AI-native customer revenue operating system
Somebody forgot to call back. The person on the floor never saw what she spent an hour reading last night. The reply came on Tuesday. None of that is carelessness — it is what happens when a business keeps its customers in six systems that cannot see each other. topAgent makes it one system, and turns the loss into a number you can move.
Never lose a customer. Never lose the context. Never lose the opportunity.
Built for India first: IST business calendars, INR, E.164 phone handling, data resident in Mumbai, and DPDP-aware consent from the first row of the schema.
The problem
Thirty years of treating customer relationships as a data-entry exercise produced a stack that records what already happened and waits for a person to notice. These are published industry figures, and they describe the market rather than this product.
58%
of revenue-team capacity goes to the tooling, not the customer
Process misalignment and integration work across dozens of disconnected applications. Uncoordinated handoffs between marketing, sales and service extend enterprise sales cycles by a further 30%.
20.6%
of revenue organisations can show AI producing verified revenue
Effectively all of them have deployed it. The gap is not the model — it is putting a language model on top of a passive database that still needs a person to notice, decide and act.
81%
of mobile lead forms are abandoned before submission
And a mobile landing page that loads slowly loses 53% of its traffic before anyone engages with it. The lead was paid for at the ad auction and lost at the form.
70%+
of high-intent leads are dropped after a single outreach attempt
A reply delayed past five minutes costs a quarter of the responses. Past thirty minutes, half of them. Intent does not wait for a working day, and neither does a competitor.
Sources cited in the TopAgent architecture blueprint and PRD v2.4.0. Industry research, not measured platform results.
The category
The CRM, the CDP and the chatbot are each excellent at the job they were designed for. None of them was designed for this one — to sense a customer, decide what to do, and do it, without a person in the middle of every step.
| Attribute | Legacy CRM | Customer data platform | Chatbot | topAgent |
|---|---|---|---|---|
| What it is for | A record of what people remembered to type in | A warehouse of behavioural events | A scripted widget in one channel | A loop that senses, decides and acts — continuously |
| How data arrives | A person logs it, or it never happened | SDK telemetry into analytical storage | A session cache that dies with the window | Streaming, from the till and the ERP as well as the web |
| Identity | Exact match on one field, usually email | Stitched on the warehouse’s schedule, in batches | None; every thread is a stranger | A live graph across device, phone, message handle and invoice |
| Ability to act | None. It waits for a human to open it | None. It exports an audience and stops | A decision tree, until the question is novel | A governed agent workforce with real tools and real limits |
| Memory | Free-text notes nobody reads twice | Structured events with no comprehension | Wiped when the window closes | Relational and semantic, surviving every channel change |
| Time to act | Hours to days — a person’s calendar | Minutes to hours — an ETL schedule | Instant, and disconnected from everything | Sub-second triggers; a conversation answered inside 30 seconds |
The mechanism
Everything else in this platform is a consequence of this loop being closed rather than open. An open loop is a dashboard: it tells somebody something and hopes.
An ad tap, a page view, a WhatsApp message, a phone call, a reception check-in, a settled invoice. Every signal is captured and durable before anything is decided about it.
Signals resolve to one person. Nine identifiers across six systems become a single customer with a history, not nine half-records competing to be the real one.
Intent, urgency, value and consent are evaluated together. The output is a ranked action with the reasons attached, not a score somebody has to take on faith.
The action is executed inside its authority: a reply, a booking, a follow-up, a briefing pushed to a showroom tablet. Anything beyond that authority stops and asks a named person.
The outcome is written back to the same graph — including to the ad platform, so the auction learns what a sale looks like rather than what a form fill looks like.
What worked changes what is tried next. The loop is the product; every other feature is a consequence of it being closed.
On the floor
Everything else on this page is a system — planes, budgets, gates, ledgers. All of it exists for a woman walking into a showroom in Jubilee Hills to be recognised, and for the person who greets her to already know what she has spent three weeks looking at.
She arrives as an ad tap, a WhatsApp message, a missed call and a walk-in — which in most businesses is four strangers in four systems. Here it is one person with one history, resolved before anybody is assigned to her.
Merged only on a certain identifier. Anything less certain personalises the session and waits for proof.
What she prefers is read off what she actually bought and what she actually said — metal, stone, category, price band — and never guessed from a description. A bill whose lines carry attributes teaches the graph; one that does not teaches it nothing, and says so rather than inventing a preference.
Inferred from purchases, conversations and browsing. Never from an assumption about who she is.
After the sale is where a jewellery relationship is kept or lost: an order, a warranty, a repair, a cleaning before an anniversary. It is triaged as relationship work rather than as a ticket queue, and the whole conversation is already there — she never has to explain herself twice.
