Put a boxed plain-English summary at the front, draft it with AI from the finished technical text only, then check every sentence against its source clause: numbers, qualifiers such as "may" or "consistent with", limits and recommendations must survive unchanged. The engineer reviews each sentence, and anything the AI added, softened or strengthened gets cut.
Plain English rarely fails loudly. It drops a qualifier, merges two findings, blurs a number, or adds reassurance the report never gave, and each version still reads well. Every example here uses a sample or blank report, never a real client's. The technical report is finished and checked first, and the summary is built last, so the AI never has a chance to shape the findings.
Where the summary box sits, and the blank report around it
The summary belongs on the first page after the title block, before the technical sections, so a client who reads nothing else still reads the limits. This is a blank skeleton to adapt to your own house style:
[Title block: project, client, date, reference, author]
SUMMARY FOR THE CLIENT (box; read with the full report; the full report prevails where they differ)
1 Instruction and scope
2 Limits of the inspection and sources of information
3 Observations (numbered, with photo and drawing references)
4 Discussion
5 Conclusions and recommendations
Appendices: A photographs, B drawings and monitoring locations, C glossaryTwo decisions here matter more than the layout. First, the summary is the last thing written, from a report the engineer has already approved. Second, the box carries its own precedence line: the summary must be read with the full report, and the full report prevails. A client can quote a summary without the limits, and that line is your protection when they do.
What the summary box holds: seven short headings
Keep the whole box to roughly 250 to 350 words. Each heading answers one question the client would otherwise ask by phone:
- What we were asked. One sentence, in the client's terms.
- What we looked at. What was inspected, when, and how, with the inspection type named ("visual inspection only").
- What we found. Facts only, with the numbers and locations from the observations section.
- What it may mean. The discussion, with the report's own uncertainty words.
- What we recommend. Actions in order, each with who, when and any condition.
- What this report does not cover. The limits, close to the report's own words.
- How sure we are. A plain statement of what is established and what is not.
Headings 6 and 7 are the ones AI drafts quietly erode. The limits shrink to "visual inspection" and the uncertainty disappears into confident phrasing. That is why the prompt below treats the limits as fixed text.
A readability score can help with the other headings, but it cannot check meaning. In Word for Microsoft 365 (Microsoft's page says this needs the Current Channel), the Editor button on the Home tab opens Document stats, which shows Flesch Reading Ease and Flesch-Kincaid grade level. Microsoft's help page suggests a reading ease of 60 to 70 and a grade level around 7 to 8 for most documents. Technical sections will sit well below that, so run it on the summary box only. A box that scores 65 and says "caused by" where the engineer wrote "consistent with" is readable and wrong.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowSix ways plain English shifts engineering meaning
These are the drifts to look for, with illustrative examples of each. Teach them to anyone checking summaries in your office.
| Drift | Technical text | Plain-English version that goes wrong | What is lost |
|---|---|---|---|
| Climbing the certainty ladder | "The cracking is consistent with differential foundation movement." | "The cracks are caused by the foundations sinking." | "Consistent with" is not "caused by", and "sinking" names one mechanism the report did not choose |
| Dropping the scope | "No evidence of distress was observed in the areas accessible on the day." | "There is no sign of distress." | The limitation to accessible areas and to one day |
| Blurring a number | "Crack widths up to 5 mm." | "Some cracks about half a centimetre or more." | The ceiling of the range; "or more" removes it |
| Merging two findings | "Clay soils may contribute. A tree stands about 6 m from the rear wall." | "A nearby tree may be making this worse." | Two separate statements fused into a cause the report never claimed |
| Adding comfort | (nothing) | "There is no immediate danger." | An opinion on safety the engineer never wrote |
| Turning a precaution into a conclusion | "Do not commission underpinning until the monitoring results are reviewed." | "Underpinning will not be needed." | The condition; the plain version reverses the meaning |
The last two rows are the dangerous ones, because they read well. A client who is told there is "no immediate danger" will act on it. If the engineer did not say so, nobody did.
A sample extract to work from
The rest of this tutorial uses an invented extract from a sample report on a two-storey house with cracking at the rear. Nothing in it comes from a real project. The clause numbers are what the tags and the ledger refer to.
