Grokbot Multi-Agent System for Post-Sales Follow-Ups

Key takeaways

Treat Grokbot as a staff of agents with tools, APIs, MCPs, and their own virtual computers — not a one-question chatbot.
For post-sales, staff specialists (follow-ups, voice, source of truth, internal radar, one bot per account on a small-to-medium book) and put a chief of staff in front so you are not managing ten chats.
Start with a 10–15 minute voice dump of the job, the broken processes, and what you want more time for. Paste it to the first bot and ask it to design the system.
Keep customer-facing sends as drafts. Blake’s rule: the pack comes back ready, and a human tweaks and hits send.
Feed Wally (or your writer bot) real emails and Slacks you have already written, plus persona rules: casual internal vs short and formal for executives.
Turn the setup into a living system: scheduled briefs, a weekly self-improvement scan limited to one suggestion, and voice learning from the delta between draft and sent.
Spend cost where it matters. Virtual-machine UI work and always-on routines are the expensive habits; a lean, specialized team is cheaper than 45 bots checking every 15 minutes.
Humans stay on the customer. Use the bots for the in-between work so you can keep the meetings that actually move the account.

Who this is for (and who should skip it)

This is for a post-sales owner — customer success, deployment, account management, or any customer-facing role after the contract is signed — who lives in meetings and loses the work between them. Blake’s stack assumes Grokbot plus Gmail, Slack, Google Calendar, Google Meet, Google Forms, and Granola notes. You do not need to be an engineer. You do need to connect those plugins, review drafts, and spend real time in the first stretch teaching the team how you work.

Skip the “one specialist per account” pattern if you are running something like 1,500 accounts; Blake says there is a better way at that scale, and he does not spell it out here. Skip full autonomy if your plan is that Grokbot never talks to you and owns the customer. Blake’s view is that customer-facing work still needs a human, and he does not let bots send on their own.

What Grokbot is (stop thinking in single chats)

Blake maps AI use on a maturity curve: chat, then copilots, then bots, then full teams of bots that take work start to finish. Grokbot sits on the last step. The interface looks like iMessage: bots on the left, pin the ones you talk to, marketplace for shared bots and plugins (Gmail, Google Calendar, Slack, and the rest). Each bot has its own computer. It can open Chrome, build a deck, join a Meet, send you screenshots or screen recordings, and — with Teach a task — watch you do a workflow once and repeat it.

They run 24/7. You can close the laptop. The hard part, Blake says, is not capability. It is deciding what you actually want done, and thinking bigger than a chatbot Q&A.

The thing probably holding back your Grokbot is you thinking about what is actually possible.

What you are trying to accomplish in this phase: get the mental model right before you name ten bots. Prepare access to Grokbot and the plugins you already live in. Good looks like a short list of jobs (not a wish list of 45 agents) and a decision that one bot will be the front door. Failure mode: a blank page, shopping other people’s bots, and never dumping how your job actually works.

Post-sales work Grokbot can take off your plate

Post-sales, as Blake uses it, is anything customer-facing after the sale. His days are emails, Slack, Teams, meetings, and conversations. About 75% of a 9-to-5 can sit in meetings if that is the part of the job you protect. The six use cases below are the ones he actually runs. He says he could have made twenty slides; these are the spine.

Use case	Bot / tool	Input needed	Output	Owner
Morning status board	Gus (chief of staff)	Overnight activity, calendar, yesterday’s unfinished work	Fires, who you meet, what you promised and did not do, how to set up the day	You read it
Call prep	Gus	The meeting, 15–20 minutes out	Who is on the call, title, last conversation, something new since then, a suggested talking point, anything to unblock	You walk in ready
Follow-up desk	Gus routes to Frankie, Wally, account bot	Granola transcript when the call ends	Draft replies, Slack updates, requested materials (ROI one-pager, forms)	You edit and send
Promise keeper	Routine + Gus	Commitments you made	A nudge so you actually do what you said	You follow through
Ask watch	Gus / specialists	Asks you made of other people	Surface when the reply never comes back	You chase with context
Account reset	Gus + account bot, Frankie, Scout	“Where are we at with [account]”	Full download: risks, people, blockers, open promises, Slack, next steps (Blake also gets a space joke)	You decide the next move

Two more patterns show up in the live workflow. Gus can join a Google Meet you choose to skip, identify as your bot in chat, and send takeaways — Blake is protective about which calls get that treatment. And a staff meeting pulls specialists into one thread when you need competing views (for example, how to spend the only free hour in the day) instead of one bot always biasing toward its own account.

