Bringing AI to your business
My own business runs on AI agents. Below is what they actually do, area by area, with links to the skills they use.
If you want the same in your business, I go through these areas with you, find what is worth handing to AI first, and help set it up. The first conversation and the written plan are free.
What the agents do, by area

Customer support
- Moved support off Zendesk to a help desk on my own server, with an AI assistant built in. No support staff.
- Finds tickets waiting on a reply, drafts the answer, updates status and tags
- Answers on WhatsApp, Telegram, Signal, SMS and more from one place, each reply approved first
- Reads its own SMS verification codes instead of asking someone to check a phone

Development
- 1,036 commits last week, mostly written by agents
- Nothing lands until its tests pass and a separate reviewer agent approves it
- Agents try to break web apps in a real browser before a release
- Pull requests opened, watched through review and confirmed deployed
- Closed apps with no API turned into tested commands by reverse-engineering their traffic
- A bad deploy of this site rolls back in about a second

Design
- Web pages explored, critiqued and polished against a quality bar, with contrast and layout checked automatically
- Social graphics, logos and app mockups; any image with text must pass a readability check
- Slide decks, every changed slide checked before moving on
- Words inside a screenshot changed so the edit is invisible

Video
- Launch and promo videos built from a product’s own site, colours and assets
- Narrated explainer videos made straight from a code change
- Motion graphics: animated type, counting stats, charts and logo stings
- Captions added to talking-head footage, on my own machine

Marketing and content
- Blog posts taken from outline to illustrated draft to published
- Every post turned into drafts for X, LinkedIn, Reddit and the right communities; nothing goes out unapproved
- Facebook, Instagram, Google and Reddit ads; new campaigns start paused with a spending cap
- Checks whether AI assistants can find and quote a product
- Landing pages built after exploring four to six design directions

Sales and outreach
- Watches Reddit, X, Hacker News, LinkedIn and more for relevant conversations and drafts a reply
- A daily LinkedIn run: find prospects, write a personal note for each, track who replies
- Every outgoing message drafted in my voice, tested against about 19,000 of my real ones
- Real phone calls by an AI voice agent: phone menus, hold music, and I can take over mid-call

Finance and accounting
- Went through my books line by line and replaced $5,509 a year of software (breakdown)
- Running costs down 61%, from $71k a half-year to $27.5k
- Bank transactions categorized in my accounting software
- Revenue and failed-payment reports from Stripe, read-only by design
- Payments are previews until I approve them, under a daily cap
- Every subscription logged with its renewal and cancel-by date

Email and admin
- Inbox taken from 424 messages to 169, with written rules on what it may never delete or send
- Calendar events created and on my phone instantly
- Flights searched and filled in up to the payment step, with my seat and bag preferences
- Logs in to sites with saved passwords and 2FA codes, and gets past CAPTCHAs

Servers and operations
- Alerts go to an agent that starts investigating instead of buzzing my phone
- Every running service compared against what should be running, so one that died quietly shows up
- Deploys, settings and logs on my self-hosted platform
- Domains, DNS and email forwarding managed; nothing bought without confirmation
- Reminders and follow-ups that outlive a session wake an agent when the email or reply arrives

Research
- Deep research across GitHub, X, Reddit, Hacker News and blogs, with sources
- Data pulled from sites with no API; whole sites turned into readable documents
- Business ideas researched for market, competitors and pricing before anything is built
- Product names screened for domain, trademark and search conflicts

Safety and access
- 163 guardrails stop an agent before it does something wrong
- Anything that spends money or messages a person waits for my yes
- Passwords and 2FA codes from 1Password, never written into files or commands
- Undoable actions run without asking; one-way actions stay gated

Memory and reporting
- A long-term memory every agent can read, and each new fact is confirmed saved
- The history of any past project rebuilt from sessions, cards, git and chats
- An end-of-day report from eight sources that separates the agents’ work from mine
That is a sample. The agents follow 276 skills in all, every one on the skills map; the hardware and models are on my setup page.
How it works
1.Map how work moves
You tell me which tools you use and which jobs come back every week. On a 30-minute call I map who does what and how often.
2.Sort every recurring job
Each job goes in one of three places: an agent can do it, a cheaper or open-source tool can replace it, or it stays with a person.
3.Get a written list
In the order I would do them, with a rough estimate of the time or money each saves, and what I have already built for the same job.
Get in touch
A few lines is enough: what your team does, the tools you pay for, and the jobs that take up the week.
