How to automate scoring & detection
Turn signals into a number or a flag — lead score, health score, risk, anomaly, forecast.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 12, 2026
In short
What scoring & detection means in practice.
Scoring is how an automation ranks and prioritises: which lead first, which account is drifting, which alert matters, which SKU will run out. It turns several signals into one number or one flag that people can act on, and it is the capability where the business most wants to see the working.
Two designs dominate. Transparent, rule-weighted scores — components you can read, weights in a table — and statistical or learned models — forecasts, anomaly detection. The first fits most business processes and should be the default; the second earns its place when there is a year of outcome data and a clear metric.
Where it shows up
26 automations use it.
Each link opens the full blueprint — trigger, every step, guardrails, template and what it costs to have it built. The step names below are the ones that do this job.
Customer Health Scoring
Score each component · Analyse ticket sentiment · Compute score and trend · Detect drops and thresholds
Finance & accountingKYC/AML Monitoring Automation
Screen the customer · Pre-assess matches · Evaluate transactions · Continuous re-screening
Supply chain & procurementLogistics Optimization with AI
Select the service · Track in flight · Detect and explain delays
HRResume Screening Automation
Score against criteria with evidence · Flag verification items · Build the ranked shortlist
Supply chain & procurementAI Inventory Optimization
Measure supplier lead times · Compute safety stock and reorder points
Supply chain & procurementAI Supplier Risk Monitoring
Measure delivery performance · Compute the composite score
InsuranceAutomated Insurance Underwriting Intake
Apply appetite rules · Indicative rating
Customer successCustomer Onboarding Automation
Track milestones from product events · Detect stalls
MarketingAI Competitor Monitoring
Detect material changes
Supply chain & procurementAI Demand Forecasting
Run the forecasting model
Supply chain & procurementAI Spend Analysis
Detect opportunities
EngineeringAI Unit Test Generation
Find untested changed functions
EngineeringAutomated Code Review with AI
Analyse the change
Security & ITAutomated Security Alert Triage
Score and classify
Finance & accountingClient Reporting Automation
Find the variances that matter
Customer successCustomer Feedback Analysis Automation
Aggregate
Customer successCustomer Win-Back Automation
Match triggers to segments
Customer successExpansion Opportunity Detection
Compute the readiness score
SalesInbound Lead Qualification with AI
Score against the ICP
InsuranceInsurance Claims Intake Automation
Screen fraud indicators
HRPayroll Automation
Anomaly review
OperationsPrior Authorization Automation
Detect auth requirement
Customer supportProactive Customer Outreach
Decide whether to reach out
SalesSpeed-to-Lead Automation
Qualify
Finance & accountingSubscription Dunning Automation
Choose the retry strategy
OperationsWeekly Reporting Automation
Assemble the metrics table
How it works
Step by step.
Components, not a black box
A score is a weighted sum of named components (usage trend, support sentiment, billing status, engagement). Each component is visible on the record, so when the score drops the reason is obvious.
Thresholds with hysteresis
Alerts fire on crossing a threshold, not on being near it, and re-fire only after a cool-down. Without this, one account can page the team every hour.
Language models for the soft signals
Tone of the last three tickets, whether a champion has gone quiet, whether a news item is material — these are judgement calls an LLM turns into a bounded score with a reason, feeding the same weighted sum.
Calibrate on what happened
Every quarter, compare scores from 90 days ago with the actual outcome — churned, converted, defaulted. Adjust weights where the score was confidently wrong, and publish the accuracy so the team keeps trusting it.
By tool
How it is built, by tool.
The step is the same whichever tool runs it; this is where it lives in each. We build in the one that fits your accounts and budget, and you own the result.
- n8nCode for weighted scores and thresholds, Sentiment Analysis / AI Agent for soft signals, HTTP Request to a forecasting service, IF for threshold logic.
- MakeMath and text functions in a Set variable module, an LLM module for judgement signals, filters for thresholds.
- ZapierFormatter numbers for simple scores, AI step for sentiment; forecasting and anomaly detection need an external service.
- CodeFeature computation in SQL or pandas, a gradient-boosting or time-series model where justified, a calibration job on a schedule.
Where it goes wrong
The failure modes we design against.
A score nobody can explain
If the CSM cannot say why an account is red, they will ignore it. Show the components.
Never recalibrated
Weights written once drift from reality within a couple of quarters. Schedule the comparison against real outcomes.
Alert fatigue
Every threshold crossing as a message equals a muted channel. Batch, cool down, and rank.
See what we automate
The automation libraryEvery process, by department, each with a price and a free template.
Or get a plan for your own process
Three sentences; one option and its price appear in minutes. No call.
FAQ
Questions people ask.
Can I get automated alerts when an account’s health changes?
Yes. A weekly (or event-driven) score with named components, a threshold with a cool-down, and an alert that names the component that moved and proposes the next action — usually with a drafted message attached.
Do I need machine learning for a lead score?
Not to start. A transparent weighted score against your ICP, calibrated quarterly against which leads actually converted, beats most ML for a small team. Move to a trained model when you have a year of outcomes and someone to own it.
How is anomaly detection different from a threshold?
A threshold compares to a fixed number; anomaly detection compares to what is normal for this entity at this time — a Monday spike is not an anomaly for a Monday. Use it when ‘normal’ varies by entity or season.