AI at AbhiKai
AI is not a feature we bolted on to a CRM. It is the reason the product exists: the work that gets skipped when a sales day gets busy is exactly the work a model can do well. This page states precisely what AI does today, what is still being built, and where it is deliberately not allowed to act alone.
Written for people evaluating whether the AI is real. Every claim below is either live in production or explicitly labelled as in development.
Running in production today
LiveLead scoring on every enquiry
When a lead arrives, the model reads what the person actually wrote, which ad or form they came from, and how quickly they responded, then ranks them against the deals this business has already won. The rep opens one ordered list instead of a table sorted by time.
Runs once per inbound lead. The score carries its reasoning in plain language, so a rep can disagree with it — and when they do, that correction is what the ranking learns from.
Call notes written from two lines
After a call a rep types or dictates a fragment — "budget ok, wants 3bhk, visiting sunday". The model turns it into a clean note, picks the right status, and sets the next follow-up date.
This is the single biggest reason CRM data goes stale: writing notes is unpaid work. Removing the typing is what keeps the pipeline honest.
Draft replies, not sent replies
A suggested WhatsApp or email message appears alongside the lead, written from that conversation and your own price list. The rep reads it, edits a word, sends.
Drafting is where a model is reliably good and the risk is near zero, because a human is still the one pressing send.
Follow-ups that surface themselves
The model watches which leads are going quiet relative to how leads like them usually behave, and pushes those to the top before they go cold — rather than waiting for a date field to expire.
Reminders based only on a due date fire too late. This is closer to what an experienced sales manager notices by instinct.
In development
In buildA conversational sales agent
An agent that answers the WhatsApp enquiry itself in seconds, asks your qualifying questions, books the appointment, and hands a warm lead to a human. Multi-turn, in the language the customer wrote in, including Hinglish and voice notes.
This is the largest inference workload on the roadmap — a live conversation per enquiry rather than a single scoring call — and the main reason our compute needs grow with our customers' ad spend.
Review responses and sentiment
Personalised replies to every Google review in your brand voice, plus grouping of reviews into themes — waiting time, staff, pricing — so the pattern behind a rating becomes visible.
Sentiment is read per review and aggregated over time, which turns free-text feedback into something a business owner can act on.
Content generation for marketing
Captions from a photo in your brand voice, an Indian festival and season calendar, and post suggestions based on what performed for you before.
Generation is cheap per item but high in volume, since a business posting daily across six channels needs hundreds of drafts a month.
How a lead actually flows through it
- 1
Ingest
A webhook arrives from a form, a Meta lead ad or WhatsApp. It is validated against a schema and written as a lead, idempotently, so a retried delivery cannot duplicate it.
- 2
Enrich and score
An inference call reads the message text, the source and the timing, and returns a priority with its reasoning. Both are stored on the lead, not recomputed on every page view.
- 3
Route
The lead is assigned by round-robin or by rule, and appears at the position its score earned in exactly one person's list.
- 4
Assist
As the rep works, further calls draft the note, the status and the next message. Each one is scoped to that lead's own history and the business's own documents.
- 5
Learn
Edits, corrections and outcomes are recorded against the lead. What converted for this business is what shapes its future rankings — not a global average across unrelated industries.
Where AI is not allowed to act alone
- Nothing reaches a customer without a human pressing send, unless you explicitly switch a specific flow to automatic.
- Every score shows why it decided what it decided, so a rep can overrule it in one tap.
- You set the prices it may quote and the promises it may not make; outside those bounds it escalates rather than improvises.
- A do-not-say list is enforced before a draft is ever shown, not after.
- Any rep can take over a conversation mid-thread, and the agent steps back until they are done.
- Every AI action is written to the activity log with a timestamp, so an automated message is as auditable as a human one.
Your data and the model
- Only the content needed for a specific task is sent for inference — the text of that lead, your own price list, that conversation's history. Not your whole database.
- Your workspace content is never used to train general-purpose models. Our inference provider is engaged under terms that prohibit training on customer data, and we do not train our own models on it either.
- Nothing is shared between businesses. What the model learns about what converts for you stays scoped to your workspace, enforced by the same row-level isolation as the rest of your data.
- AI features can be switched off entirely. The product still works as lead management without them; you lose the assistance, not the pipeline.
Why this makes us an AI company, not a CRM with a chatbot
The unit of cost here is the enquiry, not the seat. A traditional CRM's cost scales with how many people log in; ours scales with how many conversations a customer's advertising produces. A business running festival campaigns can generate more inference in a week than in the preceding quarter.
That shape is the whole thesis. It means we can charge on lead volume instead of per user — which is what makes a five-person team adopt the product at all — but it also means compute is our largest variable cost, and it grows precisely when a customer is succeeding.
It is also why the roadmap is sequenced the way it is. Scoring and drafting are single calls per event and affordable today. The conversational agent is a multi-turn workload per enquiry, and shipping it to every customer at once is a capacity question before it is an engineering one.
What it does not do
- It does not predict revenue or forecast a pipeline. Ranking which lead to call first is a much narrower and more reliable job than predicting an outcome.
- It does not give professional advice. Where a customer asks something legal, medical or financial, the agent is required to escalate to a person.
- It does not decide who to hire, fire or pay. Performance summaries are written for a human manager to read and disagree with, and are never scored automatically against a person.
- It does not make pricing decisions. It quotes only what you have entered, and it cannot invent a discount.
- It is not a general-purpose assistant. Asking it something unrelated to your leads and customers will get you a polite refusal, on purpose.
See it on your own leads
The free plan includes the AI scoring and drafting described above. No card, no demo call.