Tag Archive for: agentic AI

Is Your AI Actually an Agent? The Agentic Spectrum and What It Means for Your Family Court Case

Not every AI tool that looks intelligent is acting as an agent. Sometimes it is simply answering a question. Sometimes it is reading documents you have uploaded. Sometimes it can use tools or interact with another application. And at the other end of the spectrum, an AI system may be capable of working through an entire multi-step task with relatively little intervention from you. For litigants in person, understanding that difference matters because the more an AI system is permitted to do, the more carefully you need to think about its instructions, evidence, access, permissions and supervision.

1. What actually makes an AI system an agent?

The easiest distinction is this:

That does not mean every AI product fits neatly into one box.

Modern AI systems can behave very differently depending on what you ask them to do and what access you give them.

You might use the same AI application to ask:

That is essentially chatbot use.

Five minutes later, you might give the same system a folder containing previous court orders, position statements and hearing notes and ask it to identify every issue that remains unresolved.

Now it is doing something substantially more involved.

If it can then create a structured issues document, place it into your workspace and continue through further tasks without you manually prompting every step, the workflow becomes more agentic again.

So we should stop thinking simply in terms of:

Chatbot OR Agent

and start thinking in terms of:

How much independent action is this system taking towards the outcome I have asked for?

2. The four things to look for

A useful way of understanding agentic AI is to break it down into four practical components:

Those four components are worth understanding because each one creates both an opportunity and a risk.

3. First: the goal

A traditional chatbot normally has a very short-term objective:

answer the question in front of it.

An agent can instead be given an outcome.

For example:

That is not merely a request for information.

It defines a product.

The quality of the goal matters.

Compare that with:

The second instruction is dangerous because it begins with the conclusion you want and invites the system to select material that supports it.

That is not evidence analysis.

It is a recipe for confirmation bias.

4. Second: context — what does the AI actually know?

AI cannot properly analyse evidence it has never seen.

That sounds obvious.

But it is one of the easiest mistakes to make.

Imagine you upload:

• your statement;
• your chronology;
• your emails;
• and your account of the history.

You then ask the AI:

The system may give you a confident answer.

But if it has only been given one side’s material, it does not possess the evidential context required to make a balanced assessment.

That does not mean your documents are wrong.

It means the dataset is incomplete.

This is especially important in family proceedings where there may be:

• two competing factual accounts;
• Cafcass material;
• school records;
• medical records;
• police disclosure;
• social work records;
• historic orders;
• findings already made by the court;
• allegations that have not been determined;
• and events about which nobody yet has complete information.

Good AI use therefore requires you to understand the limits of the context.

Sometimes the correct output is not:

“This proves X.”

It is:

“On the documents currently available, this appears to suggest X, but Y material has not yet been reviewed.”

That distinction is much closer to responsible evidence analysis.

5. Third: the workspace — where can the AI actually go?

This is one of the features that makes modern agentic systems very different from earlier chatbots.

An AI system may increasingly be able to interact with the places where the work itself is stored.

Depending on the technology and permissions involved, that might include:

• documents;
• folders;
• email;
• calendars;
• spreadsheets;
• cloud storage;
• case-management systems;
• task-management applications;
• databases;
• web browsers;
• and other connected services.

This is tremendously powerful.

It also changes the risk calculation.

6. Fourth: continuous action

A chatbot normally answers and then waits.

You ask a question.

It responds.

You decide what happens next.

A more agentic system can instead continue working through a sequence.

In a family case, a hypothetical workflow might be:

Instead of asking the AI seven separate questions, you have given it one outcome and allowed it to work through the component steps.

That is a much more agentic workflow.

It is also why the accuracy of the original instructions matters so much.

An error in step one can potentially travel through steps two, three, four, five and six.

Automation can therefore scale mistakes as efficiently as it scales good work.

7. The agentic spectrum: five levels of AI use in a family court case

Instead of trying to decide whether something is “an agent” or “not an agent”, I think a much more useful approach for litigants in person is to understand the spectrum.

These examples are deliberately simplified.

The important point is that the degree of agentic behaviour depends on what the system is actually doing, not merely the name of the product.

8. The same AI tool can sit at several different points on the spectrum

This is one of the most useful concepts to understand.

People often talk about particular products as though:

“This one is a chatbot.”

or:

“That one is an agent.”

Increasingly, that is too simplistic.

The same system may behave like a basic chatbot in one task and much more like an agent in another.

The useful question is:

That question becomes especially important in legal work because permissions can carry consequences.

9. Permission is part of risk

Suppose an AI system incorrectly concludes that a particular document is irrelevant.

If it merely tells you that in a conversation, you can disagree.

If it has permission to reorganise your working folder and moves that document into an “irrelevant” archive, the error has had a practical consequence.

If it has authority to delete documents, the risk is greater again.

The same principle applies to communication.

There is a major difference between:

• AI drafting an email for you to review;
• AI creating an email draft in your account;
• AI sending an email after you approve it;
• and AI independently deciding that an email should be sent and sending it.

Those are not equivalent uses.

In a high-conflict family case, one poorly worded communication can be exhibited months later and scrutinised by Cafcass, solicitors or the court.

That makes human oversight extremely important.

10. Access to evidence is not the same as understanding evidence

This distinction is fundamental.

An agent may be capable of reading every document in a case folder.

That does not mean it understands the case in the way a human being does.

