Table of Contents
To automate shift reports and handover notes, have an on-premise LLM assemble the shift's events from the historian, alarm journal, maintenance system, LIMS and operator log into a fixed template, with a source link on every line. The outgoing operator reviews, completes and signs the draft, and the face-to-face handover still takes place.
Key takeaways
- Poor handover keeps recurring in major accidents. Piper Alpha, BP Texas City and Buncefield all involved failures of communication between shifts (Wilkinson and Lardner, 2013).
- Automate the written part, not the conversation. HSE guidance calls for face-to-face, two-way handover using both verbal and written communication (HSE).
- Structure matters more than fluent prose. Structured logbooks with required categories improved what operators recorded at BP's Grangemouth refinery (CSB, 2007).
- LLM drafts help but are not reliable on their own: in a hospital study, LLM handoff notes contained more incorrect content than clinicians' notes, 9.6% against 2.0% (Hartman et al., 2024).
- Take numbers from source systems, never from the model, and have the outgoing operator approve every draft.
Why do shift handovers fail?
The UK HSE defines the goal of handover as "the accurate reliable communication of task-relevant information across shift changes or between teams". It has three parts: preparation by the outgoing shift, a two-way exchange, and cross-checking by the incoming shift. HSE cites the 1983 Sellafield beach incident, where radioactive waste was discharged to sea after a communication failure between shifts, and the Piper Alpha inquiry, which found handover failure among the contributing factors (HSE).
At BP Texas City in 2005, where explosions and fires killed 15 people and injured 180, the night lead operator left about an hour early. The control room logbook entry read: "ISOM: Brought in some raff to unit, to pack raff with." The day board operator understood that liquid had been added only to the tower, not that heat exchangers and piping had also been filled. The CSB found that BP "did not have a shift turnover communication requirement for its operations staff" (CSB, 2007).
At Buncefield, also in 2005, the human factors analysis for the investigation concluded that shift changeover was "inadequately structured, variable and informal", that the key information to pass on had not been determined, and that passing information on verbally was the default when time or events were pressing. Control room staff ended up with a mistaken picture of which pipeline was feeding which tank (Wilkinson and Lardner, 2013).
In both cases a log existed. What failed was what went into it and whether the next shift understood it. An LLM that turns a thin log into fluent paragraphs fixes neither. One that pulls in facts the operator did not write down, and asks for what only the operator knows, can help.
What should a shift report contain?
Before AI, fix the template. BP's Grangemouth refinery studied its handovers in the 1990s and found unstructured logbooks (lined notebooks), very different writing styles, mostly historical records with little on what should be done next, and no safety issues written down. Logbooks redesigned with the operators, with required categories for safety, maintenance and technical problems, work outstanding, comments and signatures, led to more maintenance and technical problems being recorded, safety issues flagged and event times recorded more consistently (CSB, 2007, Appendix J).
A template built on those categories, with the systems that can feed each section:
| Section | What goes in it | Where the draft can take it from | Who confirms |
|---|---|---|---|
| Plant status | Units running, rates, start-ups or shutdowns in progress, equipment out of service | Historian, production reporting | Outgoing panel operator |
| Safety-critical conditions | Overrides and bypasses, safety systems unavailable, active permits and isolations | Bypass register, permit-to-work system | Shift supervisor |
| Events and alarms | Trips, upsets, standing alarms | Alarm and event journal, historian | Outgoing operator |
| Maintenance | Work spanning the shift change, new defects | CMMS work orders and notifications | Outgoing operator |
| Quality | Off-spec results, pending samples | LIMS | Outgoing operator or lab |
| Work outstanding and instructions | What the next shift must do or watch, and why | The operator only | Outgoing operator |
| Signatures | Author and receiver | — | Both |
The "work outstanding" row is what Grangemouth's logbooks lacked, and only a person can write it. The draft should prompt for it, never invent it.
The header should also flag higher-risk handovers. Wilkinson and Lardner list them: after a long absence, between experienced and inexperienced staff, during an upset, when maintenance spans shifts, and when safety systems are overridden or unavailable.
