Specialist recruitment, senior investment-professional placement
AI Inbox with Human Approval: Email, WhatsApp & Follow-Up Drafts
The four workflows run live, reading, filtering, matching and drafting every reply. No time, cost or revenue saving was measured, 'hours every day' is the documented before-state, not a measured result.
- Our role
- AI automation engineer, architecture, build and safety design
- Timeline
- Build and handover; running in production
- Sector
- Specialist recruitment, senior investment-professional placement
At a glance
- Connected workflows
- 4
- Channels unified
- 3: email, WhatsApp, follow-ups
- Approval points per draft
- 1
- Messages sent without a human click
- 0
- Follow-up cap
- 3 touches, 3-day cooldown between sends
- Engineering hazards fixed by name
- 10
- Stack
- n8n (self-hosted), Claude, Gmail, WhatsApp Business, Slack
What was breaking
The volume was not the problem. The ratio was. Inbox triage consumed hours a day, and almost none of that time was spent on messages that needed a human. Newsletters, automated booking confirmations, no-reply notifications and auto-responses all arrived looking exactly like a client enquiry until someone opened them and judged. The cost was not reading the real messages; it was reading everything else first.
WhatsApp made it worse by being easier. Candidates and clients increasingly messaged there, where nothing was logged against the CRM, nothing created a follow-up task, and a message read on a phone between meetings was a message that might never be actioned. The channel with the fastest expectations had the weakest record-keeping.
Two specific blind spots cost real money. New-business enquiries from unknown senders were indistinguishable from any other unknown email, so the highest-value messages in the inbox were the ones most likely to be skimmed past. And candidate follow-ups ran on memory, which in practice means the candidates you spoke to recently get chased and the ones who went quiet three weeks ago do not.
The constraint that shaped everything: full automation was off the table. This is a business where relationships are the product. An AI that confidently sends a slightly wrong message to a senior client does not save time, it destroys trust, and you cannot un-send it. Any solution had to make the AI useful without ever letting it speak for the business.
Hours a day spent judging noise
Every message had to be opened to find out whether it mattered. The work was triage, not recruitment.
WhatsApp was unlogged and easy to miss
The fastest-expectation channel had no CRM record and generated no follow-up task, so a message seen between meetings could simply evaporate.
New-business enquiries hid among unknown senders
The most valuable messages arrived from addresses the CRM did not recognise, exactly the messages most likely to be skimmed past.
Follow-ups ran on memory
Recent conversations got chased; candidates who went quiet weeks ago did not. Memory always prioritises the recent over the important.
Incomplete CRM records
Because logging was manual, the CRM reflected what someone had time to write down rather than what actually happened.
The obvious automation was unacceptable
Letting AI reply unsupervised risked a confidently wrong message to a senior client, an irreversible cost in a relationship business.
What we built
The architecture follows from the central constraint. Three workflows handle channels (email, WhatsApp, and the daily follow-up digest), and each of them can do everything except send. A fourth workflow, the approval handler, is the only component in the system with send permission. That means there is exactly one code path capable of putting a message in front of a client, and it only ever executes in response to a human clicking a button. The safety property is structural rather than a policy someone has to respect.
The email workflow earns its value before the AI is involved at all. A deterministic pre-filter strips quoted history, reads calendar invites, and silently closes mail from the agency's own domain, auto-replies, no-reply addresses, bulk mail, newsletters and notifications. Only what survives costs a model call. It also recovers the real person from the reply-to header when a booking tool masks them, otherwise the most human message in the inbox looks like it came from a robot.
Classification is confidence-scored, and the threshold is load-bearing. Known senders are classified as client, candidate, lead or noise, with a confidence score and a flag for whether a meeting was requested. Anything below the threshold, or anything unreadable, goes to a triage channel with no draft attached. This matters more than it sounds: a low-confidence guess does not produce a cautious draft, it produces a confidently wrong one, and a wrong draft is worse than no draft because it invites an approving click.
Unknown senders get a second, deliberately generous AI gatekeeper, because the cost of the two errors is asymmetric. Discarding a genuine new-business enquiry is expensive; surfacing one extra ambiguous message into a new-lead channel costs a few seconds. Alongside it the system shows possible CRM matches by name and company for human verification, and never auto-links them, because quietly attaching a conversation to the wrong record corrupts the CRM in a way nobody later detects.
