Two companies approve the same AI budget in the same quarter. Twelve weeks later, one has a claims-triage workflow running in production and a second one scoped. The other has a signed offer letter, a start date in November, and a status slide saying the program is on track.
Neither company made a stupid decision. They compared the same three options, ran the same rate math, and reached opposite outcomes. The difference was not the quality of the choice. It was the axis they made it on.
Almost every guide written about this decision compares price. Salary against hourly rate against monthly retainer, with a table at the bottom and a conclusion that happens to favor whoever published it. That framing is why so many of these decisions look sensible in the approval meeting and disappointing two quarters later. A $180,000 hire who ships nothing for five months costs more than an engagement that puts a workflow in production in week three, and no rate comparison will ever tell you that.
What follows is the comparison on the axes that move the outcome, applied evenly to all three models. Including where hiring wins outright, and where you should not use a fractional team at all.
What the three models actually are
Half the confusion in this decision is vocabulary. So, precisely:
An in-house AI team means permanent employees on your payroll. You own the recruiting, the compensation strategy, the management, the career path, and the retention risk. The capability lives inside the company and stays there, assuming the people do.
Outsourced AI development means a contract with an agency or development firm , either for a defined project or for a dedicated team assigned to you. The vendor owns staffing, the delivery process, and usually the methodology. You own the specification and the acceptance criteria. When the contract ends, the team leaves and the knowledge mostly goes with them.
A fractional AI team means an embedded senior team working with you on an ongoing basis, combining strategy, build, and operate roles, without permanent headcount. It is not a project contract and it is not a staffing agency. The team is already assembled and already senior, and the engagement is measured in workflows shipped, not hours billed or seats filled.
A fractional AI team is not a fractional CTO
This distinction gets collapsed constantly. Collapsing it produces bad decisions.
A fractional CTO is one person providing part-time technology leadership. They set direction, evaluate architecture, sit in board meetings, and help you hire. They are an executive, and what you buy is judgment.
A fractional AI team is a delivery unit. It ships. Strategy is part of the engagement, but the output is production systems doing work your business can measure, not a roadmap document and a hiring plan.
If your gap is direction, hire a fractional CTO. If your gap is that nothing is getting built, a fractional CTO will produce an excellent plan for a problem you already understood. Those are different purchases, and the market's habit of using one phrase for both is why buyers end up with the wrong one.
The six axes that decide this
Six axes. Most buyers compare one of them. Here is what each means and why it moves the decision.
Cost. Not the headline rate. The total cost of having the capability available for twelve months , including everything that never appears on an invoice.
Speed. Time to the first workflow running in production, not time to first commit or first standup. Nothing counts until the business is using it.
Risk. Who absorbs the loss if delivery slips, if the person you depended on leaves, if the system handles data badly, or if the whole approach turns out to be wrong.
Flexibility. What it costs to change direction in month four. Every model is easy to start. They differ enormously in how expensive they are to stop.
Organizational maturity. What the model assumes you already have internally. This is the axis that quietly decides most failures , and almost nobody evaluates it before signing.
Time-to-value. When the business feels the change, which is usually later than the technical go-live and is the only date your CFO remembers.
The comparison below holds all three models to the same standard on these six, using published market data where it exists and delivery experience where it does not. Estimates are labelled as estimates.
