Ask it about your own jobs.
Ray reads the records you already keep — projects, materials, crews, receivables — and answers in things you can click. And it is careful about what it will not do.
Ray
on this workspaceNo report to build and no filter to set. The question is the interface.
An illustration of the assistant, not a recording. What Ray answers depends on your workspace's own data.
Six things, each one a real flow.
Not a list of what AI could be used for. This is what is built, what it reads, and where it stops.
Ray, the assistant
Ask about your own workspace in plain words. Ray reads projects, materials, crews, receivables and quotations through named tools, and can draft a follow-up or open a task when you ask it to.
Rooftop trace
A vision model outlines the roof faces and the obstructions on the satellite tile the Design Studio has already fetched, so you start from a traced roof instead of a blank one. It is never asked for heights — a top-down photograph has no elevation.
Design review
A second read of a finished design before it goes to a customer — what looks thin, what was assumed, what a reviewer would ask about.
Bill scanner
Photograph an electricity bill or a vendor bill and the fields come out as fields — units, amounts, dates, line items. You check them before anything saves.
Quote audit
Reads a quotation your customer received from somebody else and scores it out of 100 against the marketplace price book and the document itself. Where there is no benchmark it says so rather than guessing.
Task categories
Files a task against the categories your team already uses. It is told to reuse an existing one first, and a bulk run may not invent any at all.
Answers the easy ones. Hands over the rest.
An enquiry that arrives at 11 pm gets an answer from what you have told it about your business — and a question it cannot answer becomes a person's job, not a dead end.
It answers from your knowledge base
A document you wrote, a product catalogue you maintain. That content is the ceiling of what it can say — it does not read the open internet and it does not improvise policy.
And it stops there
A price is not something a bot may improvise, so it does not. The conversation is flagged for a human, the right people are notified, and the customer is told a person is coming — rather than being given a number nobody will honour.
One counter, weighted by what the call really costs.
An allowance that counted requests would let one expensive call cost the same as one cheap one. So a call costs what it costs — and the number on your billing page means the same thing every month.
An everyday call
1from your allowance
A chat reply, a task category, a drafted follow-up. The fast model, and the one nearly everything runs on.
A premium-model call
10from your allowance
A rooftop trace or a design review on the strongest model available. It costs 10× because it costs us about 10× — the weight comes from the published price per token, not from a pricing meeting.
The month's allowance goes first
Purchased calls are only touched once the included allowance is gone — spending what you paid for while the free part expires would be charging you twice.
Top-ups carry over
The monthly allowance resets; anything you bought does not. A top-up never quietly raises the recurring cap either.
You choose the model
Set a default for the workspace, or pick one for a single feature, with the cost of each shown beside it. Running out downgrades the call and says which model answered — it never fails.
Free
—
Starter
—
Professional
—
Growth
—
Counted in everyday-call equivalents, not requests — so 10 of them is one premium call. Read live from the same limits the product enforces.
What the AI is not allowed to do.
The useful thing to know about a tool like this is where it stops. These are rules in the code, not intentions.
It never invents a price
Not in a chat reply, not in a quotation draft, not in a quote audit. A rate comes from a product in your own catalogue that the line actually matched, or the line comes back blank for you to fill. A plausible guess is far more dangerous than an obvious gap — the gap gets filled in, the guess gets sent.
It does not compute what the engine computes
Yield, string sizing, wind load, steel takeoff, subsidy slabs, GST splits and every balance on this site are arithmetic, done by code you can test. The model reads those results and explains them; it is never asked to produce them.
It cannot fill your workspace with categories
A bulk auto-tagging run may only reuse categories that already exist, is capped per run, and asks before it spends the quota. Left unattended, a model inventing near-synonyms turns a task board into one column per task.
It reads only what you can read
Ray runs inside your permissions, on your workspace. It is not a way around who can see which records, and your data is not used to train shared models or exposed to another Solset customer.
It does not send on your behalf
Drafts are drafts. Customer-facing messages are reviewed by a person before they go — except the WhatsApp bot answering from knowledge you wrote, which hands over the moment it is out of its depth.
There is no lead scoring or forecasting
Both are easy to build badly and we have not built them. You will not find a number on a lead that claims to predict whether it will close.
A whole workspace can switch it off. AI is per-workspace and can be turned off entirely by an admin. Nothing in the product stops working when it is — the assistant simply is not there.
Two of these are public. No account needed.
The most useful thing we can offer instead of a testimonial is a tool you can point at your own document right now.
What teams ask before switching the AI on
The rest of the record
Put one real job on a record and ask about it.
The free plan includes AI. Add one enquiry, design a roof, and ask Ray what it would cost to build. If the answer is not useful, you have lost an afternoon.