers, or see AI engineer contract rates for the AI-specific bill ranges.

Four Questions CTOs Ask Before They Open Reqs

Is it cheaper to build or buy an AI agent?

Buying is cheaper to start, and building can win later if the agent does enough work inside your own systems. I cannot give you the crossover, because it depends on a vendor price I have not seen. The build side I can price, at about $510,300 for four contract seats over two quarters at index medians, plus a permanent product owner. Ask the vendor for its number at your volume and set the two side by side.

Realistically, how many people does an agent build take?

Five seats, one of them permanent: an agent engineer, an integration engineer, an evaluation engineer, a data engineer, and a product owner. That is one agent on one workflow. Each workflow you add brings integration and evaluation work before it brings anything else.

Can contractors work on regulated data?

Contractors can usually work on regulated data, with conditions. The data stays in your environment, the contractor works under your access controls with the least access the job needs, and the approvals are signed before day one and not after the first sprint. KORE1 contractors are W2 employees of KORE1 whom you direct. Your compliance team decides what that means for your data, and this is not legal advice.

Do we really need a data scientist on an agent team?

You do not need one to build the agent, but you need one to judge it. A senior contract data scientist or ML engineer bills $107 to $132 an hour nationally, median $122 (KORE1 Contract Rate Index 2026-Q4), and the evaluation seat is where that skill earns it. A strong software engineer who has written eval sets before can fill it, so look at the track record and not the title.

Decide Who Owns It Before You Decide Who Builds It

Whichever way you go, the same question comes first. Who owns this agent in month nine, when the demo is a memory and the first odd result lands on a Thursday?

If the answer is a name, build or buy will mostly work out. If the answer is a committee, it will not matter which you pick. Put the table above next to the use case you have funded. Count the seats that already have a person in them.

If you build, price the seats by hourly bill rate before you open the reqs. Or talk to the KORE1 team and bring the workflow with you.

As of November 2026. Rates are from the KORE1 Contract Rate Index 2026-Q4, effective October 16, 2026, and the engagement figures reflect KORE1 W2 contract placements, 79% of them in California. This post is reviewed every six months. It is general information and not legal advice.