| Clause | Technical text (sample) |
|---|---|
| 2.1 | Visual inspection only, on [date], from ground level and from the internal rooms that were accessible. No part of the structure was opened up. No structural calculations were carried out. |
| 2.2 | The roof space and the foundations below ground were not inspected. |
| 3.2 | Diagonal stepped cracking was observed in the rear elevation brickwork above the kitchen window, up to 5 mm wide at the window head and narrowing to under 1 mm at ground level. |
| 3.3 | Cracks up to 2 mm wide were observed at the junction of two first-floor ceilings with the rear wall. The rear bedroom door binds at the head. |
| 3.5 | A mature tree stands about 6 m from the rear elevation. |
| 4.1 | The pattern of cracking is consistent with differential foundation movement in the rear part of the building. Seasonal movement of clay soils may be a contributing factor. Whether movement is continuing cannot be determined from a single visit. |
| 5.1 | (a) Fit crack monitors at the three locations in Appendix B and read them monthly for six months. (b) Do not commission underpinning, or repairs to the cracked brickwork, until the monitoring results have been reviewed by us. (c) Commission a camera survey of the rear drains. (d) Internal redecoration may proceed at the owner's discretion; cracks may reappear. |
Prompt: draft the summary from the technical text only
Paste the approved clauses, give the AI the seven headings, and make it show its sources. Two features do most of the work: a tag after every sentence, and a mandatory flag for any sentence it cannot trace.
You are writing the plain-English summary box for a structural engineer's report.
Use ONLY the clauses pasted below. Do not add causes, reassurance, advice or facts.
Rules
1. Keep every number and unit exactly as written.
2. Keep every qualifier at the same strength: "may", "consistent with", "cannot be determined" must not become "is", "caused by" or "will".
3. Keep every negative: "not inspected", "no calculations", "cannot be determined".
4. Keep each recommendation with its owner, timing and condition.
5. Put a source tag after every sentence, for example [3.2]. If a sentence cannot be traced to a clause, write [NO SOURCE] and explain why you wrote it.
6. Aim for a reader aged about 14. Define any technical word in a few plain words. Do not rename a mechanism (do not say "sinking" for "differential movement").
7. Use these headings: What we were asked; What we looked at; What we found; What it may mean; What we recommend; What this report does not cover; How sure we are.
8. Copy clause 2.1 and 2.2 into "What this report does not cover" in near-verbatim wording.
Clauses:
[paste clauses 2.1 to 5.1]An illustrative first draft from that prompt, abridged to the sentences that matter, with the tags it produced:
- We looked at the outside of the house and the rooms we could enter on [date]. [2.1]
- There are diagonal cracks in the back wall, up to 5 mm wide above the kitchen window. [3.2]
- The cracks are caused by the back of the house sinking. [4.1]
- A nearby tree may be making this worse. [3.5, 4.1]
- There is no immediate danger to the building. [NO SOURCE: added as reassurance]
- Watch the cracks for six months before doing any repairs. [5.1a, 5.1b]
- Underpinning will not be needed. [5.1b]
- The drains should be checked by camera. [5.1c]
The tags are the reason this draft is checkable at all. The model flagged sentence 5 itself, because the prompt made it say where each sentence came from, and the others can now be traced in seconds.
The sentence ledger: trace each claim to its clause
Open the draft beside the report and fill one row per summary sentence. The engineer does this, not the person who prompted the model. Each row asks four questions: do the numbers match, is the qualifier at the same strength, are the limits intact, and does the sentence have a source.
| # | Draft sentence | Source | Numbers | Qualifier and scope | Verdict and fix |
|---|---|---|---|---|---|
| 1 | We looked at the outside and the rooms we could enter. | 2.1 | n/a | "Visual inspection only" is missing here | Pass, but add "visual inspection only" or a pointer to the limits box |
| 2 | Diagonal cracks, up to 5 mm, above the kitchen window. | 3.2 | 5 mm matches | Location matches | Pass |
| 3 | The cracks are caused by the back of the house sinking. | 4.1 | n/a | "Consistent with" became "caused by"; "sinking" is new | Fail. Use: "The pattern fits part of the foundations moving more than the rest." |
| 4 | A nearby tree may be making this worse. | 3.5, 4.1 | 6 m dropped | Two separate statements merged; the report never links the tree to the cracks | Fail. Split: soil may contribute; a tree stands about 6 m away; its effect was not assessed |
| 5 | There is no immediate danger. | none | n/a | An opinion the engineer did not write | Cut |
| 6 | Watch the cracks for six months before doing any repairs. | 5.1a, 5.1b | Six months matches; "monthly" dropped | "Any repairs" is wider than "repairs to the cracked brickwork"; the review step is missing | Partial. Restate with monthly readings, the brickwork-only scope and "until we have reviewed the readings" |
| 7 | Underpinning will not be needed. | 5.1b | n/a | A precaution became a conclusion | Fail. Use: "Whether underpinning is needed cannot be decided until the readings are reviewed." |
| 8 | The drains should be checked by camera. | 5.1c | n/a | Matches | Pass |
Three of eight fail, one is partial, one is cut, and three pass. That spread is illustrative, not a measured error rate, but it shows why the check cannot be skipped: the three failures are all fluent sentences. A skim for "does this read well" would pass every one.
