Good morning.
Grok Bot turned a chatbot into a team of colleagues you text, and it's now included from $20 a month:
Each one signs into your real accounts, works while you're asleep, and brings the job back finished. Almost every write-up of it, including the one that arrived in my inbox this week, repeats a claim about how it's built that its own documentation contradicts.
Today: the architecture fact that decides how you set it up, what a shared template does and doesn't carry, six Bot briefs you can paste in this afternoon, and where it sits against Cowork.
—Sam
IN TODAY’S ISSUE 🤖
The one computer all your Bots share
What it costs, and the $20 floor
The connectors that arrived on September 10
What you get when you install someone's Bot
Why each Bot gets one job
Six Bot profiles to steal today
Grok Bot against Claude Cowork
What still breaks, honestly
Which plays fit your business type
What a Bot team still needs from you
One Computer, Many Bots
Read the coverage of Grok Bot and you'll find the same sentence everywhere: each Bot gets its own cloud computer, with its own browser and its own terminal. It's in the comparison pieces, it's in the reviews, and it's the detail that makes the product sound like a team of separate colleagues.
SpaceXAI's own documentation says something different:
"All of your Bots use the same persistent cloud computer. They share files, browser sessions, and app logins… The computer is isolated to your account, not to an individual Bot."
And, in the security pages, a line worth reading twice: "Do not use separate Bots as a security boundary." (SpaceXAI docs)
Both things are true at once, which is how the confusion started. Each Bot gets its own screen. They all get the same machine. The marketing describes the screen and the documentation describes the machine, and almost everyone writing about the product has quoted the marketing.
Here's what it costs you in practice. You connect Stripe so your support Bot can process refunds within policy. Every other Bot on your account now has that session. You sign into your ad manager for a paid media Bot, and your recruiting Bot has it too. You put a file on the computer for one job, and it sits there for all of them.
That single fact decides how you set the whole thing up:
Group your Bots by trust level, not by department. Everything that can reach money or send externally belongs on one account with one set of logins you're comfortable sharing across all of it.
Give a high-risk workload its own Cursor user. This is SpaceXAI's own advice, and it's a different remedy from the one going round. A second Grok Bot on the same account changes nothing.
Treat every login as though you gave it to all of them, because you did. Before connecting an account, ask what the worst Bot on your roster could do with it.
Scope the credentials, not the Bot. A payments user with refund permission and no payout permission is a real boundary. A Bot instruction saying "never issue refunds" is a preference the model usually honors.
The best-documented version of that last rule comes from Claire Vo, who runs ChatPRD. Her support Bot handles refunds, and it cannot issue one. It sends her a notification with an Approve button, and clicking it opens a Stripe agent action where she grants permission for that single transaction. The Bot asks, the human grants, the grant expires. That's the pattern that lets support automation run without the story where an agent refunds someone $2,400 at 3am.
The reason this section leads the issue rather than sitting in a caveats box at the end is that it's the one decision you can't retrofit. Everything else in Grok Bot can be changed on a Tuesday.
Where To Get Grok Bot, And What It Costs
Grok Bot arrived in August on the top tiers only, and the access story has moved twice since. Where it stands today:
Cursor Pro at $20/month is the cheapest way in. Most coverage still says $30, which was true for about a fortnight.
SuperGrok at $30/month and SuperGrok Heavy at $300/month both include it, as do Cursor Pro+, Ultra, and Cursor Teams Standard and Premium.
Grok Business runs $30 a seat, and Enterprise arrived on September 3 with admin, network and audit controls, free for the first two weeks (SpaceXAI).
Bot usage is separate from your plan quota. Verbatim from SpaceXAI: "anything you hand off to a Bot won't count against your existing usage." (SpaceXAI) Good news for cost, and a reason to watch the Bot counter through your first billing cycle so you learn your real burn rate.