Staff the team: one chief of staff, specialists behind it

Blake only talks to Gus. Gus is chief of staff. Specialists do the work; Gus routes, waits, and brings back one pack. That is a personal constraint — he is distractible — not a product requirement. Other people talk to every specialist. He tried that and did not want to manage a team of ten.

I didn’t want to manage a team of 10. So Gus manages a team of 10.

Name the seats, then name the bots however you like:

Chief of staff (Gus). The only chat you live in. Routes to specialists, compiles packs, joins selected meetings, runs staff meetings, returns one answer.
Follow-ups (Frankie). Owns the after-call work: replies, materials, the things that used to sit until 7 p.m.
Voice (Wally). Writes like you. Load past emails and Slacks so you are not starting from scratch. Teach personas: internal Slack can be lowercase and emoji; an exec email is short, formal, and to the point.
Source of truth (Trudy). Documentation and platform answers with sources. Gus goes to Trudy when the answer has to be right, not vibes.
Internal radar (Scout). Replaces juggling 30–40 Slack channels and launch emails. Daily: what is new, what you need to know, link to the channel. Some days Scout is the second bot Blake actually manages.
Account specialists (Harbor, Northwind, Brightline in the demo). One bot per account on a small-to-medium book, loaded with that account’s context. Fun names are fine. At very large books, do not clone this 1:1.
Staff meeting. A room where those bots argue on purpose, then Gus brings you the call.

In the product: bots on the left, pin the ones you use, add a label in the description (Blake labels Gus as chief of staff). Specialists talking to Gus should not ping you. Blake wants a green/blue “ready” state on the thread he is actually in — not ten interruptions. Franny, in the demo, is the bot he opens directly when he wants to watch a virtual machine build a form.

What good looks like: you manage one or two chats, Gus can currently handle on the order of 15–20 specialists before Blake would even consider another management layer, and account bots stay quiet unless Gus asks. Failure mode: you become the switchboard, or you add middle-manager bots you do not need. Blake has not needed an org chart of chiefs. If Gus were hitting a wall, he would add a layer; until then, direct routing is enough.

Run the workflow: meetings, follow-up packs, resets, staff calls

Join a meeting you cannot attend (selectively)

Tell Gus to join and send takeaways. In the demo he joins a Google Meet, clicks through mute and camera, puts in the name, and writes in chat that he is Gus, Blake’s bot, and will send takeaways. Use this when the meeting is an internal update you would otherwise miss — Blake’s example is a branding update — not as a default for every customer call. Be protective. After the meeting, Gus can send decisions, links, asks, and owners even if you dismissed a handoff because you were in a demo.

Close a customer call with a follow-up pack

When the call ends, the follow-up desk should already have the Granola transcript. In a live day this can be automatic; in the demo Blake typed that he was done with the Harbor call so you could see the routing. Gus does not pretend to be the writer, the follow-up specialist, and the account brain. He messages Wally, Frankie, and the Harbor account bot. You get one Harbor post-call pack, marked drafts only.

In the demo pack:

A Gmail draft to Maya in Blake’s voice (he is “an exclamation point person”).
A Slack draft to Alex, the AE: just chatted with Harbor, meeting new champions next week, working on the ROI PDF.
An ROI one-pager the customer asked for, built with Flylow branding, via Frankie.

Blake tweaks, then sends. That in-between work, he says, used to take at least 45 minutes. Good output is a pack you can ship after a short edit, not a bot that already emailed your customer. If the writing is off, fix the send and let the later voice-learning routine catch the delta — do not skip the human pass.

Build something in the virtual machine when the artifact is the point

For a second VM example, Blake goes to Franny: make an ROI form for Northwind. The bot opens Forms, clicks through, and drafts questions (team and role, primary use cases, what done looks like), then returns a share link, an edit link, and asks whether to send. Watch the VM if you want; if the path is a repeatable click-path you do not want guessed, use Teach a task. Speed depends on the task. This is also the expensive pattern — more on that under cost.

