The cost of an AI project comes from eight things: building the software, preparing the data, model usage, infrastructure, human review of exceptions, monitoring, maintenance, and helping people change how they work. You pay for building, data preparation, and most of the change management around launch. Model usage, infrastructure, review, monitoring, and maintenance continue for as long as the software runs, so estimate them before you approve the build.
Building the software
Build cost depends mostly on scope. It goes up with the number of systems the software reads from and writes to, the number of document types and exceptions it has to handle, how many of its results need a person’s judgment, and the security review your organization requires.
Include testing in the build estimate. Someone has to collect examples with accepted answers, run the software against them, and have the people who know the work review the differences. That takes their time as well as the engineers’ time.
Preparing the data
Find the records, confirm you have permission to use them, and check whether they are complete and current. Cleaning records, matching them across systems, and masking sensitive fields take engineering time. So does assembling test examples with known correct answers.
If the data do not exist or cannot be used, change the scope before you spend money on the build. A dashboard, a rule, or a fix to an upstream process may cost less than an AI project and solve the same problem.
Running costs
Model usage is usually billed by volume. The cost depends on how many items you process, how many model calls each item needs, and which model you use. Use ordinary code for calculations with a defined rule, reuse results when the inputs have not changed, and use a smaller model where it passes the same tests.
Infrastructure, storage, licenses, and support add a steady monthly cost. Human review is easy to leave out of the estimate. Someone has to handle the exceptions the software sends to people, and their time belongs in the estimate.
Monitoring and maintenance
Monitoring tracks inputs, correction rates, processing time, and cost per item, and someone has to act when those numbers change. Maintenance covers new document layouts, changed business rules, model version changes, and a test run after each change. Estimate both for each year you expect to run the software.
The last cost is change management. People need training on the new work, managers need to reassign time, and some roles change. Budget for it alongside the build.
Staff time and budget savings
Report the return on two separate lines. Staff time released is time the work no longer needs. Budget saved is spending that stops, such as a contract, a license, or overtime you no longer pay for.
For an insurance carrier, the StrataEdge Reconciliation Platform reduced quarterly auditing and reconciliation from five people working for a week to one person working for a day. That is staff time released. It becomes a budget saving only if spending falls, so decide in advance where the time will go, such as resolving discrepancies sooner or taking on other work. Read the insurance case.
Our AI investment review usually takes two to four weeks. It covers the work, the data, and the full cost of building and running what you choose. See our typical engagements.