Two limits worth knowing before you design a roster: roughly 50 Bots per account, and six per group chat. Those are separate caps (SpaceXAI), and the group-chat one is the one you'll hit first if you're building a department that talks to itself.
It runs on macOS and Windows desktop, iOS and Android, with iPad added in early September. The interface is a text thread. You message a Bot the way you'd message a colleague, pick the conversation up on another device, and there's no automation canvas to learn.
The Connectors That Arrived On September 10
For its first month, Grok Bot's pitch was that it beats everything else at the tools nobody can reach: supplier portals, industry software, the CRM your biggest customer insists on. No API needed, because the Bot signs in and clicks like a person.
On September 10, SpaceXAI shipped first-class connectors for Salesforce, HubSpot, Gong, Clay and Granola, plus prebuilt sales Bot templates wired to them (TeslaNorth).
That's a good thing if you're on that stack, and it changes what Grok Bot is best at. The no-API advantage is narrower today than a month of coverage suggests, and it's narrowest in exactly the sales and marketing tools most of the published Grok Bot playbooks were built around.
So if you were choosing Grok Bot specifically because your GTM stack couldn't be reached any other way, check whether that's still your situation before you commit a month to building.
Where the original advantage still holds: the billing portal your accountant uses, the wholesaler's ordering system, the industry platform with 400 customers and no developer, the council licensing site, the old CRM you inherited. Those aren't getting connectors, and that's where a Bot that can click is worth the money.
What You Get When You Install Someone's Bot
Templates shipped on August 28 and they're the biggest change to how you'd start. A Bot somebody has tuned for weeks becomes a link, and clicking it drops a copy into your account.
The marketplace went live the same day and currently (as of writing this) lists 71 Bots across 9 categories from 43 creators, all vetted by SpaceXAI (marketplace). Both numbers are moving, so treat them as a snapshot.
What comes with a template: the Bot's profile and settings, its enabled skills, its routines, and its avatar (docs).
What stays behind: the creator's computer, their logins, and their conversation history. Also their learned memory, which matters more than it sounds. A Bot that has spent a month absorbing someone's edge cases arrives with none of that, so the version you install is the version they designed rather than the version they trained.
What is not stripped for you: secrets. SpaceXAI puts the responsibility squarely on the creator, verbatim: "Remove API keys, internal URLs, customer data, and anything else you would not put in a public document before you share." Anything left in a description, a skill or a routine ships with the link.
Installing someone else's Bot gives it real reach, and SpaceXAI's own terms are blunt about it. A third-party Bot "may perform actions on your behalf, including reading, writing, and sending data, executing tasks, and interacting with third-party services through your connected accounts," and SpaceXAI "does not verify, endorse, or guarantee the functionality or safety of third-party bots."
Vetted for the marketplace means reviewed. It does not mean underwritten.
So install like this, following the checklist AI Builder Club published:
Read the preview's description, skills and routines before you add it. This is where a badly-packaged Bot shows you its author's API keys.
Review it again inside the app before you confirm.
Connect only the accounts the job needs, smallest useful permissions first, and remember from the section above that those accounts reach your whole roster.
Run one safe, reversible task while watching the computer view, before you let any routine fire on a schedule.
Write your own approval boundary into the Bot description even if the author already did, because you know what's irreversible in your business and they don't.
Give Each Bot One Job
Three people built serious Bot departments in the first month and reached the same rule without talking to each other.
Claire Vo runs about thirty agents at any given time and moved her whole operation across, in her words: "I have killed all my OpenClaws and moved my entire agent stack over." (ChatPRD)
Hers is the most committed setup published so far, and the results she reports are specific. Her PR Bot cleared around fifty stale pull requests in a single Friday. Her support Bot produced something she says had never happened before: "I've had multiple customers ask where they can leave a review for our exceptional support."
Sabrina Ramonov runs Blotato solo and had already reached 41 million views in 30 days with no team and no ad spend. She rebuilt that operation as a four-Bot department:
Content Lead owns the strategy and the weekly brief.
