Switch

Switch from Zero.

An AI CRM for scaleups with a large contact database. Belda does the work, and you decide with one key. Your Zero export comes across as a file, deduplicated by domain, name and e-mail, the field map approved before anything is written.

What changes on Monday

On Monday the replies are read for you. Every reply is read into one of twelve classes and answered in its own language; a held reply waits with the answer drafted and the rule that held it, and you approve, edit or dismiss it with one key. In Zero, a reply stops the sequence on the response types the seller ticks, and a type can start an agent that receives the contact and the reply, by its docs read whole 2 October. In Zero, an answer can be drafted by an agent from the seller's instruction, and no language rule is named in its docs, read whole 2 October.

The meeting books itself from your calendar. Two genuinely free slots go inside the answer, under your own rules for lead time, hours and meetings a day; the slot ledger refuses a double booking and the invite leaves from you. In Zero, in chat it schedules a calendar event once you approve the card.

The deal moves when the work moves. A meeting the agent booked moves the deal on its own, with the reason on the timeline, and only you move it to Won. In Zero, contacts create themselves from mail, the calendar, LinkedIn and an events tool once the mode is switched on, each mode ships off and was off in the walked workspace, and a new contact brings its company.

What you keep

Your record. Your Zero export comes across as a file, deduplicated by domain, name and e-mail, the field map approved before anything is written. Nobody you imported gets a letter until you release their market.

Your mailbox, your calendar, your notetaker and your team's chat stay yours. Letters leave from your own mailbox through the send gate, the notetaker you already pay for connects in one press, and Slack or Teams carries the card to the channel you name, where you decide it with the same keys.

Your rules, market by market. A market card sets the default language and the rules for each market, with the reason for each, and the person's own language wins over the card.

What Belda does that Zero only recorded: 12 of 19

  • Answer every reply and book the meeting from my calendar. Zero: partly, drafts without slots. Belda: Every reply read, classified and answered in the prospect's own language, with two free slots from your calendar in the reply itself, not a booking link; a yes books the meeting and the invite goes out from your calendar.
  • Guard every send with code, and let me decide with one key. Zero: partly. Belda: The model drafts; the quote guard, the letter guard, the policy check and the slot ledger stand between every draft and the prospect; you decide with one key, and a reply goes out unattended only once that kind of reply has earned it on your own record; step in once, it is pulled back.
  • Run one record from the first letter to the settled invoice, and keep what accumulates mine. Zero: partly. Belda: One record from the first letter to the meeting, the recap, the quote, the signature and the invoice, yours to export with the source and the day behind every fact, with no other CRM under it.
  • Read every reply. Zero: partly, three classes. Belda: Twelve classes. A held reply waits with the answer drafted and the rule that held it. A approves, E edits, D dismisses.
  • Answer in the prospect's language. Zero: partly, by instruction. Belda: The language is read off the reply itself and the answer is written in it within minutes; the market card sets the default and the person's own language wins.
  • Book from my calendar. Zero: partly, event by approval. Belda: Two genuinely free slots from the seller's own calendar inside the reply, under the seller's own rules for lead time, hours and meetings a day; the slot ledger refuses a double booking; the invite goes from the seller with the right colleagues on it.
  • Join my calls. Zero: partly. Belda: The notetaker you already pay for connects in one press; the transcript, the summary and what was promised land on the record, the meeting marks itself held, and the follow-up is drafted from the transcript and held for you.
  • Draft the proposal from the call, held for me. Zero: partly. Belda: After a held meeting the agent proposes the quote from the transcript and your price book, and it waits for your yes.
  • Quote and sign. Zero: partly. Belda: Pick the product and the deal adds up; the quote leaves from the deal with TCV, MRR and ARR; it goes out for signature and the deal comes back Won on its own.
  • Invoice. Zero: partly. Belda: The first invoice is drafted where your money is counted; the invoice that went out sits on the deal with whether it was paid, and the agent never upsells a customer with an unpaid invoice.
  • Tell me what to change. Zero: partly. Belda: Every number carries its denominator and says too early under the floor; the funnel from letters to meetings held, per market and per owner; what changed since last Monday, whether this list worked, and the Monday numbers as one checked sentence, with the download.
  • Reach LinkedIn and WhatsApp from the same desk, at human pace. Zero: partly. Belda: LinkedIn and WhatsApp are paired from the seller's own signed-in tab, run as lanes of the same sequence at human pace within a budget you can read, and a reply in any channel stops the sequence and lands on the desk.

