CrossNibbleBook a call

AI in production

AI solutions we have built

Most B2B companies do not have a sales problem. They have a leakage problem. Leads arrive at midnight and are seen on Monday, missed calls are never returned, half the pipeline is dead deals, and nobody can say which source is worth funding. These are thirteen systems we have built and run to fix that. Each has its own case study: the situation as the business felt it, how the system works, where AI does the work and what changed.

Client names are withheld. References are available on request.

Lead capture

High-intent signals that arrive when nobody is watching, and traffic already paid for that never turns into a named lead.

The situation

A serious enquiry lands at 11 PM on WhatsApp. The rep sees it on Monday morning. By then the buyer has messaged three other suppliers, had answers from at least one of them, and no longer remembers which company was which.

What we built

An AI agent on WhatsApp that replies in under 30 seconds, asks the qualifying questions, captures requirement and budget, and hands over to a person only when the lead is worth a call.

What changed

First response goes from hours to seconds. Qualified leads reach the rep with the conversation already started.

Read the case study →

The situation

Sales lines are busy at peak hours. Calls that are not picked up go into a missed-call list, and the list is worked through when someone has time, which is usually hours later.

What we built

The calling system linked to WhatsApp. The moment a call goes unanswered, the caller gets a personal WhatsApp message within seconds, and the AI keeps the conversation going until someone is free to take over.

What changed

Missed calls turn into live conversations. Nobody waits on a callback that never comes.

Read the case study →

The situation

A visitor reads the pricing or solutions page, scrolls to the bottom, does not fill the form and leaves. The traffic was paid for, the interest was real, and nobody knows who it was.

What we built

A chat that opens only on real intent signals: time on page, scroll depth and no clicked call to action. It asks a question about that page, answers from the company knowledge base, and asks for a phone number or email only once the visitor is engaged. Pages viewed, time spent and questions asked go to the CRM with the lead.

What changed

Anonymous traffic becomes named pipeline, and the rep gets the whole conversation, not just an email address.

Read the case study →

Data quality and qualification

A lead with no context cannot be scored, routed or prioritised. These make every record usable before it reaches a rep.

04The half-empty CRM

AI-assisted · CRM
The situation

Industry is blank on most records. City is missing. Designation says “Director” with no department. Product interest was never recorded.

What we built

Missing fields pulled automatically out of conversations and email signatures: city, state, company, role and product interest, written back into the CRM without overwriting anything a person has edited.

What changed

CRM data goes from “we will fix it later” to usable for territory routing, account lists and reporting.

Read the case study →

The situation

A new lead arrives with a phone number and a first name. No company, no role, no industry. There is nothing to score it on and nothing to route it by.

What we built

The CRM connected to third-party enrichment services. When key fields are empty, the system looks the record up by email or phone and fills in company, industry, size, role, seniority and location. It only fills empty fields and never overwrites a rep’s edits.

What changed

Leads arrive ready to score and route, ideal-customer filters work from day one, and reps stop spending five minutes researching each lead.

Read the case study →

The situation

What one rep calls qualified, another calls junk. The sales head sees lead quality swing from week to week and cannot tell whether the leads changed or the reps did.

What we built

The company’s own qualification criteria written down once, then applied by AI to every lead the same way: a score, the reasons behind it and a recommended next step. The criteria can be changed in plain language, without code.

What changed

The sales head stops arguing about lead quality, and reps spend their time on the right fifth of the list.

Read the case study →

Solutions 4 and 5 work as two layers: what can be learnt from your own conversations first, at no cost, and the paid lookup only when that is not enough.

Engagement and conversion

Most teams give up after two ignored follow-ups. These keep conversations alive without sounding desperate, and run campaigns at volume without losing the account.

The situation

A WhatsApp conversation goes quiet after three messages. The rep sends “Just checking in”. That is ignored too.

What we built

Two stages. First, a follow-up written from what was actually discussed, adding something new rather than “are you there?”, sent after a set delay with a limit on attempts. If that is ignored too, the system stops sending text and picks the most relevant case study, product video or brochure from the media library instead.

What changed

Conversations that would have died restart on their own, and some convert weeks later because the right document stayed in their WhatsApp.

Read the case study →

The situation

The same template goes to thousands of numbers. Many of them do not fit the offer, some get the message twice, and recipients start blocking and reporting.

What we built

Audiences filtered by score, status, location, product interest and source. No-repeat rules so nobody gets the same template twice, sending limits inside the platform’s safety thresholds, and audiences that refresh as new leads qualify.

What changed

The same volume with no account suspensions, and more replies because the audience actually fits.

Read the case study →

Pipeline intelligence

The CRM holds the data but rarely gives the answers. These turn it from a record of what happened into a tool for deciding what to do next.

09Pipeline buried in inboxes

AI-assisted · Email
The situation

Reps email prospects who never get added to the CRM. Decision makers copied on threads stay invisible. Deals progress entirely over email.

What we built

Team inboxes checked for outgoing emails to companies not in the CRM. Orphan threads are surfaced, and new contacts are read from email signatures and created in one click.

What changed

Teams typically find 20 to 40 percent more active conversations than their CRM was tracking. Some close within weeks because they were already warm, just unrecorded.

Read the case study →

The situation

Half the pipeline is deals nobody wants to mark lost. They sit at the same stage for months, still counted at full value.

What we built

Every deal that has not moved stage or seen real activity in 60 days is flagged and needs a decision, and the forecast is rebuilt on stage probability instead of rep optimism.