How the three models compare
| In-house team | Outsourced agency | Fractional AI team | |
|---|---|---|---|
| Cost | Highest fixed commitment. Salary plus recruiting, benefits, and management overhead. Committed whether or not there is work. | Variable and contract-bound. Predictable per project, expensive on change requests. | Fixed monthly, no employment liability. Higher than one salary, lower than a loaded team. |
| Speed | Slowest start. Months to hire, then months to ramp on your systems and domain. | Moderate. Gated by procurement, scoping, and specification quality rather than by talent supply. | Fastest start. The team exists on day one, so the constraint is your data access, not their availability. |
| Risk | Key-person risk concentrated in one or two scarce hires. You own delivery risk entirely. | Delivery risk contractually shifted to the vendor. Knowledge risk stays high because it lives outside your building. | Delivery risk shared. Creates dependency on an external team you must manage deliberately. |
| Flexibility | Lowest. Changing direction means redundancy, redeployment, or carrying idle senior salary. | Moderate. Cheap to stop between contracts, expensive to change mid-project. | Highest. Scope and stop points are contractual rather than personal. |
| Organizational maturity | Highest bar. Requires the ability to evaluate AI engineers and a technical leader to direct them. | Requires the ability to write and enforce a specification, and someone to own acceptance. | Lowest bar. Requires a named business owner per workflow and data access. |
| Time-to-value | Long. Real value typically arrives after the ramp, once domain knowledge accumulates. | Medium. Value arrives at handover, then depends on whether anyone internally can operate it. | Short. Designed to produce a working workflow early, then compound. |
| Best for | AI as a permanent core capability with an 18-month-plus horizon. | A well-specified system to build once and hand over. | Production output before the permanent shape is known. |
The largest gap in that table sits in speed and maturity, not cost. Those two rows also interact: the model that starts fastest is the one that demands least of you before it starts. That combination is why the comparison so often ends somewhere the rate math did not predict.
What twelve months actually costs
Rates are the easy part. Here is what the twelve-month cost of capability looks like for each, with the invisible line items included.
In-house. Senior AI engineers in the US now command base salaries of roughly $220,000 to $310,000, with total compensation frequently higher again (Kore1, 2026), and time-to-fill for senior AI and ML roles runs about 54 to 120 days depending on how competitive the offer is (Recruits Lab, 2026; KORE1, 2026). Around that base you add recruiter fees, benefits and employer overhead, the ramp period before the person is productive on your systems, and the management time of whoever directs them. Then there is attrition, which is not theoretical on a skill this scarce. A single senior hire is an annual commitment well above the salary line, and one person is rarely a team.
Outsourced agency. Cost arrives as a project fee or a blended rate against a dedicated team. It is the most predictable of the three on paper and the least predictable in practice, because the variance sits in change requests. Add the internal cost of writing the specification, running vendor selection , managing the relationship, and absorbing knowledge transfer at handover. A specification that was wrong is paid for twice.
Fractional. Cost is a monthly engagement fee. Our own Fractional Agentic Team starts from $8,000 per month, which is more than a single junior salary and materially less than a loaded senior team with recruiting attached. Add the internal time to point the team at the right workflows. That cost is real but small.
| Twelve-month view (estimate ranges, not quotes) | In-house | Agency | Fractional |
|---|---|---|---|
| Direct cost | One or more senior salaries plus employer overhead | Project fee or blended team rate | Monthly engagement fee |
| Acquisition cost | Recruiter fees, roughly two to four months of search (Recruits Lab, 2026) | Vendor selection and contracting | Minimal |
| Productive from | After hire plus ramp | After scoping and contracting | Within weeks |
| Cost if you stop | Redundancy or idle salary | Low between contracts, high mid-project | Notice period |
| Cost of being wrong | Highest, and slowest to discover | Moderate, discovered at handover | Lowest, discovered early |
The row worth staring at is the last one. You are not really choosing a price. You are choosing how long it takes to find out you were wrong. Every model can produce the wrong thing, and finding out in week four is a fundamentally different financial event from finding out in month seven.
How long until something is live
Every vendor on this topic quotes a timeline. Almost none of them show the weeks. Here they are, for all three.
In-house, realistically. Write and approve the requisition. Source candidates in a market where the good ones are not looking. Run the interview loop, which for AI roles usually means a technical assessment plus panel time from people who are already busy. Make the offer, negotiate, and absorb a notice period that is frequently three months. Then onboard. Then ramp on your systems, your data, and your domain, which for a senior engineer joining a business they do not know is measured in months, not weeks. First production workflow is realistically two to three quarters out , and that assumes the first hire works out.
Agency. Scoping call, then a statement of work, then contracting and procurement. Team assembly on the vendor side. A discovery phase, because they do not know your business yet either. Then build, then handover. Faster than hiring by a wide margin. The gate here is not talent supply. It is how long your procurement takes and how good your specification is, and a vague specification lengthens everything downstream.