What Zero does well: 7 of 19

  • Find the companies: A lead database its site counts in the hundreds of millions of contacts and tens of millions of companies, searched by a plain-language description turned into filters or by hand across eleven filter groups, industry, location, headcount, funding, technology, lookalikes, hiring signals and the lists already held among them, with a net-new filter that hides what the workspace has.
  • Write the first letter: A send step writes in one of three message modes, one shared template, one per contact, or written by the agent per contact from that contact's data and an instruction, in full or in named parts dropped into a shared template, with variables, spintax, a test send and drafts that wait for approval when the gate is on.
  • Send from my own mailbox: A sender per step, chosen among the members' connected mailboxes, and in the composer a from address chosen among the seller's own connected addresses and their aliases.
  • Record itself: Contacts create themselves from mail, the calendar, LinkedIn and an events tool once the mode is switched on, each mode ships off and was off in the walked workspace, and a new contact brings its company.
  • Live in Slack or Teams: One Slack connection per workspace by an administrator.
  • Bring my old CRM across in one press: Five sources on one page: Pipedrive with one login and one press, starting on its own, HubSpot with one login and an optional two-way sync that asks which values win, Attio and one more CRM through a key with its read scopes, and a file, each running in the background with a list of past imports.
  • The fifty words a CRM claims: Records in three types with one identifier each, a website domain, an e-mail, a pipeline and stage, with twenty-three property kinds and custom properties that behave like the standard ones, listed per company, contact and deal with each marked custom or filled by AI, and an archived property restored before it can be deleted, a record page of overview, activity, notes, tasks and files with a touchpoints strip, archive reversible and delete not, the delete sitting in the record's menu behind a confirmation that says it cannot be undone, as a sequence's, an agent's and a task's do, and a company deleted on 30 September leaving its task standing on the tasks page with no sign that its record was gone.

The matrix below carries every cell with the day it was read.

Others record what you did. Belda does the work, and you decide with one key.

Belda and Zero, nineteen rows

Answer every reply and book the meeting from my calendar

every reply read and classified, answered in the prospect's language, two genuinely free slots offered and booked from my calendar, the invite from me

Guard every send with code, and let me decide with one key

the model writes; the quote guard, the letter guard, the policy check and the slot guard stand between every draft and the prospect; trust is earned per reply class and one human takeover demotes it

Run one record from the first letter to the settled invoice, and keep what accumulates mine

letter, reply, meeting, recap, quote, signature and invoice on one deal that writes itself; the not-now pool, the approved answers and the record of what converted stay mine

Find the companies

the market mapped, a fit reason with evidence, names and addresses

Write the first letter

in my voice, from what we sell, every claim sourced

Send from my own mailbox

Read every reply

what it meant; hold what needs me

Answer in the prospect's language

Book from my calendar

two real slots in the reply, my rules, no double booking

Record itself

the CRM writes itself from the work

Join my calls

recap, next step, follow-up onto the record

Draft the proposal from the call, held for me

Quote and sign

Invoice

Tell me what to change

k of n with denominators; what converted; next week

Live in Slack or Teams

decide the card there

Reach LinkedIn and WhatsApp from the same desk, at human pace

Bring my old CRM across in one press

The fifty words a CRM claims

records, lists, tasks, notes, pipeline, views, export, permissions

does itpartlynonot read

Belda's column reads the product as it launches; every other cell was read from that tool's own pages or our walks, with the day. Where we could not read a fact, the cell says not read, never no.

Who Zero is for.

