AI Lead Scoring and Prioritisation Engine

The problem

Zoho CRM's built-in scoring is rule-based. Someone in a workshop decided a job title is worth twenty points and an opened email is worth five, and those numbers have not been revisited since. The result looks quantitative but is not: the weights encode a guess, not what your pipeline actually did. Two leads sit at 65 points and one of them was never going to buy.

Meanwhile the answer is already in your org. You have closed Deals with stage histories, source attribution, industry, region, deal size and cycle time. You have Leads that went nowhere and Leads that converted in nine days. That history is a training set, and nobody is using it.

The second problem is time. A score frozen at creation says nothing about a lead that has gone quiet for six weeks. Without decay, your reps work a list sorted by a number that stopped being true. And a score on its own does not tell anyone what to do — a rep looking at 82 still has to decide whether to call, send a quote, or route it to a senior closer.

What ECOSIRE builds

This is a build-to-order engine. Nothing is pre-trained and nothing downloads instantly. We build it against your data, in your Zoho org, after a scoping call and a fixed quote.

Feature extraction from your own history

We pull your historical Leads, Contacts, Deals and related activity through the Zoho CRM REST API — including Deal stage history, Lead Source, Industry, Rating, campaign attribution where you use Zoho Campaigns or Marketing Automation, and engagement signals such as Call, Task, Meeting and Email counts from the Activities modules. Where the business runs on Zoho Books, we can include commercial signals that matter for a returning buyer: prior invoice value, payment behaviour, or an existing customer flag drawn from the Books organization.

Before any modelling we produce a data-readiness report. It tells you plainly how many closed outcomes you actually have, which fields are too sparse to use, and where labels are ambiguous — for example Deals parked in a stage that means "lost" to your team but is not marked Closed Lost. If your history is too thin to support a model, we say so and propose a calibrated rule-based interim instead of selling you a model that fits noise.

The model

We train a supervised classifier on your labelled outcomes and hold out a time-based validation split, so the model is judged on periods it never saw rather than on a random shuffle that leaks the future. You get the evaluation honestly: discrimination, calibration, and which segments the model is weak on. The output is a probability between zero and one, calibrated so a 0.30 lead really does convert about three times in ten in your business — not an arbitrary points total dressed up as a percentage.

Score decay and recency

A lead's probability is recalculated on a schedule and on meaningful events. A decay function reduces the score as time passes without engagement, with the shape of the curve fitted to your observed sales cycle rather than a default half-life. New activity — an inbound call logged in Calls, a replied email, a Meeting booked, a quote sent — refreshes the score. Sales sees a list that reflects today.

Next-best-action written onto the record

Probability alone does not change behaviour, so the engine writes an action recommendation alongside it: call now, send pricing, nurture, route to senior owner, or disqualify for review. Recommendations are derived from what historically moved comparable leads forward, and each carries the top contributing factors so a rep can see why. Custom fields on the Lead and Deal hold the probability, band, decay-adjusted score, recommended action, top factors and last-scored timestamp — ordinary CRM fields, so they work in list views, filters, Kanban, workflow rules, assignment rules, Blueprint transitions and Zoho Analytics.

Delivery into Zoho

Scores reach your org through Deluge custom functions and scheduled functions calling the scoring service, with a Zoho Flow path where you prefer that orchestration. Bulk backfill runs through the Bulk Write API rather than record-by-record updates, so an initial scoring of a large database does not exhaust your API credits. We build API-limit awareness into the scheduler and document the consumption we expect at your volume.

Monitoring and retraining

Models drift. We build a monitoring job that compares predicted probability against realised conversion by cohort and raises a flag when calibration slips. Retraining is a documented, repeatable procedure — you own the pipeline and the artefacts, and you can commission a retrain or run it yourself.

Who this is for

Sales organisations with at least a year of clean closed-won and closed-lost history and enough inbound volume that prioritisation is a real constraint. Marketing teams held accountable for lead quality who need a defensible measure rather than an argument. Revenue leaders who want routing and SLA rules driven by likelihood rather than by whoever shouted first.

How delivery works

Scoping call. We look at your CRM structure, historical volume and label quality, which modules and fields are in play, your definition of a qualified lead, and how you want recommendations to appear to reps.

Fixed quote. You get a written scope covering the data-readiness report, the model class, the fields we will create, the scoring cadence, and the price and lead time. If readiness comes back weak, the scope changes before you are billed for a model.

Build. Typical lead time is two to four weeks from approval, depending on data volume, how much cleaning the history needs, and the number of modules in scope.

Install in test, then production. We deploy to a sandbox or non-production org, backfill scores there and review the distribution with you against leads your team already has an opinion about. Only then do we move to production, with the ability to disable the write-back cleanly.

Support. Handover includes training for administrators and a session with the sales team on how to read the score, the decay and the recommendation. A defect-fix and tuning window follows, with duration stated in the quote.

What we will not do

We will not promise a lift percentage. Any vendor quoting you an uplift figure before seeing your data is guessing. We will not train on a few dozen outcomes and call it AI. And we will not hide the model behind a black box — you receive the feature list, the evaluation, and the retraining procedure, because a score nobody can explain is a score nobody will trust.