Fractional. The benchmark we hold ourselves to is a first production workflow live in two to four weeks. Three specific things make that achievable, and naming them beats asserting the number. The team already exists, so there is no assembly phase. Senior judgment is present from day one, so the scoping conversation and the build conversation are the same conversation. And the engagement starts on one narrow workflow instead of a platform, so there is something small enough to finish inside a month.
What breaks the benchmark matters just as much, because the number is conditional and any vendor who tells you otherwise is selling. Two to four weeks holds when the workflow is scoped, has a named business owner, and the data is accessible. It does not hold when nobody internally can decide what "correct" looks like, when data access needs a security review that has not started, or when the actual request is to rebuild a platform instead of shipping a workflow. In those cases no delivery model is fast, and the honest answer is that the constraint is not the team.
Who owns the risk when it breaks
This is the axis a CFO asks about first and the market writes about last. Four distinct risks hide inside it, and the three models trade them against each other instead of eliminating any.
Delivery risk. In-house, you own it completely. If the hire underperforms or the approach is wrong, that is your loss and your remediation. An agency contract shifts some of it, but read what the contract actually says, because most of them commit to effort and deliverables, not to business outcomes. A fractional engagement sits in between, and the honest framing is that the risk is shared, not transferred.
Key-person risk. This is where in-house looks worst and rarely gets scored. If your AI capability is one or two people and one leaves, the capability leaves. On a skill this scarce, with recruiters actively working your team, that is a live risk and not a tail one. Agencies and fractional teams both spread it across a bench.
Knowledge risk. The mirror image. Agency work concentrates knowledge outside your building, and at handover you receive a system plus documentation and hope the documentation is good. In-house keeps knowledge inside by default. A fractional engagement should be structured so knowledge transfers continuously instead of at an exit event, and a provider who cannot explain how that works has told you something important.
Dependency risk. The one to hold against fractional specifically. An embedded external team that ships everything can become the only group that understands how your AI systems work. That is a genuine exposure. Documentation standards, internal pairing, and a deliberate build-up plan manage it, but none of that happens automatically, and any provider who does not raise it is not being straight with you.
What each model assumes you already have
Every model is described as though the buyer is ready for it. They are not interchangeable in what they demand.
Hiring in-house assumes you can evaluate AI engineers. Most companies cannot yet, and the ones that can usually know it. The market is full of candidates whose experience is calling an API rather than building a system, and telling the difference requires someone who has built one. It also assumes a technical leader who can direct the work once it exists. Hiring a senior AI engineer into an organization with nobody to direct them produces an expensive, frustrated person who leaves.
Using an agency assumes you can write and enforce a specification. The model works when you know what you want built. It fails when discovery is really happening inside the build phase, because you pay for that discovery at delivery rates and receive it as change requests. It also assumes someone internal owns acceptance and can say no.
Using a fractional team assumes you can name an owner and grant access. The bar is lower, which is the point, but it is not zero. Someone in the business has to own the workflow and be able to say what a correct output looks like. Data access has to be grantable inside the engagement's timeframe, not after a six-month review. If neither is true, a fractional team will spend its first month doing your internal alignment, and you will pay delivery rates for it.
A short self-diagnosis. Can you tell a real AI engineer from a good API integrator? Can you write a specification precise enough to hold a vendor to? Can you name, today, the one workflow that matters most and the person who owns it? Your answers point at a model more reliably than any rate comparison does.
Five situations, and what we would tell you
Abstractions are easy to agree with. Here are five concrete calls, two of which point away from us.
A Series B software company where AI is becoming the product. Hire, and start now. When the AI capability is the thing customers buy, it belongs inside the company permanently, the domain knowledge compounds, and you can afford the ramp because the horizon is years. Use external help to bridge the hiring gap if you need to, but the destination is in-house.
A mid-market manufacturer automating three back-office workflows with no internal AI capability. Fractional. The work is a portfolio of workflows rather than one project, there is nobody internally to direct a hire, and the value arrives early enough to fund the next step. Hiring here means spending two quarters and a senior salary to discover what a fractional team would have told you in month one.
A company with a fully specified system to build and a technical lead who can own it. Agency. This is precisely the case the model is built for. The specification exists, acceptance criteria are clear, someone internal can enforce them, and there is a defined end. Do not retain an ongoing team for work that has a finish line.
