Insights

Food-production CAPA: from a finding to effectiveness review

Author: Piotr AdamskiPublished: 2026-10-02Updated: 2026-10-10

A food manufacturer's first automation should shorten the path from a record to a check. Collecting the documents for one batch, identifying missing records or preparing training after an instruction changes are useful starting points. Set a clear boundary: a model does not independently establish product safety or release a batch to market.

A polished summary cannot replace a missing measurement. If a system invents a plausible value or links similar batch identifiers, it can make incomplete evidence look complete. Start with data, responsibility and a way to verify the result. This guide presents Brillnet's recommendations for quality and operations managers. Legal obligations are identified separately using official EU sources.

Separate obligations from organisational choices

Article 5 of Regulation 852/2004 establishes HACCP based procedures for businesses within its scope. Establish that scope for the actual activity, including the separate treatment of primary production. Owning an application does not prove that the obligation has been met. [1]

The European Commission places primary food safety responsibility with the business operator. Its guidance explains traceability to the immediate supplier and next recipient, with the retail final consumer exception and sector specific requirements. [2] An automation supplier does not take over that responsibility by writing a report.

Commission Notice 2022/C 355/01 addresses Good Hygiene Practices and HACCP based procedures. The official Codex Alimentarius list includes General Principles of Food Hygiene, CXC 1-1969. These are useful reference points; they do not replace establishing the rules and conditions for a specific operation. [3][4]

We recommend using AI to prepare information under the team's control. We do not suggest that the law requires AI, a particular cloud or one product. A project should support the established food safety system, using approved methods and retaining responsibility with the appropriate people.

Select a workflow where an error can be stopped

Checking whether a batch information pack is complete is a useful first candidate. The system reads an agreed list of required records, retrieves linked documents and highlights gaps. Its output is a list for review. It is not a statement that food is safe or an automatic shipment decision.

Preparing training after a cleaning or product changeover instruction is revised is another candidate. AI can help simplify language, draft questions and identify the relevant source location. A technologist or hygiene owner checks that the material describes the actual equipment, activities and rules at the site.

Do not start by generating a complete HACCP plan from a product description. That output may overlook the actual line, storage conditions, working practices or a hazard specific to a raw material. Hazard analysis requires process knowledge. A model can organise approved material, but should not replace the team assessing hazards.

Build a minimum data pack for one batch

Choose one product type and one stage. Establish where the batch number comes from and how it links to raw materials, packaging and process records. You need confirmed identifiers. Similar file names and nearby timestamps are not sufficient to establish that a document belongs to a product.

Field	Record	Response to a gap or conflict
Product and batch	Confirmed code and batch identifier	Clarify before further assessment
Process stage	Line, operation and covered period	Check the document scope
Source document	Form, version and storage location	Identify the missing source
Result	Recorded reading, unit and time, without additions	Assign clarification to the responsible person
Criterion	Approved assessment basis	No automatic assessment without it
Deviation	Original account and linked record	Review and establish the action
Decision	Person, date and approved scope	Retain a pending status until resolved

Not every record concerns a critical control point. The control type and assessment method depend on the established system and specific analysis. A tool must not label every measurement as a CCP because that simplifies its interface. It first reproduces the relevant classification from checked documents.

A fictional example: a bakery changeover

This example is fictional and does not describe a Brillnet customer. A bakery makes two products on one line. The next order requires a different set of ingredients and a different label. The quality team wants to collect line preparation, material issue and labelling check records more quickly.

The automation receives an approved list of records and the order number. It finds a preparation form, but the identifier field contains the previous order number. Instead of accepting the document, it highlights the conflict and requests clarification. It does not insert the expected number using the production schedule.

The pack includes a photograph of a label. A model may draft a reading of its content, but comparison with the approved specification requires checking the reading and the applicable version. An omitted ingredient or a character obscured in the photograph could matter. An unreadable image must produce an explicit limitation.

The authorised person assesses the evidence and determines the next action under the site's procedures. Our proposed automation ends at an organised pack with visible gaps. Cleaning parameters, allergen management practices and assessment criteria remain the site's approved arrangements. They are not created from a model's general answer.

Traceability needs confirmed relationships

Consider a separate scenario in which a supplier reports a problem with a particular raw material batch. The team must establish which production orders used it and which deliveries relate to the products under assessment. The first automation task is to prepare the data for that assessment, showing the source of each relationship.

Do not link identifiers solely through text similarity. “A-102” and “A102” may describe the same batch or different batches. The data owner must confirm the rule. Repacking, splitting batches, returns and material reintroduced to the process also need consideration where those operations occur at the site.

