Pyramid AI

Most "AI for planning" today is a chat wrapper bolted onto a spreadsheet. It hallucinates numbers, makes confident claims it cannot verify, and either cannot take actions on the plan at all, or takes them silently with no way back.

Pyramid AI is the opposite of that. It is an autonomous planning agent that reads the actual plan, runs computations against it, and proposes specific, applicable actions. When it says "Wk 14 is overbought versus forecast", it has already read the columns and done the math. When it suggests an edit, it has already simulated the impact, and the planner can apply it, dismiss it, or revert it after the fact.

We are building it because retail merchandising is one of the highest-leverage workflows in the business, and the existing tools force a choice between speed and rigour. The agent gives planners both.

The copilot streams over Server-Sent Events from a Cloudflare Worker, which iterates a tool-calling loop over OpenAI Chat Completions for up to eight turns per user message. Tool calls emit structured events the client renders as tool cards in the chat, each with its own affordances: apply, dismiss, revert, open.

Four tools pause the stream instead of chaining into more text. propose_edits and commercial_action wait for an Apply / Dismiss decision. propose_plan waits for Approve / Deny. run_code pauses so the browser can actually execute the snippet in a sandboxed Web Worker, then re-opens the stream with the real stdout as a structured tool result. The model sees its own previous tool_call plus the actual output, and continues reasoning in the same response.

This is what makes the agent feel autonomous without being reckless. It can take real actions, but mutating actions always end the turn, and the human is always the one who commits.

The run_code tool is the heart of why Pyramid AI gets numbers right. The agent writes JavaScript, the browser runs it in a hardened Web Worker (3-second hard timeout, no fetch, no DOM, no cross-origin access) with the plan exposed as a frozen object, and the captured stdout is fed back into the model as a tool result.

A question like "which stores are overbought on linen tops in Q3" stops being a confident-sounding paragraph and becomes: the agent writes a five-line filter and aggregation, runs it against the live plan, and answers from the actual numbers. If the planner does not believe the answer, the code is sitting right there in the chat.

Grid is the working plan: a virtualised spreadsheet with inline editing, drill by time / product / location, multi-column pinning, undo/redo, filters and CSV/Excel import-export. Direct Import lets a planner link a local folder once via the File System Access API, and reload a single CSV in a click whenever a downstream ETL overwrites it.

Assessment is a structured plan health check. The agent grades the plan 0 to 100 across five dimensions, surfaces prioritised risks, and prepares applicable edits the planner can drop straight into the grid.

Analytics is a charts dashboard the agent can deep-link into. Scenarios is a Kanban-style review board for what-if drafts, with revenue, margin and risk chips on each card. Collaborate is the approval queue for pending edits across teammates, with per-persona avatars.

A read-only share view renders an ROI-led snapshot of the plan and a chosen scenario for stakeholders who do not need the whole workspace.

Tool cards collapse to one-line chips once applied or dismissed. Applied edits store the previous values and offer a one-click revert. Heavy detail (rationale, impact preview, comparison tables) is hidden behind a "Show detail" chevron by default. The agent narrates one short sentence in the user's language before each tool call ("Looking at the Wk 14 demand spike…"), never the tool name.

No auto-navigation: assessments, filter changes, charts and selections render inline with a "View in X" link if the user wants the full view. The agent never pulls the user between tabs without permission.

Pyramid AI is in internal alpha. We are working with a small group of merchandising teams to harden the tool set, the prompt and the import pipeline. If your team plans in spreadsheets today and you want a closer look, get in touch.