Why two similar projects quote differently
Two businesses asking for demand forecasting can receive quotes that differ by a factor of several. That is not necessarily one supplier overcharging - the underlying work genuinely differs that much.
The differences are predictable, and knowing them lets you estimate roughly where your own project sits before you ask anyone.
Data condition is the biggest factor
| Data situation | Effect on cost |
|---|---|
| One system, clean, outcomes recorded | Lowest |
| Several systems, joinable by a shared key | Moderate |
| Several systems, no shared key | High - matching is a project |
| Outcomes not reliably recorded | Very high, or not feasible yet |
| Free text needing structuring | High, and often the whole first phase |
| Historical data unavailable as it stood | Limits what is possible at any price |
The join problem is the one that most often surprises people. A business with data in a CRM, an ERP and a website that cannot be reliably connected has a data engineering project before it has a modelling one.
Integration is the second
A model producing a number in a report is far cheaper than one whose predictions must appear inside an existing system at the moment of decision.
- A report or dashboard - lowest integration cost
- A scheduled file or feed into another system - modest
- Predictions written back into an existing application - significant
- Real-time scoring inside a customer-facing flow - highest
- Changes required to a third-party system you do not control - potentially blocking
That final row can dominate everything. If acting on the prediction requires a change to a system whose vendor charges for it and schedules it next year, that constraint sets the project timeline regardless of the modelling.
What usually costs less than expected
The modelling itself is frequently a modest share of the total. The techniques are well established, the tooling is mature, and for most business problems the work is applying known methods carefully rather than inventing anything.
Compute costs are also usually modest for typical business data. Training on a few million rows of tabular data is not expensive, and the figures that alarm people generally come from a different kind of work entirely.
Getting a useful estimate
The most effective thing you can do to get an accurate quote is let a supplier look at a sample of real data early, under an appropriate agreement.
Without it, any responsible quote includes contingency for the problems that data usually contains. With it, the estimate narrows considerably - and occasionally the answer is that the data will not support the idea, which is worth knowing for the cost of a short engagement rather than a build.
The model is rarely the expensive part. Getting the data into a state where a model is possible usually is.