Resolves routine service end to end. A defect claim, a return or sentiment turning negative goes to a person.
The person on the floor is briefed before the customer reaches the counter: who is walking in, what she has been looking at for three weeks, what she will ask about the setting. The graph is worth nothing if it stops at a dashboard nobody opens mid-conversation.
Copilot, not autopilot: drafts and dossiers, and zero unmonitored sends to a customer.
The asset
Models are procured. Channels are rented. The graph — every signal this business has ever received about every person it has ever served, resolved and kept — is the only part that compounds, and the only part that does not leave when a vendor does. It is held across eight planes.
Phone and email hashes, ERP account code, message handles, device cookies
The root every other plane hangs from. Get this wrong and everything downstream is about the wrong person.
Clickstream, WhatsApp transcripts, call recordings, showroom check-ins, ad impressions
So nobody is ever asked to repeat something they have already said. That is the whole of it.
Settled revenue, margin contribution, returns, open quotations, carts, store receipts
Value tiers and discount authority come from here — from bills, never from a derived guess.
Funnel stage, product affinity, design preferences, budget band, search velocity
What she is actually shopping for, expressed precisely enough to act on.
Preferred channel, the hours she replies, language and tone
The right message at the wrong hour is a worse outcome than no message.
Conversion probability, churn risk, twelve-month lifetime value, next-best-action
Where attention goes first. A rep has forty good hours a week and this decides how they are spent.
Open tickets, repairs, returns, CSAT and live sentiment
A marketing message must never land on somebody with an unresolved complaint. This is the plane that stops it.
Tier progress, household edges, anniversaries, gifting occasions
High-consideration purchases repeat on a multi-year clock. Somebody has to be holding that clock.
Identity
A buyer browses anonymously on a phone, messages from a number, clicks a link in an email, and pays at a counter under a tax identifier. Four systems, four strangers. Resolving them is the hardest thing this platform does and the thing everything else depends on.
Certain identifiers — a phone number, an email, an account code on a bill — merge immediately and permanently. Everything less certain is held as a probability, attached to the live session so the conversation can still be personal, and never written as a merge until something certain arrives. A merge is irreversible, so it waits for proof.
And when two signals contradict each other — two people, one household tablet — the graph partitions them back apart rather than averaging two customers into one who does not exist.
Resolution thresholds
A behavioural event never creates a person. Promotion to a named customer requires recorded, lawful evidence — a form, a login, a purchase — and never an inference.
The workforce
A single agent with every tool and the whole history is how you get drift, cost and a confident answer nobody authorised. Each agent here sees only what its task needs and holds only the tools its task permits. Every one has a ceiling and a trip-wire, and both are written down.
Ingest, parse and qualify every inbound signal — ad, form, message, call — inside thirty seconds.
Brief the person on the floor: who is walking in, what they looked at, what they will ask.
Carry a prospect across a multi-week consideration arc without becoming a nuisance.
Turn digital interest into a confirmed appointment against a real associate’s roster.
Triage service after the sale — orders, warranties, repairs, cleaning — as relationship work.
Turn a sentence of commercial intent into an audience, a message and a channel mix.
Audit the funnel continuously and write the morning brief a director would otherwise build by hand.
The limits
The question enterprise buyers actually ask is not what the agents can do. It is what they cannot do, who decided that, and how it is proved after the fact.
Level 1
Drafts recommendations. A person dispatches every single one. Nothing reaches a customer unread.
Level 2
Handles standard qualifying interactions inside a defined envelope and escalates every exception.
Level 3
Runs complete qualification, booking and routine follow-up. Still cannot touch price, margin or credit.
Standard retail pricing executes autonomously. One to five per cent needs a floor manager, on their phone, inside three minutes. Beyond five per cent needs two signatures. An agent can never move a margin, waive a tax or extend credit — not at any autonomy level.
The reasoning, the context it saw, the tool it called, what came back, and the name of whoever approved it. Written once, kept seven years, mutable by nobody. An audit trail somebody can tidy afterwards is not one.
Permission is recorded per channel and per purpose, with who asked and when. It is never inferred from a reply, a purchase or a visit. Every suppressed send records the rule that suppressed it, so “we reached 240 of 3,100” is a question with an answer.
Customer records and embeddings sit in AWS Mumbai, built to India’s DPDP Act alongside GDPR and CCPA. Identifiers are masked before any transcript reaches a model, and every reveal of a real one is audited.
Latency
A reply delayed past five minutes costs a quarter of the responses; past thirty, half of them. A voice agent that pauses for more than a second and a half stops sounding like a conversation. These are not aspirations written on a wall — they are budgets the system is built against and measured on.
