Words, pixels and sound waves are now one vocabulary. A practical map of multimodal models, what changed in the last year, and where a small business should start.
A few years ago, AI that wrote and AI that saw were different products with different vendors. That separation is gone. The current generation of foundation models reads text, looks at images, listens to audio, watches video, and can produce any of the four on request. People call this multimodal, and the practical effect is that more of your raw material, receipts, memos, voicemails, photos of a broken part, can go into a machine and come out as structure.
What each modality is actually for
Text is still the backbone: drafting, summarising, code, conversation. If your AI usage is only text, you are using a quarter of the toolbox.
Image work splits into two halves, and the underrated one is reading. Generating a marketing picture is fun; photographing a delivery note and having it arrive as a line in your spreadsheet is what pays. A phone photo of the worn belt on the shop machine, filed as a maintenance ticket before you walk back to the office, saves the walk.
Audio moved fastest in the last twelve months. Transcription that handles an accent and a bad phone line, meetings summarised while you are still in the room, voice interfaces that answer a customer at 6 p.m. when nobody is left to pick up. For anyone whose business runs on phone calls, this is the highest-leverage modality per dollar.
Video is the hungriest for compute and the least mature for generation, but understanding has quietly arrived: describe what happened in a clip, find the moment someone hands over a package, flag the safety lapse on a jobsite camera. Generation is fine for rough social cuts and still unreliable for anything where hands or on-screen text have to look right.
Why the release cycle matters to a non-engineer
The labs now compete on combinations: can the model watch a screen recording and name the mistake in it, read a photographed spreadsheet and write the formula, take a voice command plus a photo and act on both at once. Each jump makes something you tried six months ago work that didn't before. Here is a cheap habit worth more than any newsletter: keep a list of the three things that broke when you last tested AI for your work, and re-run them against a new model when one lands. That is how people quietly got good at this while everyone else was arguing about which chatbot is smarter.
Where a very small business should start
Pick the pile. Every owner has one: the stack of photos of paperwork, the voicemail backlog, the invoices that get retyped by hand. Multimodal pays off first there, because the input is already media and the output is already data, and nobody has to learn a new process. A shop that photographs each job's completion sheet, has the model pull out customer, date, materials and dollars, and drops them into a running tab, has built itself an automated books feed without deciding to learn bookkeeping.
Two cautions from real deployments. Extraction can be confidently wrong, so a human glance stays on anything that touches money. And a model that can see is a model that can be fooled: instructions hidden inside a photographed document are a live attack style, which is why our lessons keep a person between any AI that reads the outside world and anything it is allowed to touch.
Kit for this lesson
If your day runs on calls, the proper headset is the unglamorous first purchase. A noise-canceling mic makes every transcript you feed the model cleaner, which is most of the quality battle.
View on Amazon
The laptop members most often bring to the Labs. 24GB of unified memory runs mid-size local models comfortably, and it does it in silence in front of a client.
View on Amazon
If you would rather not carry your AI box in a bag: small, quiet, and about $600 of Ryzen AI will run a 7-to-8-billion-parameter model locally, no cloud required.
View on AmazonAs an Amazon Associate, AI Universe City earns from qualifying purchases. Prices link to Amazon and may change; the picks above are what we actually use in the Labs.