The AI Implementation Team for a $200M Company: Four Contract Seats, Not a Department

Last updated: October 28, 2026

By Colin Boothe, CIO at Foretopia

A mid-market AI rollout needs four seats, not a department, and they are a business owner who stays an employee, an ERP-savvy automation developer, a data person, and a reviewer of record for AI-written code. The last three work well as contract seats. KORE1-placed software engineers bill $91 to $120 an hour, median $105 (KORE1 Contract Rate Index 2026-Q4, n=16).

Picture the kickoff call for a funded AI initiative at a $200M distributor. Fourteen names on the invite. The CFO, the CIO, two directors, someone from legal, someone from HR, and a handful of people who got added because they asked what the call was about.

Four of them can actually do something. Maybe three.

KORE1 partner Devin Hornick sees the same pattern from the staffing side, where every new AI search seems to arrive with another layer of managers attached to it. On a recent internal call he said, “With AI, I’m seeing that so many more managers are involved… HR kind of ran things. HR’s there and… they just sit there on the calls now.” Lots of chairs. Not many hands.

So here is the team I would build instead. Four seats. One stays on your payroll, and three you can rent. Disclosure up front, because it matters. KORE1, whose site you are reading, places contract engineers, so telling you to rent three seats is convenient for them. My consulting group at Foretopia sells implementation work, which makes me biased in the other direction. If you already run a strong platform team, skip most of this. If you want the service version, our AI implementation services page covers it, and the contract staffing page covers how the seats get filled.

Which four seats does a $200M company’s AI rollout need?

A mid-market AI implementation team has four seats. A business owner who owns the outcome. An automation developer who builds inside the ERP and the apps you already run. A data person who cleans what the model reads. A reviewer of record who signs for any code an AI wrote. Only the first needs to be an employee.

Not a department. Not a Chief AI Officer, a center of excellence, or a six-person pod with its own Slack channel. A department is what you build when AI is the business. For a company doing $200M in distribution, manufacturing or wholesale, AI is a set of jobs inside the business you already run.

Here is what each seat costs. The first row is a salary you already pay. Everything else comes from the KORE1 Contract Rate Index at 40 hours a week for 13 weeks, where observed cells are our own W2 contract placements started between 2023 and 2026 and modeled cells use BLS wages plus a 24 percent employer burden. Bill rates include payroll taxes, insurance and benefits.

Read the rows under each seat as a choice, not a stack. You pick one automation developer, one data person, one reviewer.

The data analyst row looks odd, and it is. Eleven placements, a low end of $69, a high end of $116, and a median of $115, which tells you that most of them landed near the top while a few sat far below, so use the median as where the typical placement landed and treat the low end as a warning against shopping on the number alone.

Now the arithmetic. Take the ERP consultant at $47,200, the mid-level data engineer at $57,200 and the senior reviewer at $62,400, all full time for a quarter, and the three contract seats come to about $166,800. Make the reviewer half time, which is plenty for a team shipping a few changes a week, and the total drops to about $135,600. That is arithmetic on index cells, not a quote for your scope.

Want those seats priced against your ERP and your start date? Send the seats you need, your ERP platform, the start date and the length of the engagement through the contract rate quote request. If you would rather see every role first, the 2026 Contract Rate Card has them, and the index methodology shows how each cell was built.

Why should the business owner stay an employee?

Because this is the one seat where leaving takes the knowledge with it.

The owner is whoever’s number gets better or worse when the process works, whether that is the head of operations, the controller or the VP of fulfillment, and they decide which process goes first, what counts as correct, and who gets told when it breaks. Contractors are great at doing a defined thing. They are a terrible place to store the definition.

People keep telling me AI adoption is a technology problem. It isn’t. ERP isn’t complicated either. Your org chart is, and nobody on the org chart wants to be the person who signs for the result.

The numbers lean the same way. NBER Working Paper 34836, “Firm Data on AI”, was revised in October 2026 and surveys nearly 9,000 CFOs, CEOs and senior executives across seven countries. Executives report essentially no net effect from AI on firm employment over the last three years, and productivity growth that rose by about 0.2 percent a year. The paper does not say why the gains are so small. My bet, from the inside of a lot of these projects, is ownership, meaning plenty of spend and no named person who has to answer for what it produced.