A scaleup that wants a lead database of hundreds of millions of contacts, a multichannel sequencer, LinkedIn and WhatsApp through an extension, and its notetakers connected. It has no reply desk and does not book from a host calendar.

What Zero does well.

Find the companies

does it

A lead database its site counts in the hundreds of millions of contacts and tens of millions of companies, searched by a plain-language description turned into filters or by hand across eleven filter groups, industry, location, headcount, funding, technology, lookalikes, hiring signals and the lists already held among them, with a net-new filter that hides what the workspace has; a result grid of industry, founding year, headcount with its growth, location, contacts and jobs to reveal, website and last funding, its count capped at two thousand, a saved search re-opened on 30 September reading two thousand and more and opening a row into a drawer over the grid with the company's facts in four cards and a find-people control with role chips; the rows picked import with the contact linked to its company, deduplicated on the domain and the LinkedIn profile, and can go straight into a sequence; its marketing page claims a fit and timing score with the reasons under the number, while the trial's results showed the count and the columns and no fit reason per company, the nearest being a fit property on the record filled by hand or by an AI column that shows its reasoning; a template agent, when a company joins the target list, researches and adds up to three contacts with a one-line reason to reach out on each, added and switched on by a person; a street address, a fit sentence with its source, and the source and freshness of the database were not read, the one independent customer account read, of 10 September 2026, saying the agent proposes that a deal is stale and never that it is strategically wrong and warning of a pipeline the agent calls healthy after the champion has left, and its weekly of 20 September 2026 says a person can be found by name or by connection, which none of its 98 docs pages describes; its own chat, asked on 30 September from the home composer's chip for a list of at most three companies from a saved search the workspace already held, could not read which companies the search holds from the chat without importing or revealing them and stopped rather than guess, so a list built from a sentence needs a spend or a hand; by the workspace's own chat threads of 26 to 28 September re-read on 30 September, the chat reviews a saved search's filters and proposes narrower ones with the trade-off named, changing nothing until told, says the guide names no upstream provider, no refresh interval and no last-checked date per field, does not describe what revealing a contact exposes or saves, and does not define whether an address marked verified was checked at the moment of reveal or read from a stored status; a lookalike search starts from example companies or a company list at one of three closeness settings and rests on company descriptions, guaranteeing no match on size, place, funding or investors; a saved search saves itself, is public or private, and is cloned rather than edited when a colleague wants a separate version, what happens when two people edit one at once documented nowhere; a web-signals feature asked about by name had no guide the assistant could find; and a Get started card walked on 30 September describes the ideal customer and ranks accounts by fit as strong, partial or poor; a web-signals entry sits under its settings in the workspace by its own assistant's navigation read on 30 September, with no docs page behind it, no documented rule for how a visit becomes a company, no place among the touchpoint strip's documented series and no agent trigger for a visit, the one documented visit-to-company mapping in its docs belonging to an analytics integration that matches on the mail domain, while its entry of 20 September 2026 announced website visits as a report type; its Get started page, captured again on 30 September as the lane's round four began, offers a guided setup that promises a target account list ranked by fit in three grades, strong, partial and poor, from a description of the ideal customer typed in, and a customers list that keeps itself current, the page naming no source and no reason behind a grade, so a fit reason with its source stays unread; the lead search's filter popover, opened on 30 September as the lane's round four went on, keeps the structured filters one glyph behind the sentence prompt, in sections for about, location, headcount, industry, funding, acquisitions, tech, lookalikes, the seller's own lists, hiring signals with matching job ads and contacts with matching employees, the page holding its saved searches by name and day beneath the prompt and its suggestion pills for the profile above them, the search control reading disabled with nothing chosen; its own product page's lead-search demo, read in its scripted data on 30 September as the source lane's round four read the site, scores one company through evidence blocks of three parts, a source naming the provider and the channel, a finding stated with its figures and a decision saying which way the score moved, the page's own markup naming the same three, while the four outside sources the demo names, a company-data provider, two funding databases and an applicant-tracking system, appear among none of the integrations its plans page lists and no page states the product's live sources, so the fit reason with its source stays the site's picture and not the walked grid's; its own public intro demo of June, heard and read frame by frame on 30 September as the source lane's round four read the vendor's videos, shows a spoken search for venture-backed companies in one city with a few people in go-to-market roles arriving as a card whose criteria chips spell out the agent's reading of the sentence, the venture-backed phrase turned into a total-funding filter at or above nought, so any funding at all, visible and corrected by nothing, the grid heading its results with a round count of ten thousand that reads as a ceiling, the filter categories as tabs each badged with its count of active filters, and the rows selected into a list or a sequence from one bar, the track saying that contacts the workspace already holds are left out by themselves and every touch is logged; a company search pressed on 30 September as the lane's round five began turned a one-sentence description into two filter chips, a country and a headcount band, each removable, announced two thousand and more results first and listed them in a twelve-column table with a growth arrow and a signed share on the headcount and two per-row controls that name their act, find contacts and find jobs, beside a docked chat that read the results back, said the filters matched the country and the size and not the industry, proposed a keyword filter and changed nothing; the search saved itself under a title the product wrote, public by default, and opened a chat thread of its own before any save was pressed, three objects from one press, two of them named by the product; no price appeared before or after the press and the credit ledger read the same before and after, so a company search costs nothing and a customer learns that from the ledger in the settings and not from the page; a saved companies search opened on 30 September by its own row, as the lane's round five walked it, and switched by the tab strip over the grid to its contacts tab listed people under the criteria as chips, one removable countries chip and an add-filter control, with a capped count line of two thousand and more first and seven columns, the contact, the company, the title, the location, the education and the dates since the current position and the current company began, and a public switch beside the save on the header, neither pressed, while the companies results of the same search carried no count line in that read and a similarity tab the