Three Administrative AI Projects for Health Teams. None Touches a Clinical Decision.

All three of these are administrative. None touches a clinical decision. That is the point. ## 1. Forecast stock for one product group Take twelve months of dispensing and procurement records for a single group of medicines. Build a forecast from actual consumption rather than from historical order quantities, which is what most facilities are effectively doing now. Compare against what you ordered over the same period. The gap between consumption and ordering is usually where both the expiry and the stock outs are coming from. One product group, one facility or one cluster. Prove it, then widen. ## 2. Look at your scheduling honestly Pull appointment and capacity data for a quarter. Where did patients wait longest, and was capacity genuinely full at those times or unevenly distributed? Most organisations find the constraint is not total capacity but how it is allocated. That is a scheduling problem and it responds well to modelling. It also responds to simply seeing the numbers laid out, which many organisations have never done. ## 3. Claims and billing patterns For private providers and HMOs, take a year of claims and cluster them. Look for the patterns that sit outside the normal distribution. You are not accusing anybody. You are producing a shortlist for human review, which is what these tools are for. ## The rule that comes first Write down which AI tools are approved and state plainly that patient identifiable data does not go into any tool the organisation does not control. Then tell the administrative staff, not just the clinical leadership, because the records are usually being handled by people who were never in the policy meeting. BROAI in healthcare is not a theoretical exposure. Patient data is the most protected category under the Nigeria Data Protection Act, and an administrator reformatting a patient list in a free chatbot has created a breach while trying to be helpful. Provision matters more than prohibition. If the approved route is slower than the unapproved one, people will keep choosing the unapproved one. This is the Scan and Score phase of SWITCH, the framework I use with clients. Pick one of the three. Give it an owner and a number, whether that is stock outs avoided or waiting time reduced. Review it in thirty days. If you want this worked through with your own facilities, systems and data protection obligations rather than as a generic session, the [speaking and training page](/speaking) sets out the formats, or reach me at me@martins.com.ng or through the [work with me page](/work-with-me).