WhatsApp bot to collect photos from staff and score them
A bot can collect photos from a sales team on WhatsApp and score each one before it reaches you. Here is how that works, and what I have and have not built.
A WhatsApp bot collects photos from staff by giving each person one small job per message, saving every image against a record, and scoring it with fixed rules before any AI opinion is involved. The staff keep using WhatsApp the way they already do. You get a folder of photos that have already been checked, instead of a chat you have to dig through.
One honest line first. The bot I run for Project One Motors goes the other way: a salesman asks for a car and it sends the photos out. I have not shipped a bot that collects and scores photos from staff. What follows is how I would build that version, using the same rules that keep the outgoing bot from lying. I say which is which as I go.
What problem does a photo-collecting bot solve?
Anywhere a team is supposed to send you photos, the same things go wrong. The picture arrives without saying which car, unit or job it belongs to. It is blurry, or shot into a window, or the same bumper five times. Nobody tells the sender until you ask for a retake two days later, when the car has gone.
A bot fixes the timing. The check happens while the person is still standing next to the thing, so the retake costs ten seconds instead of a callback.
How does the collection flow work?
The whole design is a short conversation with one job per message:
- The staff member sends a reference first: a chassis number, a stock code, a job number. The bot confirms it against your real list and says what it found.
- The bot asks for one shot at a time from a fixed list, for example front three-quarter, rear three-quarter, driver side, interior, odometer.
- Each photo is saved against that reference and that slot, with the sender and the time.
- When the list is complete the bot says so and closes the record. If a slot is missing it names the slot.
The bot only listens to a whitelist of staff numbers, the same way mine ignores everyone who is not on the sales team. Anything from an unknown number is dropped.
How do you score a photo without a human looking at it?
In two layers, and the order matters.
The first layer is plain code and it does most of the work. It can measure things that do not need judgement: the resolution, how sharp the image is, how dark or blown out it is, whether it is a near-duplicate of a photo already in the record. Each check has a threshold you set, and a photo that fails gets an instant reply such as "too dark, try again facing the window". These checks are deterministic. The same photo scores the same way every time, and you can explain every rejection.
The second layer is a vision model, and it is only allowed to answer narrow questions about what is in the frame: is this a rear three-quarter, is the odometer readable, is a person standing in the shot. It classifies. It does not decide what car this is, because the reference number already told the bot that.
The score itself can be as simple as a checklist: slots filled, checks passed, retakes needed. A single number out of 100 feels precise and hides which part failed. I would report the failed checks, not a grade.
What would it cost?
There is no published price for this exact bot, because I have not built it and I do not quote what I have not built. The closest published reference is the AI automation Build package: AED 7,500 setup plus AED 1,250 a month for one workflow on one channel, with a stated cap on changes. A collect-and-score bot would be scoped as one workflow of that kind, after a call about what you need scored. Software, API and WhatsApp provider charges are billed separately.
Which WhatsApp route should a business use?
An unofficial WhatsApp library is the fast way to build, but a bot on one and a bot on the official Cloud API carry different risk, and if replies cannot afford to go down that is the first conversation to have. The route Meta officially supports for businesses is the WhatsApp Business Platform, and it has rules that shape a photo bot: a business can reply freely for 24 hours after a user's last message, and outside that window it can only send pre-approved templates. For a bot that mostly answers staff who message first, that window suits the job.
Platform rules change and I would re-check Meta's documentation at the start of any build. What I would not do is pick the route before I know how many people will message the bot and whether it must message them first.
What is the risk with this kind of bot?
- A score that staff learn to game. If the rule is "sharp enough", people will send sharp photos of the wrong thing. That is why the reference number and the slot list come first.
- Privacy. Staff photos can include customers, plates and locations. Decide who can see the folder before you build the bot.
- A bot that is confidently wrong. It should say "I could not read that" rather than pass a photo it is unsure about.
Frequently asked questions
Can a WhatsApp bot collect photos from many staff at once? Yes. Each conversation is separate and every photo is saved against a reference, so ten people sending at the same time do not mix records.
Does the bot need AI to score photos? Not for the first layer. Sharpness, brightness, resolution and duplicates are ordinary code. A vision model is only needed for questions about what is in the picture.
Where do the photos end up? Wherever you choose: a folder, a cloud bucket or your own system, keyed by the reference number. That is a build decision, not a limit of WhatsApp.