Meni: from asking questions to automating the work
TLDR: Managers wanted oversight into how things were performing, and enterprise customers wanted the same visibility into every branch. I designed Meni, an AI assistant that turns Menaia’s existing data into instant, grounded answers and pinnable dashboards, without anyone building a single report.
Managers wanted oversight into how things were performing, and our enterprise customers wanted that same visibility into how each of their branches stacked up.
The typical answer was to leave Menaia entirely: export the data, stand up a reporting tool, build your own boards from scratch. But Menaia already had all of that data. It just wasn’t available to the user in an intuitive, friendly way.
Skip straight to the shipped design↓A chat box on its own would mean asking the same question every Monday morning. The goal was never a chatbot. It was giving the answer somewhere to live.
Chat is how you ask the first time. Dashboards and playbooks are how you stop asking. The recurring questions become a view you return to, and the recurring work becomes a queue waiting on your approval.
Once we knew what to build, the harder questions were about trust: who Meni was really for, and what it should never be allowed to do.
Permissions first
Every read and write had to respect the same role- and branch-level permissions as the rest of Menaia, so cross-org and cross-branch data could never surface in an answer.
Built for non-technical users
The people asking Meni questions aren’t writing SQL or building custom reports. The interface had to feel like asking a coworker, not operating a BI tool.
Designing around model limits
Different LLMs vary in tool-calling reliability, reasoning depth, and context window. The experience had to hold up across that range, not assume the ceiling of one model.
Privacy & security by default
Every number Meni states has to trace back to a real record, and anything it reads is treated as information, never as an instruction to act on.
With those guardrails set, the assistant needed enough depth to actually be useful, not just a friendlier search bar.
- Grounded, never guesses. It has to fetch real records with its read tools before answering or proposing anything, so every number and name comes from your actual data, not the model’s imagination.
- Page-context aware. It knows which page or record you’re on. Say “summarize this lead” or “draft a follow-up about this estimate” and it uses what you’re viewing, no ID required, across leads, estimates, clients, projects, jobs, invoices, ride-alongs, and the schedule.
- Agentic. In a single turn it can chain multiple tool calls, up to about 12 steps: find a lead, read its activities, read prior communications, then draft a follow-up.
- Every write is approval-gated. Meni never changes anything silently. A write request renders an editable approval card, and the change only happens once you approve it, edit a field, deselect an item, or decline.
Interaction & Experience
None of this works unless it’s trustworthy by default, not by exception.
- No silent writes Every create, change, or send is human-approved through a card before anything happens.
- Permission-scoped Reads and writes enforce your role and branch server-side, so cross-org and cross-branch data never leaks.
- Grounded It never states a number, name, date, or price it didn’t pull from a real record.
- Injection-resistant Data it reads, like lead notes or memos, is treated as information, never as instructions to act on.
- Observable Every run is traced (Langfuse) and continuously evaluated against a test suite covering grounding, action discipline, and prompt injection.
Meni shipped as three connected surfaces, one for each half of the loop above: the chat you ask in, the queue where drafted work waits on your approval, and the dashboard you build by pinning what you ask about.
Everett is your strongest branch this month, closing 64.4% of estimates, 22 points above the group average.
Starter prompts adapt to the page you’re on, and every answer is grounded in the record in front of you, cited back to its source.
Everything Meni drafts lands in one queue as an editable card, whether you asked for it directly or a playbook produced it overnight. Adjust a field, dismiss an item, or decline; nothing sends until you approve it.
Ask a question, turn the answer into a chart, and pin it. Dashboards build themselves out of the questions you actually ask.
Building business intelligence inside Menaia let us retire the third-party BI embed our customers had been paying for on the side.
The cost saving was the easy win. The more interesting result was behavioural. Traffic to Meni jumped 70% only once it stopped being a chatbot, once the same assistant also gave people a dashboard they could shape themselves and automations they could set for their own work.
It’s clear that users weren’t looking for another chat tool. They were looking for somewhere to automate their work and connect the system they were already in.