A polished engineering report still needs a sound engineering basis.
An AI assistant could help assemble a first draft from project records, summarise comments or organise information for review. The question is whether the people responsible can explain and check what it produces.
The Institution of Civil Engineers’ 2026 Code of Professional Conduct addresses AI directly. It applies to ICE members across grades, roles and locations. This article focuses on its practical implications for civil engineering teams.
01 / PROFESSIONAL CONDUCT
Understand what the code says
The AI-related guidance covers retained accountability, understanding tool limitations, supervision of AI-assisted work, disclosure of the level of AI use, proportionate justification and responsible data selection. It also addresses digital privacy, confidentiality and security. Read the full code and guidance notes.
Our practical approach below translates those themes into decisions about everyday work. It is not an ICE-approved checklist.
Agree what AI may prepare, what the engineer must check and who can authorise the result.
02 / COMPETENCE
Choose work you can oversee
Begin with a task whose output your team can assess independently. For example, an assistant could prepare a draft narrative from approved calculations and an agreed design brief.
Specify the limits. The assistant may organise the supplied information and flag gaps. It should not silently introduce a new design assumption or present a calculated conclusion that nobody has established.
The reviewer needs both the relevant engineering knowledge and an understanding of how the selected tool can fail. Training should include missing context, plausible but unsupported statements and the difference between extracting a result and verifying it.
A senior job title alone does not establish that someone has the time or expertise to oversee a particular AI-assisted task. Make the review responsibility specific to the work.
03 / EVIDENCE
Keep the evidence visible
For a report-drafting workflow, the reviewer should be able to move from each important statement to the information supporting it. Ask for source references and clear flags where the inputs are missing or inconsistent.
From a draft to a defensible output
Source basis
Approved drawings, calculations and assumptions are identifiable and current.
Technical checking
A competent reviewer checks meaning, units, omissions and unsupported conclusions.
Accountable decision
The responsible engineer resolves issues and authorises the output through the practice’s process.
A fluent answer is the beginning of the review, not evidence that the engineering is correct.
Illustrative checks for a report assistant using approved project information. Technical checking and authorisation must suit the actual task.
Check meaning as well as transcription. A value copied correctly can still be attached to the wrong element, revision or load case. References also need checking: a citation can point to a real document without supporting the statement beside it.
Use the practice’s checking arrangements and decide what record is needed to show how issues were resolved. Where appropriate, retain the relevant inputs, draft, corrections and approval.
04 / TRANSPARENCY
Explain AI’s contribution
Build a consistent way of recording AI use into the workflow. An entry might explain which task the system supported, what information it used, what it produced and how the output was checked.
Use that record to support the disclosure required by the code. Agree who needs the information, when to provide it and how to describe the contribution accurately in the particular engagement.
For example, “AI assisted the preparation of a draft narrative from the listed approved calculations; the responsible engineer checked and approved the issued report” is more informative than a vague reference to AI somewhere in a general policy. Use such wording only when it accurately describes what happened.
Check client requirements and the terms of the appointment too. A disclosure process should fit the work and the obligations that actually apply.
05 / INFORMATION
Protect the information
A design brief can contain confidential client information even when it contains no personal data. Establish which records may enter the service and whether project permissions are maintained.
Review storage, processing, retention, access and training arrangements for the actual product and account being used. Keep access proportionate to the task and include external integrations in that review.
Our guide to where your information goes when you use AI provides questions to take to a supplier. Where personal information is involved, use the ICO’s UK AI and data protection resources alongside the project assessment.
06 / A CONTROLLED TRIAL
Test before extending the role
Test a defined task with representative information you are authorised to use. Include an obsolete revision, an ambiguous statement and an incomplete input to see whether the workflow identifies problems.
- Can reviewers trace important statements to valid sources?
- Do they find unsupported conclusions and missing context?
- Does total preparation, checking and correction time improve?
- Can users pause the process and continue without the service?
Agree success and stop criteria before the trial. Record the outcome, the remaining limitations and the person responsible for the next decision.
This is how we apply BUILD™: understand the process and its consequences, test a focused improvement, then develop it on the strength of the evidence.