AI for Speech Therapy Reports: Draft, De-Identify, Review

De-identify before you draft. Replace the child's name, age, school, dates and places with placeholders, keep test names, scores and items out, let AI structure only your own findings into the report headings, then put the real details back yourself and review every sentence against your notes before you sign.

Most advice on AI and reports starts with the drafting prompt. That is the wrong end. The prompt is the easy part; the order of steps is what keeps a child's identity, a test publisher's material and your professional judgement where they belong. This tutorial puts the protective steps first and the drafting second.

Follow me on Instagram@sagnikteaches

Step one: decide what the assistant may never see

Write the exclusion list before you open any chat. It should be short enough to remember.

Connect on LinkedInSagnik Bhattacharya

• Identity: name, nickname, date of birth, exact age, address, school, teacher, siblings, family names.
• Identifiers by combination: an unusual condition, a rare event at home, a distinctive local place, a combination of age and class.
• Test material: item content, protocols, record forms, scoring tables, and usually the scores themselves.
• Raw records: audio and video of the child, photographs, scanned forms.
• Third parties: names of other professionals, referrers and family members.

The test material point deserves a moment. Standardised assessments are usually sold under terms that protect their content. Pearson's guidance on appropriate use of its clinical assessment content, for example, lists scanning or photocopying stimulus images and uploading stimulus materials to a shared drive or website as not allowed, and says professionals may not store or share screen captures that include test content. Pasting items or protocol wording into a chat tool runs into the same problem, and the scores are the child's personal data on top. Read your own publisher's terms, and treat any doubt as a reason to keep it out. The assistant never needs the content of a test to improve the wording of a sentence about what a result means; it needs only a placeholder.

Subscribe on YouTube@sagnikteaches

Step two: a de-identification routine you can repeat

Treat de-identification like a checklist, not a feeling. A quick routine for a working copy, in about ten minutes:

1. Copy your findings into a new blank document. Never edit the clinical record.
2. Use find-and-replace for the name and nickname, replacing with [child].
3. Replace pronouns only if the assistant needs gender for the wording; otherwise use "they" and add the right pronoun yourself at the end.
4. Replace age with [age band], such as "early primary school age", or delete it.
5. Replace the school, teachers and every place name with [school], [teacher], [town].
6. Replace every date with [date] or a relative phrase such as "at assessment".
7. Replace every score with [result] and every test name with [test].
8. Read the whole copy again as a stranger would. Could a person in this community recognise the child from the story alone? If yes, generalise further.
9. Keep a private key listing what you replaced, in your own system, not in the chat.

The final read is the step people skip, and the one that matters most. A sentence such as "[child] recently moved after a family bereavement and attends the only [school] in [town]" is still identifying in a small place even with placeholders. Rewrite it as "[child] has had a recent change in circumstances", or drop it if it is not needed for the findings.

A filled-in before and after

Before (capitals stand in for the real name, school and test): "FIRSTNAME, 4;7, assessed on 12 March at SCHOOLNAME. Expressive language standard score 71 on TESTNAME. Mum reports he rarely speaks at nursery since his sister was born. Observed to use 2-word combinations in play with the blocks."

After: "[child], [age band], assessed on [date] at [setting]. Expressive language result: [result] on [test]. Parent reports [child] speaks less at nursery since a recent change at home. Observed to use two-word combinations in play."

The "after" keeps everything the assistant needs to word a paragraph on expressive language, and gives away nothing that lets someone find the child. Notice that "a recent change at home" generalises the sibling detail. Whether to mention it in the final report is your clinical decision, and you add it back in your own document if relevant.

Step three: choose the tool and settings before you paste

Even a de-identified outline is clinical information, so choose the account with care. The cleanest option is a business plan from a main vendor, where the terms say business content is not used for training by default, set up under the practice's own login with the retention and sharing settings checked. Vendor terms change, so read the current ones and note the date. A personal free account is a poor choice for practice work.

• Confirm the plan's training setting, and turn off chat history if you do not need it, noting that retention windows differ by vendor and setting.
• Use one fresh chat per report, so details never drift from one child to another.
• Do not connect the tool to your email, drive or clinical software for this job.
• Delete the chat when the report is final, if your policy says so.