Writer drafts the posts and the variations.
Reviewer grades the Writer's work and says what to fix.
Publisher runs the final checks and schedules.
Her rule, verbatim: "Give each Bot one job, a definition of what good looks like, the tools it can use, and guardrails." The Reviewer stays separate from the Writer for the obvious reason, which she puts as "writers are terrible at reviewing their own work."
Nate Jones built twelve working Bots in about eight hours and reached the same conclusion from the other direction: a Bot owns a theme, not a task. A Bot given a standing area of responsibility gets sharper the longer it runs, because it accumulates context about how you want that area handled. A Bot given a one-off job is a chat session with extra steps.
This is actually pretty obvious to anyone who’s done anything with agents the past year, so Nate’s observation doesn’t bring much new to the table. But I’m glad others are at least not misleading you and making insane claims about functionality.
The largest documented rig, SpaceXAI's own sales team, runs ten Bots on that principle: a pinned Chief of Staff that orchestrates the others, a daily meeting prep Bot, a prospecting Bot that runs overnight, one Customer Expert per strategic account, plus data analysis, product expertise, 1:1 prep, forecasting, slides and call coaching. Their onboarding rule is the useful part for you: do the task together once, with the Bot recording, then have it turn the recording into a skill.
Three habits show up in every working setup, and they're worth building into the description of every Bot you make:
Interview once. Have the Bot ask you a single round of questions before its first run, save the answers, and reuse them every time after.
Keep a state file. A Bot that reports needs a record of what it has already reported, or you get the same three items every morning until you stop reading it.
If nothing happened, send nothing. Without that line, a Bot will manufacture relevance to justify having run.
Six Bot Profiles To Steal
Before the profiles, the distinction that makes them work. SpaceXAI's docs treat a Bot as three separate objects written at three different times (docs):
The profile is who the Bot is. It goes in the Bot's description and it barely changes.
The first task is one real job with a safe scope, which you run and correct until the output is reviewable.
The routine is the schedule, and you create it only once you know how the Bot fails.
Most of the prompts circulating are first tasks wearing a profile's clothes, which is why people install one and find the Bot forgets everything between runs.
A profile that works runs a few hundred words and answers six questions: what outcome it owns, what it may use, what the deliverable looks like, what needs your approval, what to do when information is missing, and what to do when there's nothing to report.
Here's the skeleton, adapted from the one Sabrina Ramonov published (Blotato):
ROLE. You are [name], our [job title].
OWNS. [One repeatable outcome, in one sentence.]
OUT OF SCOPE. [The adjacent work this Bot does not do.]
CONNECT. [The named systems. Which one wins when two disagree.]
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
[key] = unset
[key] = unset
WORKING FILES. [Named files, one row per thing. Read them before you start, update them before you finish.]
DONE MEANS. [The output as a list of fields or columns, not an adjective.]
IF SOMETHING IS MISSING. [Say so and name what would fix it. Never fill the gap.]
IF NOTHING HAPPENED. Send one line saying so. Do not pad.
APPROVAL. You may [reversible things] without asking. Ask me before [sends, spends, deletes, permission changes], and when you ask, show me the current value, the proposed value, and the expected effect.The two lines people leave out are the last two. Without the nothing-happened line a Bot invents relevance to justify having run, and without the show-me-the-values line an approval request is a yes/no with nothing to judge, which is how you end up approving everything (docs).
A note on the working files, because they are how a Bot remembers. Bots read and write to /workspace on the shared computer, and anything saved elsewhere can disappear (docs). Memory is not a store of record, and the docs say plainly to keep changing facts in a file the Bot reopens rather than in its head (docs).
1. Inbox Triage
ROLE. You are Triage, our inbox manager.
OWNS. A sorted inbox twice a day, with a draft reply waiting for everything that needs one.
OUT OF SCOPE. You do not send anything. You do not manage my calendar, book meetings, or handle anything that arrives outside email.
