Yes — for part of the job. The useful question is not whether AI can review a RAMS but which part of reviewing a RAMS is mechanical and which part is judgement, because the answer is different for each and getting it wrong in either direction is expensive.
We build a tool that does this, so treat what follows as an interested party being specific about the limits rather than a neutral survey.
What the mechanical part of a review actually is
A large share of what a reviewer does is comparison and completeness checking. Does every section exist? Is every field populated? Does every activity in the scope produce a hazard, and every hazard trace back to an activity? Does every hazardous substance have a COSHH reference? Does every item of plant have an operator competency? Is every residual risk score filled in? Does the muster point match the site, and does any text in the file refer to a different project?
None of that requires professional judgement. All of it requires reading every line of a thirty-page document with equal attention — which is the thing humans reliably stop doing at about page eleven, on the fourth document of the day, with a start date on Monday.
What automated review does well
- Structural completeness: missing sections and unpopulated fields, every time, in the same way.
- Cross-document consistency: dates, revision numbers, site names and named individuals that disagree with each other between pages or between the RAMS and its attachments.
- Scope-to-hazard coverage: activities in the scope with no corresponding hazard, which is the highest-value single check available.
- Vague control language: flagging "take care", "be vigilant" and "as required" wherever they appear as a control.
- Hierarchy-of-control ordering: identifying hazards where PPE is the only control offered.
- Copy-and-paste residue: a muster point that reads as though it came from a previous site, a revision history that is not specific to this project.
- Consistency across a portfolio: the same checks applied to every subcontractor's document, rather than a stricter reading on a Monday than on a Friday.
In other words, it reliably finds most of the categories listed in twelve RAMS mistakes that fail inspection — which is most of what causes a rejection. What it will not find is a gap that only somebody who has been to the site would recognise as a gap.
What it cannot do
Decide whether the method is safe
A model can tell you a lifting operation has controls listed. It cannot stand under the load, look at the ground conditions and the proximity of the façade, and judge whether the plan is appropriate. That is engineering judgement about a physical situation the document only partially describes, and it belongs to a competent person.
Know your site
The document is all the tool has. It cannot know that the muster point moved last week, that the survey found ACMs in the ceiling void, or that the neighbouring unit runs a nursery. It can only check the document against itself and against what a document of this kind should contain.
Judge competence
It can check that a competency is stated and that a certificate number appears. It cannot judge whether the person is actually competent for this task on this site, which is what regulation 8 of CDM 2015 is asking about.
Approve anything
No automated output is an approval, a certification or a sign-off. It has no legal standing. It is a reading aid that produces a list of things worth looking at.
Where the responsibility sits
Exactly where it sat before. The employer's duty to make a suitable and sufficient assessment under regulation 3 of the Management of Health and Safety at Work Regulations 1999 is not affected by the tools used to prepare it. The contractor's duty under CDM 2015 regulation 15 to plan, manage and monitor is unaffected. The principal contractor's coordination duty is unaffected.
Which means "the software did not flag it" is not a defence, in the same way "the consultant did not mention it" is not a defence. Using a tool changes how thoroughly the first read happens; it does not move the duty. Any vendor implying otherwise is selling you a liability, not a product.
The two ways to get this wrong
The first is over-trust: treating a clean automated report as evidence that a document is compliant. A document can pass every structural check and still describe a method that will hurt somebody, because the checks are about the document and the danger is on the site.
The second is under-use: insisting that only a full human read has any value, and therefore continuing to give document six of the day the attention document six of the day actually gets. Consistency is a real safety property, and it is the one humans are worst at — how long a RAMS review should take works through where the hour actually goes.
A sensible division of labour
- Automated pass first, on the author's side, before submission — catching the mechanical gaps while they are cheap to fix.
- Author corrects and re-runs, so the document that reaches a reviewer is structurally sound.
- Competent person reads for method: is this the right way to do this work, here, given what else is happening?
- Principal contractor reviews for site compatibility and coordination, which no tool can assess.
- Named individual signs, and the sign-off records their decision — not the tool's.
Used that way, the automated pass does not replace anyone. It removes the part of the work that was being done badly because it is tedious, and hands the human the part that needs a human.
Questions worth asking a vendor
- What happens to my documents, where are they stored, and for how long?
- Is my content used to train models? It should not be.
- Does the output claim to certify or approve anything? It should not.
- Can I see the reasoning behind a finding, or only the verdict?
- What does it explicitly not check?
On our side those answers are published: documents are encrypted in transit and at rest, deleted after 90 days, never used for training, and every report carries a statement that it is not an approval. What critical means in a report shows how findings are ranked, and our limitations page lists what Valify will not catch — which is the page worth reading first.
This article is general guidance on UK construction documentation practice. It is not legal advice, a compliance certification or a substitute for review by your own qualified health and safety personnel, who remain responsible for deciding whether any particular document is adequate. See our disclaimer.