There are only three real paths to building your first AI feature: hire and build in-house, bring in fractional or embedded specialist talent, or hand the whole thing to an agency. Every founder I talk to has already ruled one of these out for the wrong reason, in-house because it seems slow, agencies because they seem expensive, fractional because it seems like it won't stick. All three are viable in the right situation, and picking the wrong one for your situation is one of the most expensive mistakes a company building its first AI feature can make.
The three paths, side by side
| Dimension | In-house hire | Fractional / embedded | Agency |
|---|---|---|---|
| Speed to start | Slowest, weeks to months to hire | Fast, days to a couple weeks | Fast, but scoping and contracting add time |
| Cost structure | Highest fixed cost, ongoing | Mid cost, scales with need | Often lowest cash cost per project, but IP and continuity cost later |
| Who maintains it after launch | The team you hired | Your team, with the specialist's knowledge transferred | Nobody, unless you pay for an ongoing contract |
| Institutional knowledge | Stays with you | Mostly stays with you if transfer is designed in | Leaves when the contract ends |
| Best fit | Core, permanent AI capability | First build, senior judgment needed, team ramping | Bounded scope, proof of concept, no plan to own it long-term |
Why 'what's cheapest' is the wrong first question
Every path has a real cost, the question is which cost you're willing to carry. In-house carries the cost upfront and continuously, in salary, whether or not the feature succeeds. An agency defers cost into the future, in the form of institutional knowledge that walks out the door and a codebase your own team has to reverse-engineer if the feature becomes important enough to bring in-house later. Fractional sits in between: you pay for senior judgment when you need it, and if you design the engagement to transfer knowledge, your team owns the result afterward. The right question isn't which path is cheapest, it's which cost you can least afford: cash now, or ownership later.
When in-house is right
In-house makes sense when the AI feature is, or will become, core to your product, when you expect to iterate on it for years rather than ship it once, and when you're prepared to pay ongoing salary whether or not any given quarter's iteration succeeds. It's the only path that reliably builds durable institutional knowledge inside your team, which matters enormously once the feature is load-bearing for the business.
- The feature is core to your product roadmap for the next several years, not a one-off.
- You have or can get senior judgment on staff, or via fractional support, to check the build.
- You're prepared to carry the cost through the inevitable false starts before the system is good.
- Retention and continuity of the people who understand the system matters to you long-term.
When fractional or embedded is right
Fractional and embedded engineering exists to solve one specific, common problem: you need senior AI judgment now, but you don't yet have the scope, budget clarity, or conviction to make a permanent senior hire. An embedded specialist can set the architecture and evaluation methodology for your first system, work alongside your existing team, and hand off a system your people can maintain, at a fraction of the time-to-start and often the total cost of a full-time senior search.
- You need senior-level architecture and eval decisions made correctly the first time.
- You're not ready to commit to a full-time senior hire, whether for budget or scope-certainty reasons.
- Your existing team can maintain a well-designed system once it exists, they just can't design it from zero yet.
- You want the option to convert to a full-time hire later, once you know the role and scope.
When an agency is right, and when it quietly isn't
Agencies are genuinely the right call for bounded work: a proof of concept to test with investors or an early customer, an MVP you need to see working before deciding whether to invest further, or a well-defined feature you plan to fully outsource on an ongoing contract basis. Where agencies quietly go wrong is when 'quick outsourced MVP' silently becomes 'the core AI feature the whole product depends on', at which point you own a system nobody on your team designed or fully understands, built by people who are no longer in the room.
- Good fit: an MVP or proof of concept with a clear, bounded scope and a decision point after delivery.
- Good fit: work you have no intention of owning or iterating on internally, ever.
- Bad fit, retroactively: a 'quick outsourced build' that becomes permanent product infrastructure without a deliberate decision to keep it that way.
- Bad fit: anything where a wrong architectural call is expensive to discover after the agency has moved on to its next client.