Human-only vs. AI-only vs. the third way: a cost model

When a client asks "why not just hire an agency?" or "why not just let AI build it?", we stop arguing philosophy and draw three columns. This post is those three columns, with the arithmetic we actually use. The numbers below are representative of the engagements we see; yours will vary, but the shape of the comparison rarely does.

01The human-only model

A traditional delivery team for a mid-sized product: one tech lead, two to four engineers, a part-time designer and PM — call it six people across six months. Most of their hours go to work that is inherently repetitive: boilerplate, test suites, data plumbing, environment wrangling, and the meeting overhead of coordinating six calendars. You are paying senior rates for a mix of judgment and mechanical work, because the traditional model has no way to separate the two.

02The AI-only model

A fleet of autonomous agents is nearly free per hour — and that is exactly the problem. With no human judgment in the loop, the failure modes are confident wrong architecture, security blind spots, and codebases that decay into unmaintainable sludge. In our assessment work we regularly meet teams that saved 80% on generation and then spent 300% on remediation. The unit cost is not the cost; the outcome is the cost.

03The third way

Our model: one senior engineer directing agents plus one product-minded human, for weeks instead of quarters. Agents absorb the mechanical work at near-zero marginal cost; humans spend their expensive hours exclusively on judgment: architecture, domain modeling, review, and trade-off decisions. The delivery quality is human-grade because the sign-off is human; the speed is machine-grade because the grunt work is machine-done.

04When each model wins

• —Human-only still wins for: novel systems research, deeply regulated builds where every decision needs an accountable human author, and organizations that cannot expose code to any AI tooling.
• —AI-only increasingly wins for: throwaway prototypes, one-off scripts, and glue code where remediation cost is trivial because the artifact is disposable.
• —The third way wins for: the enormous middle — products that must be good, on a calendar and budget that traditional delivery cannot meet.

If you take one thing from this model: stop pricing hours and start pricing the division of labor. The question is never "humans or AI" — it is "which tasks, specifically, get which treatment". Getting that split right is the whole game.