Which parts can an LLM draft?
Divide the work: plant systems supply the facts, the model selects, groups and phrases them, and people add intent and sign.
The model is useful for:
- grouping hundreds of alarm journal entries into a few events with start and end times;
- summarising the free text of work orders and lab remarks;
- carrying open items forward from the previous report and asking whether each is closed;
- harmonising notes written in Arabic, French and English.
The model should not:
- write any value, unit or timestamp itself;
- decide that a condition is safe or drop an item from a safety section;
- write the instructions to the next shift.
How does an automated handover pipeline work?
- Agree the template with operators for each unit, as Grangemouth did, with required sections.
- Collect the shift's data at a fixed time before the change through read-only accounts: historian tags, alarm journal, CMMS, LIMS, permit and bypass registers, and electronic log entries. Our guides to integrating AI with SCADA and OT systems and to text-to-SQL on production databases cover safe read access.
- Pre-process in code. Compute values, filter chattering alarms and keep every fact as a record with its source ID and timestamp.
- Generate a structured draft. Ask for output that follows a JSON schema matching the template; on-premise servers such as vLLM can enforce a JSON schema on the output (vLLM documentation). Require a source ID on every item. Vendor guidance on reducing hallucinations recommends allowing the model to say it does not know, restricting it to the documents provided, and having it back each claim with a quote or citation and retract the claim if it cannot (Anthropic documentation).
- Check the draft in code. Every number must match its source record, every open item from the previous report must be carried or explicitly closed, and no required section may be empty.
- Have the outgoing operator review, complete and sign it. The approval logic is covered in human-in-the-loop for AI agents.
- Hold the face-to-face handover with the report on screen, and let the incoming shift cross-check it on displays and in the field.
- Archive both the draft and the signed version. The differences show where the model goes wrong.
The goal is to spend preparation time on checking, not to shorten the conversation. Wilkinson and Lardner note that a handover on a 12-hour shift may last up to 30 minutes, and that it should be protected from interruptions.
How accurate are LLM-drafted handover notes?
We found no published evaluation of LLM-drafted shift reports in process plants. The closest evidence comes from hospitals. Hartman and colleagues built a pipeline around a fine-tuned Llama-2 7B model and studied it on 1,600 emergency department records. Three physicians scored 50 of them: usefulness 4.04 out of 5 against 4.36 for physician-written notes, and patient safety 4.06 against 4.50. The LLM notes had more incorrect content (9.6% against 2.0%), though none was rated a critical safety risk. Automated text-similarity metrics rated the LLM notes better, and the authors stress the need for a physician-in-the-loop design (Hartman et al., 2024).
Two lessons carry over. Generic text metrics do not measure safety, so test with operators. And structure on its own works: in nine pediatric residency programs, the I-PASS bundle of standardised oral and written handoffs and training was associated with 23% fewer medical errors and 30% fewer preventable adverse events, without lengthening handoffs (Rosenbluth et al., 2018).
Two known model limits also apply. Summarisation models are "highly prone to hallucinate content that is unfaithful to the input document" (Maynez et al., 2020). And language models use information at the start or end of a long input better than information in the middle (Liu et al., 2023), so pasting a whole 12-hour alarm journal into one prompt is a poor design. Filter first and generate section by section. Our guide to evaluating an LLM on your own documents covers building a test set.
What can go wrong, and how do you control it?
| Risk | Plant example | Control |
|---|---|---|
| Altered value | Draft says column pressure was stable when the historian shows a rise | Values inserted by code from the source; automatic comparison of every number |
| Omission | A bypassed trip left out of the summary | Safety sections fed directly from the bypass and permit registers, not filtered by the model; carry-over check against the previous report |
| Vague wording | "Brought in some raff" | Template fields for equipment tag, state and required action |
| Instructions hidden in text | A contractor's note containing text addressed to the AI | Treat all free text as data; OWASP describes indirect prompt injection through external content such as files (OWASP LLM01:2025) |
| Signing without reading | Every draft signed unchanged within seconds | Show source links; track edit rates per operator and section |
| Personal data | Names, injuries, absences | On-premise processing, access control, audit log |
Shift reports name people and sometimes record injuries. According to the CMS guide to Algerian data protection law, Law 18-07 requires security measures commensurate with the risks, transfers of personal data abroad need the authority's authorisation, and Law 25-11 adds an automated logbook of processing operations and breach notification within five days. Confirm with legal counsel how this applies to your reports. Our guide to securing an on-premise LLM covers access control and logging.
