Plain English in, a real report out — with the working shown. It reads and only reads, it obeys each person’s own permissions, and it runs on your Anthropic key rather than a per-seat charge from us.
SimplCRM’s assistant answers questions about your own workspace in plain English. It has three tools and all three only read: it can list what can be reported on, run a report, and find records. It cannot create, update, delete or send anything. Reports run under the asking person’s permissions, the tool calls behind each answer are shown, and the workspace supplies its own Anthropic API key — so questions are billed to and processed under your account, with a monthly token cap you set.
The first four sit on the empty screen as buttons, because the hardest question to type is the first one. It will take any of them.
It groups open deals by stage, totals the value, and shows the table. The figures come from the same reporting engine the Reports screen uses, so they agree with what you would get by building the report yourself.
A filter on last activity, listed as records rather than summarised — the kind of question that is obvious to ask and tedious to build, which is most of what an assistant is for.
Money grouped and totalled to the paisa, in Indian digit grouping. The answer is a sentence and a table, not three paragraphs of preamble.
Not a data question at all. The assistant carries notes on how the product works, so the software can answer questions about itself instead of sending somebody to a help site.
Leads grouped by source. The kind of question a marketing spend is argued about with, and the kind nobody builds a report for because the argument is already over by then.
The service desk keeps a first-response and resolution clock per case. Breaches are a measure the reporting layer already carries, so this is one question rather than a screen to go and read.
Every tool the assistant has is read-only. Asked to update a deal or send an email, it says what it cannot do and points at the screen where you can. This is a deliberate ceiling rather than a stage we have not reached yet: an assistant with write access to your invoices and your payroll is a different product, and the day it misreads a request there is nothing to undo it with. Reading is where the value is anyway — the questions people cannot be bothered to build a report for.
The assistant does not write SQL and never sees your schema. It works through the same typed reporting layer the Reports screen uses, where objects, fields and measures are keys into a map the product declares. A name that is not in that map is an error rather than an escape problem, which is what makes handing the shape to a language model reasonable in the first place.
Create a key at console.anthropic.com, paste it into Settings, choose a model and set a monthly token cap. Anthropic bills you for what you ask, at their prices, with no markup and no per-seat AI charge from us. Your key is encrypted with your workspace’s own key, exactly as the mail passwords and bank details are. Turn the assistant off and nothing is sent anywhere.
It is told never to answer a numeric question from memory — it has not seen your workspace before, and a plausible number nobody checked is how an assistant over business data becomes worse than no assistant. When a report comes back with nothing in it, the answer is that nothing matches.
The second way this goes wrong is somebody leaving a conversation believing a lead was reassigned when it was not. Asked to do something it cannot, it says so plainly and names the screen where it can be done.
A stranger’s phone number, next quarter’s weather, a snippet of code — one line saying that is not something this assistant does. It is a tool inside a CRM, not a chatbot that happens to live there.
In SimplCRM it answers questions about your own records — “how is the pipeline looking this month”, “which deals have gone quiet for more than 30 days”, “what did we invoice last month, by customer” — by running a real report and showing you the table it read. It also answers questions about how the product works, so “how do I raise a credit note” is a question you can ask the software itself.
No, and it is built so that it cannot. The assistant has three tools and all three only read. There is no tool that creates, updates, deletes or sends, and there is not going to be one: an assistant that can change the books is a different product with a different conversation about liability, and “tidy up the pipeline for me” is not a request anybody could take back.
No. Its reports run under the permissions of the person asking, in the same query path the report screen uses. A salesperson who cannot see another branch’s deals cannot get them out of the assistant either — the rows are simply absent rather than filtered out afterwards.
Yours. You create a key at console.anthropic.com, paste it into Settings, and Anthropic bills you directly for what you ask. There is no per-seat AI charge from us and no markup. You can also set a monthly token cap so a curious afternoon cannot become a surprise.
It is instructed never to answer a numeric question from memory, and it shows the tool calls it made under every answer, so the figures can be traced to the report that produced them. If a report comes back empty it says so — “nothing matches that” is a useful answer and a plausible total nobody checked is worse than none.
To Anthropic, under your own API key, along with the rows the report returned so the answer can be written. Nothing is sent until you ask a question, and the assistant is off until somebody turns it on and supplies a key.
Import a month of deals, paste an Anthropic key, and see whether the answers hold up.