lane's read of the vendor's own demos had named appeared on none of three states, the index at rest, the companies results and the contacts results, so it sits behind a typed search or has left this build; and the settings' integrations pane that day ended on an event card and an analytics card, each a purpose sentence and one connect control, with none of the settings rail's anchors naming signals, the web, visits, tracking, analytics or events at a laptop width, so the web-signals entry its assistant had placed under settings has no open door in the walked trial workspace, a plan or flag gate or a surface behind a connected analytics tool; the vendor's own marketing pages, read anonymously in their scripts by the source lane on 30 September as its round five ran, load the vendor's own visitor snippet on every page, a script from a host of the vendor's own, two cookies, a consent mode granted or denied read from the banner and sent again on change, an identify call keyed by any e-mail the site is given and a track call per event, every analytics event the site captures mirrored into its website-visitor signal and the two cookies forgotten on denial, so the client half of the signal whose workspace half the walked integrations pane ended on as an analytics card and an event card runs live on the vendor's own front door, the vendor its own customer there, while the walked trial workspace showed no open door for it; and the lead-search mock's markup names each row's selection state, seven of seventeen rows carrying a deselect label as the rows pre-selected at a score of eighty-two and up and ten of seventeen a select label, so the pre-selection the earlier rounds read off the picture is confirmed in the markup on two pages, a marketing mock still and not the walked grid; the vendor's own intro demo of June, every second of both demos classed by surface by the source lane on 30 September as its round five ran and the seven surfaces its round four had not described read at full size, shows the lead search's contact preview as a panel over the results, the person's picture, name and one line of title at company, then the positions as a timeline grouped by employer, a mark and the employer's name over one row per role with its date range and its duration and a hairline joining the roles, the education in the same shape, a column of social profiles and a footer offering to add the person to a list or a sequence, a career read as a timeline being what a seller reads to judge seniority, the walked saved search carrying the education column so the preview stands in the build too, the panel itself a picture of the demo and not of the walked grid; the vendor's own visitor script, the client of that website-visitor signal, read in full by the source lane on 30 September as its round five read the careers pages beneath their markup, keeps a visitor cookie for a year and a session cookie that rolls every half hour, holds every event while the consent answer is unknown, releases them with their identifiers on a yes and empties them and deletes both cookies on a no, sends the events in small batches to a host of the vendor's own with the page's address stripped of nineteen kinds of sensitive query key before it leaves the browser, passwords, tokens, codes, sessions and e-mail among them, together with the host, the path, the browser's language, the referrer and its domain and the session and device identifiers, adds to a page view the title and any of twenty-two campaign or click identifiers found in the address and to a page leave the previous path, the seconds spent and the deepest scroll, and captures no click, no form, no fingerprint beyond a browser-automation flag and nothing that names the visitor's company, so whatever turns a visit into a company happens on the vendor's server from the request, or from an e-mail a form hands to identify, by a rule its pages do not publish, the script being the vendor's own install on its own pages while the customer's install stays unread; the hiring-signals filter of its lead search takes five documented inputs, open roles, departments, locations, when the postings went up and how many, with no title, seniority or posted-within field named as such, and its agent door's company search tool takes three filter arguments, for the company, the contacts and the jobs, whose job-filter fields are printed on no page, the schema reaching a connected client alone, so the mapping from the filter's words to the tool's fields is a live read and not a documented one, by its docs' answer endpoint asked anonymously on 30 September as the lane's round five ran, with the raw pages beneath it; its website-visitor signal is named on none of the ninety-eight docs pages, the changelog entry of mid-September naming it twice, a website-visits report for workspaces with it switched on and its traffic-source classes, with no install, no tag, no site key and no consent mode documented anywhere public for a customer's site, a shipped and report-gating feature with no docs page ten days after its only changelog mention, the endpoint answering off target about an analytics integration and flagging the substitution itself, by its docs' answer endpoint asked anonymously on 30 September and 1 October as the lane's round six ran, nineteen of nineteen answered at the first attempt and every absence checked by grep on the raw pages beneath it; the lead table of the vendor's own marketing pages, counted figure by figure by the source lane on the night of 30 September as its round six ran over files the weekly refresh found unmoved, carries eighteen bare figures of forty-one, the seventeen headcounts and the filter count with no window, source or prior on the cell, and its one transformation on the cell is the score that ranks every row of seventeen, a why-now line standing on seventeen of seventeen rows with a time word and on six of them with a figure, an age or a window, while the frame carries what the cells lack, one line above the table with a count against its total at a threshold and the next verb and the score's sort beside it, so the table is never unframed, each result row naming in its own parts a company cell, a score cell, a why cell, a check cell and a row of source chips beside stacked marks and a logo; a marketing mock still and not the walked grid, which showed the count and the columns and no fit reason per company, and whether the app's own result table carries the why-now line and the per-row sources by default or as columns a customer adds the lane leaves as an open question for its walk; the lead-search page of the vendor's help centre, checked by the source lane on the night of 30 September as its round six ran over pages the sixth refresh found unmoved, states the result grid's default columns in one sentence, nine for companies and seven for contacts, and the vendor's own screenshot on that page keeps them in the documented order, the first six of the nine in the first fold, the company with its logo, the industry, the founding year, the headcount with a growth arrow, the location and the two action columns that find the contacts and the jobs, with the website, the description and the funding off-screen to the right, six of six visible columns in place, two filter pills reading as a sentence above the grid, a selection toolbar with its count and one action and twenty rows in the fold, so the help and the picture agree column for column, the identity first, then the three numbers a seller ranks by, age, size and growth, then the place, then the two things the row can still fetch; the results have no object in the interface file, three tools of the agent door reaching them; and the number the grid ranks companies by, the headcount with its growth arrow, is no field of the company object in the file, whose eighteen fields hold no headcount, and no built-in column on the columns page, but an enrichment property under a custom column's identifier reached by listing the columns first, so what the screen shows first the interface names last.