A detail I would repeat on every build like this: the signature and booking link are appended by code, not by the model. Ask an AI to include a signature and it will eventually invent one, or omit the booking link in precisely the message where a meeting was requested. Deterministic things belong in deterministic code; the model's job is the prose.
The follow-up digest replaces memory with pacing rules. It finds candidates who have gone quiet, drafts a personalised follow-up referencing the most recent real topic from their CRM history, and respects explicit caps: at least three quiet days before chasing, no more than fifteen days cold if never chased, a maximum of three follow-ups until the contact re-engages, and a three-day cooldown after each send. The digest lists up to ten pending drafts, which keeps a daily review a two-minute job rather than an aspiration.
Finally, the Slack approval surface is one card per message offering approve, edit or discard. Approving an email sends the live Gmail draft, including any edits made in Gmail itself, so the recruiter can refine in the tool they already use. Every card is replaced with the outcome and the name of the person who acted, so the channel becomes an audit trail rather than a pile of stale buttons.
One component may send, and only on a human click
Three workflows draft; a fourth sends. Concentrating send permission in a single path makes 'the AI cannot message a client' a structural fact rather than a rule to trust.
Filter deterministically before spending a model call
Newsletters, auto-replies and own-domain mail are closed by code. The AI is reserved for genuinely ambiguous messages, which is both cheaper and more accurate than asking a model to recognise noise.
Below the confidence threshold, draft nothing
A low-confidence classification produces a confidently wrong draft, and a wrong draft invites an approving click. Triage with no draft is safer than a plausible guess.
Make the gatekeeper generous where errors are asymmetric
Missing a real new-business enquiry is expensive; one extra ambiguous message in a review channel costs seconds. The filter's bias should follow the cost of being wrong.
Suggest CRM matches, never auto-link them
Silently attaching a conversation to the wrong contact corrupts the record in a way nobody detects later. Possible matches are shown by name and company for a human to confirm.
Deterministic content belongs in code
Signatures and booking links are appended programmatically. A model asked to include them will eventually invent a signature or omit the link in the one message that needed it.
Non-negotiables
- The AI never sends. It may read, classify and draft, but a human clicks before anything reaches a client or candidate.
- No message may be lost silently. Every inbound message ends up processed, flagged for a human, or raising an error alert.
- Everything is logged to the CRM as a note, so the record reflects what happened rather than what someone had time to write down.
- Failures must be visible. A dedicated error channel exists so a broken run is noticed rather than discovered weeks later.
- Settings live in one place per workflow (domains, labels, channels, models, booking link and confidence threshold), so operational changes need no engineering.
The pipeline
Drawn as text and shapes rather than shipped as an image, so the step labels stay readable and downloadable.
How one message becomes an approved reply
Deterministic filtering runs before any model call, confidence gates the draft, and the only component that can send sits behind a human click.
- Trigger
- Branch / filter
- Error handling
- Data store or CRM
- AI / model call
- Processing step
- Human checkpoint
- Delivered output
- Primary path
- Conditional or exception path
- Loops back
Four workflows, one send permission
Three workflows draft across three channels. The approval handler is the only component permitted to send, so there is a single path to a client's inbox.
- Processing step
- Human checkpoint
- Branch / filter
- Delivered output
- Data store or CRM
- Error handling
- Primary path
- Flow
- Conditional or exception path
How the work runs
Each stage, what it does, and the specific failure it is built to prevent.
The email assistant runs hourly and filters before it thinks
Each run fetches up to twenty new inbox emails. A deterministic pre-filter strips quoted history, reads calendar invites, and silently closes own-domain mail, auto-replies, no-reply addresses, bulk mail, newsletters and notifications. Only survivors reach a model, which makes the system both cheaper and more accurate than asking an AI to recognise noise.
Reserves model calls for genuinely ambiguous messages, where judgement is actually needed.
Senders are matched to the CRM, with masks unwrapped
The sender is looked up in the CRM, and where a booking tool has masked the real person the workflow recovers them from the reply-to header. Without that step the most human message in the inbox, someone who just booked a call, appears to come from a robot and gets filtered as noise.
Known senders are classified with a confidence score
Messages from recognised contacts are classified as client, candidate, lead or noise, with a confidence score and a flag for whether a meeting was requested. The classification drives both the draft's content and whether a draft is produced at all.
Unknown senders pass a deliberately generous gatekeeper
A second AI check evaluates unknown senders, biased toward surfacing rather than discarding. Genuine new-business enquiries appear in a new-lead channel with a qualifying draft and possible CRM matches shown by name and company for human verification, never auto-linked, because a wrongly attached conversation corrupts the record invisibly.