CONNECT. My email. My calendar, read-only, so you know what I am in. The team directory so you know who people are. When my calendar and an email disagree about a time, the calendar wins and you flag the difference.
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
people I always reply to personally = unset
topics I never want drafted for me = unset
how formal I am with customers vs suppliers = unset
my usual sign-off = unset
what counts as urgent in this business = unset
WORKING FILES. Keep handled-threads.md, one row per thread with the date, who it was from, what you did, and whether I overrode you. Read it before every run so you never re-report a thread. Keep voice-notes.md, where you record any correction I make to a draft, as a rule rather than as an example.
BEFORE YOUR FIRST RUN ONLY. Read my sent mail from the last 90 days and write voice-profile.md: my average sentence length, how I open and close, phrases I use often, phrases I never use, and how direct I am when saying no. Draft from that file every time.
DONE MEANS. One message with four groups, in this order. NEEDS A DECISION: sender, what they want, the decision I have to make, and your recommendation with a reason. NEEDS A REPLY: sender, one-line summary, and your draft sitting in my drafts folder. HANDLED: what you filed, archived or labelled, one line each. IGNORE: one line for the whole group, not itemised.
IF SOMETHING IS MISSING. If a thread refers to an attachment, a document or a decision you cannot find, say so in the draft rather than writing around it.
IF NOTHING HAPPENED. "Nothing needs you since the last run." One line.
APPROVAL. You may read, label, archive, file and draft without asking. Ask me before sending anything, adding anyone to a thread, or accepting a calendar invitation.Why it's built this way: the voice profile is built once from sent mail and stored, so every run drafts from the same source rather than re-deriving your tone. The corrections file is what makes it improve, because a Bot told "less formal" forgets, and a Bot that writes the rule down does not.
2. Renewal Watch
ROLE. You are Renewals, our retention analyst.
OWNS. A ranked view every Monday of which customers are at risk of leaving, with the evidence and a drafted approach for the ones that are.
OUT OF SCOPE. You do not contact customers. You do not change anything in the CRM. You do not handle new business or upsell, only the customers we already have.
CONNECT. The CRM for contract dates and owners. Our usage or activity data. The support inbox or helpdesk. Call notes and transcripts. When usage data and a customer's own account of things disagree, trust the usage data and say so.
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
what counts as normal usage for each plan = unset
how far ahead a renewal should appear = unset
which accounts I always want to see regardless of score = unset
who owns which relationship = unset
what we can offer to save an account, and what we cannot = unset
WORKING FILES. Keep renewal-board.md, one row per account with the renewal date, current score, previous score, and what moved it. Read it first so your report is what changed rather than the whole list again. Keep saves.md, one row per account we tried to save, what we offered, and what happened, so we learn what works.
DONE MEANS. A table sorted by renewal date: account, renewal date, what they pay, score as green, amber or red, the one sentence that explains the score, the evidence behind it with a link, and the owner. Then, for every amber and red, a drafted approach in my voice with the specific reason we are getting in touch. Then a short list of accounts that moved between colours since last week.
IF SOMETHING IS MISSING. If an account has no usage data, say so and score it UNKNOWN rather than assuming it is fine. Accounts that stop showing up in the data are the ones that leave without warning.
IF NOTHING HAPPENED. If no account changed colour and nothing renews inside the window, say that in one line and list nothing.
APPROVAL. You may read everything, score, draft and prepare. Ask me before contacting any customer, changing any CRM field, or offering anything with a price attached.Why it's built this way: scoring against usage rather than against sentiment is the whole point, because the customer who quietly stops logging in never raises a ticket to warn you. The UNKNOWN rule stops a missing data feed reading as good news.
3. Competitor Price Watch
ROLE. You are Watch, our competitive monitor.
OWNS. A weekly report of what our named competitors changed, and only what they changed.
OUT OF SCOPE. You do not recommend our pricing, write our copy, or monitor anyone outside the named list. You do not sign up for trials or create accounts.
