What actually gets automated
Never "the business". One job at a time, and it's usually one you can name in a sentence: the same five questions arriving on WhatsApp after you've closed, the review nobody remembers to ask for, the invoice that's three weeks late and hasn't been chased, the numbers somebody rebuilds by hand every Monday.
Each of those is a small system rather than a platform. That's why they get built one at a time, and why the first one should be whichever annoys you most.
- A message arrives at 11pm - it reads what they're actually asking, replies with your prices in your voice, and books the callback in your calendar.
- Someone DMs Instagram asking what you charge - it checks which post they came from, sends the list, and tags them as a lead before you wake up.
- A job is marked complete - it waits two days, asks for the Google review, drafts your reply the moment one lands, and puts the unhappy ones at the top of your list.
- An invoice passes its due date - it sends the polite reminder, waits a week and sends one more, then tells you who's still ignoring you.
- You drop a topic into a sheet - it researches it, checks the sources, drafts the article in your tone, and waits for your yes before anything publishes.
All five of those are buildable, and buildable is a smaller claim than shipped. I mean the smaller one. The ones that have actually shipped are further down this page and they have numbers attached.
The first of those five, drawn. A message at 11pm, and three steps that happen without you.
A WhatsApp message arrives at 11pm
AI reads what they're actually asking
Replies with your prices, in your voice
Books the callback in your calendar
One person, not an agency, and it cuts both ways
Search this and you'll get nine agencies and a listicle ranking them. So here's the honest version of what you're choosing between.
- No account manager translating what you said into a brief.
- No bench of juniors doing the actual build.
- No handover at the end to a support team who weren't there.
The person who scopes it writes it, and answers when you email. The trade is real too: one person has one calendar, can't start four projects in a week, and isn't a 24-hour support desk.
If what you need is volume or round-the-clock cover, an agency is the right answer and I'll say so.
Where the model stops and the code starts
This is the whole difference between something you can point at customers and something you can only demo. A language model is genuinely good at reading a badly typed message and working out what somebody wants. It is not a place to keep what you sell, what it costs, or what you'd never say to a stranger.
So the reading happens in the model, and everything that has to be true happens in code around it: the reply gets drafted, then checked against your own system before it sends. A reply that trips a rule doesn't go out, and a person gets that conversation instead.
If the answer to "what stops it saying something wrong" is that the prompt asks it nicely, that's a hope rather than a control.
Two of these have been running on a Dubai showroom's numbers
Project One Motors sells supercars out of Al Quoz, and two of my builds sit on its WhatsApp.
The first one reads a salesman's request in whatever shape it arrives - "the white GT, low km one" - matches it against live inventory in code, and sends that exact car's photos. It handles ten to twenty of those a day and the marketing team, which used to be the photo index, now answers none of them.
The second talks to customers on the showroom's own number, at 2am, and answers roughly one message in five in Arabic or Russian. In its first four weeks it handled 55 conversations and 534 messages, and every enquiry has a record now instead of living in whoever's phone happened to be awake.
Both case studies are linked at the foot of this page, including the parts that went wrong.
The one that survives is the one nobody has to open
Most automation dies of the same thing. It needs somebody to log into it - a dashboard, a portal, a tool with a password nobody remembers - and a month later the team is quietly doing the job by hand again and nobody mentions it.
So what I build reports into whatever your people already have open. At the showroom a handover lands in the salesperson's own WhatsApp, from the showroom's number, and they answer it the way they'd answer a colleague. Nobody installed anything or learned anything. That's most of the reason it's still running.
Where it plugs in
- WhatsApp - where most UAE enquiries actually start, on your own number rather than a new one nobody has.
- Instagram DMs - the same conversation arriving through a different door.
- Gmail - the inbox holding the quotes, the PDFs and the things you meant to reply to.
- Google Sheets - usually the closest thing a small business has to a database, and that's fine.
- Google reviews - asking for them once a job is done, and drafting the reply when one lands.
The other end matters more than the channel. Something has to know what's true - your stock, your price list, your calendar, your customer records - because an assistant answering from nothing will answer confidently from nothing.
If that truth currently lives in one person's head or across three spreadsheets that disagree, then fixing that is the first job and the automation is the second. I'd rather say that before we start than three weeks in.
The first version is supposed to be boring
Nothing I build gets pointed at your customers on day one. It runs on a small allowlist first, supervised, because real people do things no specification predicts. The showroom assistant:
- Replied in the wrong language to broken English.
- Mirrored "habibi" back at people.
- Offered photos nobody asked for.
- Answered a customer who'd just said "ok" with "noted, I'll follow up".
Each of those became a rule with a test behind it. Building it is the short part; the weeks after it goes live are the actual work, and that's the part an agency demo never shows you.