What 209 QHSE software profiles state about AI features

What the records show

1. Most records state no AI feature. 74 of 209 (35%) describe at least one AI feature of the product. 136 records contain an AI word or a related technology word; we read each one, and 62 were excluded for a stated reason, 54 of them because the only AI words in the record belong to our own disclosure. The remaining 73 of 209 (35%) contain none of these words.
2. Nearly all of the profiles with a statement name a function. 68 of 74 (92%) of the profiles with a statement name at least one function. 6 say that AI is part of the product, a plan or a workflow and name nothing it does. 13 profiles describe three or more functions.
3. The functions stated most often are drafting and suggestions, extraction and classification, assistants and agents. By profiles: drafting, summaries, translation and suggestions (31), search, classification, extraction and tagging (25), assistants, chat, question answering and agents (20), computer vision and image or video analysis (13), analytics, pattern and anomaly detection, insights (12), predictive analytics or risk scoring (6), speech, voice input and transcription (6). The counts describe what the records contain; they do not show which functions the products offer or use.
4. Most profiles with a statement have a conditional one. 46 of 74 (62%) of the profiles with a statement have at least one the record calls conditional, and 29 of 74 (39%) name a plan, an add-on, a licence or usage credits. 6 have a statement that the record says is not established, such as use without a connection, and 4 have one marked preview, new or roadmap.27 of 74 (36%) have statements that are all documented and carry no preview mark.
5. A review step is reported by fewer than half of the profiles with a statement. 22 of 74 (30%) of the profiles with a statement have a record that reports a review, approval or override step for the AI output, and 4 records say that the vendor's pages do not describe one. The other records either leave the point out or tell the buyer to ask for a review step, which this page does not count as a statement.
6. Data use is reported less often than a review step. 9 of 74 (12%) of the profiles with a statement have a record that reports a data-use statement: 6 about training on customer data, and others about the model provider, where processing happens or switching the function off. 3 records say the vendor's pages do not state it. 48 of 74 (65%) report neither a review step nor a data-use statement.
7. Shares differ between categories, on small counts. Of the 12 categories holding at least 5 profiles, the share with an AI statement is lowest in Construction Safety (1 of 7, 14%) and highest in ESG & Sustainability (17 of 30, 57%). The categories hold 5 to 43 profiles each, so adding or correcting a single record changes a share noticeably; the spread describes this sample only.

What counts as an AI statement

A record counts when it uses an AI word about the product: AI, artificial intelligence, machine learning, generative, large language model, copilot, chatbot, a named model provider, neural network or computer vision. Words such as automatic, predicted or smart do not count on their own, and neither does a technology that the record never calls AI, or a function that it names only in a question or an instruction to the buyer. Each statement is assigned to the function that the record's wording describes. A profile can appear in several rows, and a record that describes one feature in several places is counted once for that function.

Status and conditions

Each statement has the status the record gives it. Documented means that the record lists the function without a condition. Conditional means that the record's status for it is conditional, or, where the statement comes from a feature list or a note rather than a capability row, that the record ties it to a plan, an add-on, a licence or the scope of a quote. Not established means that the record says the function, or its use without a connection, is not established or was not described. A statement can also be marked preview, new or roadmap, and it can name the condition that applies. The columns are not exclusive: a profile with several statements can appear in several of them.

The conditions named, by profile: a plan, package, tier or edition: 9; an add-on, an optional product, a separate licence or a separate scope: 18; usage credits: 1; another condition, such as a request, a region or an app version: 2. A profile can name more than one. By statement rather than by profile, 55 of the 135 statements are documented, 74 conditional and 6 not established. A condition is the vendor's own, as the record has it; whether it applies to a given quote has to be confirmed with the vendor.