Mechanical checks that catch what reading misses
- Numbers. List every number and unit in the summary, then find each in the source. Any number that appears only in the summary is an error, and so is a source number that vanished (the 6 m in sentence 4).
- Qualifier ladder. Rank the certainty words in the source and check the summary stays on the same rung (table below).
- Negatives. Search the source for "no", "not", "without", "cannot" and "none". Each must reappear in the summary or be covered by the limits box.
- Scope words. "In the areas accessible", "on the day", "visual" and "single visit" must survive somewhere in the box.
- Recommendations. Each keeps its owner, timing and condition. "Until we have reviewed" is a condition and cannot be dropped.
- Compare after edits. When the AI revises an approved draft, run the Compare command on Word's Review tab with the approved version as the original, so every changed word shows as a revision mark. Microsoft's help page says the differences appear in a new copy and the original stays unchanged, so the approved text is never overwritten.
| Rung | Technical wording | Plain wording that stays on the rung |
|---|---|---|
| 1 | "Cannot be determined" | "We cannot tell yet" |
| 2 | "May", "might", "could" | "May" |
| 3 | "Consistent with", "appears" | "Fits with", "looks like" |
| 4 | "Likely", "probably" | "Probably" |
| 5 | "Was observed", "is", "caused by" (only where established) | "We saw", "is", "is caused by" |
Words such as "because", "due to" and "caused by" jump from rung 3 to rung 5 in a single step, so search for them first.
The corrected summary box, filled in
After the ledger fixes, the box for the sample extract reads like this. It is illustrative, and every sentence now traces to a clause or to the limits.
Summary for the client. Read this with the full report. Where they differ, the full report applies.
What we were asked. To look at cracks at the back of the house and say what they may mean. [1]
What we looked at. The outside and the rooms we could enter, on [date]. This was a visual inspection only. We did not open up any part of the building and did not see the roof space or the foundations. [2.1, 2.2]
What we found. Diagonal cracks in the back wall above the kitchen window, up to 5 mm wide and narrowing to under 1 mm at ground level. Cracks up to 2 mm wide where two upstairs ceilings meet the back wall. The rear bedroom door binds at the top. A mature tree stands about 6 m from the back wall. [3.2, 3.3, 3.5]
What it may mean. The pattern of cracks fits part of the foundations moving more than the rest. Clay soil that shrinks and swells with the seasons may be contributing. We have not assessed whether the tree affects the ground. One visit cannot tell us whether the movement is still happening. [4.1]
What we recommend. First, fit crack monitors at the three marked places and read them monthly for six months. Second, do not commission underpinning or repair the cracked brickwork until we have reviewed the readings; whether underpinning is needed cannot be decided before then. Third, have the back drains surveyed by camera. Redecorating inside is your choice, but the cracks may come back. [5.1]
What this report does not cover. The roof space, the foundations below ground, any testing, and any structural calculations. It describes the building on the day only. [2.1, 2.2]
How sure we are. We are clear about what we saw. The pattern fits foundation movement, but one visit cannot confirm the cause or show whether it is continuing. [3, 4.1]
The source tags stay in the working copy for the checker and come out of the client version. Save both: the tagged copy is your record of how the summary was checked.
Limits wording and glossary entries the AI may not rewrite
Two parts of the report should be protected text. Your firm's limits paragraph is worded with your insurer and your solicitor in mind, so the model copies it and never paraphrases it. A blank to adapt, which you should replace with your own approved wording:
LIMITS OF THIS REPORT (firm's standard wording; do not edit)
This report is based on a visual inspection of the parts of the building that were accessible on [date].
No part of the structure was opened up, and no testing or structural calculations were carried out unless stated.
Areas not inspected: [list].
The report describes the condition on the day of inspection and is prepared for [client] only.The glossary is the second protected area. A glossary entry is plain English in miniature, with the same drift risk, and a poor one can undermine a careful sentence elsewhere. Let the AI propose entries, then vet each one. Illustrative before-and-after pairs:
| Term | AI first draft | Problem | Vetted entry |
|---|---|---|---|
| Settlement | "When a building sinks into the ground." | Suggests failure, and a sinking building | "Downward movement of a building's foundations as the soil beneath compresses. It can be small and finished, or continuing; the report says which, where it can." |
| Differential movement | "When one part of a building moves." | Loses "differently from another part" | "Different parts of a building moving by different amounts." |
| Lintel | "A beam above a door." | Misses windows and what it carries | "A beam over a window or door opening that carries the wall above it." |
| Crack monitor | "A device that measures cracks." | Overstates; it shows change over time, not cause | "A small gauge fixed across a crack that shows whether it opens or closes over time." |
Keep vetted entries in a firm glossary file and paste from it. A new report starts from the approved list, not from a fresh model attempt each time.