What changed

A forecast the founder can trust, and no dead deals to hide behind.

Read the case study →

The situation

The CRM holds three years of deals. Nobody can say why deals from one source close three times faster than another, which industries the team actually wins in, which owners convert best in which segments, or where in the funnel deals die.

What we built

An analysis layer on top of the CRM: lead-to-won funnel with drop-off at each stage, forecast weighted by historical conversion, win and loss by source, owner, industry and deal size, and quarter and Indian financial-year filters.

What changed

The Monday questions answer themselves, and marketing knows which sources to fund and which to cut.

Read the case study →

Team performance

Coaching that runs on memory produces uneven reps. These give sales leaders evidence instead of impressions.

The situation

Managers hear fewer than 2 percent of recorded calls. Coaching runs on memory and impressions.

What we built

A call analyser on top of the existing dialer. It pulls the recordings already being made, transcribes them with rep and customer separated, handles Hindi, English and mixed calls, scores each call on seven parameters from greeting to closing, flags risks, intent and competitor mentions, and writes a short summary.

What changed

A manager can review 50 calls in 20 minutes. Coaching runs on evidence, new reps ramp faster, and compliance issues surface in days rather than at the end of the quarter.

Read the case study →

The situation

The sales head cannot tell who is slow to respond, who converts and who lets leads go cold. The data exists, but it is spread across four different tools.

What we built

One scorecard per rep: leads owned, conversion rate, average response time, pipeline value, and how much of the messaging is AI versus the rep. Managers see their own team and reps see their own leads.

What changed

Coaching stops being “I feel you are slow” and becomes “your average response time is four hours and the team’s is forty minutes”.

Read the case study →

How they are engineered

A production system, not a demo

The model is one component, not the product. Around it sits the part that keeps the system working when real users arrive: data flow, state, error handling, fallbacks, access control and audit logs. The AI does the language work. The software around it does the rest.

CRM records, call recordings and customer messages move through controlled paths with access boundaries, not into prompts logged somewhere nobody can see. Each part is built so that ten times the volume does not mean a rebuild. Plenty of AI tools work at 100 leads a day and fall over at 1,000.

Running cost is designed down before go-live, not after. The model runs only when no cheaper, rule-based step can do the job, so cost per transaction falls as volume grows instead of becoming the line item leadership questions in month three.

Rules first, AI second

Simple rules, lookups and conditions handle 70 to 80 percent of decisions. The model is called only when real judgement is needed.

The smallest model that works

Classification, extraction and short replies run on lighter models. Premium models are kept for the tasks that need them.

Cache and reuse

A question answered once is not paid for twice. Embeddings and structured knowledge are reused, not regenerated.

Tight prompts

Short, specific instructions, with no context the model does not need.

Batch where speed does not matter

A job that does not need an answer in two seconds does not pay two-second prices.

Running cost and data

Your environment, your AI account, no markup

We build the system in your own environment. You pay only for the AI it uses, directly to the AI provider, with no markup from us.

01Your environment

The system is set up in your own cloud account. Conversations, recordings and CRM data move between your systems and the AI provider you choose. They do not pass through our servers or a separate chatbot platform.

02Pay as you go for AI

The AI runs on your own account with Google (Gemini) or OpenAI. You pay them directly, only for what is used, at their published rates.

03No markup

We do not resell AI usage or charge per message, per conversation or per seat. Our fee is for building and improving the system, and it is agreed upfront.

04Minimal by design

Rules and plain code do most of the work, and the model is called only where judgement on text is needed, so the AI bill stays small as volume grows.

WhatsApp and calling charges, where they apply, are billed by your provider as they are today.

Your existing stack

A layer on top, not a platform to move to

Nothing here asks you to replace what you use. If it has an API or a webhook, it connects. You keep control of the underlying systems, and we add the automation and AI on top.

CRM

Zoho, HubSpot, Salesforce, Pipedrive, Freshworks, or anything with an open API.

Calling

Your dialer and recording stay as they are: Exotel, MyOperator, Knowlarity, Ozonetel, CallHippo, Servetel and others. We work from the recordings and logs they already produce.

WhatsApp

Any approved provider, including the Meta Cloud API, Gupshup, Interakt, Wati, AiSensy and Twilio.

Email, ads, forms and website

Anything that exposes an API or webhook.

Beyond the thirteen

What else we build

The thirteen above are repeatable. Beyond them we build custom tools, workflows and integrations whenever a business has a specific problem worth solving. Most projects take two to eight weeks.

Extensions to existing systems

Custom workflows inside your CRM, calling routes and IVR logic your dialer does not support, custom WhatsApp flows, and two tools that should talk to each other finally doing so.

Calculators and sales tools

ROI and pricing calculators that capture leads on the website and help reps in the meeting, plus quote builders and proposal generators that do the maths for the buyer.

Dashboards and reports

KPI views for founders, sales heads and operations, with daily and weekly digests sent to email, Slack or WhatsApp, pulling from systems that never shared one view.

AI agents for repetitive work

Reading inbound RFQs, summarising long email threads, classifying support tickets, drafting first proposals. Built once, running every day.

Lead capture and scoring

Multi-step forms built around your funnel, scoring models specific to your ideal customer, and routing by territory, vertical, deal size or product line.

Recognise two or three of these?

We start with a 30-minute call to map where leads come from and where they drop off. If there is a fit, you get a short written diagnosis of the two or three problems worth fixing first, with a rough investment range. Then we build one or two, prove them, and expand. No annual contract before the first system goes live.

Start with an AI auditBook a call