The owner’s first week is short. Pick one process. Write down what done looks like. Name the person who hears about it when it breaks. One page, not a charter.

What does the ERP automation developer actually build first?

Nothing, for about a week. They turn on what you already own.

Most of the apps you pay for now ship AI features, and in my experience a lot of them are sitting switched off or half configured while the company shops for something new to buy. NetSuite has SuiteAgents and an MCP connector. MCP, the Model Context Protocol, is an open standard for connecting AI applications to systems like your ERP. Think ramen. Every bowl needs noodles and broth, then you add whatever you want to it. The standard handles the noodles and the broth. What goes in the bowl is the developer’s job. The explainer on the NetSuite MCP connector and SuiteAgents walks through it, and AI in ERP: what’s real and what’s hype sorts the demo from the shipping product.

Second, simple and well-scoped workflows. Here is my test. If you could not hand the task to a new intern with a one-page instruction sheet, it is not ready for an agent. Matching vendor invoices to receipts and flagging the misses passes, and so does turning a week of call notes and emails into a summary that a project manager can scan in two minutes before the next meeting. Replacing the saved search you export to a spreadsheet every Monday passes, and a prompt can finish that one in seconds. “Optimize our supply chain” fails, loudly.

Small jobs first. Then you have a track record, a list of what broke along the way, and a reason for the next budget request that does not rest on a vendor demo. The longer version of this lives in how to build your first agentic AI workflow and in the broader practical AI for business operations guide.

Which developer? Pick the ERP consultant if your first five jobs live inside one ERP. Pick the software engineer if they cross systems, say the ERP, the warehouse system and a storefront. The same person can usually do both, but the rate card prices them differently, and now you know why. If you want that seat filled by someone who has already done it, NetSuite MCP and agentic AI consultants and AI business operations consultants are the two placements we run.

Who cleans the data the model reads?

A person, on purpose, before the model sees anything.

A model reads whatever you hand it, and that includes the three spellings of one customer, the item records with units of measure buried in the description, and the general ledger account called “Misc” that quietly holds a surprising share of last quarter. It will not complain. It will answer confidently from all of it, and finding that out in a board meeting sucks. That is the actual risk, and it is a data problem wearing an AI costume. I wrote the longer version in using AI to fix the data your AI needs, and the same logic applies if you plan to build an internal AI assistant on your own company data.

Which kind of data person depends on the mess. A data engineer fits when the work is fixing pipelines and deduplicating at the source so the cleanup sticks, while a data analyst fits when the work is reconciling and defining, as in agreeing on what “active customer” means before anything gets built on top of it. A data scientist fits only if you are building models, and most $200M companies in their first year are not. That is my read, not an index finding.

One more thing. The first cleanup is the big one. After that it is upkeep, and upkeep is where this seat starts to look like a permanent job. Hold that thought for the conversion section.

What does a reviewer of record do for AI-written code?

They sign for it. That is the whole job, and it is harder than it sounds, because signing means you can explain every line to someone who was not in the room when it was written.

At Foretopia we run a written policy for AI-assisted code on regulated work, the kind where an auditor is allowed to ask where a line of code came from. Stripped to the parts that travel:

• A named person is the author of record. AI output is draft material, never a deliverable. If that person cannot explain a passage to an auditor without mentioning the tool that produced it, the passage gets rewritten, not kept.
• Generation runs against an approved specification, not a prose prompt. “Build me a script that does X” is a wish. A spec lets you trace a requirement to the code to the test.
• Review before merge is mandatory, documented, and done by someone other than the author. The record shows who, when and what scope. The reviewer looks hard at platform-specific correctness, governance and resource limits, deprecated interfaces, and what happens on the error path.
• Tests are written by a human. Never by the same means as the code they exercise, because a test generated the same way as the code can share its blind spots.
• Tools are recorded and disclosed on request. And if a client’s quality system bans AI-assisted development, the ban wins.

You may not need all of it. A $200M distributor does not need a validation protocol for a saved-search tweak. Keep two rules, though, which are that every change has a named owner and that a second person who did not write the change looks at it before it ships to anyone. Our guide to AI governance for mid-market companies covers the lighter version. A code review checklist with 21 checks turns those two rules into something a reviewer can work down.