read on 26 September 2026; the database count from its site read 27 September 2026 and re-read by the market ear on 30 September 2026; the filter groups, the grid, the cap and the fit property from the walks of 27 and 28 September 2026; the template gallery from the teardown's walk round 2 of 30 September 2026; the saved search's drawer from the teardown's design round 2 and the weekly's search by name or connection from its changelog read, both of 30 September 2026; the customer account's line on fit from the teardown's source round 2 of 30 September 2026; the chip's stop from the teardown's play round 3 and the Get started card from its walk round 3, both of 30 September 2026; the search review, the provider and freshness answers, the lookalike and saved-search rules and the web-signals gap from the workspace's own chat threads of 26 to 28 September 2026, re-read in that walk round; the web-signals entry and its gaps from the teardown's ask round 3 of 30 September 2026; the guided setup's fit grades from the Get started page re-captured by the teardown's walk round 4 of 30 September 2026, read by the Zero lane's round 4 of 30 September 2026, as it began; the lead search's filter popover from the teardown's walk round 4 of 30 September 2026, read by the Zero lane's round 4 of 30 September 2026, in its later landings; the lead-search demo's evidence blocks and the unlisted sources from the teardown's site read round 4 of 30 September 2026, read by the Zero lane's round 4 of 30 September 2026, in its site landings; the sentence's reading shown as criteria chips, one loose filter, the round ceiling and the one bar to a list or a sequence from the teardown's video demos round 4 of 30 September 2026, read by the Zero lane's round 4 of 30 September 2026, in its video landings; the search press's chips, its self-saved public search and thread, its per-row controls and its cost of nothing from the teardown's play round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, as it began; the saved search's contacts tab, the missing similarity tab and the integrations pane's closed door to web signals from the teardown's walk round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, in its walk landings; the vendor's own visitor snippet on its front door and the lead-search mock's selection labels from the teardown's site read round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, in its site landings; the contact preview as a career timeline grouped by employer from the teardown's video demos round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, in its video landings; the visitor script's whole client, what it sends, holds, forgets and never captures, and the company match left to the server from the teardown's careers read round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, in its careers landings; the hiring-signals filter's five inputs and the search tool's unprinted job-filter fields from the teardown's ask round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, in its ask landings; the website-visitor signal on none of the ninety-eight docs pages, its report switched on by it and no install or consent mode documented from the Zero lane's ask round 6 of 30 September and 1 October 2026, in its ask landings; the lead table mock's eighteen bare figures of forty-one and the frame that carries what the cells lack from the teardown's site read round 6 of 30 September 2026, read by the Zero lane's round 6 of 30 September 2026, in its site landings; the grid's documented defaults kept in order on the vendor's own screenshot, the results with no object in the interface file and the headcount as an enrichment column and no field from the teardown's docs read round 6 of 30 September 2026, read by the Zero lane's round 6 of 30 September 2026, in its docs landings

