Upbuild
Upbuild / AI Insights
AI marketing insights , one thread
One chat over every ad and analytics account your login can reach. Ask in plain words, point at an account or a dashboard with @, and get the table and the chart. Then save the answer as a data stream, or share the thread.
The @ picker
The composer takes plain questions. Typing @ opens a picker over everything the organisation holds: dashboards, data streams, accounts, platforms and the individual platform tools, filtered as you type by name, by id or by stream key, grouped and chosen with the arrow keys. What you pick travels with the question as a chip, so "the EU account" is never guessed at.
Point at a dashboard and its streams and their latest rows go with the question. Point at an account and its roster line goes. Up to twelve mentions ride with one question.
A question about numbers is read live from the platform, up to four reads per question. A dashboard or a stream is opened for its rows only when you pointed at it, or as a last resort when the live reads were refused.
Ask anything about your data. Type @ to point at a dashboard, stream, account, platform or tool.
Where the number comes from
Upbuild does not keep a shadow copy of your accounts and write invented SQL against it. A question becomes a call to the platform's own typed tool, with the arguments checked against the connector's own tool catalogue before the call is made. What comes back is what the platform would show in its own interface, and the arithmetic on top is done on those rows and no others: a figure the rows do not contain is not estimated, it is said to be missing.
Platforms have limits: TikTok gives daily rows for at most 30 days per ask and hourly rows for one day, Snapchat 32 days and seven. An impossible window is refused before it is sent, and a shape the platform rejects is rewritten and asked again inside the same question.
Every advertising tool is read-only. There is no path from a chat to a budget, a bid or a tag, so the worst a badly aimed question can do is return rows you did not want. Every call is written to the audit log with its tool and its arguments.
Every change your people made, newest first. Every live read a chat makes is a row here, with the tool it called and the arguments it sent.
| Person | Action | What |
|---|---|---|
| amelia@lumen.co09/09/2026 18:21 | insight.ask | job ij-3bkxq2fa |
| amelia@lumen.co09/09/2026 18:21 | insight.tool | meta_insights level campaign |
| amelia@lumen.co09/09/2026 18:22 | insight.tool | gads_campaign_performance |
| amelia@lumen.co09/09/2026 18:24 | insight.save_stream | meta_objective_daily, loose |
| amelia@lumen.co09/09/2026 18:26 | insight.share | with tomas@lumen.co as viewer |
Saving and sharing
Say "save that as a data stream" and the read you just made becomes a named stream: the exact tool and arguments that produced the number, never a shape nobody ran. It lands in the organisation's scratch area, or on a dashboard you can edit, as a draft nobody has published yet. Once it is published there it refreshes with that dashboard, on its cadence, every 30 minutes, hourly or daily.
Later edits may add metrics, dimensions or a longer window. Anything that would hand back less than the stream returns today is refused with the reason, and the new shape is run for real and compared column by column before the save is allowed.
Share a chat with named colleagues as viewer or editor, or open it to everyone in the organisation. A viewer reads the whole thread, questions, tables and charts, with no composer. The organisation is the outer wall.
The table the answer drew, and the stream it became, in a thread anyone you shared it with can open.
| Objective | Spend (USD) | Impressions | CTR (%) |
|---|---|---|---|
| OUTCOME_LEADSact_4418260931 | 36,280 | 4,109,442 | 1.42 |
| OUTCOME_SALESact_4418260931 | 28,915 | 2,884,017 | 1.88 |
| OUTCOME_TRAFFICact_5507213864 | 19,640 | 6,220,905 | 0.94 |
| OUTCOME_AWARENESSact_5507213864 | 10,365 | 9,013,770 | 0.31 |
Saved as the data stream Meta Objective Daily
How it goes
Three steps, in this order. The first one happens once, the second one takes a sentence, and the third is what stops the same question being asked again next month.
Google or Microsoft, per-user OAuth. The chat reaches the accounts your own login reaches and nothing else, and an admin can narrow that further inside a connection, account by account, so one person sees three clients where another sees forty.
Type the question the way you would ask a colleague, and attach up to five files if the answer needs your own numbers beside the platform's. The chat runs on Gemini 3.8 Flash, and the line under the composer counts down the monthly question allowance your own admins set.
Save the answer as a named data stream, land it on a dashboard as a draft, or share the thread with the people who will ask the same thing next week. The answer stops being a message in someone's chat history and becomes something the organisation owns.
The four questions people actually ask
An agency reporting on forty clients and an analytics lead covering eighteen markets ask the same four things. Here they are, answered by how the product is built rather than by how it is described.
It is the platform's own number, fetched through the platform's own typed tool with arguments validated against the connector's catalogue. The chat computes from the rows that came back and is told never to estimate a figure the rows do not contain. "I do not have that data, here is the read that would" is a sanctioned answer, and it is used.
Every identifier in an answer is copied from the context rather than written from memory, which is why account ids, stream keys and tool names render as mono chips you can check. Every live call is audited with its tool and arguments. And every advertising tool is read-only, so a demo can never move a budget.
Yes, in two ways. The thread can be shared with named colleagues or with the whole organisation, and it keeps its tables and charts. The read behind it can be saved as a data stream with a name, put on a dashboard as a draft, and refreshed from then on without anybody asking again.
The account passport, checked before every live read rather than described in a policy. Access starts at what your own OAuth login can see, is narrowed by per-account grants inside a connection, and is enforced again when an answer is saved. A chat never leaves its organisation, and files you attach are treated as data, never as instructions.
Get in touch