Account reset when you feel underwater

A short trigger is enough if you have taught the shape of the answer. “Where are we at with Harbor” is Blake’s stress phrase. Gus knows to collect a full picture (Harbor, Frankie, Scout), wait on the stragglers, and return one pack: what is at risk, who the people are, blockers, open promises, what Slack is doing, what to do next. The space joke is optional personality. You do not need it. You do need a fixed output shape so one front-door bot can serve a chaotic day.

Staff meeting when one bot’s bias is the risk

When the question needs several perspectives — what to do with the only free hour, whether you are always favoring one customer — start a staff meeting. Tell them to disagree and fight for your time; uncritical agreement is not useful. In the demo, Harbor pushed SSO, Northwind and Brightline stood down, Frankie said protect the customer-facing send that needed to go out, Scout agreed. Gus came back with a sequence: SSO first, send Maya the next-step email (already drafted), park the rest. Grokbot can juggle this while other packs are still running.

Start with a voice dump, not a blank bot

The intimidating part is the empty system. Blake’s fix is context, not more marketplace browsing. Take 10–15 minutes. Pace with a phone voice memo, or use Grokbot’s voice-to-talk if you are not a pacer. Talk through the job: who you are, what you like, what you are struggling with, which processes are broken, what could be better. Paste the long transcript into the first bot — it does not have to be named Gus yet — and ask it to build a system that fits you.

In Blake’s case the bot reflected the job back: protect calendar time for live customer conversations; take Slack-checking off the plate. That is the point. The useful system is the one that matches how you actually work, not a generic agent org chart. If you get stuck, ask Grokbot. You can tell it not to stop until it figures out the solution. Push it on “how can we make this better?” You do not need all the answers in your head before you start.

Prompts, bot instructions, and workflows

These are the messages and rules Blake actually used or described. Cleaned of filler; not rewritten into a different method.

Join a meeting and return takeaways (demo):

Can you join the internal learnings call for me right now? Send takeaways when it's done.

What Gus posted in the meeting chat (demo):

Hey everybody, this is Gus, Blake's bot. I'm going to sit here and send takeaways.

Kick the follow-up desk after a named account call (demo; in production this can fire from the Granola transcript automatically):

I'm done with the Harbor call.

Standing rule on that pack (teach this once, keep repeating it):

Harbor post-call pack is drafts only. Never send email or Slack on your own. Draft in Gmail and Slack, build any materials they asked for, and bring the pack back to me.

Build a form in the VM (demo):

Make me an ROI form for Northwind.

Account reset trigger (demo):

Where are we at with Harbor?

Staff meeting (demo):

Start a staff meeting to talk about what I should spend the next and only free hour of my day today.

Reconstructed from the speaker’s description — voice dump to the first bot:

Hi, my name is [name]. This is my job. Here's what I like about my job. Here's what I'm struggling with right now. Here are the processes that I think are broken. Here are the things I feel like we could do better.

I need a system that's going to work for me. Help me build it. Protect time for the customer conversations I should be in. Take the in-between work and the channel-checking off my plate.

Reconstructed from the speaker’s description — teach the account reset shape:

Sometimes I feel underwater with an account. When I ask where we are at with [account], I need one pack: risks, people, blockers, open promises, what is happening in Slack, and what to do next. Collect from the account bot, follow-ups, and Scout. Do not ping me from every specialist — route it through the chief of staff. [Optional: open with a space joke.]

Reconstructed from the speaker’s description — staff meeting ground rules:

Always disagree. Fight for my time. Do not all bias toward your own account. Come back with the one or two priorities that make the most sense.

Reconstructed from the speaker’s description — Wally voice and personas:

Here is a batch of emails and Slacks I have actually written. Match this voice.

Internal Slack: casual, often all lowercase, emoji is fine, not formal.
Exec email: short, to the point, formal, still recognizably me.

Reconstructed from the speaker’s description — self-improvement limits:

Once a week, run a self-improvement scan. (1) System audit: if I am still doing work manually that I never asked you to take, flag it. (2) Voice learning: compare what you drafted with what I actually sent, send the delta to Wally, update the rules.

Send me only one automation suggestion per week. If I reject it, give me a different one. Do not propose ten new bots.

