Reports and Analysis in the AI Era: A Practical Approach
TL;DR: Upload your data + provide its structure to a Claude Project → ask any question in plain English → get answers, charts, insights. You can even ask “what questions can you answer with this data?” and the AI tells you what’s possible. This complements traditional dashboards, it doesn’t replace them.
We track AI costs across multiple cloud providers, AWS, GCP, OpenAI, Claude workspaces, and various internal tools. The team built a nice dashboard for this. It works well for what it was designed to show.
But I wanted to try something different. What if the report itself could answer questions I haven’t thought of yet?
I uploaded a JSON file to a Claude Project and started asking:
“What’s our total AI spend?”
$100,452.55
“How much did we spend in December and what was the main driver?”
$131,544.80. OpenAI was the largest contributor at $39,576.88.
“Give me a graph of AI cost over time.”
[Chart generated showing the December spike]
Then I asked something you can’t ask a dashboard:
“What type of questions can you answer based on the data you have?”
The AI responded with a full breakdown: overall spending, provider comparisons, monthly trends, per-workspace costs, hosting vs AI cost separation, daily granularity… It told me what was possible with this data.
This is fundamentally different from a dashboard where someone has to anticipate what views you’ll need.
What Makes This Different
A dashboard shows you what someone designed it to show. Even with fast development cycles, someone still needs to decide what views to build, what filters to include, what drill-downs matter.
With this approach:
No anticipation needed. You don’t need to predict what questions will come up. The data is there, the AI understands its structure, you ask what you need when you need it.
You can ask about the data itself. “What can I learn from this?” is a valid question. The AI examines the structure and tells you what’s queryable. A dashboard depends on documentation and shared understanding of what it contains.
It improves as AI improves. When Claude gets better at analysis, visualization, or reasoning, your report project gets better automatically. No rebuild required.
Follow-up questions are natural. “What drove the December spike?” → “Show me OpenAI breakdown by project” → “What’s the trend for the top 3 projects?” The conversation flows. Each follow-up doesn’t require a new view to be built.
How to Set This Up
Here’s exactly what I did.
The Data
I had a large JSON file (ai_costs.json) with our cloud cost data. The file is too large to fit in the AI’s context window entirely. That’s fine.
The Key Instruction
This is the core of the approach. In the Claude Project, I provided this instruction:
So I uploaded the file ai_costs.json to the project. It is large and you cannot read and keep it in your context in one go, but I checked its structure using one of these JSON to Go online tools and it is like below. Use this structure to see how the file will be to write python script and read the file when I have questions about cloud costs.
Then I pasted the structure:
type AutoGenerated struct {
AiCost struct {
AvailableMonths []string `json:"available_months"`
AwsAi struct {
Accounts struct {
DailyTrend []struct {
AccountID string `json:"account_id"`
AccountName string `json:"account_name"`
AiCost float64 `json:"ai_cost"`
Currency string `json:"currency"`
Date string `json:"date"`
HostingCost float64 `json:"hosting_cost"`
TotalCost float64 `json:"total_cost"`
} `json:"daily_trend"`
// ... rest of structure
} `json:"accounts"`
} `json:"aws_ai"`
GcpAi struct { /* similar structure */ }
Openai struct { /* similar structure */ }
ClaudeMonthlyByWorkspace []struct { /* workspace costs */ }
// ...
} `json:"ai_cost"`
}
You can generate this structure from any JSON using tools like json-to-go. The AI now understands the shape of your data without needing to hold it all in memory.
How It Works
When I ask a question, the AI:
- Looks at the structure to understand what fields exist
- Writes Python code to query the actual JSON file
- Executes the code
- Returns the answer (or generates a chart)
The structure acts as a map. The AI navigates to exactly what it needs.
Asking What’s Possible
This is the part I find most interesting. You can ask:
“What type of questions can you answer based on the data you have?”
