Do the math before you build

Salaries are the visible part. The simulator also counts cloud and GPU compute for inference, LLM API tokens, the vector database, observability and CI/CD tooling, and the maintenance team you need from year two onward. On the mid-market profile those non-salary lines add roughly $36k per year before anyone has touched the model.

The line teams forget most often is maintenance. LLM models move every three to six months, so the work does not stop at go-live: monitoring, retraining, migrations and security updates continue for as long as the assistant runs.

The simulator defaults to 6 months for a mid-market build and 9 months for enterprise, covering hiring, architecture, MVP, testing and rollout. Hiring alone runs 45 to 60 days for the profiles involved. With iAdvize, go-live is 15 days.

That gap is not neutral. Every month without the assistant is a month of incremental margin you do not capture, which the simulator shows as opportunity cost. On the default mid-market assumptions it is the single largest line in the comparison after the build team itself.

This is where the two scenarios separate hardest. Building means a mobile SDK to write and a native UX to adapt (3 to 6 months), infrastructure to adapt per region including data residency (1 to 2 months), and a new module per additional use case (2 to 4 months).

On iAdvize those extensions are part of the subscription. The mobile SDK is ready to integrate, a new market deploys in hours, and new use cases arrive through the product roadmap. Over three years that difference is worth $180k to $420k on the build side.

The engineering costs use loaded market salaries for senior AI/ML profiles, $90k to $150k per year (talent.io, 2024). The execution-risk framing uses the MIT / Fortune 2025 finding that 95% of enterprise generative AI pilots fail. The uplift assumptions come from A/B tests run across iAdvize clients: +5% on conversion on average and +3% on average order value, with top performers at +11% and +10%.

Every input is editable. If your salary bands, traffic or margin differ, change them and the whole analysis recalculates. The model is deliberately conservative on the build side.

No, and the comparison is usually more useful at that point, because you now have real numbers instead of estimates. Set the development timeline to what is left rather than the full project, and put your actual team cost in. Sunk cost does not belong in the decision: what matters is the cost to finish and maintain, against the cost to switch.

Building is the right call in one specific case: an existing dedicated AI team plus technical requirements the platform genuinely does not cover.