Measurement and business impact

Blake has not calculated an exact time-saved number. The qualitative checkpoint is the shape of the day. Before: roughly six hours of meetings plus three to four hours of the work those meetings create — nine- or ten-hour days. After: the same volume of calls, and the in-between work close enough to an eight-hour day that it feels like there is room to do more. Follow-up packs replace a stretch he puts at least 45 minutes. Call prep replaces the scramble 15–20 minutes before you join.

On the customer side, the impact he describes is not a retention percentage. Account bots surface micro-ships and issues he would have missed. He can send a tailored note while the window is open, get ahead of problems because Harbor (or the equivalent) is always watching that book, and spend the recovered hours on the live conversations he already believed moved the account. Customers feel more tailored because he has the time and energy to make them feel that way. Managing the bots themselves is heavy in the onboarding stretch and light once the system is quiet unless a routine or a specific trigger needs him.

Pitfalls and guardrails

Do not let it send. Drafts only. Blake is explicit: he never wants Grokbot sending on its own.
Do not put a bot on every call. Meeting join is powerful and easy to overuse. Be protective.
Do not skip voice training. If you do not load real writing, you will edit every draft forever. Wally only knows you if you show it.
Do not over-index on routines. A check every 15 minutes that you can ignore still burns prompts in the background. Three of those is hundreds of messages a day.
Do not default to the VM. Building a full UI form in the virtual computer costs more than gathering the content and using Google Forms AI. Use the computer when the click-path is the job; do not use it as a party trick.
Keep the team lean. You do not need 45 bots. Specialize. Blake would rather manage Gus than a crowd.
Cap self-improvement. Unlimited suggestions overcorrected into ~10 new bots. One suggestion per week, taken or pushed back, is the rule he kept.
Do not clone account bots at huge scale. Medium-to-small books: yes. ~1,500 accounts: find another pattern.
Humans stay in the customer conversation. Asked how far away we are from never talking to the customer, Blake’s answer is: far. Take the work that does not need a human; make the remaining hours more impactful.
Big voice dumps can ramble. Blake does not see much hallucination in Grokbot, but if you paste a massive blurb, ask it to distill to the specific pieces you need, then do the deeper analysis.
Prevent sprawl on day one. Interdependencies get messy if you delete bots later. Set routines and ownership now — Blake’s analogy is doing the chores when you first move into a house. A dedicated “tech debt” bot that catalogs unused agents and interdependencies is an option he floated, not a product default.
Other agents inside the VM as a router. Someone asked about parking Claude Code, Cowork, or similar inside one VM and using Grokbot as the router. Blake has not seen that done and would not claim it works. If you have invested heavily elsewhere, that is a conversation to have with Grokbot about consolidating — not a recipe.

Keep your team lean. You don’t need 45 bots.

7-day implementation plan

Day 1

Record the 10–15 minute voice dump. Paste it into your first Grokbot. Ask it to design a system around the job you described: protect customer time, name the in-between work, propose a chief of staff plus the fewest specialists that cover follow-ups, voice, source of truth, internal radar, and your real accounts. Connect Gmail, Google Calendar, and Slack from the marketplace. Write the drafts-only rule before anything else can send.

Days 2–3

Stand up Gus (or whatever you name the front door) and Wally. Feed Wally a batch of real emails and Slacks and the two personas (internal vs exec). Add Frankie for follow-ups. Pin Gus. Label roles in descriptions. Tell specialists to talk to Gus, not to you, unless you open them directly. If you already have Granola, point the follow-up desk at those transcripts. Do not add a routine that runs every 15 minutes.

Days 4–5

Create one account specialist for a live account, not a hypothetical. After one real call, run the follow-up pack: draft email, draft Slack to the AE or internal owner, any artifact the customer asked for. Edit and send yourself. Teach the account-reset trigger and the output shape (risks, people, blockers, promises, next steps). If you repeat a click-path, use Teach a task rather than hoping the VM guesses. Skip VM form-building unless the form is the actual deliverable and you accept the cost.