Write the first letter

does it

A send step writes in one of three message modes, one shared template, one per contact, or written by the agent per contact from that contact's data and an instruction, in full or in named parts dropped into a shared template, with variables, spintax, a test send and drafts that wait for approval when the gate is on; a per-contact edit survives a later template change and the seller keeps or takes the new; the voice comes from context documents, the company, the sales motion, the customer profile and the mail tone, that sign-up drafts from the website and a person edits, attached where a prompt is written and applied automatically nowhere, a personal and a workspace context with no documented precedence when they disagree and no history to restore; sequences are built by chat and wait for a save, with no documented list of which settings the chat set and which it left for the seller, and a message prepared before the contact's record changes goes out as prepared unless a person refreshes it; a claim with its source was not read, and its own assistant said Zero does not know the company's positioning or tone by itself; a sequence asked for in one sentence from the home composer's chip on 30 September was created as an inactive draft with no steps at all, opened in its builder with the chat docked beside it and its start step already carrying ten starts a day, weekdays nine to five and ten minutes between sends, so the sentence made the shell and not one letter; a template edited after a step has gone out changes nothing already sent and reaches the messages not yet prepared, a per-contact edit flagged when the source changes, by the workspace's own chat threads of 26 to 28 September re-read on 30 September; a settings pane walked on 30 September lists each context document under a workspace tab and a personal tab with a mark naming the agents that read it and the day it was last edited, so which agent's writing a document shapes and how old it is are read off the row; one of those rows opened on 30 September, as the lane's round four went on, is an editor titled edit context that holds several topics as tabs, company information, investor information and the mail writing style, with a prose field beneath, a switch that keeps a context personal where everyone in the workspace can use it by default, an invitation to ready-made prompts and cancel and save controls, the pane's four rows, a company overview, the sales motion, the ideal customer profile and example customers, each reading no agents and an edited-ago age under a workspace tab and a personal tab with an add-context control.

