CMS Is Using Predictive Analytics to Score Paying Parents

An official UK Government transparency record confirms that the Child Maintenance Service is using a machine-learning system to predict whether currently compliant paying parents may miss a future payment.

This is confirmed by the Government itself

The Department for Work and Pensions has published an official Algorithmic Transparency Record titled:

“DWP: Child Maintenance Service (CMS) – Predictive Analytics Compliance”

It confirms that machine learning is being used to predict which compliant CMS cases may become financially non-compliant.

What exactly is CMS doing?

CMS operates a predictive analytics system designed to identify paying parents who are currently considered compliant but who the model predicts could miss an upcoming payment.

According to DWP, the system analyses current and historic information held within its CMS2012 system.

DWP says the objective is to identify cases at risk of a payment breakdown early enough for caseworkers to intervene before payments stop.

Who is being assessed?

One of the most important points in the transparency record is that this is aimed at cases which are currently compliant.

In other words, the system is not simply identifying someone who has already failed to pay.

It is attempting to predict whether someone who is currently compliant may become non-compliant in the future.

What information can be used?

The Government transparency record identifies a wide range of CMS information feeding into the model.

Your communications with CMS can form part of the data

The transparency record specifically refers to information concerning communications exchanged between paying parents and CMS.

It also lists service requests such as changes of circumstances, employment changes and liability changes.

This creates an important transparency question:

Could the frequency or type of contact a person has with CMS affect how their case is scored?

The published record identifies communications as part of the dataset but does not provide the public with a complete list of every engineered model feature, its weighting or exactly how an individual communication affects a particular person's score.

What technology is being used?

This system is live — not merely experimental

The Government record lists the system's phase as “Production”.

DWP says it was rolled out to a wider group of operational caseworkers in March 2025.

The model is run at the beginning of each month and around 100,000 cases are processed.

The resulting prediction is then available within the CMS2012 system used by caseworkers.

What does CMS receive about a parent?

The published record says each prediction can contain:

• A risk score
• A High, Medium or Low risk level
• The date the prediction was created
• An expiry date
• Information supporting the prediction
• Potential actions for the caseworker

These predictions are displayed to CMS caseworkers as part of the case-management process.

DWP itself identifies the risk of incorrect predictions

Published performance

DWP published the following results based on data from May 2025:

• Accuracy: 0.86 / 86%
• Precision: 0.79 / 79%

DWP also says a newly trained model will not be accepted unless its precision is at least 0.62 / 62%.

What about discrimination and bias?

DWP states that the model dataset does not contain sensitive attributes.

Its Equality Assessment says protected characteristics are not impacted and that the machine-learning system has been trained to reduce or eliminate bias.

However, the transparency record also says:

This does not prove that discrimination is occurring. But it is a legitimate transparency issue when assessing how confidently differential effects can be ruled out.

Does the computer make the final CMS decision?

That distinction matters.

The published information does not establish that CMS is allowing an algorithm to automatically impose enforcement action or automatically change somebody's maintenance liability.

However, the system can determine which cases are highlighted as being at greater risk and therefore which cases receive earlier attention from caseworkers.

Right First Time UK believes there should therefore be clear accountability showing that the required human review is meaningful and that staff are not simply relying upon the automated score.

Questions that now need answering

• Are parents told when their case has been processed by this predictive system?
• Are parents told whether they have been classed High, Medium or Low risk?
• Can a parent see their actual risk score?
• Can a parent see which factors contributed to that score?
• Can a parent challenge inaccurate information before it affects how their case is handled?
• Are caseworkers recording that they independently checked the prediction?
• How often do caseworkers disagree with the algorithm?
• How many parents identified as high risk actually remain fully compliant?
• What specific interventions are triggered by a High-risk prediction?
• Are parents contacted differently because of an algorithmic prediction?
• Have any enforcement processes indirectly followed from a case first being prioritised through the predictive system?
• How does DWP assess whether historic errors already held within CMS2012 could contaminate the prediction?

Bad data in = potentially bad prediction out

This is particularly important for CMS because the machine learning model depends upon information already held within the Child Maintenance system.

If underlying information about a case is wrong, disputed, incomplete or out of date, an obvious question is:

Could that incorrect information also influence the predictive risk score?

The Information Commissioner's Office says organisations must take reasonable steps to make sure personal data is not incorrect or misleading and must carefully consider challenges to its accuracy.

Can you ask DWP what information it holds about you?

You have a right to ask DWP for copies of personal information it holds about you through a Subject Access Request (SAR).

Where relevant, a person may wish to ask DWP whether information connected with the Predictive Analytics Compliance system forms part of their personal records.