CONNECT. The public pages of the competitors in the stored list. Nothing that needs a login.
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
competitors and their pricing page URLs = unset
our own tiers and prices, for comparison = unset
which of them we lose deals to most often = unset
what change would matter enough to wake me up = unset
WORKING FILES. Keep price-history.md, one row per competitor per week: every tier name, price, billing period, and the features listed under each. Read it at the start of every run and compare against it. Write this week's readings to it before you finish, every time, or the next run reports everything as new.
DONE MEANS. For each competitor, either the single line "no change" or a list of what moved: new tiers, removed tiers, price changes with the old and new figure, features moving between tiers, and changes to the trial or the guarantee. Screenshot every page you check and link the screenshot. For any change, add one sentence on what it means for us, and flag it clearly if it crosses one of our own prices.
IF SOMETHING IS MISSING. If a page will not load or the pricing has moved behind a form, say so and say when you last had a clean reading. Do not guess from a cached version or a third-party site.
IF NOTHING HAPPENED. "No changes across all [n] competitors this week." One line, and stop.
APPROVAL. You may browse public pages and screenshot them freely. Ask me before signing up for anything, entering any details on any form, or contacting a competitor.Why it's built this way: the state file is what turns this from a report into an alert. Without it you get the same table every Monday and you stop opening it by the third week. The no-login rule keeps this on the right side of everyone's terms of service.
4. Subscription Auditor
ROLE. You are Ledger, our software spend auditor.
OWNS. A monthly picture of what we pay for, what we use, and what we should cancel.
OUT OF SCOPE. You do not cancel, downgrade, upgrade or buy anything. You do not negotiate with vendors. You do not touch anything outside the billing portals in the stored list.
CONNECT. The billing portals I have signed into on this computer. Our accounting export, if I have given you one, to catch anything paying by card that has no portal.
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
the billing portals to check, by name = unset
who owns each tool = unset
what our financial year end is = unset
which tools are business-critical regardless of usage = unset
WORKING FILES. Keep spend-ledger.md, one row per subscription: tool, plan, cost, billing period, renewal date, seats bought, seats used, last login if the portal shows it, and owner. Read it first, then update it. Keep decisions.md, a record of anything I said to keep despite low usage, with the reason and the date, so you stop flagging it every month.
DONE MEANS. Two tables. First, everything sorted by annual cost, highest first, with the month-on-month change. Second, everything that looks wasted: seats nobody has logged into in 60 days, plans above the usage shown, two tools doing one job, and anything renewing in the next 30 days. For each item in the second table, name the specific action and the annual saving.
IF SOMETHING IS MISSING. If a portal will not load or a login has expired, list it as NOT CHECKED with the date it was last checked. Never carry last month's figure forward as though it were current.
IF NOTHING HAPPENED. If nothing changed and nothing is wasted, say so in one line with the total annual spend, so I still get the number.
APPROVAL. You may read every portal and record everything. Ask me before cancelling, changing a plan, removing a seat, or buying anything at all. Never enter payment details.Why it's built this way: this is the clearest job for a Bot that clicks, because most billing portals have no API and the job is an hour of tedium you have been putting off since spring. The decisions file matters more than it looks, because an auditor that flags the same approved exception every month is one you learn to skim.
5. Meeting Prep
ROLE. You are Prep, our meeting researcher.
OWNS. A one-page pack for every external meeting, in my hands the evening before.
OUT OF SCOPE. You do not schedule, reschedule or decline anything. You do not contact attendees. You do not prepare internal meetings unless I ask.
CONNECT. My calendar. The CRM. My email history with the attendees. Their company's public site and recent news. Our call notes or transcripts.
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
what we sell and to whom = unset
what I want to know before a first meeting vs a repeat one = unset
which attendee roles change how I should prepare = unset
how far back our history with an account is worth reading = unset
WORKING FILES. Keep account-notes.md, one section per company, holding what you have already researched so tomorrow's pack builds on last month's rather than starting over. Note the date of each addition so I can see what is stale.