read on 26 September 2026; the message modes, the drafts and the context documents from the docs read 27 September 2026 and the step inspector walked 28 September 2026; the assistant's words from the walk of 27 September 2026; the contexts' rules and the chat-built sequence's settings from the teardown's ask round 2 of 30 September 2026; the sequence made from the chip from the teardown's play round 3 of 30 September 2026; the template-edit rule from the workspace's own chat threads of 26 to 28 September 2026, re-read in its walk round 3 of 30 September 2026; the context pane from the teardown's design round 3 of 30 September 2026; the context editor from the teardown's walk round 4 of 30 September 2026, read by the Zero lane's round 4 of 30 September 2026, in its later landings

Send from my own mailbox

does it

A sender per step, chosen among the members' connected mailboxes, and in the composer a from address chosen among the seller's own connected addresses and their aliases; one connection covers mail and calendar for either of the two big providers; an agent sends only from its owner's mailbox; a later mail step can go as a reply inside an earlier step's thread by a setting its API names, and the sender's own mailbox name is a variable a step's text can carry, a mail step carrying no copy, blind copy or attachment field and addressing the contact's own address alone, and the sequence's sending window falling back to universal time when its zone is set wrongly, the mail never going out in the recipient's zone, by its docs' answer endpoint on 30 September, which found no unsubscribe header or link by default and no custom tracking domain either; a daily starts cap on new enrolments, its hint naming fifty a day, a minimum number of minutes between automated sends, active days and a sending window in one time zone per sequence, the walked sequence set to ten a day, weekdays nine to five and ten minutes apart; no daily limit per mailbox across sequences and hand-written mail, no warm-up and no rule for a provider's rejection were read, and its own chat said on 30 September that no bounce or complaint threshold pauses a sequence, and that what puts a contact into a sequence's needs-attention group, what clears it and whether it pauses the contact are documented nowhere, the group naming no reason; a sequence made from the chat on 30 September opened with ten starts a day, weekdays nine to five and ten minutes between sends already set before anyone touched it, and a mail step with no mailbox chosen is flagged in the builder as missing its configuration; by the workspace's own chat threads of 26 to 28 September re-read on 30 September, a step without its own mailbox falls back to the sequence's, activation refusing until every mail step resolves to one, and whether a teammate's connected mailbox can be chosen is documented nowhere; the daily cap counts new starts alone and not the follow-ups due from earlier cohorts, so a day's sends can pass it; a later mail threads onto an earlier step's only when that setting is on and the earlier mail actually went; a bounce is an enrolment outcome of its own counted among the errors, with no documented retry or suppression after it; and a sequence enrols contacts alone, never a company, copying them into a source list of its own, so a change to the list they came from moves nobody; a selection in any list goes into a sequence from the list's floating toolbar in one press, an enrolment and not a start, since the sequence's start settings may hold the contacts for approval and an active run takes its new enrolments in the background under its caps, no single press from a selection to a send being promised, and a dynamic list's later matches enrol themselves nowhere, the action reaching the contacts selected at the press alone, by its own assistant and its docs' answer endpoint on 30 September; its docs describe one panel on a sequence's start step for the rules that live above the steps, who enters, automatically as the daily capacity allows or held for approval, with the daily starts, the approval of model-written messages when a step drafts, the shared schedule of active days, a sending window and its zone, the minimum minutes between automated mails, the step defaults a new message step inherits, ending the run for a contact who responds with the response types chosen there and skipping a step for a contact with no reachable address on that channel, and the thread's visibility to the whole workspace or to the sender alone, a block the endpoint itself left out, and of the stop rules a reply is the one default, a booked meeting is no documented stop rule and an unsubscribe is a per-contact act across all messaging and not a rule of the sequence, the label the live editor shows on that panel matching no word on the page so that the two are one panel by inference alone, by its docs' answer endpoint asked anonymously on 30 September as the lane's round five ran, with the raw pages beneath it; one enrolment lifecycle carries three vocabularies joined by no page, the drafts tab's seven groups that say what the contacts need from the seller, waiting for setup, needing attention, preparing, needing review, approved, active and completed, with a retry of the preparation on the attention group, the interface's seven statuses with one-line meanings, awaiting the worker, awaiting approval, active, paused, completed, stopped early such as by a reply, and cancelled when removed from the source list, and the door's own verbs on the audience and on a message revision, none of the seven group names appearing on the interface page and none of the seven status ids on the sequences page, so which status holds a contact while its messages are prepared is a reading and not a documented fact, the endpoint's own being the one awaiting the worker, by its docs' answer endpoint asked anonymously on 30 September and 1 October as the lane's round six ran, nineteen of nineteen answered at the first attempt and every absence checked by grep on the raw pages beneath it.

