Your previous AI implementation didn’t deliver. What went wrong and how to fix it

The five most common causes

1. Tool first, problem second. The company buys access to an AI chat and looks for something to use it for. The result: a curiosity, not a saving.
2. No metric. Nobody decided before launch what should change — handling time, headcount, number of cases. A few months later there’s no result to show the board.
3. A general model instead of company data. A public model doesn’t know your procedures, documents or offer, so it gets specific tasks wrong.
4. AI next to the process, not in it. Employees copy content into the chat and back again. Real savings only appear when AI is connected to systems: the calendar, CRM, document system.
5. Too much scope at the start. An “AI across the whole company” project rarely succeeds. One well-chosen process does.

How to start again

• Choose one process with the most repetitive work: document analysis, tenders, customer registration, answering the same questions.
• Set a metric before launch and record the baseline.
• Build the solution on company data in a controlled environment — without sending documents to public models.
• Connect AI to the system where the work happens.
• Measure and decide: expand, change or stop.

Examples from Coblit deployments

• Tenders in a pharmaceutical company: a dedicated AI model downloads notices from an aggregator, scores their fit and fills in the documents. Handling tenders now takes 3 people instead of more than 10.
• Patient registration in a clinic: AI talks to the patient, collects details, runs a preliminary interview and books the appointment — reception is no longer a bottleneck.

Both projects started with one process and a specific metric. See how we run an AI pilot.