Every CRM has an advanced filter builder, and almost nobody uses it. The gap between the question in your head — "which bigger deals have gone quiet?" — and the six dropdowns that express it is just wide enough that most people never cross it. Closing that gap, and then going one step further into an agent that can act, is the most interesting engineering inside Autometa CRM. This post explains how both work, and why the safety design matters more than the intelligence.
From question to filter
Natural language search translates a plain-English question into the CRM’s real query model. Type "show SaaS leads above $15k inactive for 10 days" and the engine resolves each fragment against your actual schema — industry field, deal value, last-activity timestamp — and produces a live smart filter you can see, adjust and save. The output is not an AI-generated answer you have to trust; it is an ordinary filter you can inspect, built faster than you could click it together.
That inspectability is deliberate. The translation is visible, so a wrong guess — the model reading "SaaS" as a tag when it is an industry value — is obvious and fixable in one click, and the corrected filter teaches the workspace’s vocabulary for next time.
An agent that plans before it touches anything
Search answers questions; the autonomous agent does work. Built on LangGraph, it runs a deliberate loop: detect the intent behind a request, plan which tools to use — search, field updates, drafts, task creation — and then execute step by step. Ask it to "find deals stuck in Negotiation over 30 days and draft check-ins", and it decomposes that into a query, a review pass and a batch of drafts, each step visible as it happens.
Why human-in-the-loop is non-negotiable
The agent’s most important feature is the one that slows it down: approval gates. Anything that changes records in bulk or leaves the building — sending messages, reassigning owners, updating stages en masse — pauses for a human yes. The agent presents what it intends to do, on which records, and waits.
This is a design position, not a limitation. An agent that occasionally does the wrong thing autonomously destroys more trust than a hundred right actions earn back. Draft-first, approve-to-send is how automation earns its way into a team’s workflow — the machine prepares, the human decides, and over time the approvals get faster because the plans keep being right.
Calls become records on their own
The same intelligence layer runs on voice. Calls made through the built-in cloud telephony are transcribed automatically — speech-to-text with sentiment analysis and key-takeaway summaries attached to the record. The rep hangs up and the call log, summary and suggested next step are already there. Multiply that by every call a team makes in a quarter and "the CRM is always out of date" quietly stops being true.
Scoring that learns your pipeline, not a generic one
Predictive lead scoring classifies deals as Hot, Warm or Cold using win signals from your own pipeline — response cadence, stakeholder count, stage velocity, channel engagement — rather than a vendor’s one-size-fits-all model. Scores surface directly on the lead with the contributing signals visible, so a rep can disagree intelligently. The point is not to replace judgment; it is to make sure the next hour goes to the right record.
The privacy line under all of it
Everything above runs under enterprise zero-data-retention terms: your customer records, notes and emails are never used to train public or foundational models. Model calls are scoped to the task at hand and covered by the workspace’s tenant isolation and RBAC — the agent can only see what the person invoking it can see. Intelligence with someone else’s data is a liability; intelligence with your own, under your permissions, is leverage.
Safe autonomy, defined
Our working definition of safe autonomy is simple: visible plans, inspectable outputs, human gates on anything irreversible, and privacy terms you could read to a customer without flinching. The models will keep getting better on their own. The guardrails are the product decision — and they are why teams actually turn the agent on.