AI Construction JHA and Pre-Task Planning Automation

Why use AI if a person still has to review the plan?

Because review is not the only work in the process.

A foreman or safety lead may spend time collecting repeated project details, finding the current form, checking whether required sections are filled, comparing today's plan with yesterday's version, routing corrections, and entering the approved record into another system. Those are preparation and coordination tasks. They can create friction without removing the need for human judgment.

A governed AI workflow may assist with:

• collecting task, crew, location, equipment, material, access, and sequence details;
• retrieving information from approved, current sources;
• normalizing field input into a consistent draft structure;
• identifying blank fields, conflicting details, and missing attachments;
• proposing questions for a qualified reviewer to resolve;
• comparing versions and highlighting what changed;
• routing a draft to named reviewers;
• preserving correction and approval history; and
• reconciling the approved record with the intended system of record.

The value is not “automatic safety.” The value is better-controlled preparation, review routing, version handling, and recordkeeping.

JHA, JSA, and pre-task plan: start with your own definitions

Contractors do not always use these terms the same way.

A job hazard analysis, or JHA, generally breaks work into steps, identifies hazards associated with those steps, and documents controls. A job safety analysis, or JSA, is often used in a similar way. A pre-task plan may cover the work planned for a shift or operation, the crew, changing conditions, coordination, controls, and briefing.

A toolbox talk, inspection, observation, incident record, permit, and stop-work process serve different purposes. Combining them into one generic “safety form” can blur authority and create bad records.

Before adding AI, define:

• what each record or meeting is called;
• when it is required;
• who supplies the field facts;
• who reviews and corrects the record;
• who has approval or permit authority;
• how workers participate and raise concerns;
• what triggers a new version; and
• which system holds the official record.

Those answers may vary by employer, project, trade, owner, insurer, union, and jurisdiction. The workflow has to follow the contractor's real requirements—not a generic internet template.

Freeze the identity of the task before drafting

A useful draft needs more than a project name and a work description.

The intake should identify the project, exact work area, date or shift, trade, crew, work sequence, tools, materials, equipment, access, adjacent operations, environmental conditions, energy sources, applicable permits, and emergency conditions. It should also identify which details came from a person in the field and which came from a company or project record.

Worker and foreman input comes first. AI should not manufacture jobsite context from a template or treat yesterday's conditions as today's facts.

If task identity is incomplete, the system should stop and ask for missing information rather than fill gaps with confident-sounding language.

Use approved sources and show where information came from

A contractor may need to draw from company policy, the project safety plan, scope, schedule, equipment records, safety data sheets, permits, weather information, qualifications, prior observations, and other approved records.

The workflow should make source control visible:

• source name;
• document owner;
• revision or effective date;
• project or jurisdiction, when applicable;
• status as approved, superseded, or uncertain;
• conflicts with other records; and
• facts that are still missing.

AI-generated text should not be presented as an authoritative source. Vendor marketing should not be treated as independent safety guidance. When sources conflict, a named person should resolve the conflict before the record advances.

Let AI draft candidates—not final safety decisions

A preparation layer may propose work steps, possible hazards, candidate controls, follow-up questions, and missing-information flags. Each proposal needs review against the actual task and conditions.

The AI must not be assigned authority to:

• decide that all hazards have been identified;
• select final controls;
• determine whether a person is competent or qualified;
• approve a permit;
• confirm regulatory or contractual compliance;
• verify field conditions it cannot observe;
• conduct the real crew briefing;
• decide whether a worker understood the plan;
• override a stop-work concern;
• direct emergency response; or
• release a crew to begin work.

Those decisions belong to the accountable people and processes defined for the employer, project, task, and jurisdiction.

Put a hard human review gate in the workflow

The draft should not move directly from AI output to an official record.

A named reviewer should be required to resolve:

• missing task or crew details;
• unsupported or outdated sources;
• generic hazards or controls that do not match the work;
• conflicting records;
• high-consequence activities;
• unclear permit or qualification requirements;
• changed conditions;
• worker questions and objections; and
• any uncertainty the system cannot safely resolve.

The gate should record who reviewed the draft, what changed, when it changed, and why. Silence or inactivity should not count as approval.

Preserve versions, corrections, and worker input

A safe recordkeeping workflow does not silently overwrite the field record.

Each revision should preserve the earlier version, the changes made, the person who made them, the reason, the timestamp, and the current status. The system should distinguish a draft, a returned correction, an approved record, a superseded record, and a voided record.

The real crew briefing still has to happen. The process should capture questions, added hazards, control changes, acknowledgments, refusal, and stop-work paths as the contractor's requirements direct. A signature can document an action; it does not prove understanding or prove that conditions are safe.

Reset the plan when conditions change

A pre-task plan is not “done” for the day if the work changes.