Days 6–7

Add one or two routines you will actually read — morning brief, unfinished promises, a Friday dashboard — not a nest of ignored checks. Schedule the Wednesday self-improvement scan with a hard cap of one suggestion, plus voice learning from draft-vs-sent. Run one staff meeting on a real priority conflict and require disagreement. Review cost: kill extra bots, move “build this in the UI” jobs to cheaper paths where you can, and confirm Gus is still the only chat you live in. If you are migrating from OpenClaw or another stack, export the context you care about, hand it to Grokbot as a fresh brief, and optionally staff-meet on what to keep versus rebuild. Blake does not offer a technical import path.

Make it a living system, then get back to the customer

Grokbot will do a lot on its own — Blake puts that at about 90%. The extra you invest should be the parts you refuse to get wrong: voice, drafts-only, a quiet front door, and a weekly loop that notices both leftover manual work and the sentences you keep fixing. You invest in it the way you invest in a coworker. Then you spend the recovered hours in the meetings you already knew were the job.

Watch the walkthrough to see Gus join a Meet, assemble a Harbor pack, build the Northwind form, and run the staff meeting in real time — then start with one chief-of-staff bot and the voice dump, not a roster of 45 agents.

FAQ

Do I need a Grokbot for every account?

On a small-to-medium book, Blake’s favorite pattern is one account specialist with all of that account’s context — “your own employee” dialed into Harbor, Northwind, or whoever. If you are at something like 1,500 accounts, he says there is probably a better way and does not recommend cloning the pattern. Start with the accounts that actually create follow-up and risk, not a bot for every logo in the CRM.

Grokbot vs OpenClaw (or Claude Code / Cowork) — what is actually different?

Blake is not a deep user of those tools. What he claims Grokbot does well: every bot has its own computer, you are not carrying a Mac Mini to keep an agent running, VM work is a first-class path, and the product is built for a staff of agents rather than a single assistant. On migration, he does not know a technical import path. His method is dump the context, let Grokbot rebuild, and optionally run a staff meeting on what was working and what to improve.

Should Grokbot email my customers without me?

Not in this method. Blake’s follow-up packs are drafts only. He never wants the system sending on its own. Asked how far away we are from Grokbot handling the account so he never talks to the customer, he said we are far: humans are still what you want on customer-facing work. Use the bots for the pile between meetings.

Do I have to manage ten specialist chats?

No. That is the point of the chief-of-staff pattern. Blake only talks to Gus; Gus talks to Frankie, Wally, Trudy, Scout, and the account bots. Specialists should not ping you when they are collaborating. If you like talking to specialists directly, you can — he found it distracting. Add another management layer only if one chief of staff is actually hitting a limit; he has been fine into the 15–20 range and has not needed the extra org chart.

How much time does this save, and how much time does managing bots take?

No precise timesheet. Anecdotally, Blake went from six hours of calls plus three to four hours of follow-up work (nine- to ten-hour days) to the same calls with the admin close enough to an eight-hour day. A single follow-up pack replaces work he puts at least 45 minutes. Managing bots takes a lot of time at the start — it feels like onboarding a team, including correcting voice — and much less once the system is taught to stay quiet except for specific triggers and routines.

Why not run routines every 15 minutes “just in case”?

Because they still cost. Blake says the product was built to be cost-efficient if you use it strategically. VM walkthroughs cost more than giving Google Forms AI the content. Always-on routines are easy to ignore and expensive to leave running. Three 15-minute loops become hundreds of messages a day. Prefer a morning brief, a promises check, a Friday dashboard, and one weekly improvement pass.

How do I keep Grokbot from drifting off my voice or forgetting instructions?

Load real writing on day one, then run voice learning: the delta between draft and sent goes back to Wally and updates the rules. Each bot is its own individual with a number, history, and change logs if you want traceability. Blake’s easier path is to ask Grokbot for an analysis of what it has done, what changed, and what may have been forgotten. He also says you do not have to manage context as tightly as on other LLMs — it is meant to know what to remember — but the weekly scan is how he keeps investing in the parts he cares about.

What if two bots disagree, or I start accumulating bot sprawl?

Blake does not describe specialists fact-checking each other’s homework. Each bot has its own verification loop; Gus’s job is the broader view across them. For sprawl, do not wait until you are deleting interconnected agents. Put ownership and routines in place at the start, keep the roster specialized, and consider a bot whose job is unused agents and interdependencies. Call the herd on purpose; do not grow to 45 and hope.