| What you are deciding on | In-house team | Agency (us) |
|---|---|---|
| Time to first result | Three to six months before the first hire is productive: search, notice period, onboarding, then the project. | Weeks. The mini-audit runs in days and a first build is typically live in four to eight. |
| Cost shape | Fixed and ongoing. Two senior specialists plus tooling is a substantial annual commitment whether or not there is a queue of work. | Variable. Fixed price per project, a retainer only if you want one, nothing between projects. |
| Cost at high volume | Cheaper once the pipeline is continuously full. This is the real argument for in-house. | More expensive per unit of work at that volume — you are paying a margin. |
| Domain knowledge | Deep and compounding. They know your systems, your exceptions and your politics, and it gets better every month. | Learned during the audit and inevitably shallower. We compensate with pattern knowledge from other builds; it is not the same thing. |
| Breadth of exposure | One company’s problems. Skills can drift if the work is repetitive. | Many companies’ problems. Our advantage is having already seen where this class of project fails. |
| Continuity risk | Concentrated in individuals. One resignation can strand a system nobody else understands. | Concentrated in the relationship. Mitigated only by documentation and handover, which is why we treat both as deliverables. |
| Hiring risk | Real and often underestimated. Assessing senior AI engineers is difficult if nobody senior in-house can assess them yet. | You are assessing delivered work instead of a CV, and the first engagement is small enough to be a test. |
| Who owns the result | Unambiguously you. | Also you — code, configuration and data — but only if the contract says so. Ours does; check that any agency’s does. |
Both columns are written from the same set of projects: engagements we have run, and clients who left us to build a team, or arrived after trying to.
You need something in production this quarter and a hiring round finishes after that.
This is your first or second automation and nobody internal can yet interview for the role.
The work is periodic — a few projects a year — and would leave a permanent hire underused.
You want the first project to also produce the documentation and the internal capability, which is a deliverable we scope explicitly.
Automation is becoming core to the product rather than to the back office. Do not outsource the product.
You have a continuous pipeline that would keep two or three people busy all year — at that volume, in-house is simply cheaper.
The domain takes months to learn and is genuinely unusual. A resident team will outperform any visitor.
Regulation or client contracts prevent external access to the systems involved. Say so early; it changes the whole shape of the answer.
A useful in-house unit is not one person. It is at least a senior engineer who can build and own production systems and someone who can specify the work with the business, plus tooling and model budget. Below that, one person becomes a single point of failure and spends half their time on work that is not engineering.
Against that fixed annual commitment, agency work is priced per project. The crossover is not a headcount number, it is a utilisation number: if you can keep that unit genuinely busy, in-house wins on cost within the first year. If the honest answer is three or four projects a year, the fixed cost sits idle for months and the comparison inverts.
The hybrid that works in practice, and that we are happy to be part of: an agency delivers the first one or two projects while you hire, the internal team takes over operation with the documentation and monitoring from those builds, and the agency stays for peaks and unfamiliar territory. Nobody waits, and the capability ends up inside the company.