Children's Home Software: Which Systems Add AI and What to Check First

Of the children's home systems whose own pages I read in November 2026, Mentor Software and OVcare name AI features, while Lindi IMS names none. Before switching any on, a registered manager should check what data the feature sends out, who reviews its output, and whether the supplier says so in writing.

Notice the word "pages". Every claim below comes from what suppliers publish about themselves, and several of the headline features are still labelled limited rollout or coming soon. That gap between a brochure line and a switch you can flip tomorrow is where managers get caught, and closing it takes written answers from the supplier, not a demo.

Follow me on Instagram@sagnikteaches

What three suppliers say about AI, and what they leave out

A fair comparison starts with the label on each feature, not the adjective. Here is what the three pages I read stated, in my own words.

Connect on LinkedInSagnik Bhattacharya

Two things stand out. First, the same word, "AI", covers very different risk. An assistant that finds a paragraph in your own safeguarding policy and shows you the source is a smaller step than a tool that reads three weeks of daily logs and drafts a paragraph about a child. Second, a supplier with no AI is not behind. For a small home, tidy records and a good inspection pack matter more than a chat window.

Subscribe on YouTube@sagnikteaches

Mentor's pricing page is also worth a look on a different point: it publishes monthly starting prices by type of service (children's homes, supported accommodation and 18+ services), says a pre-registration home uses it free until inspection, and says longer contract terms save 10 to 20 per cent. That is unusual transparency, but the plan lists say nothing about AI, so ask whether the assistant comes with every plan or costs extra, and whether it is switched on during the free period.

Sort every AI feature into one of three risk bands

Before you open any supplier's settings, sort what is on offer. I would use three bands, because the checks differ.

Band one: finds things in your own documents

Policy question tools such as OVinsights, or an AI-enhanced search of a knowledge base. The input is a staff member's question and your policy library, not a child's record. The main risks are a wrong or out-of-date answer, and staff treating the answer as the procedure. The supplier says staff should check the guidance against their service's procedures, which is the right instruction.

One detail changes the checks for OVinsights in particular: its page says staff ask their questions through WhatsApp, by text or voice, and that managers can see the questions asked so they can spot training needs. That puts part of the workflow on staff phones. Before switching it on, decide whether staff use a work phone or their own, write a rule that a question never includes a young person's name or details, and tell staff that a manager can read what they asked. "What is our procedure when a young person comes back late and will not say where they were?" needs no name in it, and staff can be trained to phrase every question that way.

Band two: summarises records

Report summaries, "ask a question of the care record", end-of-day insights. Now child-identifying information goes into a model. The risks are omission (a quiet but significant entry is dropped), invention (a detail that is not in the record), and tone (a neutral note turned into a judgement).

Band three: judges or scores

Placement impact support, risk assessment support, trend analysis. These produce views about a child or about a mix of children. OVcare describes its own versions as limited rollout, with reviewers making the decisions. I would treat anything in this band as a pilot with a named reviewer, never as a default setting.

The pre-switch checks, in the order I would run them

1. Get the feature's status in writing. Ask the supplier to email you, per feature, whether it is generally available, limited rollout or in development, and the date it went live. A demo of something in development is not a product.
2. Ask where the data goes. Name of the model provider or providers, country of processing, whether the supplier or provider keeps prompts and outputs, for how long, and whether either uses them to train models. Ask for the answer in the contract or a data-processing schedule, not a chat message.
3. Ask what the AI can see. The whole care record, or only the screen you are on? Can it see records of children in other homes in the same account? For a multi-home provider that last question matters most.
4. Ask about switching it off. Per home, per user role, per feature. If you cannot disable it for one home while you pilot in another, that is a finding.
5. Ask for the audit trail. When an AI summary is created, is it logged with who asked, when, and which records were used? When a person edits it, is the edit logged?
6. Check the privacy notice for young people. Your own notice needs updating. Draft the plain-language version for children and the version for placing authorities and parents.
7. Run a data-protection impact assessment with your data-protection lead before switching on any band two or three feature. It is not a formality: it is where you record the lawful basis, the risks and who signed off.
8. Decide the review rule. One sentence: "No AI output is saved to a child's record, sent to anyone or used in a report until a named staff member has read the source records and corrected it."

Take the time on step two. It is common for a supplier to say a feature is "secure" and mean the connection is encrypted, which is a different question from whether a third-party model provider retains the text. Ask the specific question and keep the specific answer.

A worked pilot: a five-bed home and one band-two feature

Illustration only, with invented figures. A five-bed home has a registered manager, a deputy and nine support workers on rotating shifts. The manager spends roughly six hours a month preparing the quality review and the pack for the independent visitor: reading daily logs, incident entries and key-work notes, then writing a narrative. The supplier offers an AI report summary in limited rollout.

The manager sets a four-week pilot with these rules.

• One feature only: report summaries, for the manager and deputy only. Support workers do not get access.
• Records covered: daily logs, not medication records and not anything flagged as a safeguarding referral.
• Every summary is read against the source records before use. The manager keeps a simple tally of what she had to fix.
• At the end of week four she decides on the numbers below.