Ask your insurer and professional body whether they have views on AI tools in clinical documentation, and write down what they said and when. For a checklist you can use on any vendor, see AI note-taking tools for therapists: a confidentiality checklist.

Step four: draft from your findings, not from the assistant's ideas

Now the drafting. A good report prompt gives the assistant a structure and strict limits.

You are helping a speech and language therapist structure an evaluation report. Use ONLY
the notes between the markers. Do not add findings, causes, diagnoses, predictions or
recommendations that are not in the notes. Keep all square-bracket placeholders exactly as
they are. Write under these headings: Reason for assessment; Background (brief); Assessment
methods; Findings: understanding of language; Findings: spoken language; Findings: speech
sounds; Findings: social communication; Summary; Recommendations. If a heading has no
supporting notes, write "No information provided" under it. After the report, list every
sentence you were unsure about under "Questions for the clinician".
START
[child], [age band], referred because of concerns about speech clarity. Assessed on [date].
Methods: [test] and play-based observation. Understanding: follows 2-step instructions in
play, not reliably for 3-step. Spoken: mostly 3-4 word sentences, some grammar errors with
past tense. Speech: substitutes /k/ with /t/, final consonants often missing. Intelligibility
about half to unfamiliar listeners per parent. Social: good eye contact, takes turns in play.
Recommend therapy for speech sounds, parent programme, review in [period].
END

Notice "No information provided". That instruction guards against the assistant filling an empty section with plausible text, the single most dangerous habit of these tools in clinical writing. Notice too the request for a question list: it makes gaps visible instead of hidden.

An illustrative extract of the output, under Findings: speech sounds: "[child] substitutes /t/ for /k/ and frequently omits final consonants. Intelligibility to unfamiliar listeners is reported by a parent as about half." What you would fix: check that "about half" is attributed to the parent, as the notes say, and not stated as a measured result; and check that the assistant has not added "which is typical for this age" or any developmental comparison you did not supply.

Common drafting faults

• Developmental claims you did not make. "Typical for a child of this age" appears often. Delete unless it is yours.
• Diagnostic labels. A label in the draft that you did not write needs your explicit decision, not acceptance.
• Parent report written as fact. Keep "parent reports" attributions.
• Recommendations that expand. The draft proposes extra sessions or referrals. Reduce to what you recommended.
• Smooth transitions that overstate links. "Because of this, [child] struggles at school." You have no evidence of cause. Remove.

Step five: put the details back, then review against the evidence

Return to your clinical system and open your own report template. Paste the structured draft into it, then replace every placeholder using your private key and the source records. Enter test names, scores and any standardised results yourself, in the format the publisher and your service require. Only now does the document refer to a real child.

Review has three passes, and each takes real time:

1. Evidence pass. For each sentence, find the note, observation or result behind it. Mark anything you cannot trace. Delete or evidence it.
2. Strength pass. Check that each claim is no stronger than its evidence. "Difficulty" is not "severe impairment". "Parent reports" is not "observed".
3. Family pass. Read it as the parent would. Is every term explained? Is anything alarming without being useful? Would you say each sentence to them across the table?

Then sign, save the version in the record, and add a line noting that AI was used to structure a de-identified draft, who reviewed it and when. If any step cannot be done, the report is not ready.

What lives where: a map of the whole job

The easiest way to keep the steps straight is to ask, for each piece of information, where it is allowed to be.

Pin that table beside your desk. It turns a vague worry into a yes-or-no question for every line you are about to paste.

A second case: a child with a stammer and a mixed picture

Not every report fits the tidy headings above. Suppose a child comes in with concerns about fluency and some word-finding difficulty. Your notes are less structured: "Repetitions on first sounds, about 1 in 10 words in conversation, blocks when under time pressure. Aware of it, avoids reading aloud. Word-finding pauses in picture naming. Parents anxious. Strong understanding."

The de-identified outline is shorter and the sensitivity is higher: the child's own awareness and avoidance are personal in a way a vocabulary score is not. Two adjustments help. First, leave out any detail that is not necessary to the wording, such as the parents' state of mind, and add it in your own document if you judge it relevant. Second, tell the assistant explicitly that fluency language is clinician-owned: "Do not suggest techniques, causes or advice to the family. Do not use the word 'cure' or any outcome language."

