Planning: building confidence in AI forecasts

The estimated construction completion date (ECCD) was already part of every project. The goal was a predicted ECCD (pECCD), generated by AI from each project's constraints, that finance could plan around.

• Limited visibility into budgets. Every invoice, whether for quotable labor, estimated materials, or a project's appropriated amount, is billed against a project budget. That budget reflects what the build makes possible: the homes a new line reaches, the services Comcast can offer them, and related government funding and taxes. Finance had no straightforward way to see whether a job was tracking above or below budget, or when a budget would reach its limit.
• Understandable caution about AI. In late 2024, the team had little experience with AI tools, concerns about accuracy were common, legal and compliance had open questions, and the change itself was significant for users.

• Cost against budget at every level. Finance leaders could see each job's position against its budget by division, region, market, and area (DRMA), open or close invoices as a budget approached its limit, and plan resources accordingly.
• Explaining the predictions. A guide on Vox described how predictions were made and which factors influenced them.
• Transparency in the interface. The design showed what drove each prediction, a choice that took more development time but built trust.
• Highlighting bottlenecks. The guidance also showed where subcontractors were slowed down, so their Comcast counterparts could use Sphere data to help.

The finance team proposed Planning, adopted it, and became its advocates. Clear visibility into bottlenecks built confidence in the predictions, and simpler resource planning encouraged adoption. My role ended before the first satisfaction survey results were available.

Early adoption was slow, reflecting real concerns about AI. Working sessions with the product team helped identify where users felt uncertain, and we addressed those concerns through clear visual cues and greater transparency. I built much of the supporting guidance in WordPress, which allowed the developers to stay focused on the product.

People are more likely to trust AI predictions when they can see what drives them, and to adopt them when they make everyday work easier.