Read that table honestly. A fabrication in week one and tone changes every week are why the rule is "read the source". Ten minutes saved across around 20 summaries is 200 minutes, a little over three hours a month, and that is before counting the time to keep the tally. The saving is real but modest, so the decision is not obvious. If the supplier's contract is clear on data and the manager is comfortable, keep it for report drafting only. If any week produces an invented fact that a tired reviewer might miss, the home switches it off.

Sample wording: the questions to send a supplier

Copy and adapt this email. It is short on purpose, because suppliers answer short emails.

Subject: AI features: written confirmation before we switch anything on

Hello [first name],

Before we enable any AI feature in [system name], please confirm in writing:

1. Each AI feature available to us, and whether it is generally available, limited rollout or in development.
2. The model provider(s) used, and the country where our data is processed.
3. Whether prompts, records or outputs are kept by you or the provider, for how long, and whether they are used to train any model.
4. Exactly which records each AI feature can read (current child only, whole home, whole account).
5. Whether we can disable each feature per home and per user role.
6. Whether AI use is recorded in the audit trail, including who requested it and which records were used.
7. Where this is covered in our contract or data-processing terms.

We will not switch any feature on until we have these answers.

Kind regards,
[your name], Registered Manager

What a good reply looks like, and what a weak one looks like

A sample strong answer to question 3 reads: "Prompts and outputs are processed in region X, not stored by the model provider beyond the request, and not used for training; see schedule 2, clause 4." It is specific, points to a clause and has a limit on retention.

A weak answer reads: "Your data is always secure and never shared inappropriately." That sounds reassuring and answers nothing. Send it back with the same question numbered again. If the second answer is the same, you have learned something about the supplier.

A middle answer is also common: "We use a leading provider and data is not used for training." Good as far as it goes, but ask for the provider's name and whether you can see the terms. A claim you cannot check is a promise, not evidence.

Staff guidance for the first month

Features fail in practice less often because of the software and more often because of habits. Give staff one page. Here is a filled-in version to adapt.

AI in our recording system: what you may and may not do

You MAY: ask the policy assistant a procedure question and check the source link it shows.
You MAY: ask the registered manager or deputy to use a summary to prepare for a review.
You MAY NOT: paste any child's name, details or record into any AI tool outside our system.
You MAY NOT: save, send or quote an AI summary as fact without reading the records it came from.
If a summary is wrong: tell the manager and note what was wrong so we can report it to the supplier.
If you are unsure: do not use it, and ask.

Where AI adds risk that a paper-light system does not

Three realistic failures are worth rehearsing in a team meeting, with an example for each.

The smoothing effect. A daily log says "refused to come downstairs for dinner, shouting about a phone call, staff gave space". A summary reads "had a difficult evening". The detail that mattered for a pattern, the phone call, has vanished. The fix is to make summaries link back to dated entries and to read the source whenever a summary feeds a review.

The second is false neutrality. A model told to produce a balanced summary may present a serious allegation and a minor grumble as equally weighted. A human reader knows which one triggers an escalation. Keep safeguarding entries out of any AI pilot until you have lived with the tool for months and your policy says otherwise.

The confident invention. A tool asked "when did the last missing-from-home episode happen?" returns a date. If the date is wrong and nobody checks the log, an inspector's question gets the wrong answer. The check is cheap: open the entry the answer cites.

Two sample outputs, and what you would fix in each

Examples help more than principles when you train a team, so here are two illustrative outputs of the kind a band one and a band two feature might give. They are invented to show the review habit, not copied from any product.

A policy question (band one). A support worker asks: "What do I do if a young person is missing and it is past curfew?" The assistant replies with three bullet points and a link to "Missing from Home Procedure, version 4". The review: open the link, check it is the current version and that the bullets match the numbered steps. In this sample the assistant has left out the step about informing the placing authority. The fix is to correct the document so the step is a heading in its own right, because tools tend to skip text buried in the middle of a long paragraph. The tool has just told you something about the quality of your policy.

A daily-log summary (band two). Here is a before and after.

Source entries (three days, abridged)
Mon 18:40  Came home late from college. Said nothing about the day. Ate dinner in room.
Tue 20:15  Phone call with sibling ended with shouting. Went to room. Staff checked at 20:30, calm.
Wed 07:30  Refused breakfast. Said "nobody listens". Attended college after key worker chat.

Draft summary
"Had a settled few days, attending college and eating regularly."

Almost everything in that summary is wrong in emphasis. The corrected version a reviewer would write keeps the sibling call, the refusal of breakfast and the sentence "nobody listens", and suggests a key-work conversation about contact with family. The lesson to teach staff: a summary of a difficult week that sounds calm should make you suspicious, not relieved.

What to do if a feature misbehaves after you switch it on

Have the plan before you need it. Switch the feature off for the affected home, note the date and a description of the fault, and keep the example record so the supplier can reproduce it. Then review whether any AI-drafted text has already gone into a record or a report, and correct it. Report the fault in writing to the supplier and ask for their response in writing. If personal data may have reached the wrong place, follow your own data-breach procedure and speak to your data-protection lead at once, because deadlines for notifying start from when you become aware.