The baseline they replace is an industry average inbound response of four point two hours, by which point the customer has been sold to by somebody else.
Engineering budgets
Sized for five thousand events per second per tenant, because festival trade does not arrive at the average rate — and a queue that grows is acceptable where a lead that vanishes is not.
The closed loop
The reason a jewellery business cannot tell which campaign paid for itself is that the click happens online and the sale happens at a counter, three weeks later, under a different name. Closing that gap is the single highest-value thing this platform does.
The click identifier is captured at first touch, because it cannot be reconstructed afterwards.
The tap opens a WhatsApp thread rather than a form. Nothing is abandoned at field four.
Digital interest becomes a confirmed visit against a named associate on a real roster.
The till writes the sale back to the same person, with its line items and its margin.
The platform hears that this click became revenue, so it optimises for buyers instead of form fills.
Attribution is append-only: snapshots are added and never edited, so the answer to “what produced this sale” cannot drift as people work the record.
Who it is for
Departmental incentives in a retail business are structurally misaligned: marketing is paid for volume, sales ignores anything that needs six weeks, service is a cost centre, and technology resists all three. One layer, reading the same graph, is what reconciles them.
Growth is capped by headcount. Revenue leaks between departments nobody owns.
An operating layer that senses and acts on every opportunity without another hire.
Spend buys clicks. Nobody can tie a click to a showroom sale three weeks later.
First-party signals open a conversation on tap, and the settled invoice goes back to the ad platform.
Follow-up is inconsistent. Leads are abandoned after one or two attempts.
An algorithmic cadence that works a prospect for weeks and never forgets one.
Support is a cost centre with a backlog, and service disputes quietly end relationships.
Routine enquiries resolve themselves; the rest arrive with their history already assembled.
The floor has no idea what the customer spent an hour looking at online last night.
High-intent digital leads arrive as booked visits, with the associate briefed before the door opens.
A thicket of point-to-point integrations, each one a rate limit and a security review.
One composable layer above the systems of record. It replaces none of them and reads all of them.
What it costs
Priced on what the platform can count: connectors live, channels feeding them, how long history is kept, which analytics are reachable, how many agents run and at what autonomy, and how many automations are active. Not on how many of your colleagues log in — a product whose argument is that the machine does the follow-up cannot then bill you per person doing follow-up.
One showroom, getting every enquiry answered inside the SLA instead of the next working day.
$99per workspace, per month
or $1,128.60 a year — $94.05 a month, saving $59.40
Most workspaces start here
Several stores, real money in paid acquisition, and a return-on-spend number somebody has to defend.
$249per workspace, per month
or $2,838.60 a year — $236.55 a month, saving $149.40
A group running its revenue on this, writing its own adapters against the published contract.
$399per workspace, per month
or $4,548.60 a year — $379.05 a month, saving $239.40
List prices, per workspace, excluding Indian GST. Annual billing takes 5% off the twelve-month total. The platform refuses on connectors, lead channels, retention, analytics, automations: a workspace over its limit cannot add another, and every refusal names the plan that would allow it. single sign-on, custom adapters, the agent limits are written down rather than enforced, because the work behind them is not built yet. Nothing commercial ever refuses an enquiry: once ingestion has accepted a lead it is kept, whatever plan the workspace is on.
Start here
For enterprises
topAgent replaces no system of record. It sits above the ERP, the storefront, the till and the ad accounts, reads all of them, and gives the business one customer instead of six. A pilot starts on one showroom and one channel, and the first thing it proves is how many leads are currently being lost.
Talk to us about a pilotFor investors
Models are procured; the event graph is not. Every interaction compounds into institutional memory that the next interaction is measurably better for — and that asset does not transfer when a vendor changes. The platform is multi-tenant from the first migration, with a live reference deployment in fine jewellery.
Request the architecture briefFor partners
Messaging, telephony, POS, commerce, ad platforms and models all sit behind one contract with one test suite. Integrating is implementing that contract, not negotiating a place in the product — and no product logic anywhere branches on whose name is on the connector.
See the connector contractFor engineers
Identity resolution that must not merge two sisters. An append-only attribution record. A consent ledger that survives a merge and an unmerge. A scheduler that is a table with a lease, not ten thousand cloud timers. If you think correctness is a feature rather than a tax, this is a good place to work.
See what we are buildingMeta, Google, WhatsApp, the telephony carrier, the storefront and the till are adapters. Swap any of them and nothing in the product changes — the customers, the conversations and the revenue stay exactly where they are, because they were never anywhere else.