It may not automatically know that:

• a later order superseded an earlier one;
• a disputed allegation was never proved;
• a phrase in a Cafcass report records what somebody said rather than an independent finding;
• a child’s words were reported second-hand;
• a document is incomplete;
• a chronology was written by one party and is not itself proof of the events within it;
• a communication needs to be understood against a wider pattern;
• or a document that appears minor is legally or evidentially significant.

This is why instructions about evidential status matter.

In family litigation, those categories should not be casually collapsed into each other.

11. Why this matters particularly in domestic abuse and coercive-control cases

Pattern evidence matters in domestic abuse cases.

Coercive or controlling behaviour may consist of repeated actions which appear relatively minor when isolated but take on a different significance when considered cumulatively and in context.

AI can potentially assist by locating repeated events across large bodies of material.

For example, it might help identify:

• repeated changes to arrangements;
• conditions imposed on communication;
• recurring threats or financial pressure;
• patterns around handovers;
• repeated references to monitoring or surveillance;
• persistent interference with communication;
• or inconsistencies between contemporaneous communications and later accounts.

That can be useful.

But there is an equally important safeguard.

You should not tell an AI system:

That instruction risks turning evidence review into evidence selection.

A safer approach would be:

The human can then consider the pattern against the legal framework.

AI should help expose the evidence.

It should not manufacture the finding.

12. More autonomy means more need for supervision

There is an understandable temptation to think that better AI means less human involvement.

In legal work, I think the opposite principle is often safer.

The more consequential the task, the more carefully human oversight should be designed.

That is broadly consistent with the direction being taken across the legal system.

Current judicial guidance emphasises personal responsibility for AI-assisted material, the need to protect confidential information and the risks associated with inaccurate or fabricated AI output.

The Solicitors Regulation Authority has likewise warned that AI does not remove professional responsibility and has specifically highlighted inaccurate legal material, confidentiality risks and the need for appropriate human oversight.

Litigants in person are not regulated solicitors, of course.

But the underlying practical principle remains highly relevant:

“The AI did it” is not a sensible case-management strategy.

13. Litigants in person are a distinct AI issue

This is now being recognised at institutional level.

The Civil Justice Council’s work on AI in the preparation of court documents has specifically recognised that litigants in person present distinct and evolving issues.

That makes sense.

A lawyer using AI ordinarily has legal training, professional duties and some ability to test what the system produces.

An unrepresented litigant may be using AI precisely because they do not have access to that professional assistance.

Some may rely very heavily on the output.

That creates a difficult access-to-justice tension.

The people who may benefit most from AI assistance may also be among those least equipped to detect a sophisticated AI error.

That is why AI literacy matters.

Not technical literacy in the sense of learning computer science.

Practical literacy:

• What has the AI actually seen?
• What has it not seen?
• What can it do?
• What can it change?
• What assumptions is it making?
• Where did this conclusion come from?
• Can I trace this statement back to evidence?
• Can I verify this law?
• Am I still making the important decisions?

14. A practical example: preparing for a hearing

Imagine you have a hearing in three weeks.

You have:

• three previous court orders;
• two witness statements;
• a Cafcass report;
• approximately 200 emails;
• a contact log;
• school correspondence;
• and your notes from earlier hearings.

Here is how the spectrum might look.

None of those uses is automatically good or bad.

What matters is whether the task is appropriate, whether the system has the right information, whether the permissions are proportionate and whether the result is properly supervised.

15. Do not automate the wrong thing

AI agents can make repetitive tasks significantly easier.

That does not mean every repetitive task should be automated.

Some activities deserve deliberate human involvement precisely because they involve judgement, emotion, safeguarding or important consequences.

I would be particularly cautious about automating:

• decisions about whether an allegation should be made;
• decisions about whether contact is safe;
• interpretation of a child’s wishes and feelings;
• responses to safeguarding disclosures;
• communications written during conflict;
• decisions about withholding evidence;
• final witness evidence;
• and strategic decisions about what orders should be sought.

AI may assist with the information surrounding those decisions.

It should not quietly become the decision-maker.

16. The future skill is supervision

There is a tendency to talk about AI skills as though the important thing is learning clever prompts.

Prompting matters.

But I think the deeper skill is going to be supervision.

Can you define the task?

Can you provide the correct context?

Can you decide what the system should and should not be permitted to do?

Can you recognise a weak output?

Can you identify missing evidence?

Can you distinguish a source from an inference?

Can you intervene when the system goes in the wrong direction?

Can you verify what matters?

Those are management skills.

And increasingly, they are becoming AI skills too.

17. What this means in practice for litigants in person

You do not need to understand computer science to use these tools intelligently.

You do need a mental model.

When you use AI on your case, think about four things:

If you cannot answer those questions, you probably should not be giving the system greater autonomy yet.

Conclusion: it is a dial, not a switch

The useful question is no longer simply:

“Is this an AI agent?”

A better question is:

“How agentic is this workflow, and how much control have I handed over?”

At one end, AI may simply answer a question.

Further along, it can analyse your documents.

Further still, it can interact with your digital workspace.

And increasingly, it can carry out multi-step work towards a goal with much less human intervention.

That has enormous potential for litigants in person.

It could help people manage evidence, orders, chronologies, deadlines and enormous quantities of information that would otherwise consume days or weeks.

But autonomy changes risk.

The more an AI system can do, the more carefully we must decide what it should do.

In family justice, the objective should not be autonomous litigation.

It should be something much more useful:

technology doing more of the mechanical work while the human being retains control of evidence, judgement, safeguarding and decisions that matter.