A round of fixes to decision-table Agent Blocks, spreadsheet imports, and how WhatsApp billing explains itself when purchased credits run out.
A cluster of fixes across three areas: Agent Blocks backed by decision tables, spreadsheet imports into Records, and WhatsApp billing clarity.
Agent Block, decision-table accuracy. A column set to fill silently from the match now writes correctly even when the table pins it to a single value before the respondent has answered anything; previously it was skipped unless an already-asked column held a value. A silently-filled column is now only written when the questions already answered determine a single row, so every respondent with the same answers gets the same value instead of an arbitrary first match. Table-backed questions now validate the respondent’s answer against the actual valid choices, on web and WhatsApp alike, instead of silently accepting anything or skipping ahead. When an agent block captures several fields in one reply, all of them now land in a single submission instead of splitting across several.
Agent Block, exports and the submissions table. Agent block answers used to sit at the very bottom of the export column picker, under the block’s full instruction text, easy to miss in a long form. Each answer now appears where its block sits in the flow, with its own icon and a short spreadsheet header. Large queued exports without a manual column selection now include agent answers too, and the submissions table preview shows what the respondent actually answered, agent-collected turns included.
Cleaner Records imports. A spreadsheet with two columns sharing the same name used to crash the import preview; repeated headers now get renamed automatically (V2, V2 (2)) and import as separate fields. A column containing only 0 and 1 used to import as a checkbox with no way to fix the guess; those columns now import as numbers unless the values are spelled out (true, false, sí, no), and every mapped field now has a type dropdown so you can correct it either way.
Clearer WhatsApp billing. A workspace with plan credits left but no purchased credits is now told, plainly, why WhatsApp won’t send: in the inbox composer, the usage page, and a low-balance email, in English and Spanish. Plan credits cover AI and integration tools; WhatsApp needs purchased credits specifically.
Agent Block fixes on Pro, Business, and Enterprise. Import fixes on all plans. WhatsApp billing clarity on Business and Enterprise.
A Records-table column set to fill silently from the match is now written whenever the table pins it to a single value, even before the respondent answers anything. Previously it was skipped unless an asked column already held a value, so a block whose only mapped column is auto-filled ended on its first turn with the value missing. The editor warning for these columns is now based on the actual rows, so a column that holds one value everywhere is no longer flagged.
A Records-table column set to fill silently from the match is now only written when the questions the respondent answered pin it to a single value. Previously an ambiguous match wrote whichever row came first, so every respondent with the same answers got the same arbitrary value. The block editor now warns when a silently-filled column is not determined by the questions asked, so the author can ask for it instead.
Agent block answers were listed at the very bottom of the export column picker, under the block's full instruction text, so in a long form they looked missing. Each block now appears where it sits in the flow, named by the block, with its own icon in the picker and short spreadsheet headers like "agent_matrix > Sede". Large queued exports that run without a column selection include agent answers too, and the preview column in the submissions table now shows what the respondent actually answered, including answers collected by an agent block.
When an agent block captured several fields in one reply (the answer plus any auto-filled or inferred fields), each field was saved at the same moment and each one created its own submission. A single respondent could end up as four submissions holding one field each. The fields are now saved one after another, so they all land in the same submission.
Agent Block questions backed by a Records table now check the respondent's answer against the actual valid choices before accepting it, on web and WhatsApp alike. A typed answer outside the table, or a question asked mid-flow, no longer gets silently captured or skips ahead. Rejected answer combinations now clear and re-ask the conflicting field instead of resubmitting the same invalid combination.
An Excel or CSV file with two columns of the same name (for example two columns called V2) crashed the import preview. Repeated headers are now renamed (V2, V2 (2)) and imported as separate fields.
A spreadsheet column containing only 0 and 1 was imported as a checkbox field, which rewrote the values as true and false, and there was no way to correct the guess. Those columns now import as numbers unless the column also spells the values out (true, false, sí, no), and every column mapped to a new field has a type dropdown in the mapping step. Checkbox fields also accept the ways a spreadsheet writes them: 1, 0, yes, no, sí.
A workspace that still has plan credits but no purchased credits is now told why WhatsApp will not send, in the inbox composer, the usage page, and the low-balance email, in English and Spanish. Plan credits cover AI and tools; WhatsApp needs purchased credits. The email also warns as the purchased bucket runs low, and the bot still sends exactly one closing notice to the contact then stays quiet while inbound keeps arriving.