Review steps and data use

Two further points are counted from the same records. A review step counts when the record reports that the vendor's page, policy or product describes a review, approval, sign-off or override of the AI output. A data-use statement counts when the record reports what the vendor says about training on customer data, which model provider or location processes it, or whether the function can be switched off. A record that only asks the buyer to require a review, or lists data handling as a question, is not counted, because that is our wording and not the vendor's. Where a record says that the vendor's pages do not describe the point, that is counted separately.

A review step is reported in 8 of the 12 categories with at least 5 profiles, and a data-use statement in 5. Those counts are small, and a record that is silent may belong to a vendor that does document the point elsewhere. The figures show how often our records hold the statement, which is what a buyer reading them would find.

By category

The charts below use the 12 categories with at least 5 profiles, which together hold 180 of the 209 profiles. The remaining 12 categories (29 profiles) are merged into a single row of the table and the CSV, which keeps any row from describing just one or two products. A profile is placed in one category only, as the directory shows it.

In the last column, only profiles with an AI statement are counted. A category's figures say what its records hold; they say nothing about how many of its products use AI or how any of them treat data.

Records read and not counted

136 records contain an AI word or a related technology word. 74 are counted as statements and 62 are reviewed exclusions, each with a reason written down. A record can have more than one reason, so the rows add up to more than 62. An automated check scans every record for these words and stops the build when it finds one that is on neither list.

How to read these numbers

• Vendor wording, untested. Each function, status and condition here comes from a vendor page or a reviewer's note. We ran no feature, tested no output and checked no model, so nothing is verified and nothing measures how well a feature works.
• A gap in the record is not a gap in the product. A record with no AI statement may belong to a product that has AI features. Records spend their words on price, trials, integrations, offline use and much else, so an AI feature appears in one only when a source page or a note brought it up.
• A count of records is not a count of features. A profile that names one function and a profile that names five both count once in the headline figure. The function rows count profiles, not products, and the statement count is a count of rows in our table.
• Conditional is the record's word. It shows that the record attaches a condition, not that the condition is unusual or that the feature is unavailable. Review and data-use rows count what the record reports; a buyer still needs the contract wording.
• Records differ in depth. Some records were read against many pages and some against few, and a longer record has more room to mention AI. The counts reflect each record as it stood when checked and say nothing about what a product offers now.
• The sample is ours. The 209 profiles make up our research set. They are not all QHSE products, and nothing here ranks, scores or recommends any of them.

What to ask when a record or a quote mentions AI

1. Which AI functions are in the quoted edition, which need an add-on or usage credits, and which are preview, new or on a roadmap?
2. What does each function do, in the vendor's words: draft, classify, extract, answer, detect or predict? Ask for the list in writing, not the label.
3. What happens to the output before it becomes a record: who reviews it, can the reviewer reject or edit it, and is the change logged?
4. Is customer content used to train a model, which provider processes it, where, and can the function be switched off by role or for the whole tenant?
5. Does the function work without a connection, and what does the mobile app do when it is offline?
6. What evidence can the vendor show on records like ours: which sample, what error types and what was missed?

The guide to AI in QHSE software sets out the wider evaluation. Record the answers in the RFP template and test them against the demo checklist, using one real task on your own data with the review step on show.

How the figures were computed

Profile records hold no structured AI field. We therefore read the record text once and entered every AI statement in a table, beside the exact wording that supports it. One script computes every count from that table plus each profile's category and writes the page data and the CSV in the same run. If quoted wording disappears from a record, the script refuses to run until the record has been read again. The table holds 135 function statements, 26 review rows and 16 data-use rows. A statement takes the status the record gives its capability row; one taken from a feature list or a caution was given one of the same three words under the rule above. Where a profile carries no category of its own, the catalogue's category is used, and percentages are rounded to whole numbers. The records were last checked on 3 October 2026.

The counts use no scores, ratings, prices or payments, and no product is named or put in an order. The methodology describes the research set, and the same records underlie the capability map and the 2026 report on vendor transparency.

Disclosure. A vendor in the set pays for sponsored placement on this site. The site's editor works there. The same reading rules applied to its record, which is counted with the rest and is not named here.