AI where it earns its place — and rules everywhere else
A lot of what gets sold as AI in a CRM is a rule with better marketing. We use models for the handful of jobs they are genuinely better at, and we say plainly which parts of your process should not have a model near them.
Runs on your data · Never used to train shared models · Every answer traceable
The second answer scored 61%. Below your threshold, so it stopped and fetched a person instead of improvising a price.
The list on the right is why clients trust the list on the left
Every vendor will tell you what their AI can do. Fewer will put in writing what they refuse to let it do.
What we use a model for
- Answering repeat questions outside working hours
- Turning a twenty-minute call into three lines and a next step
- Scoring which enquiries look ready to buy, with the reason attached
- Drafting a reply the rep edits before it goes out
- Pulling fields out of a forwarded PDF for a human to confirm
What we will not let it do
- Quoting a price or approving a discount on its own
- Making the final call on a loan, claim or refund
- Writing to your books or moving money
- Replying to a complaint without a person reading it
- Anything where a wrong answer is expensive and nobody would notice
Narrow work, done properly, beats broad work done badly
These are the places a model genuinely outperforms a rule. Everything outside them is better served by logic you can read.
The nine-at-night enquiry that used to wait until morning
Answers the questions your team answers forty times a week — availability, timings, documents, directions — and hands over the moment the conversation moves somewhere it should not go alone.
Confidence below your threshold routes to a person instead of guessing.
Incoming
“Do you have anything under 80 lakhs in Wakad?”
Answered in 4 seconds
- Matched 3 units against live inventory
- Replied with options and floor plans
- Created lead #5104, assigned to Wakad pod
Notes that actually get written, because nobody writes them
Transcribes the call, writes the summary a manager would write, extracts what was promised and logs it against the record — before the rep has walked back to their desk.
The rep edits before it saves — the draft is the time saving, not the accuracy.
Input
22-minute site visit call · Rohit Sharma · Vertex Motors
Logged automatically
- Wants delivery before Diwali; exchange valuation pending
- Action: send finance options by Thursday
- Next step: showroom visit, Saturday 11 AM
Which forty of these four hundred are worth calling first
Scores enquiries on budget fit, source quality, response behaviour and how close the language is to buying — and shows the reasoning, so a rep can disagree with it.
Every score carries its reason. A score without one is just a number nobody trusts.
Signals
Meta lead · budget stated · replied twice in 10 min · asked about possession
Scored
- Intent 87 / 100 — call today
- Because: budget stated, fast replies, possession question
- Comparable leads closed in 18 days on average
The PAN card someone photographed at an angle
Reads forwarded documents — identity proofs, purchase orders, bank statements — and fills the fields for a human to confirm rather than making you retype them.
Nothing saves without a person confirming it. Extraction removes typing, not judgement.
Received on WhatsApp
purchase_order_scan.jpg · 2.1 MB
Extracted for review
- PO number, date, line items, tax split
- Matched to an existing quotation
- Two fields flagged as low confidence
The settings your compliance team will ask about
These are exposed in the admin screen, not buried in a contract. Change them yourself, any time.
- Confidence threshold
- 75%
- Anything the model is less sure of goes to a person, not out the door.
- Always escalate
- price · refund · complaint · legal
- Keyword list you control. These never get an automated answer.
- After-hours autonomy
- Restricted
- Full answers during working hours; acknowledge-and-queue outside them.
- Data retention
- 30 days
- Transcripts purge on schedule. The extracted record stays, the raw text does not.
- Model training
- Disabled
- Your conversations are never contributed to a shared or public model.
- Answer audit
- Every response
- Model version, inputs seen and confidence stored against the record.
Start with the questions your team answers forty times a week
The front desk is usually the fastest payback, and the easiest to switch off if it does not suit you. Tell us what people ask you most, and we will show you what it would handle.
- Confidence threshold you set, not us
- Human handoff built in from the start
- Your data never trains a shared model
- Any job type switchable off, instantly