The team should re-evaluate the record when there is a change in:

• task or work sequence;
• crew or supervision;
• tools, equipment, or materials;
• access or work area;
• nearby operations;
• weather or environmental conditions;
• energy sources;
• permits;
• identified hazards; or
• selected controls.

The contractor's actual trigger list may be broader. The workflow should make it easy to stop, revise, re-review, re-brief, and preserve the prior version. It should never allow AI to waive that reset.

Reconcile only the approved record

Once the responsible people complete review, the workflow may prepare the approved version for the intended system of record. That handoff still needs verification.

Check the destination record's project and task identity, status, version, reviewer identities, timestamps, attachments, signatures, and system identifiers. Confirm that the write did not create a duplicate, attach the record to the wrong project, drop a correction, or mark an unapproved draft as final.

If reconciliation fails, keep the approved source record available and route the exception to a named person. Do not hide a failed write behind a success message.

Test the workflow without live worker or jobsite data

Before using a workflow in production, test it with synthetic projects, crews, tasks, records, images, and conditions.

The pilot should test:

• role-based permissions;
• missing and conflicting inputs;
• source freshness;
• duplicate prevention;
• version history;
• exports and retention;
• outages and offline fallback;
• failed routing or reconciliation;
• incident and stop-work escalation paths;
• model failure and unsupported output;
• decommissioning; and
• rollback to the contractor's current process.

The goal of a pilot is to expose weak points and authority gaps before live data or live operations are involved. It is not evidence of compliance, safety performance, or customer results.

A practical authority matrix

| Workflow action | AI may assist | A named person must own | |---|---|---| | Gather task context | Prompt for and organize supplied facts | Verify actual work and field conditions | | Retrieve records | Search approved, permissioned sources | Approve source set and resolve conflicts | | Prepare a draft | Structure content and propose questions | Identify hazards and select final controls | | Check completeness | Flag blank or inconsistent fields | Decide whether the record is adequate | | Compare versions | Highlight additions, removals, and conflicts | Approve corrections and changed-condition response | | Route review | Notify assigned reviewers and track status | Review, reject, correct, or approve | | Support briefing records | Prepare the current approved version | Conduct the briefing and address worker input | | Reconcile records | Prepare or execute a controlled handoff if approved | Verify the official record and handle exceptions | | Authorize work | No | Accountable project and employer roles only |

Questions contractors should answer before adding AI

• What is the current JHA, JSA, or pre-task planning process from field intake to official record?
• Which steps are repeated administrative work, and which require safety judgment or legal authority?
• Who owns each decision, correction, briefing, approval, permit, stop-work action, and release to work?
• Which sources are approved, current, and accessible under the right permissions?
• What data is too sensitive or unnecessary for the workflow?
• What conditions force the plan to be revised and the crew to be re-briefed?
• What happens during an outage, failed write, model error, or conflicting source?
• How can the contractor return to the current process without losing records?
• What evidence must be collected before any production decision is made?

Frequently asked questions

Can AI create a construction JHA?

AI can help prepare a draft from supplied task context and approved sources. Accountable people must verify actual conditions, identify hazards, select controls, correct the draft, involve workers, and approve or reject the record according to applicable requirements.

Is an AI-generated JHA OSHA compliant?

A generated document does not establish compliance. Compliance depends on the actual work, conditions, employer duties, applicable requirements, worker participation, selected controls, implementation, and competent review. Any compliance conclusion requires qualified legal and safety review.

Who reviews and approves a JHA or pre-task plan?

That depends on the employer, project, task, trade, jurisdiction, and applicable requirements. The workflow should name the responsible employer, safety, supervisory, competent- or qualified-person, permit, and worker roles rather than assigning authority to AI.

Can AI identify jobsite hazards from photos or video?

AI may surface possible hazards or questions, but an image can omit critical context, and a model can miss hazards or produce false matches. Photo or video analysis cannot replace field observation, worker input, or required qualified review.

When should a pre-task plan be revised?

The team should re-evaluate it when the task, crew, tools, equipment, materials, access, sequence, weather, nearby work, permits, hazards, or controls change. The final trigger list must match the contractor's actual requirements.

Does AI replace a competent person or safety manager?

No. AI may assist with preparation and checking, but it cannot assign competence, fulfill employer duties, verify every field condition, make final safety decisions, or assume a named person's authority.

Can AI authorize a crew to begin work?

No. Work authorization remains with the accountable people and processes defined for the employer, project, task, permits, and jurisdiction.

Map the workflow before choosing the tool

The first decision is not which AI model or software to buy. It is where the current JHA and pre-task planning process depends on repeated intake, manual routing, disconnected records, unclear ownership, and weak change control.

Map that process first. Then decide whether a limited, human-controlled AI preparation layer belongs anywhere in it.