read on 26 September 2026; the sender, the cap and the window from the step inspectors walked 28 September 2026; the aliases and the agent's mailbox from the docs read 27 September 2026; the defaults and the chat's answer from the teardown of 30 September 2026; the reply-in-thread setting and the mailbox variable from the teardown's API read and the needs-attention answers from its ask round 2, both of 30 September 2026; the mail step's fields, the zone rule and the unsubscribe and tracking gaps from the teardown's ask round 2 of 30 September 2026; the shipped defaults from the teardown's play round 3 and the missing-configuration flag from its walk round 3, both of 30 September 2026; the mailbox fallback, the cap's scope, the threading, the bounce and the enrolment rules from the workspace's own chat threads of 26 to 28 September 2026, re-read in that walk round; the one-press enrolment that is no start and the dynamic list that feeds no sequence from the teardown's ask round 4 of 30 September 2026, read by the Zero lane's round 4 of 30 September 2026, in its ask landings; the start step's one panel of six rule blocks, the booked meeting that is no stop rule and the unsubscribe that is a contact's act and not the sequence's from the teardown's ask round 5 of 30 September 2026, read by the Zero lane's round 5 of 30 September 2026, in its ask landings; the enrolment lifecycle's three vocabularies, seven groups, seven statuses and the door's verbs, joined by no page from the Zero lane's ask round 6 of 30 September and 1 October 2026, in its ask landings

What Belda does that Zero records.

Three of the nineteen.

Answer every reply and book the meeting from my calendar

Zero: partly

Every reply read, classified and answered in the prospect's own language, with two free slots from your calendar in the reply itself, not a booking link; a yes books the meeting and the invite goes out from your calendar.

read on the product as it launches, re-read 1 October 2026

Guard every send with code, and let me decide with one key

Zero: partly

The model drafts; the quote guard, the letter guard, the policy check and the slot ledger stand between every draft and the prospect; you decide with one key, and a reply goes out unattended only once that kind of reply has earned it on your own record; step in once, it is pulled back.

read on the product as it launches, re-read 1 October 2026

Run one record from the first letter to the settled invoice, and keep what accumulates mine

Zero: partly

One record from the first letter to the meeting, the recap, the quote, the signature and the invoice, yours to export with the source and the day behind every fact, with no other CRM under it.

read on the product as it launches, re-read 1 October 2026

What accumulates from week one

The record writes itself from every reply and every call: the class each reply fell in, the slots offered, the meeting held, the recap, with the source and the day behind every fact. Every number carries its denominator and says too early under the floor. Every market rule is logged with its reason, and the record of what converted builds list by list.

What it costs to stay

Staying costs the hours. Somebody reads every reply, picks every slot and writes every answer. Belda's price is on its own page.

See the difference on your own list.

Bring your list. Watch the agent read it, write to it, and hold what needs you.

Sign up

Belda is the CRM that runs the whole revenue motion, built for agents from day one. The agent executes. The seller approves.