Keep a short log of every fault: date, feature, what went wrong, what you did, and the supplier's reply. At your next policy review that log is the best evidence of whether the tool earns its place.

Fitting the decision into your inspection and review cycle

An inspector will not ask whether you use AI. They will ask how you know what is happening for each child, how you oversee records and how you learn from incidents. Any AI feature you adopt has to improve those answers, not replace them. A practical test: could you explain to an inspector, in two sentences, what the tool does and who checks it? If not, hold off.

For the monthly independent visit and the periodic quality review, AI is most useful as a drafting aid. The visitor's report starts from your records, the manager adds judgement, and the final text is human. If you want a starting structure for that report, this template shows how visit notes become a draft with AI.

Which home should look at which option

These are leanings, not rules.

• A solo home that is pre-registration. A supplier that is free until inspection, as Mentor states for pre-registration homes, lets you learn the system before you carry costs. Leave AI off during set-up. You have enough to build without it.
• A home with an established record system and few queries. If staff mainly ask where things are in the policy library, a Band one tool such as OVinsights-style policy search is the safest first step.
• A group with several homes. Insist on per-home switches and an answer about cross-home visibility. The risk of one home's records reaching another's query is the first thing to rule out.
• A provider happy with Lindi or a similar simple system. You may need no AI at all. If you want it later, ask the supplier for a roadmap in writing and compare it against suppliers that already ship it.

The adult-care version of the same exercise is in this care home software AI checklist, which is worth reading if your group runs both kinds of service. For a staged plan to roll anything out, the 90-day plan for registered managers is a good frame, and for how the privacy side should read in plain words see children's data in AI tools.

How to tell whether the feature is worth keeping

After the pilot, ask four questions and write the answers down.

1. Did it save time after the checking? If the net saving is under an hour a month, the extra risk is hard to justify.
2. How many errors did reviewers catch, and of which kind? Invented facts and omitted entries weigh more than clumsy phrasing.
3. Would a tired colleague on a busy shift have caught them? If not, restrict the feature to the manager.
4. Does the supplier's written answer on data still stand up when you read it a second time?

Decide keep, restrict or switch off, and put the date of the next review in the diary. Features change quickly, and a supplier's status page that said "limited rollout" in November can say something different in spring.

If you would rather map this out with someone, that is what my AI implementation consultation covers, through the contact page. Otherwise the email and staff page above are enough to begin.

Questions managers ask before switching AI on

Can a children's home use ChatGPT instead of AI built into its software?

You can, but it moves a child's information to a tool outside your care-record system, usually with different retention and training terms. Many homes ban it for anything identifying a young person. Built-in features at least stay inside the supplier's contract and access controls, which you can read and audit.

Should young people be told when AI summarises their records?

Yes, in plain words suited to their age and understanding. Add it to the home's privacy information, explain it at admission and in key-work sessions, and record any objection. Whether a child can object depends on the lawful basis you rely on, so ask your data-protection adviser.

Is an AI summary allowed to replace a manager's own review?

No. A summary is a draft to speed reading. The registered manager still reads the underlying records, corrects errors and signs off the report. Write that into your policy and keep a note of what was changed, so inspectors see a person made the judgement.

What if the supplier will not tell me where AI data goes?

Treat that as a no for now. A supplier who cannot say which model provider processes records, where, and whether data is retained or used for training has not finished its own homework. Switch the feature off and revisit at your next review.

Further reads

• AI Scribe Security: Audio, Retention and BAA Vendor Questions — Security questions that transfer to any AI feature.
• Safeguarding Policy Template for a Children's Club: Draft It With AI — A safeguarding policy structure you can adapt.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in a supplier's privacy terms.
• Questions to Ask an AI Vendor Before You Sign Anything — A general question list for any AI supplier.
• Patient Data and AI: A Confidentiality Checklist for Small Practices — A confidentiality checklist from a health-care setting.
• AI Chatbot for Healthcare: A Small Clinic's Data and Consent Checklist — Consent wording ideas that carry over to care settings.
• Home Care Agency Software Pricing: What the AI Features Add — What home care software costs on its own, what the AI features add on top, and how to price a 15-person agency three ways.
• Fostering Agency Software Checklist: What to Check and Where AI Fits — A buyer's checklist for small fostering agencies: records, panel papers, supervision logs, access controls, exit terms and the narrow places AI belongs.
• CQC Statement of Purpose Template: Draft Yours With AI and Check It — Draft a CQC statement of purpose with AI from facts you supply, then check every line against your registration, sign it off and set review dates.
• Ofsted Supported Accommodation: Registration Prep Checklist With AI — A document-by-document prep checklist for registering supported accommodation with Ofsted, with the drafting jobs AI can speed up and what it must not touch.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Mentor Software, OVcare and Lindi IMS product, AI-features and pricing pages (viewed 4 November 2026); vendor wording is paraphrased as published.