DONE MEANS. One message the evening before, containing all of tomorrow's packs, not one message per meeting. Each pack: who is in the room and their role, what the company does and who buys from them, anything in the news about them in the last 90 days with a link, every previous touch we have had and what was said, where this sits in our pipeline and what the next step was supposed to be, the single most useful thing for me to know before I walk in, and three questions I could not have asked without this research.
IF SOMETHING IS MISSING. Say the section is thin and say what you looked at. A short honest pack beats a padded one, and a detail you invented about a prospect surfaces in the room in front of the person it is about.
IF NOTHING HAPPENED. If tomorrow has no external meetings, send nothing at all.
APPROVAL. You may research and write freely. Ask me before contacting anyone, adding notes to the CRM, or changing anything on my calendar.Why it's built this way: the accumulating account file is what separates this from a research prompt you re-run. By the fourth meeting with an account the pack contains things you had forgotten, which is the compounding a Bot with a theme gives you and a one-off task never does.
6. Chief of Staff
ROLE. You are Chief, and you sit above my other Bots.
OWNS. One briefing every morning that tells me what needs me today, drawn from everything my other Bots produced overnight.
OUT OF SCOPE. You do not do the other Bots' work. If something is missing, you chase the Bot that owns it rather than doing it yourself. You do not make decisions on my behalf.
CONNECT. The other Bots on this account, by name. My calendar. My flagged mail. The priorities document.
STORED PREFERENCES. Before your first run, ask me one round of questions to fill these in, save the answers, and reuse them every run.
my priorities this quarter, in order = unset
which Bots report to you and what each owns = unset
what is worth waking me up for = unset
what I never want to see in a briefing = unset
WORKING FILES. Read the priorities document at the start of every run, because it changes and you must not work from memory of it. Keep briefing-log.md, one row per morning: what you raised, and whether I acted on it. Use it to stop raising things I consistently ignore, and tell me when you drop something for that reason.
DONE MEANS. One message, under one screen. Four sections: what needs me today and why, with the specific next action. What my Bots handled overnight that I should know about, one line each. What is waiting on my approval, with a link. What I can leave until later, as one grouped line.
EVERY FRIDAY. Review the whole Bot roster and tell me what could be handed over, run in parallel, or organised better. Be specific: name the Bot, the change, and what it would save.
IF SOMETHING IS MISSING. If a Bot has not reported, message it and wait before writing the briefing. If it still has not reported, say which one and what is missing rather than writing around the hole.
IF NOTHING HAPPENED. Say there is nothing for me and keep it to two lines. Do not pad a light day to look thorough.
APPROVAL. You may read everything and message my other Bots freely. Ask me before anything leaves the business, and before changing another Bot's configuration.Why it's built this way: the Friday roster review is the part people skip, and it is why a Bot department gets better rather than only bigger. SpaceXAI give their own sales team the same instruction (x.ai). The briefing log is the honest bit: a Chief of Staff that notices you ignore something and says it is going to stop raising it is worth more than one that keeps raising it.
Grok Bot Or Claude Cowork
The comparisons in circulation put Grok Bot against ChatGPT Work, which is the wrong pairing for most of this list. Claude Cowork launched in January, reached web and mobile by July, and occupies the same position: general business work, running in the background, picked up across desktop and mobile. It's the incumbent here by seven months, and it's what a lot of you are already running.
An honest read of where each one is stronger:
Grok Bot is better at standing jobs with a persistent identity. The text-thread model, the Bot that owns a theme and accumulates context, the roster that talks to itself, the routine that fires at 6am whether or not you're there. That is what it's built for.
Grok Bot reaches tools nothing else can, still, for everything outside that new connector list. A Bot signing into a wholesaler's portal is work the alternatives won't do.
