5 Questions Before Your Enterprise AI Project
Over the past year we delivered more than 20 enterprise AI projects. We also saw just as many projects that were well thought out before kickoff as projects that started without a clear plan. The outcomes were dramatically different. This article summarizes the five questions every enterprise should answer before starting an AI project.
1. Is the scenario right?
Good AI scenarios have three traits: they happen frequently, they have a standard answer or a clear quality bar, and mistakes are affordable. Customer-service Q&A, knowledge search and form filling are typical first-win scenarios. Conversely, letting AI make strategic decisions or directly operate core production systems is something the first batch of projects should avoid.
2. Is the data ready?
About 70% of the work in an AI project is data work. Before kickoff, check three things: are documents digitized, are formats consistent, and where are the confidentiality boundaries? In one equipment-group project, nearly 40% of 30,000 technical manuals were scans and images. Document parsing alone consumed one third of the project schedule.
Rule of thumb: if your knowledge still lives in employees' chat histories, fix knowledge management before talking about AI.
3. Who will use it?
An AI system is not delivered to the IT department; it is delivered to frontline business teams. Define seed users at kickoff. In every project we ask the client to provide 10 to 30 frontline users for the pilot and collect feedback weekly. One after-sales team reached 92% weekly active usage because users were involved from day one.
4. How will success be measured?
AI projects without metrics always end in disputes at acceptance. Put three categories of metrics into the contract: efficiency (response time, hours saved), cost (repetitive manual work reduced) and quality (manual sampling during the pilot, with an agreed answer-accuracy threshold).
5. Who should build it?
Eighty percent of an AI project is engineering, not model work. Model selection is only about 20%; the rest is data governance, system integration, permissions and compliance, and operations. A team with strong software-engineering fundamentals that is willing to work on-site will beat a team with the biggest model parameters on a slide.
If you are evaluating an AI project, talk to us. We offer a free scenario assessment and can build a POC prototype within two weeks.