The AI examines the structure and tells you:
- Overall spending and trends
- Breakdown by provider (AWS, GCP, OpenAI, Claude)
- Monthly and daily granularity
- Per-account and per-project costs
- Hosting costs vs pure AI costs
- Workspace-level Claude spending
- Month-over-month comparisons
The report describes itself. You don’t need documentation explaining what the dashboard shows, you ask the data directly.
The Flexibility
Here’s what I’ve asked so far:
- Total AI spend → instant answer
- December breakdown by provider → table with numbers
- Main cost driver → analysis with reasoning
- Cost over time → generated chart
- Which Claude workspace costs the most → ranked list
- Compare November vs December → comparison with percent changes
Each question gets answered. If I ask something the data can’t answer, the AI tells me what’s missing.
Tomorrow if I want to know something new, daily AWS trends for a specific account, correlation between workspace activity and cost, whatever, I just ask. The data hasn’t changed, but my access to it is unlimited.
The Drawbacks
This approach has limitations. Traditional BI tools like Looker and Tableau solve problems that this method doesn’t:
No central repository. Looker gives you one place where all reports live. Anyone can browse, discover existing dashboards, see what’s available across the organization. With Claude Projects, each “report” is a separate project. There’s no gallery, no easy discoverability across teams.
Limited governance. Who has access? What’s the source of truth? Version control? Audit trails? Traditional BI tools have this built in. Claude Projects sharing exists, but it’s not as elaborate.
Reproducibility varies. If I ask “December costs by provider” today and next week, I might get slightly different formatting or presentation. Sometimes this flexibility is useful, the analysis adapts to new context and focus. But not everything needs flexibility. Dashboards are deterministic, same view every time, which matters for operational reporting.
Not ideal for operational dashboards. If 50 people need to check the same 5 metrics every morning, a dashboard is better. You don’t want everyone having separate AI conversations for routine lookups.
Data freshness. Dashboards can connect to live data sources. Here, you upload a JSON file, it’s a snapshot. You need to re-upload for fresh data.
The tooling for this approach is early. For it to work at scale, we’ll likely need solutions purpose-built for AI-based reporting, the equivalent of what Looker and Tableau are for traditional dashboards. Central repositories, proper sharing, governance, live data connections.
When to Use What
This isn’t about replacing dashboards. It’s about having both tools available:
Use traditional dashboards for:
- Known, repeated, operational views
- Metrics that many people check routinely
- When you need deterministic, identical output every time
- Organization-wide reporting with governance requirements
Use AI-based analysis for:
- Exploration and ad-hoc questions
- “I wonder if…” investigations
- Follow-up questions that weren’t anticipated
- When you want to ask “what can this data tell me?”
- One-off deep dives that don’t justify building a new dashboard
The dashboard answers “what are the numbers?” The AI conversation answers “what do I want to know about the numbers?”
Claude Projects can be shared with team members. Individual conversations within a project can be shared too. When an analysis produces something useful, it can be documented and referenced. It’s not as robust as Looker’s sharing and organization model, but the capability exists.
The Bigger Picture
This fits into a pattern I’ve been writing about: putting AI in front of every function and action.
In How We Are Building an Agentic Office, I described agents as the interface layer, handling customer support, server management, incident response. The agent sits between people and systems.
This is the same principle applied to data and reporting:
Traditional: Human → Dashboard → Pre-defined views
AI Era: Human → AI → Any question the data can answer
The dashboard isn’t wrong. It’s useful for known, repeated views that many people need. But for exploration, for follow-up questions, for “I wonder if…” — having the AI as the interface layer opens up what’s possible.
We’re not replacing dashboards entirely. We’re adding a layer that handles the questions dashboards weren’t built to answer.
Try It
- Export data you work with (JSON or CSV)
- Generate the structure using json-to-go or similar
- Create a Claude Project
- Upload the file and paste the structure with instructions like I showed above
- Ask: “What type of questions can you answer based on the data you have?”
- Then ask your actual questions
The interesting moment is when you realize you’re having a conversation with your data, not looking at a fixed view of it.
Posted on January 22, 2026