Phone: 0850 346 41 09 | Weekdays 10:00–18:00 · Technical support 24/7
Anomaly detection means finding the records in transaction, meter, sensor or log data that break from normal behaviour. The systems we build learn what normal looks like for each user, device or subscriber from your history, score every new record, and send the suspicious ones to the right person with the reasons attached, at a volume your team can review.
A fraudulent expense claim, an unmetered connection, a pump that is starting to fail: each shows up in the data as a small deviation long before anyone notices it on the ground. Among thousands of ordinary records, nobody can find it by hand, and fixed threshold rules either flood you with alerts or miss the cases nobody anticipated.
In machine learning anomaly detection the system is not looking for a known pattern but for a break from the usual behaviour of that particular user, device or subscriber, which is why it can surface case types a fixed rule would miss. It connects to your ERP, meters, sensors, access control or system logs without replacing any of them. Your audit and maintenance people stop sampling at random and spend their time on the records that genuinely look wrong. If you also need forecasting or classification on the same data, see our AI integration service.
Each user, device or subscriber gets its own baseline. Comparing against its own history rather than one threshold for everyone cuts false alarms.
Records are scored as they arrive. High-risk activity triggers an immediate notification, or is held for approval before it completes.
Amount, time, location, frequency and links between records are assessed together, catching combinations that look odd even when each value on its own seems fine.
An alert does more than say "suspicious": it shows which fields drove the score, so the reviewer can record a decision with a reason behind it.
The trade-off between missed cases and false alarms is set by how many alerts your team can properly review each day.
Reviewers mark each alert as confirmed or false. That feedback is used to retrain the model at regular intervals.
Payment fraud often starts with a fake bank-account change request sent by email. SmartMail business email assesses every incoming message for identity, structure and content, and flags look-alike domains and payment-detail changes with a reasoned warning. You can buy SmartMail online through Smart360.
We agree which data to watch and what counts as a problem. Past confirmed cases are reviewed if you have them; if not, we choose methods that need no labelled examples.
Normal behaviour is learned from a trouble-free period, so seasonality, month-end peaks and shift patterns are not mistaken for anomalies.
The model is run over historical data and its alerts are reviewed with your team to set the sensitivity.
The system is connected to live data and alerts reach the right person by email, dashboard or your existing workflow. Feedback keeps the model current.
A system that cries wolf every day is soon ignored. So we do not aim to flag the most anomalies; we aim for a volume your team can genuinely review, with alerts that are worth opening. If machine failures are the concern, we apply the same approach in our predictive maintenance service, starting with a pilot on one to three critical machines.
A utility flags any account whose monthly reading halves. Holiday homes, seasonal businesses and closed shops keep the list long, so field crews end up checking it at random. An anomaly model knows each account's own pattern, so a seasonal dip is not treated as an alarm; what it pushes to the top is an account that is still trading but whose consumption has suddenly fallen, with the reasons beside the alert (the same period last year, a comparison with similar accounts). Crews start with the highest scores, and every inspection result is fed back to the model. The same approach works for expense claims, stock movements and machine sensor data.
For more detail, read our guide to how anomaly detection works, our article on bank detail change email fraud and our guide to storing sensor data in a time-series database. All related articles are on the AI guides page.
Tell us what you want to catch and in which data; we will assess whether your data is suitable and send the pilot scope in a written proposal.
Anomaly detection means automatically finding the records in transaction, meter, sensor or log data that break from normal behaviour. The system learns what normal looks like for each user, device or subscriber from your history, scores every new record and sends the suspicious ones to the right person with the reasons attached. Fraud, leakage and failing machines all show up first as small deviations.
First a separate baseline is learned for each user, device or subscriber from a trouble-free period, so seasonality, month-end peaks and shift patterns are not treated as anomalies. Each new record is scored as it arrives, with amount, time, location and frequency assessed together. The alert shows which fields drove the score, so the reviewer can decide with a reason.
Typical uses are suspected unmetered or illicit use in meter data, irregular payments and expense claims, shrinkage in stock movements, quality deviation on production lines, early fault signs in machine sensor data, out-of-hours or out-of-zone entries in access logs, suspicious access in system logs and fleet fuel consumption deviations. The system connects to these sources without replacing them.
The goal is not to flag the most anomalies but to produce a volume your team can genuinely review. Each record is compared with its own history rather than one threshold for everyone, and sensitivity is tuned to how many alerts your team can handle each day. Reviewers mark alerts as confirmed or false, and that feedback retrains the model at regular intervals.
Yes. If you have confirmed past cases we review them; if not, we choose methods that need no labelled examples, because the system looks for deviation from normal rather than a known pattern. The model is first run as a pilot on historical data, its alerts are reviewed with your team and the sensitivity is set before it goes live.
Anomaly detection looks for deviations in any kind of data, from transactions and meters to stock and logs. Predictive maintenance applies the same approach to machine sensor data to see failures coming. Our predictive maintenance projects start with a pilot on one to three critical machines, with value visible in metrics within the first 30 to 60 days.
There is no off-the-shelf price. We first look at which data you have, which deviations you want to catch and whether the data is usable. After this needs analysis you receive a written proposal setting out scope, stages and price. Projects usually start with a pilot on historical data, and you decide on go-live after reviewing its alerts with your team.
Have a different question? Ask Us