Two extra checks when the stakes are high
The reader read-back. Give the finished box to someone outside engineering and ask them to say, in their own words, what is wrong, how worried they should be, and what they must do. If they say "the house is unsafe" and the report says "monitor for six months", the box needs another pass. This is a test of what a client takes away, which the ledger does not measure.
The blind reverse check. In a new chat, paste only the summary and ask the model to list every technical statement it implies. Compare that list with the report. Any implied statement the report never made is drift. It is a useful second line for long summaries, and it never replaces the engineer's own sentence-by-sentence read.
For what AI should never be trusted with in this work, ChatGPT for structural engineers: what is safe and what never is draws the line clearly, and the sibling workflow for field notes is in turning inspection photos and notes into a draft.
Time, cost and privacy, with illustrative numbers
Say a 14-page report needs a summary. By hand, writing it takes the engineer about 45 minutes. With AI, pasting the clauses and running the prompt takes about 5 minutes, the ledger and mechanical checks take 25 to 30 minutes, and the final read takes 5. The net saving is perhaps 5 to 10 minutes a report, which is small. The real gain is a consistent structure and a recorded check. If the check starts taking longer than writing it yourself, write it yourself: the summary exists to be right, not to be fast.
On privacy, draft summaries in a business workspace rather than a personal account. ChatGPT Business, Claude Team and Microsoft 365 Copilot do not use business content for training by default; a consumer plan usually needs a setting switched off, and that setting gets renamed from time to time. The first two list at $20 a seat a month billed annually or $25 monthly, with a two-seat minimum. Remove the client's name, the address and any identifying detail before pasting. The summary needs the findings and the clause numbers, not the people. Keep the prompt, the tagged draft and the ledger with the job file as a record of how the summary was produced and checked, and ask your professional indemnity insurer whether it has views on AI-assisted documents before you roll this out across a team.
If you write fee proposals as well as reports, the scope and assumptions you quote should echo the limits above, and AI for consulting engineers: fee proposals and reports, not calculations shows how to keep the two consistent.
Questions engineers ask about AI-written summaries
Should I tell the client AI helped draft the summary?
Say so in your terms of engagement, in one plain sentence: you use drafting tools under your own review and you take responsibility for the content. Ask your professional indemnity insurer and your solicitor for wording they are happy with. What matters most is that the engineer reads and owns every sentence, whatever helped draft it.
Can the summary be sent without the full report?
It should not be. The summary is only accurate alongside the limits and discussion in the report, so put a line in the box saying it must be read with the full report and that the report prevails where they differ. Send both together, and keep the summary on the first page so a reader cannot miss that rule.
What reading level should the summary aim for?
Microsoft's Word help page suggests a Flesch Reading Ease score of 60 to 70 and a grade level around 7 to 8 for most documents. Technical content may not reach that, so aim for short sentences and defined terms, and treat the score as a prompt to simplify. A readable summary can still be wrong, which is why the ledger check exists.
Which AI tool should I use for this?
Any capable assistant can draft the text, so the choice matters less than two things: a business plan whose privacy defaults do not train on your content, and your checking routine. Leave the client's name and address out of what you paste. The summary needs the findings and the clause numbers, not the people.
Further reads
- Structural Engineer Fee Proposal: Scope and Assumptions Drafted With AI — Scope and assumptions that match the limits in your report.
- Thermal Imaging Survey Report: Turn Scans Into Plain-English Findings — Another report where findings need plain-English wording.
- Mould Inspection Report Template: Findings, Photos and Limits With AI — Findings, photos and limits written for non-specialist readers.
- Home Inspection Report Template: Build a Comment Library With AI — A comment library you can adapt for client-facing wording.
- Damp Survey Report: Write Findings and Quote Lines With AI — Turn survey findings into plain statements and quote lines.
- Sewer Camera Inspection Reports: From Defect List to Customer Summary — From a technical defect list to a customer summary.
- Foundation Repair Leads: Screen Crack Photos With AI Before the Visit — Six intake questions, a four-photo request, an AI summary that describes but never diagnoses, an engineer-referral rule, reply scripts and the lead-cost sums.
- Home Energy Audit Report: Turn Findings Into a Client Report With AI — Turn audit notes into a client report with AI: a findings table, priority order, plain-English summary and savings as ranges, with every figure checked.
- How Engineering Firms Are Using AI: Six Jobs and Their Limits — Six jobs engineering firms give AI, rated for effort and risk, with the evidence behind each and the uses to leave alone.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft's help pages for Word's readability statistics and its document compare feature, and vendor privacy defaults recorded for ChatGPT Business, Claude Team and Microsoft 365 Copilot, read in early November 2026. The report extract and summaries are invented samples, not from any real project.