The filter's bias follows the cost of error: missing a real enquiry is expensive, one extra ambiguous message is not.
Low confidence produces triage, not a draft
Anything below the confidence threshold, or anything unreadable, goes to a triage channel with no draft attached. This is deliberate: a low-confidence classification does not yield a hedged draft, it yields a confidently wrong one, and a wrong draft is more dangerous than none because it invites an approving click.
Removes the failure mode where a plausible-looking wrong draft gets approved on autopilot.
The draft is written, then code finishes it
The model drafts in the recruiter's tone using a prompt specific to the channel. Then code, not the model, appends the real signature and adds the booking link whenever a meeting was requested. Deterministic content belongs in deterministic code; a model asked to handle it will eventually invent a signature or omit the link in exactly the message that needed one.
Eliminates invented signatures and missing booking links, two failures that read as carelessness to a client.
WhatsApp is filtered for the hazards that channel actually has
The WhatsApp workflow drops the recruiter's own outgoing messages, group chats, duplicate and replayed deliveries, reactions, stickers and one-time passcodes. Greetings are logged but never treated as leads, and media-only messages are logged without a blind draft. Every message is written to the CRM record with a follow-up task.
An echo guard stops the recruiter's own replies re-entering as new inbound messages, the classic self-reply loop.
The daily digest replaces memory with pacing rules
Each morning the system finds candidates who have gone quiet and drafts a personalised follow-up referencing the most recent real topic from their CRM history. Pacing is explicit: at least three quiet days before chasing, no more than fifteen days cold if never chased, a maximum of three follow-ups until the contact re-engages, and a three-day cooldown after each send. Up to ten pending drafts are listed, which keeps the review a two-minute job.
Caps and cooldowns prevent automated follow-up from becoming automated pestering.
Review links display a page; sending needs a second action
Corporate email scanners open every link in a message, which would have triggered sends from the digest. So a review link only displays a confirmation page; the send requires a separate form submission, with a signed token checked twice. This is the kind of hazard you only find by thinking about who else opens your links.
A security scanner pre-opening links cannot cause an outbound message.
The approval handler is the only thing that sends
A click sends the approved draft through Gmail or WhatsApp, opens an edit window, or records a discard, then logs the send to the CRM. Approving an email sends the live Gmail draft including any edits made in Gmail itself. Every card is replaced with the outcome and the name of the person who acted, turning the channel into an audit trail.
Every click is acknowledged instantly, before any work, so Slack's retries cannot cause a duplicate send.
Nothing fails silently
The processed label is applied last, so an interrupted run simply retries the message rather than losing or double-processing it. Failures get an error label a person can remove to retry, and every error alert in the WhatsApp path includes the sender's number and full message text so the team can reply by hand, because webhooks fire once and cannot be replayed.
A crash mid-run cannot lose a message or send one twice; an unretryable webhook still reaches a human.
What changed
| Aspect | Before | After |
|---|---|---|
| Inbox triage | Every email opened and judged by hand. | Noise filtered silently; only real messages surface, already classified. |
| Replying | Written from scratch. | A draft ready in the recruiter's tone; approval takes seconds. |
| Unlogged and easy to miss. | Every message logged on the CRM record with a follow-up task and a draft reply. | |
| Enquiries from unknown senders | Easy to overlook among other unknown mail. | Flagged in a new-lead channel with a qualifying draft and possible CRM matches to confirm. |
| Candidate follow-ups | From memory, and often generic. | A daily digest of personalised drafts with explicit pacing and caps. |
| CRM record keeping | Manual and incomplete. | Every message, draft and send logged automatically, with the actor's name. |
| Control and safety | Full automation was too risky to consider at all. | Nothing sent without a human click; every failure visible in a dedicated channel. |
What it runs on
Every tool listed with the job it actually does in this build.
Orchestration
n8n (self-hosted)
Orchestration for all four workflows, on persistent Postgres so deduplication memory and in-flight state survive restarts. Self-hosted because client and candidate message content should not pass through a third-party automation tenant.
AI
Claude
Two models with separate prompts per channel: a fast one for classification and gatekeeping, a stronger one for drafting. Splitting them keeps per-message cost low while the drafts stay good enough to approve unedited.
Integration
Gmail API
Reading the inbox, label management for processed and error states, and holding drafts in the real thread so the recruiter can edit in the tool they already use.
