How is ROI for an AI or IT system implementation typically calculated? It usually starts with three numbers: development cost, licenses, servers. Add them up, subtract from the projected savings, and the numbers work on paper.
The problem is that these are exactly the line items no one forgets to calculate: they appear in the first commercial proposal or the first budget draft. But there are items that surface later, and they can significantly affect a project’s economics. These costs most often become visible only after the pilot is running and the team attempts to scale it. It turns out there are also process changes, staff retraining, preparation of data that no one had previously labeled, and equipment deployment across dozens of locations instead of one.
In my experience, hidden costs can easily double the original budget, and in some cases barely affect it at all. The difference comes down entirely to what “implementation” means for a specific business.
The material below expands on one point from How to Measure ROI From AI Implementation. If that article provides the full calculation framework, this one covers the part where errors most often occur: the costs the original estimate simply did not account for.
What Companies Typically Include in the Calculation
When a company sits down to calculate ROI for an IT system implementation, the expense list is usually closed after a single pass:
- Development
- Model training
- Servers
- Software licenses
The logic is straightforward: these are the line items that come up with a vendor from the first conversation, are asked about in the first brief, and appear in the first estimate.
Development, because someone has to write the system. Model training, because the system does not function without it. Servers, because computation has to run somewhere.
The problem begins where these three items are treated as the complete list. Breaking down only the infrastructure component to its full extent, rather than stopping at the first figure, quickly complicates the picture.
System development might nominally cost one million. Deployment: another $200,000-$300,000, because the servers required have GPUs, not standard hardware. If data cannot be kept in the cloud, a regulated industry such as healthcare or finance may require its own data center, power for those servers, and staff to maintain them. If the task requires cameras or sensors on site, that is a separate line item for equipment procurement and installation. Multiply that by the number of locations.
Strictly speaking, this is not yet a hidden cost. It is the same list, simply carried through to completion. The bulk of hidden cost begins where infrastructure ends and everything around it begins: people, processes, data, organization.
Categories of Hidden Cost in AI Implementation
Personnel not accounted for in the initial estimate. The first thing usually missed: a system cannot be trained by one person alone, even someone well versed in AI. An engineer is required to design the architecture and data structure, train the model, and validate the result. But if the task involves a specialized domain — labeling X-rays, for example — this is not sufficient.
A domain expert is also required: someone who can identify which region of a scan indicates a rib fracture and which indicates pneumonia. Without this person, there is no one to label the data. Training the model requires labeling tens of thousands of scans, followed by validation — confirming that a model trained on one set of scans performs without errors on other data and produces a stable, predictable result.
Hiring this combination of skills often becomes a separate budget line that is rarely accounted for at the outset. At the briefing stage, “model training” is discussed as a single line item, not as the multiple specialists of varying expertise, working over months, that stand behind it.
Data preparation is a separate, substantial undertaking. Model training requires data that someone has already labeled: marking where on a scan a fracture is located versus a hairline crack, which test results are acceptable and which are not. This is manual work performed by a domain expert, not a programmer, and its scope is difficult to estimate in advance — it depends on how much data actually needs to be processed, which only becomes clear within the specific project.
Deployment is a matter of scale. The cost of deploying the same system varies dramatically depending on where it will run. Cloud-based, single-location deployment is one scenario. Deploying across 10 to 100 to 500 subdivisions is a different one entirely. Physically installing equipment across 100 to 200 to 500 locations is yet another scenario, and cost does not scale linearly with deployment logistics — who installs the equipment on site, who configures it, who coordinates this across every location.
Regulatory constraints multiply infrastructure requirements. In industries such as healthcare, data cannot simply be sent to a public model: GDPR, HIPAA, and internal client policies directly prohibit it. This leaves one option: deploying models locally, with dedicated GPU servers, a data center, or dedicated infrastructure at the client’s site. This brings power costs for those servers and staff to maintain them. A company that planned to rely on a cloud AI subscription discovers at this stage that it needs its own server infrastructure — an entirely different order of cost.