An illustrative output: "[child] shows repetitions of initial sounds in conversation and occasional blocks, noticed more under time pressure. [child] is aware of the difficulty and avoids reading aloud. Pauses were observed during picture naming. Understanding of language is a relative strength." Check the sentence on blocks: "noticed more under time pressure" is in your notes, so keep it. Check that no frequency has been invented, and that "1 in 10 words" is kept only if you measured it and say how. Add the real figure and method yourself from your own sample.

A plain-language summary for the family, drafted from the report

Many clinicians send a one-page summary alongside the full report. Once the report is signed, you can repeat the routine: de-identify the findings and recommendations, then ask for a family-friendly version. Give strict limits, because families read these closely and sometimes share them.

Rewrite the following findings and recommendations as a one-page summary for a parent.
Reading level about age 11. Warm, calm, factual. Use only what is in the text. Do not add
reassurance, predictions or advice. Explain any technical term in brackets the first time.
Keep placeholders exactly as written. End with the sentence: "Please ask me about anything
that is unclear."

What you fix: the assistant often adds a reassuring sentence, such as "Many children make good progress". You have not said that, and it can read as a promise. Delete it. The summary also tends to flatten uncertainty ("it may be that") into confident statements, so reinstate your hedges where the evidence was limited.

A worked example with the time and the sums

Say a solo therapist writes five evaluation reports a month. Without AI each takes about 2 hours 15 minutes from notes to signed PDF: 15 minutes pulling notes together, 80 minutes writing, 20 minutes formatting, 20 minutes checking. That is 11.25 hours a month.

With the workflow above: 15 minutes pulling notes, 10 minutes de-identifying, 10 minutes prompting and reading the draft, 20 minutes restoring details and formatting, and 45 minutes reviewing in three passes. That is 100 minutes, or 1 hour 40 minutes a report, and 8.3 hours a month. The saving is about 35 minutes a report, nearly three hours a month on five reports.

These are illustrative figures, and the honest lesson is in the review line. A therapist who cuts review to 15 minutes would appear to save more and carry the real risk. The point of the order (de-identify, draft, restore, review) is that the saving comes from the middle, and the protection comes from the ends.

The village-school line and the reused chat

An illustrative case. A therapist de-identified a report but kept the line "the only child in the village school using AAC". It read as harmless, since the name was gone. In a small community it identified the child at once. The fix is not a better tool, it is the stranger's read in step two: ask whether someone local could recognise the person from the story alone.

A second easy mistake is reusing a chat. A therapist wrote a second report in the same chat as the first, and the assistant carried over a phrase about "difficulty with past tense" into a child who had no such difficulty. The reviewer caught it only because the evidence pass required a note behind every sentence. One report, one chat.

When the report goes to a school or another professional

A report often travels: to a teacher, a paediatrician, a funding body or a new clinician after a move. Each reader needs a different cut of the same findings, and each cut is another chance to over-share or distort. AI can produce the cuts, but the rule is the same. Start from the signed report, de-identify, draft the shorter version, restore, review, and send only to the person the family has agreed may see it.

Check three things on every cut. First, consent: does the family's agreement cover this recipient and this level of detail? Second, relevance: a teacher needs strategies for the classroom and a plain account of strengths and needs, not family background. Third, tone: a school may read a deficit-heavy summary as a label. Ask the assistant for a strengths-first version, then check that it has not softened any need that matters for support.

An illustrative request for a teacher summary: "From these de-identified findings, write half a page for a class teacher. Three headings: What [child] does well; What gets in the way; What helps in class. Use only the text provided and no suggestions that are not already listed." If your recommendations did not include classroom strategies, the third heading will be thin, and the right response is for you to add your own strategies, not to let the assistant invent them.

Limits and honest costs of the approach

This workflow adds steps, and some clinicians will find it slower than they hoped. That is a fair reading. For very short, routine letters, it may be quicker to write by hand. The method pays for itself on longer evaluation reports with repeated structure, where the assistant's strength (organising and rewording stable material) matters most.

• It does not replace clinical writing skill. A weak formulation stays weak after tidying.
• It does not make a risky tool safe. De-identification reduces exposure, it does not remove it, and a vendor policy can change.
• It needs discipline. The routine only works if every step happens every time, including when you are tired.
• It can lull you. A polished draft invites a lighter review. The three passes exist to counter that.