WhatsApp Business API
Inbound webhooks and outbound sends on the channel with the fastest response expectations and previously the weakest record-keeping.
Delivery
Slack
The approval surface: one card per draft with approve, edit and discard, plus separate triage, new-lead and error channels so different situations reach different attention.
Signed links and a confirmation page
Protects the digest's review links from email security scanners that open every URL, with the token checked twice before any send is possible.
Data
Recruitment CRM
The system of record. Every message, draft, send and discard is logged as a note, and identity matching reads from it rather than from a separate contact list.
What stopped being manual
Snippets are illustrative and generic. No client data appears in any of them.
Deterministic noise filtering before any model call
Replaces opening every message to find out whether it matters. The work removed is not reading the real emails. It is reading all the other ones first.
javascript
// Code closes what code can recognise. The AI is for ambiguity.
if (fromOwnDomain(msg)) return close("own domain");
if (isAutoReply(msg)) return close("auto-reply");
if (isNoReplyAddress(msg)) return close("no-reply");
if (isBulkOrNewsletter(msg)) return close("bulk");
if (isNotification(msg)) return close("notification");
msg.body = stripQuotedHistory(msg.body);
// A booking tool masks the real person, recover them.
const person = msg.replyTo ?? msg.from;The confidence gate
Replaces the assumption that an AI classification is usable. Below the threshold the system deliberately produces no draft, because a confidently wrong draft is more dangerous than no draft at all.
javascript
const { category, confidence, meetingRequested } = await classify(msg);
if (confidence < THRESHOLD || !readable(msg)) {
return toTriageChannel(msg); // NO draft attached, on purpose
}
const draft = await writeDraft(msg, category);
// Deterministic content never goes near the model:
draft.body += signature();
if (meetingRequested) draft.body += bookingLink();Follow-up pacing instead of memory
Replaces chasing whoever you happened to remember. Explicit caps and cooldowns mean the quiet candidates get followed up and nobody gets pestered.
javascript
const dueForFollowUp = (c) =>
daysSinceLastContact(c) >= 3 && // let it breathe
(c.everChased ? true : daysCold(c) <= 15) &&
c.followUpCount < 3 && // hard cap
daysSinceLastFollowUp(c) >= 3; // cooldown
// Draft references the most recent REAL topic from CRM history,
// so the follow-up reads as a continuation, not a template.Idempotent processing and single-path sending
Replaces the hope that nothing crashes mid-run. The processed label is applied last so an interruption retries safely, and drafts are located by thread at send time because Gmail changes draft IDs after edits.
javascript
try {
await handle(msg);
await applyLabel(msg, "processed"); // LAST, so a crash just retries
} catch (err) {
await applyLabel(msg, "error"); // a human removes it to retry
await alertErrorChannel(err, msg); // never fails silently
}
// At send time, find the draft by thread, not by a stale id.
const draft = await findDraftByThread(threadId);
if (!draft) return "already sent"; // so nothing goes out twiceWhat it was worth
The change is in what the recruiter's attention is spent on. Before, the day began by reading everything to discover what mattered. Now the messages that need a human arrive pre-classified with a draft attached, and the ones that do not arrive at all. Approving a good draft takes seconds; the system's value is that it made judgement the only remaining task.
Two gaps closed that were pure leakage. WhatsApp went from an unlogged channel where a message could evaporate to a fully recorded one with a CRM note and a follow-up task on every message. And new-business enquiries from unknown senders (the highest-value, most-skimmed messages in the inbox) now surface in their own channel with a qualifying draft ready.
The design decision I would defend hardest is that the AI never sends. It is tempting to read that as timidity, but it is what made the project possible at all: the client would not have accepted any system that could message a senior contact unsupervised, and they were right not to. Concentrating send permission in one workflow behind one human click turned an unacceptable risk into a tool they actually use. The lesson generalises: in relationship businesses, the constraint is not what AI can draft; it is what it is allowed to do unsupervised.
On measurement, I should be straight: there is no measured saving here. 'Hours every day' is the documented before-state, not a verified result, and nobody instrumented the after. What is verified is the system's scope, its approval gating, its logging, and ten named engineering hazards that were found and fixed. I would rather report that honestly than attach a percentage nobody counted.
Messages sent without a human click
Enforced structurally: only the approval handler can send, and only on a click.
Connected workflows across 3 channels
Email, WhatsApp and follow-ups, with one shared approval surface.