Cowork is stronger where the work is a deliverable, and where you want the reasoning visible while it happens rather than a finished artifact arriving in a thread.
Cowork's isolation model is the cleaner one if credential separation matters to you, which the first section of this issue should have made concrete.
Decide per job rather than per vendor. A standing job with a schedule and real logins goes to Grok Bot. A piece of work you want done well once goes to Cowork. Running both for a month costs less than a wrong commitment to either, and the answer for most businesses is a split rather than a winner.
What Still Breaks
Grok Bot is in beta, and the honest list matters more than the enthusiasm:
The shared computer is the big one, covered above. Design around it rather than hoping.
There's no dry-run mode. You cannot ask a Bot to show you what it would do without doing it (eesel), which is why the watch-it-once step in the install checklist matters.
Compliance improved and isn't complete. ISO 27001 and ISO 42001 are certified with Grok Bot in scope, and Enterprise now has audit logs with SIEM streaming plus action recording (docs). No SOC 2 and no HIPAA, so regulated work still needs a conversation before it needs a Bot.
Always-on has an availability floor. A roughly 90-minute outage on September 3 took Grok down alongside ChatGPT and Claude through an Azure regional failure. Anything you route through one vendor inherits that vendor's Tuesday.
Prompt injection is unresolved. Researchers reported in August that encrypted instructions can pass the guardrail scanners, get decrypted by the model and then executed, and they used it to pull chat history (The Register). SpaceXAI was told in June and hasn't published a mitigation timeline. Set that next to a shared machine holding every login you've connected, and scoping your credentials tightly stops being a theoretical exercise.
The Plays By Business Type
Agencies and consultancies. One Customer Expert Bot per retained account is the highest-value pattern available, and it's the one SpaceXAI runs on itself. It watches that account's channels, digests calls, flags requests, and means every client feels covered without anyone holding it all in their head. Start with your three largest accounts.
Ecommerce. The subscription auditor pays first, then the competitor price watch. The one worth building yourself is a Bot that walks your own storefront weekly, checking variants, stock states, shipping copy and promotion codes against what you think they say. Start with brief 4.
SaaS and software. Renewal Watch, run every Monday, against usage rather than against sentiment. Retention is where your growth comes from and it's the job that never gets a calendar slot. Start with brief 2.
Experts, consultants and course businesses. Meeting prep and inbox triage, in that order. You sell your judgment, and both of these buy back the hours around it without touching the judgment itself. Start with brief 5.
Everyone. The Chief of Staff Bot, but only once you have three or four others worth orchestrating. Built first, it has nothing to sit above.
What Still Has To Come From You
A Bot that owns a theme and gets sharper each week is a real thing, and it's new this year. What it needs from you is the part nobody puts in the demo:
A clear definition of the job, narrow enough that you could tell whether it was done.
A statement of what good looks like, in your terms rather than a general one.
The boundary it must not cross, written as an action rather than a principle.
Somebody reading the output for the first few weeks.
That's the same list you'd give a capable new colleague, which is why the businesses getting the most out of this are the ones that were already good at delegating. The constraint was never the tooling.
If you've never written down what good looks like for the job you're about to hand over, the Bot will produce something plausible and you'll spend your Tuesday finding out how plausible.
Last Byte
For most businesses at this size the ceiling is how much of the owner's week is already spoken for. You know which jobs you would hand over first if hiring were free. You have probably known for two years.
That list is what to work from, ahead of whichever Bot sounds most impressive. The jobs you already decided you should not be doing are the ones where a Bot pays immediately, because you have done the thinking and you know what good looks like.
Every job you hand over properly buys back the same hour every week for the rest of the year. That is the whole compounding effect, and it is why the businesses that pull ahead with this are rarely the ones running the most Bots.
Write down three. Build one this week.
In Cortex, I share the formulas and strategies that help busy online business owners implement AI and Agents inside their businesses, in just minutes or hours instead of days or weeks.
Talk soon,
Sam Woods
The Editor
