Custom Dashboards & AI Summaries

D-Tools had a single static dashboard that tried to serve sales, operations, and service teams with the exact same widgets. Because it didn't match anyone's actual daily routine, people barely looked at it. Managers just exported raw data into Excel every Monday morning and built their own charts. At the same time, leadership was excited about an AI-first interface where people would ask questions in a chat box. The real challenge was figuring out how to give teams useful day-to-day tools while introducing AI in a way that felt practical, not gimmicky.

Objectives

• Give teams starter templates matching their specific role so they stop exporting to Excel
• Let users rearrange widgets, resize cards, and save multiple views for different routines
• Keep AI summaries strictly tied to visible numbers on the screen so users can verify every claim

• Scope the feature

  I focused the release around role-based templates that helped users start quickly without locking them into rigid layouts. Users could start from a template, build from a blank dashboard, add widgets, rearrange sections, and choose visualization styles.

  The goal was to make dashboards flexible enough to replace spreadsheet exports without forcing every team into the same view.

• Build trust with data

  We kept AI strictly tied to verifiable numbers on the screen. The model wasn't allowed to draw speculative conclusions; it could only summarize the visible data on the active dashboard. To reduce unsupported answers and keep users confident in the numbers, I worked with engineering to constrain the LLM context window exclusively to the JSON payload of the user's active widgets.

Internally, the initial push was to bypass dashboards entirely and build an AI prompt box. But during user sessions, people froze in front of the blank input. Integrators don't think in open-ended prompts; they want to glance at a screen and immediately spot red flags or overdue shipments. Without visual reference points, an AI assistant felt untrustworthy. Once managers could see their numbers, an AI summary that highlighted week-over-week anomalies actually saved them time.

Interviews backed that up, so we landed on a clear split:

• Dashboards: show current status, surface anomalies, and give operators an anchor
• AI summaries: pull out key highlights and generate quick updates for stakeholders

I prototyped the full flow in Figma and tested it with five target users, walking them through picking a template, rearranging cards, and generating a report.

Iteration: more data density, more control

Our early prototypes were pretty minimal, but users immediately asked for more data density. They wanted quick-hit KPI counters for the numbers they check every single morning, and they wanted to drag those cards right to the top of the page. Adding that level of customization stretched our timeline a bit, but it made the difference between something operators would actually leave open on a second monitor and another tool they'd ignore.

• 4/5 users said templates matched how they think about their work
• All 5 participants customized and saved a new dashboard without prompting

These sessions gave us the confidence to cut complex formula widgets from the initial release and focus solely on template flexibility.

It’s easy to get swept up in technology trends, especially when executive teams are excited to ship something with "AI" on the landing page. But our user research gave us clear cover to push back. Integrators didn't want a chatbot; they wanted to know if their morning deliveries were late and whether their service jobs were profitable. By treating AI as a tool that summarizes visible chart data rather than an all-knowing chat assistant, we built something operators could actually trust and use every day.