Questions you may consider including in a SAR

1. Has my CMS case been processed through the Predictive Analytics Compliance model?
2. Please provide any risk score recorded in connection with my case.
3. Please provide any High, Medium or Low risk classification associated with my case.
4. Please provide the creation and expiry date of any prediction.
5. Please provide the information or factors presented to caseworkers as contributing to the prediction.
6. Please provide any potential actions generated or displayed alongside the prediction.
7. Please provide caseworker notes showing how the prediction was reviewed or verified.
8. Please provide records showing whether any action taken on my case was influenced by or connected with the predictive score.
9. Please provide any audit trail identifying when predictive information concerning my case was accessed.

Whether every requested item must be disclosed will depend on data-protection law and any applicable exemptions, but asking specifically for the information can help establish what personal information DWP holds.

Data protection and profiling

The Information Commissioner's Office provides guidance covering automated decision-making and profiling.

Profiling broadly includes automated processing of personal information used to evaluate or predict aspects of a person's behaviour, reliability or economic situation.

Special protections apply to decisions that are made solely by automated means and have legal or similarly significant effects.

DWP's published position in relation to this CMS system is that the predictive score is not itself the decision and that a human caseworker makes the decision about how the case is managed.

That makes the quality and independence of the human review especially important.

Impact assessments

The official transparency record identifies several assessments connected with the project.

The transparency record provides DWP's summary of these assessments. Anyone investigating the system may also wish to ask whether the complete assessments can be published or disclosed.

An external technology supplier was involved

The official record identifies:

Tata Consultancy Services Limited (TCS)

DWP says TCS provided technical expertise supporting the design, build and assurance of the tool and supports the live training and inference pipelines.

DWP states that the supplier does not have direct access to personal customer data in the live environment.

However, DWP says information such as logs or screenshots may be supplied to the contractor where needed to resolve incidents or problems.

Where is the model data stored?

According to the transparency record:

• The system uses AWS infrastructure.
• Data is stored in AWS S3.
• DWP says the stored data is encrypted.
• DWP Digital Children data engineers have access to the dataset.
• The stated retention period for the training dataset is one year.
• Caseworkers do not receive the raw model dataset.
• Caseworkers instead see the predictions presented in CMS2012.

If the underlying CMS information is wrong

A predictive system cannot correct a person's case simply by analysing it.

If you believe CMS holds incorrect information about matters such as:

• your income;
• payments;
• arrears;
• employment;
• liability;
• changes of circumstances;
• payment method;
• communications; or
• other factual case information,

raise the underlying error directly with CMS and keep copies of the evidence showing why you believe the information is wrong.

If you disagree with an actual CMS decision

If CMS has made a formal decision concerning your maintenance calculation which you believe is wrong, the normal challenge procedure may involve asking for a Mandatory Reconsideration.

GOV.UK says this can be requested where, for example, you think an error has been made, important evidence has been missed or you disagree with the reasons for the decision.

There are time limits, so always check your decision letter and the current official guidance.

What Right First Time UK believes should now be published

Genuine algorithmic transparency should allow the public to understand not simply that a system exists, but how it affects real people.

We believe further information should be published including:

• The number of cases rated High, Medium and Low risk each month.
• The percentage of High-risk cases which actually subsequently miss a payment.
• The number of false-positive predictions.
• The exact interventions carried out for each risk category.
• How often caseworkers disagree with the prediction.
• Whether a risk prediction has ever contributed to moving a case towards enforcement.
• A complete list of the features used by the model.
• The importance or weighting of those features.
• How disputed CMS data is treated by the algorithm.
• How corrected information is fed back into future predictions.
• Whether parents are routinely informed that their information has been used to create a predictive risk score.
• The complete Data Protection Impact Assessment, subject to any genuinely necessary redactions.
• The complete Equality Assessment and subsequent reviews.
• Independent testing of the system's accuracy and possible differential impacts.

Why this matters

Technology can potentially help public bodies identify problems earlier and provide a better service.

But when predictive technology is used by a public authority with significant powers over people's finances, transparency, accuracy and meaningful human oversight are essential.

A parent who has paid correctly should not be treated as though they have done something wrong simply because a computer predicts that they might not pay in the future.

Likewise, caseworkers need reliable information rather than automated assumptions based upon inaccurate or disputed historical data.

Predictive analytics should assist accurate decision-making — not amplify existing errors.

Official sources and further reading

Right First Time UK – CMS Information Hub

We are continuing to investigate the use of technology, data, decision-making systems and administrative processes within the Child Maintenance Service.

If you have experienced incorrect CMS data, unexplained decisions, repeated calculation errors or believe automated systems may have affected the way your case was handled, keep your records and evidence.