What about notes in Arabic, French and English?
In Algerian plants, an operator may write a note in Arabic with French equipment names, next to an English OEM alarm text. Decide the language of the report, keep tags, values and alarm texts verbatim, and keep the original note attached to any translation. Our guide to documents that mix Arabic, French and English covers normalisation, digits and testing.
How do you know if it works?
Run the drafts in parallel with the existing log for several weeks before relying on them, and track:
- the share of items operators edit or delete, by section;
- omissions found by the incoming shift;
- open items correctly carried forward;
- time spent preparing the handover;
- what operators on both sides say about it.
What we have seen at KABAS AI
DALILE, our AI assistant, runs on the client's own servers, generates reports with data, charts and analysis, connects to SQL databases, SAP, Oracle and LIMS, and lets teams comment on and approve documents together, with access controls and audit logging. Those are the building blocks a shift report pipeline needs, and they keep shift data inside the site.
When we scope this kind of workflow with our agentic AI team, our advice is to start with one unit and the logbook its operators already use, agree with operations and HSE which sections the model may never filter, and connect the maintenance system read-only first. Our article on connecting an AI agent to the CMMS covers that step.
Sources
- UK HSE, Human factors: Safe shift handover
- US Chemical Safety Board, Investigation Report: Refinery Explosion and Fire, BP Texas City, March 23, 2005 (2007)
- Wilkinson and Lardner, Shift Handover After Buncefield, Chemical Engineering Transactions 31, 2013
- Hartman et al., Developing and Evaluating Large Language Model–Generated Emergency Medicine Handoff Notes, JAMA Network Open, 2024
- Rosenbluth et al., I-PASS Handoff Program: Use of a Campaign to Effect Transformational Change, Pediatric Quality and Safety, 2018
- Maynez et al., On Faithfulness and Factuality in Abstractive Summarization, arXiv 2005.00661, 2020
- Liu et al., Lost in the Middle: How Language Models Use Long Contexts, arXiv 2307.03172, 2023
- vLLM documentation, Structured Outputs
- Anthropic documentation, Reduce hallucinations
- OWASP GenAI Security Project, LLM01:2025 Prompt Injection
- CMS Expert Guide to data protection and cyber security laws: Algeria
Frequently asked questions
Can an AI write the shift report without the operator?
It can write a draft, not the final report. The model can gather and group events from plant systems, but it does not know what the outgoing operator intends the next shift to do, and it can omit or misstate facts. The outgoing operator should review, correct and sign every report before it is published.
Does an automated shift report replace the face-to-face handover?
No. UK HSE guidance describes good handover as face-to-face and two-way, using both verbal and written communication. The written report supports that conversation and gives the incoming shift something to cross-check against; it does not replace it.
Which data sources should feed the draft?
Typically the process historian, the alarm and event journal, the maintenance system (CMMS), the LIMS, the permit-to-work and bypass registers, and the operators' own log entries. Each should be read through a read-only account, and every item in the draft should link back to the record it came from.
How do we stop the model from inventing values?
Do not let it write numbers. Have code copy values, units and timestamps from the source systems into fixed fields, let the model only group and phrase events, and run an automatic check that every number in the text matches its source. Items without a source reference should be flagged, not published.
Can this run on site without internet access?
Yes. Open-weight models can run on servers inside the plant, so shift reports, which often contain names and incident details, never leave the site. In Algeria, personal data in these reports falls under Law 18-07 as amended, so confirm the requirements with legal counsel.



