What vision and audio AI can do for a small business
We build two kinds of work: machines that look (from photos or a camera) and machines that listen (from recordings or live speech). These are realistic jobs for a small shop, warehouse or office:
- Look: recognise products on a shelf or counter from photos or a camera.
- Look: count goods, boxes, pallets or items in a picture.
- Look: check products for defects and quality.
- Look: read text from documents, invoices and vehicle plates.
- Listen: turn speech into text from recordings or live talk.
- Listen: voice commands for an app or a device.
- Listen: transcribe customer-service calls and meetings so you can search them.
- Listen: detect certain sounds, for example an alarm or an unusual machine noise.
AI is never perfect. Results depend on your real photos and recordings, so we test on your samples first and only then say what we can commit to. See vision AI and audio AI.
Our starting prices
"From" is the lowest figure. The final price is fixed in writing after a 30-minute call and depends on the factors in the next section.
| Type | What it does | Price | Tested before we fix it |
|---|---|---|---|
| Vision AI | Recognise products, count stock, spot defects, read documents and plates. | from $2,000 | Tested on your own photos. |
| Audio AI | Speech to text, voice commands, call and meeting transcription, sound detection. | from $800 | Tested on your own recordings. |
Both come with code and accounts in your name and 30 days of free bug fixes after launch.
What changes the price of AI software
For AI software, price depends on data more than on the number of screens or pages. The table lists the factors. It has no numbers because nobody can give an honest figure before seeing your data.
| Factor | Price goes up when | Ask |
|---|---|---|
| Data | Real photos or recordings are few, too similar or not yet labelled. | Do you already have real samples, including hard ones? |
| Accuracy you need | You need very few mistakes. Each step up in accuracy usually takes more data and effort. | What does a mistake cost: which is worse, a miss or a false alarm? |
| Real conditions | Poor light, changing camera angles, background noise, regional accents. | Where will it run and in what conditions? |
| Where it runs | It must run on the device itself, with no internet. | Is there a stable connection where it will be used? |
| Number of things to recognise | Many products, many voices, many kinds of sound. | Does the list of things to recognise change often? |
| Speed | Results must be almost instant instead of processed later. | How soon do you need the result? |
| Connections and privacy | A link to software you already use, or a rule that data must not leave your company. | Where are the data stored and processed? |
AI on the device or on a server
One big choice is where the AI runs. Each way has trade-offs, and the choice affects both the build price and the cost afterwards.
| Point | On the device (phone, camera, local machine) | On a server |
|---|---|---|
| Internet | Works without a connection. | Needs a stable connection. |
| Data | Photos and sound can stay on the device. | Photos and sound are sent to a server to be processed. |
| Power | Limited by the device. | Can run heavier models. |
| Updates | Each device must be updated. | Updated in one place. |
| Cost afterwards | Little server cost, but the device must be strong enough. | A server cost that usually grows with use. |
Depending on the job, we state where the AI runs and what that costs in the quote.
Accuracy: why we test on your real samples first
No system is perfectly accurate. An accuracy figure is only true for the data of whoever measured it, not necessarily for your photos or recordings. With speech, regional accents and background noise also change the result. So we work in this order:
- A 30-minute call: you tell us the job you want the machine to do.
- You send a few real photos or recordings, including hard ones.
- We test on those samples.
- We tell you plainly what is achievable, and only then what we can commit to.
- The price and scope are fixed in writing before we start building.
What to prepare before the first call
It does not need to be perfect. With the items below, the free 30-minute call is enough to tell whether your job suits AI. If the AI will sit inside an app, see also what a mobile app costs.
- One sentence describing the job you want the machine to do.
- A few real photos or recordings, including hard ones such as poor light or noise.
- Who does this job today, and how.
- Which mistake costs more: a miss or a false alarm.
- Where your data is kept and whether there is a security rule or NDA.
- The equipment you have: cameras, phones, computers or a server.
- The software the AI result needs to go into.
- The date you want it working.