You spent two weeks on the assessment. Every assumption documented, every cost modeled, every scaling threshold footnoted. Twenty airtight pages. The stakeholder meeting opens and the first question is: "So what happens to the budget if traffic triples in Q3?"
The answer is on page 14. You know it is on page 14. And in that second you finally accept the brutal law of corporate physics: nobody is going to read page 14. Not the PM, not the VP, not even the engineer who reviewed it. They do not want homework. They want a dial they can turn.
Here is the thing technical people get wrong about those tweak requests. "Can we see it monthly instead." "Can we exclude enterprise." "Is it too much effort to add last year." That is control. Every stakeholder walks in with a mental model and will keep pulling at your artifact until it matches the picture in their head. In the old workflow, you paid for that instinct with your evenings: every tweak meant leaving the meeting, fighting the BI tool, re-exporting, re-pasting, re-sending. You were not an engineer anymore. You were a screenshot mule with a backlog.
Put the old loop next to the new one:
| The old tweak loop | The new one |
|---|---|
| Leave the meeting, reopen the BI tool | Stay in the room |
| Rebuild, re-export, re-paste, re-send | Say the tweak out loud |
| Half a day per request | Five seconds per request |
| You ration tweaks and resent them | You invite them |
Text-based AI made this worse, not better. The wall of prose now generates in seconds, so there is more of it, and the number that matters is buried in paragraph nine. At that point the AI is just a slower search engine.
You pay for that wall twice. Once in tokens. Again in reading time. And the cruelest part: you finish all three pages just to discover the AI answered the wrong question. Guess what gets blamed. Your prompt.
A chart cannot hide like that. Wrong picture, one glance, one sentence, fixed.
There is a reason people reach for a picture when they actually need to understand something. AI chat started as a wall of text; meetings needed a screen. Text does not carry a meeting. Text kills it.
Here is the version of that meeting I actually live now, using mcp-dashboards, a free MCP server that renders live charts, dashboards, and diagrams inside the AI chat itself.
The question comes: "what if traffic triples?" I do not flip pages. I type one sentence into the chat everyone is looking at: triple the traffic assumption and redraw the cost curve. Five seconds later the redrawn chart lands in the chat. The follow-up comes, of course it comes, "and if we stay on one region?" One more sentence. Redraw. "Exclude the enterprise tier." Gone.
Watch what just happened to the room's economics. The cost of a tweak fell from half a day to five seconds, so I stopped rationing tweaks. I stopped resenting them. Bring your tweaks, bring your boss's tweaks. Each one gets executed live, in front of everyone. Half the time, people watch the result of their own request land and admit it changed nothing. The argument that used to survive three email rounds dies in the room, on screen, in seconds.
The dial is not mine, either. Anyone in the room can turn it. The VP asks in her words, the PM in his, and the AI serves them the way it serves me, because it already has the numbers. Hook it into your own systems and it answers from your own data, with credentials sitting in local presets on your machine, never in the chat. The key never reaches the model and never appears on screen.
Nobody breaks the format. Nobody argues from imagination. Every ask comes back as a picture in the same clean layout, built for exactly these meetings. Nobody queues behind whoever owns the BI seat. The show stops running on text.
The document did not disappear from this workflow. It became what it always should have been: the archive, not the interface. The interface is a conversation with pictures in it now, and the answer stopped hiding.
Your stakeholders were never going to read the spec. Stop writing them a longer one. Hand them the dial, keep your evenings, and let the meeting argue with the chart instead of with you.
Give them the dial
One entry in your MCP config file (claude_desktop_config.json):
"mcp-dashboards": { "command": "npx", "args": ["-y", "mcp-dashboards", "--stdio"] }
Restart Claude and type:
Build me a dashboard from these numbers [paste yours]: revenue by month, top products, the conversion funnel, and a churn KPI.
The dashboard renders in the reply, and every tweak after that is a sentence executed live in front of the room. Runs on your machine, and no third-party chart service ever sees your data. Full setup in the README.
GitHub: github.com/KyuRish/mcp-dashboards · npm: npmjs.com/package/mcp-dashboards
Connect with me on LinkedIn.
Top comments (0)