If you find yourself skipping a step, do the report by hand that day. A late report is a nuisance, and an unreviewed one is a risk.

How to build this into a routine

• Save the prompt and the exclusion list as a template document next to your report template.
• Keep a log: date, report reference (not the child's name), tool, reviewer.
• Once a quarter, have a colleague check three reports against the evidence pass.
• Rerun the prompt on a fake case occasionally to see whether the tool behaves as it did, because vendors change models and behaviour.
• Reread your professional body's guidance twice a year.

If you also send parents practice sheets, the same principle applies: type the goal, never the child. For clinics writing funding-scheme reports, the NDIS reports and scheduling tutorial adds the privacy-principle steps. For redaction generally, see how to redact personal data from documents. The wider picture is in the AI guide.

A sign-off box you can paste at the foot of the working copy

De-identified working copy. Drafted with an AI writing tool from clinician notes only. No names, dates, scores, test content or recordings entered. Restored and checked by [clinician] against source records on [date]. Version [number].

That box lives in your internal file, not the report the family receives, unless your policy says otherwise. It answers, in one line, the question a reviewer or complaints officer would ask first.

Questions about AI and evaluation reports

Is removing the child's name enough to de-identify a report?

No. Age, school, rare circumstances, place names, dates and the combination of findings can identify a child in a small community. Replace or generalise all of them, and keep the key to reverse the changes in your own system, never in the chat.

Can I put test scores into an AI tool?

Keep them out. Scores are personal data, and many test publishers restrict copying items or protocols. Write [test result] placeholders in the draft and insert the real scores yourself in your own document, checking the publisher's terms for what you may reproduce.

Who is responsible for the report if AI wrote part of it?

You are. The clinician who signs it owns every statement. If you cannot trace a sentence to your notes or observation, delete it or find the evidence. Record in your file how AI was used and who reviewed the draft.

How long should review take?

Budget real time. For a typical evaluation report, plan on at least as long for review as you saved on drafting. A review that takes five minutes is usually a skim, and a skim is how invented findings reach a family.

Should the family be told AI was used?

If identifiable information is never put into an AI tool, there may be little to tell, but check your professional body and privacy notice. Many clinics state generally that clinicians may use writing tools and that all documents are clinician-checked and signed.

Further reads

• ChatGPT for Speech Therapy Practices: Safe Admin Uses and Hard Limits — Safe admin uses and hard limits for speech therapy practices.
• Will AI Work for a Small Speech Therapy Practice? Where to Start — Where a small speech therapy practice should start.
• ChatGPT for Tax Preparers: How to Redact a Return Before You Paste — A redaction method from another profession that transfers well.
• What Does AI Cost a Speech Therapy Practice? Three Budget Levels — What AI costs a practice at three budget levels.
• Telling Patients You Use AI Note-Taking: Consent Wording That Works — Consent wording for AI in clinical documentation.
• AI Scribe for Occupational Therapists: A Buyer's Checklist — Buying questions for a clinical scribe.
• Starting a Speech Therapy Private Practice: AI and the Admin Stack — The five admin systems a new speech therapy practice needs, which parts AI can draft, what each costs, and the order to build them for your first 30 clients.
• AI for Group Therapy Practices: Notes, Consent and Review Steps — Write one note per member from a group session without leaking other members' details: the rule, the consent wording and the review steps.
• AI in Occupational Therapy: Funding Letters and Parent Updates — A private-practice OT workflow for evidence-led funding letters and parent summaries that the therapist signs off.
• AI in Speech and Language Therapy Clinics: Booking and Waiting Lists — A clinic-owner workflow for approved appointment messages, cancellation offers and a carefully managed waiting list.
• NDIS Occupational Therapy Report Template: A First Draft With AI — Report headings, an evidence table, a de-identified drafting method and a review checklist for an AI-drafted NDIS OT report you stay accountable for.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business data page and Anthropic privacy centre article on model training (both checked 4 November 2026: business and commercial plans are not used for training by default); Pearson Clinical, Appropriate use of Pearson Clinical Assessment content (checked 4 November 2026); other test publishers' own terms of use, to be checked by the reader; professional-body guidance.