Before you invest in generative AI, decide what you need it to produce. Then check whether AI can help, how people will review the work, and what it will cost to operate after launch.
What the application will do
An application might draft a document, compare records, recommend an action, or call another service. Specify the inputs, the output, and what the software may do without approval.
For example, drafting a customer email needs access to the customer’s records, and sending it is a separate step that may need approval. Decide who reviews the draft and how corrections are handled.
Start from what you need the software to produce. A rule, calculation, or short script may handle some steps. If several applications must work together, define what each passes to the next, how failures are reported, and who handles exceptions.
Data and access
Find out whether the required records exist, whether they are current and complete, and whether the team has permission to use them. Price the work of obtaining, cleaning, and connecting the data.
Preserve source information so reviewers can check an answer. A person using the model should not be able to see records they could not open otherwise.
Define which records the software can read, which systems it can change, and who approves those permissions. Agree on how corrections and updated records reach the people and applications using them.
Build and operating costs
The initial cost includes design, data preparation, integration, testing, training, and deployment. Ongoing costs include model usage, infrastructure, licenses, monitoring, exception review, support, and maintenance.
Compare the total with the current process. Measure time released, errors reduced, and additional work the team can complete. Report staff capacity and hard budget savings separately.
Budget for testing changes to models, inputs, and business rules. Test a replacement model against accepted examples before changing production. Check software licensing terms for automated access and usage-based charges.
The people who run it
Decide who in the business owns the application and which team will run it. Involve the people who understand the work in defining review criteria and testing outputs.
Staff need to know how their work changes, which decisions they make, how to report a problem, and when to fall back to the manual process. Managers need to plan how work will be assigned.
Give technology teams clear responsibility for integrations, permissions, monitoring, and maintenance. Train people with the records and decisions they will encounter, including incomplete data and incorrect outputs.
A first project
Pick one thing you need the software to produce, measure how the work performs today, and decide how you will test it. Agree on the data the client will provide, who will review the work, and what would cause the team to continue, change, or stop the project.
We help make those decisions and carry out the work. Talk with us about your AI project.