Approval point for every draft
One Slack card offering approve, edit or discard.
Follow-up touches maximum
With three quiet days before the first and a three-day cooldown between sends.
Named engineering hazards fixed
Replay, retry, duplicate-send, link-scanner and identity-collision failures.
Emails fetched per hourly run
Filtered deterministically before any model is called.
Handed over
- Four connected n8n workflows: an email assistant, a WhatsApp assistant, a daily follow-up digest, and the approval handler that is the only component permitted to send.
- Deterministic pre-filtering that strips quoted history and silently closes own-domain, auto-reply, no-reply, bulk, newsletter and notification mail before any model call.
- CRM identity matching with reply-to recovery for booking-tool masked senders, and possible-match suggestions shown for human verification rather than auto-linked.
- Confidence-scored classification into client, candidate, lead or noise with meeting-intent detection, and a triage path that attaches no draft below the threshold.
- A deliberately generous second gatekeeper for unknown senders, routing genuine enquiries to a new-lead channel with a qualifying draft.
- Channel-specific drafting prompts, with signature and booking-link insertion handled in code rather than by the model.
- A daily follow-up digest with explicit pacing: three quiet days minimum, a fifteen-day cold limit, a three-touch cap and a three-day cooldown, listing up to ten pending drafts.
- Signed review links that display a confirmation page only, so email security scanners cannot trigger a send.
- Slack approval cards with instant acknowledgement, outcome replacement showing who acted, and separate triage, new-lead and error channels.
- Idempotent processing (processed labels applied last, error labels a human can clear to retry, drafts located by thread at send time) plus full CRM note logging of every message, draft and send.
Questions about this build
What teams usually ask when they recognise their own process here.
Because the cost of being wrong is asymmetric and irreversible. In a business built on relationships with senior people, one confidently wrong automatic message can lose a client, and you cannot un-send it. The time saved by skipping approval is seconds per message; the downside is a relationship. Concentrating send permission in a single workflow behind a human click is what made the client willing to adopt the system at all, a tool they use beats a more automated one they switch off.
No, because approval is not the expensive part; discovery was. The old cost was opening every message to find out whether it mattered and then composing from scratch. Now noise never surfaces; what does surface is already classified, and a good draft is waiting. Clicking approve takes seconds. The automation removed triage and composition and left only judgement, which is the part worth a human.
It produces no draft. Anything below the confidence threshold, or anything unreadable, goes to a triage channel for a human to handle from scratch. That is a deliberate inversion of the obvious design: a low-confidence classification does not generate a cautious draft, it generates a confidently wrong one, and a wrong draft is more dangerous than an empty queue because it invites an approving click from someone moving fast.
Several guards, because there are several loops. An echo guard drops the user's own outgoing messages so replies do not re-enter as new inbound ones. Persistent deduplication memory drops duplicate and replayed webhook deliveries. Anything older than forty-eight hours is ignored, so a backlog replay cannot trigger a flurry of stale sends. And own-domain mail is closed outright in the deterministic filter.
Because a model asked to reproduce a signature will eventually invent one, and a model asked to include a booking link will eventually omit it in exactly the message where a meeting was requested. Those are deterministic requirements with a single correct answer, which makes them code's job. Give the model the part that needs judgement, the prose, and handle the rest where it cannot drift.
Because corporate email security scanners open every link in every message before the recipient sees it. A link that sends on load would have fired sends the moment the digest hit a scanned inbox. So the link renders a confirmation page and the actual send requires a separate form submission with a signed token checked twice. It is a good example of a hazard you only find by asking who else opens your links.
By never auto-linking. When an existing contact writes from a new address, identity detection searches the CRM by name and company and surfaces possible matches for a human to confirm. Automatic linking would be convenient and occasionally wrong, and a conversation silently attached to the wrong person corrupts the record in a way nobody discovers until it matters. Suggesting is cheap; guessing is not.
Not measurably, and I will not pretend otherwise. The before-state of spending hours a day in the inbox is documented, but nobody instrumented the after, so there is no verified hours-saved or cost figure. What is verified is the architecture, the approval gating, the complete logging, and ten specific engineering hazards found and fixed. If I were starting this again I would add simple counters from day one (drafts approved unedited, messages reaching triage, time from arrival to send) because those would have made the case far better than an estimate.
Capabilities this used
Read next
All ai automation workHave a process that works like this one did?
Book a strategy call. We will walk the workflow with you and say plainly which parts are worth automating and which are not.