Why contract? Independence is the point. The reviewer should not be the person who prompted the code, and a contractor from outside the build is independent by default. It is also the priciest of the three seats, a median of $120 an hour on seven senior placements, because you are paying for judgment and not keystrokes. Seven is a small sample, so treat the $91 to $138 range as wide.

Funded initiatives you cannot push into 2027

Some projects can slip a quarter, and some can’t, because a license renews, an auditor shows up, or someone promised the board a date in a meeting nobody wrote down.

Those are the ones getting staffed, and Devin Hornick has been watching the pattern from our side. “Companies may not approve three or four permanent hires,” he wrote in a recent email to our content team, “but they may absolutely approve the resources or project team necessary to complete a $300K or $500K initiative that has already been funded.” A frozen headcount line and a funded project are different budget lines. Contract seats fit the second one. The true cost of an AI pilot breaks down what that second line usually has to cover.

Then there is the ratio. The Bain Technology Report 2026, published in September under the title “The AI-Native Enterprise: Absorption Is the New Advantage,” says companies it works with “have found that for every dollar spent on technology, four are spent on people and process.” Four to one. The budget math behind that four-to-one ratio shows which of those people seats have an end date.

Read that against your own budget. If the AI line is mostly licenses and the people line is a hopeful footnote, you funded the wrong half.

When should a contract seat convert?

Later than you think, and not for every seat.

The mechanics first. Devin explained it plainly: “If this person completes 1,040 hours, then they are eligible to convert over to full time with that particular company at no additional charge from KORE1.” That is 26 weeks at a full schedule. Before that mark, a direct hire through us is 20 percent of base salary, with a 60-day replacement guarantee.

Across KORE1’s contract-to-hire placements, 37.5 percent convert to permanent employment (n=552). The median contract assignment runs 116 days (n=1,887), which is shorter than the 26 weeks it takes a full-time contractor to reach 1,040 hours. So most assignments end before the free-conversion window opens, and if you already suspect you will want the person, run that arithmetic before you pick a start date and not after the invoice for month four arrives.

Seat by seat, here is how I would think about it. The automation developer is the likeliest conversion, once the workflows start to pile up and someone has to own every one of them for the long haul. The data person is a maybe, because the first cleanup is a project and the upkeep is a job. The reviewer almost never converts, and shouldn’t. The seat works because that person sits outside the build.

Speed is not the constraint. Offer-to-start runs a median of three days across 1,997 KORE1 contract placements. Finding and vetting the right person takes longer, so start that search before the kickoff call, not after.

Things people ask before they staff the seats

What would the three contract seats cost for one quarter?

About $166,800 at median bill rates, with the ERP consultant, mid-level data engineer and senior reviewer all full time for 13 weeks. A half-time reviewer brings it near $135,600.

Those figures are arithmetic on KORE1 Contract Rate Index 2026-Q4 cells, not a quote for your scope. The ERP and reviewer cells are KORE1 observed, and the data engineer cell is modeled from BLS wages with a 24 percent burden.

Can one contractor fill two of the seats?

Short answer: the developer and the data person, sometimes. The developer and the reviewer, never. The reviewer has to be independent of whoever wrote or prompted the code, or the review means nothing.

Do we really need a data person if our ERP data looks clean?

Probably, but size it small. Run the model against a real saved search, count the wrong answers, and hire to that number.

Data that looks clean in a report is often clean only in the columns the report uses. A few weeks of an analyst checking what the model actually reads is cheap insurance compared with a quarter of confident wrong answers.

Realistically, how fast can the contract seats start?

Three days from offer, based on a median across 1,997 KORE1 contract placements. Sourcing and vetting take longer, so begin the search before the project kicks off.

Is the business owner the CIO?

Usually not. It is whoever’s number improves when the process works, such as the controller, the VP of operations or the head of fulfillment. The CIO sponsors the work. The owner answers for the result.

Ready to staff the three contract seats? Put the seats you need, your ERP platform, a start date and a length into a price for the three seats, or talk to a recruiter if you would rather start with a conversation. If you want a second opinion on the four-seat setup first, hit me up on LinkedIn, or see how we run these engagements on the AI rollout services page. Figures here come from the KORE1 Contract Rate Index 2026-Q4 and BLS Occupational Employment and Wages, May 2025. As of October 2026, refreshed quarterly with the index.