A company that needs one integration shipped in six weeks and then nothing. Contractor or agency, not a retained team of any kind. Bounded work with no follow-on does not justify an ongoing engagement, and any provider pushing you toward one for this shape of work is optimizing for their revenue, not your outcome.
A leadership team that cannot agree on which workflow matters. None of the three yet. Every delivery model will happily start, bill you, and build something, and none of them can fix an internal disagreement about priorities. Do the discovery work first and pick a delivery model afterwards, when you know what you are buying. If that is the honest description of where you are, an AI Readiness Snapshot is a free thirty-minute call that will get you to a decision faster than a procurement process will.
If the second scenario sounds like your company, that is the shape our Fractional Agentic Team is built for. Embedded senior people, an ongoing engagement rather than a project, and a first production workflow in weeks rather than quarters. And if you read the third or fourth scenario and recognized yourself instead, hire an agency. That is the right answer, and we would rather you got it right than got it from us.
Questions to ask before you sign
Internal questions first, because they determine which vendor questions matter.
- What does this capability need to look like in eighteen months, and does this model get us there or away from there?
- Who owns the workflow after it ships, and do they know that yet?
- What does it cost us to reverse this decision in month four?
- If this produces a wrong output in production, who notices, and how?
Then, by model.
Hiring. What specifically will this person have shipped by month six? Who reviews their technical decisions? What happens to the roadmap if they leave in month nine?
Agency. Who exactly is on the team, and will those people still be on it in month three? What is the change-request process and what does it cost? What is handed over at the end, and who on our side receives it? What does the contract commit to, effort or outcome?
Fractional. Which named senior people are on this engagement? What has this team put into production, for whom, and can we speak to them? How does knowledge transfer to our people while the work is happening? What are the stop points, and what happens to the systems if we stop?
Red flags, and how to change your mind in month four
Stated evenly, because all three models have them.
Hiring red flags. A requisition written around a title rather than a shipped outcome. An interview loop with nobody in it who has built a production AI system. A plan that depends on one hire succeeding. Compensation benchmarked against general software engineering instead of the actual market for the skill.
Agency red flags. A fixed-price bid with no discovery phase, which means the risk has been priced into the change requests rather than removed. A proposal team you never meet again after signing. "AI" that turns out to be a thin wrapper over a model API with no evaluation, no monitoring, and no failure handling. Reluctance to name which specific engineers will do the work.
Fractional red flags. No named senior people. No production track record you can verify. A monthly fee with no accountability for outcomes, which is staff augmentation with better branding. And no answer to the dependency question, which is the one you should ask hardest.
On reversal. Structure the first engagement so that month four is a decision point rather than a sunk cost. In practice that means one workflow instead of a platform, documentation as a deliverable instead of a promise, and infrastructure in your accounts instead of the vendor's. Do that and every path stays open. A fractional engagement can hand over to an in-house team you hired in the meantime. An agency project can become the foundation an internal team maintains. The trap is not choosing wrong. It is choosing in a way that makes the next choice expensive.
Key takeaways
- The decision is usually made on rate and usually decided by speed, risk, and organizational readiness. Compare all six axes, or you are comparing the cheapest column of a spreadsheet.
- Time to first production workflow is the number that matters, and it is the one nobody decomposes. Ask any provider to show you the weeks.
- Each model demands something of you before it works. Hiring demands the ability to evaluate AI engineers. An agency demands a specification. A fractional team demands a named owner and data access.
- In-house wins when AI is a permanent core capability. An agency wins when the scope is genuinely fixed and someone internal owns acceptance. A fractional team wins when you need production output before you know your permanent shape.
- Structure the first engagement so month four is a decision rather than a sunk cost, and the cost of being wrong drops for every model.
The three options are not better and worse. They are answers to different questions, and the question you are asking is rarely the one the rate comparison answers. Work out which of the six axes is binding for your company this quarter, and the choice usually makes itself.
If that answer points toward embedded senior capacity rather than a permanent hire or a fixed project, the Fractional Agentic Team is where to start. If you are still deciding which workflow matters, start with a free AI Readiness Snapshot instead.