Minor line items also accumulate. Additional items that rarely appear in the initial estimate surface separately: software licenses for specific access, hiring specialists the company does not have on staff, a dedicated or faster internet connection (for example, if the system requires real-time video streaming from a site and the existing connection cannot handle it). Individually, each item is small. Together, they add up to a meaningful budget increase that no one anticipated at the start, because each one is tied not to the technology but to the specifics of the client’s business.
Organizational resistance is a further risk to keep in mind. There is another category of hidden cost not directly tied to money: resistance within the client’s organization, which costs time.
Daron Acemoglu calls this “creative destruction.” In any organization, there are groups of people who stand to lose from changes that benefit the company overall. In industries where procurement volume is central to the business, for example, a reduction in purchasing volume may theoretically be disadvantageous to a specific department or individual within the company, and project progress may slow for organizational rather than technical reasons.
This is a theoretical risk — it is not confirmed on every project, but it is worth keeping in mind when planning timelines, particularly if the project directly reduces something that affects specific people’s interests.
How Much This Changes the AI Implementation Budget
There is no fixed-percentage answer to “how much exactly.” Based on my experience, hidden costs can easily double the original estimate.
The range is genuinely wide. In some cases hidden costs barely materialize at all — if the system runs through cloud access and data does not require local storage, additional line items may simply not arise. Minor items such as internet connectivity or licenses rarely determine a project’s outcome.
The most expensive failure is underestimating infrastructure as a whole, when the deployment involves building dedicated capacity for model training. If the solution required its own servers, a data center, and staff to maintain them from the outset, instead of a cloud subscription, this is effectively a different project altogether, in which the budget changes wholesale rather than increasing by a percentage.
How to Check Your AI Implementation Cost Estimate and Avoid Hidden Cost Risk
Before entering an AI implementation cost figure and presenting the expected ROI to leadership, the estimate should be run through several questions.
Who will train the model? If the data is domain-specific (medical scans, test results, industry-specific documentation), both an engineer and a domain expert are required to label the data. This is either a separate hiring line item or client staff time diverted from their regular work. Neither typically appears in the first estimate.
Where will the data physically be stored and processed? If the industry is regulated (healthcare, finance, any work involving personal data), the answer “in the cloud” may not be legally available. This brings its own servers, data center, power, and maintenance staff.
To how many locations does the solution actually scale? The cost of a single-site pilot and the cost of deploying to a hundred sites are not related by simple multiplication. Logistics, coordination, and on-site equipment configuration scale independently of the technology itself.
Are there people or departments within the organization for whom the project’s success means a loss for them? Not always, but if the project directly reduces procurement volume or someone’s current responsibilities, this can slow implementation for organizational rather than technical reasons.
What happens if the underlying prices change? Savings calculated at today’s cost of raw materials, supplies, or labor are not fixed. If the thing the system saves on becomes more expensive, the savings increase. If it becomes cheaper, the savings disappear. This is why the payback period should honestly be presented as a range rather than a single figure.
If any of these questions currently has no answer, that is not a reason to halt the project. It is a reason not to put a final ROI figure into the presentation until the answer exists.
Hidden costs do not invalidate the point of calculating ROI, and they do not mean AI projects are inherently unpredictable. This is a separate category of cost, as real as development and servers — it simply does not come to mind in the first meeting with a vendor.
If an estimate is already prepared and looks convincing on paper, before presenting it to leadership, it should be run through the five questions above: who trains the model, where the data is stored, how many locations the solution scales to, who inside the company might slow the project down, and what happens to the economics if the underlying prices change.
If at least two of these questions remain unanswered, the ROI figure is not ready yet. It becomes reliable exactly when the answers exist.
If you need help with implementation or consultations for your specific case, contact me via SOFYCOD corporate web site Contact Author, and I can consult you about any questions related to development and implementation of Computer Vision systems.