Withdrawal, notification and customer communication must follow the applicable obligations and established process. [2] In the initial project, a model should not independently send those communications. The team needs to inspect the scope, original records and intended recipients before taking action.

Retain the substance of deviation records

A deviation account must preserve what was actually observed. If an operator recorded that a reading was missing, do not change it into a result within limits. If a record was completed later, preserve that fact. A neat form must not remove the distinction between a measurement taken on time and information reconstructed afterwards.

AI can propose a structured account covering the place, time, product, observation, action taken and missing evidence. The original text should remain available. The case owner approves the revised account and identifies what has been confirmed. Avoid generating broad lists of possible causes that have no link to the actual records.

Apply the same discipline to corrective action. Recording that an employee was trained describes a completed activity; it does not establish that the cause was removed. In our recommended workflow, each approved action has a way to check effectiveness later. The task closure date and the effectiveness result remain separate records.

Training should lead to an observable action

When an instruction changes, choose situations employees will encounter on their shift. These may include an unavailable approved label, an unreadable batch number, a missing record or a mismatch between an order and the issued raw material. A question should test what to stop, whom to inform and where to record the event.

Check translated terminology against language used at the workstation. A literal equipment translation may be unfamiliar to an operator. A language version must preserve the meaning and identifier of the approved instruction. When the source changes, the owner assesses whether each training version in use needs updating.

CrewShift, known in Polish as Zmianowo, is developing training and tests from internal documents with human approval through early access. This can be considered as support for preparing material. Completing a test does not automatically demonstrate qualification, practical competence or authorisation to work independently.

For a practical assessment, the responsible person should be able to observe the intended task and record the outcome using the site's method. Training content and work authorisation serve different purposes. Keeping that distinction visible prevents a convenient learning record from being used as evidence it cannot provide.

Protect information and limit integration access

Use documents approved for processing and accounts with limited permissions. Supplier, recipe and customer information may require protection whether or not it contains personal data. Establish the service provider, processing locations, retention and deletion arrangements before a regular data flow begins. Include intermediate copies and extracted text in those decisions.

OWASP describes prompt injection, where input content can influence a model's behaviour. [6] A supplier message and its attachment are data for this workflow. They cannot authorise deleting a register, retrieving unrelated recipes or sending documents to an address written in the attachment.

We recommend a defined set of API or MCP operations enforced by the application. Reading documents and saving a draft pack is a different access scope from modifying a recipe or releasing a batch. The initial integration does not require those broader permissions. Service failure should allow the established manual process to continue.

Measure useful results

Measure the time from identifying a batch to preparing the complete pack, including review. Count missing records found before the decision and incorrect document assignments separately. Do not use the number of automatic summaries as a food safety measure. More generated text can simply mean more text for someone to check.

Use completed cases approved for evaluation in a short pilot. Include an unreadable document, a missing unit, conflicting identifiers and a record for the previous order. A correct refusal to assess can be the expected outcome. Recognising missing information may be more valuable than adding another paragraph.

NIST AI RMF is a voluntary AI risk management framework. [5] Our recommendation is to record potential harm, a stopping condition and the response owner. Pause the pilot if the tool assigns unrelated records to a batch or hides deviations. Time savings cannot compensate for that loss of reliability.

Keep a history of corrections and failures

A document may be corrected after a pack has been prepared. Record the new version, date and reason. The person who reviewed the earlier result needs to know that its basis changed. Silently replacing an attachment makes the previous decision harder to explain, even when the current record is accurate.

Specify the response to partial retrieval. If one system responds and another is unavailable, the pack must not appear complete. It should identify the missing source and unassessed scope. A case number and status let the team return to the task without starting the investigation again.

Maintenance includes updating the record list after process changes, reviewing new form formats and looking after the integration. Assign an owner. Otherwise, automation will compare today's production with yesterday's criteria. Before expanding scope, check whether correcting the pack consumes the intended benefit. The pilot should answer that question without assuming a predetermined percentage saving.

Prepare a concrete implementation brief

A starting pack contains one workflow, a deidentified document example, a list of necessary records and the person approving the result. Describe the source of batch identifiers, exceptional operations and manual fallback. Add the current task duration and the problem you want to remove. This supports pricing and reviewing a defined piece of work.

Pulsar GRC is developing the organisation of requirements, controls, risk, corrective actions and evidence using customer documents through early access. It can be considered for organising those relationships after confirming the required scope. We do not present it as a HACCP guarantee or a tool that independently decides whether a product is safe.

Brillnet builds narrow tools and services with controlled integrations. A useful first output may be a batch pack with missing records, visible sources and approval retained by quality. Assessing that result provides the basis for expanding the work.

Describe a food production workflow to Brillnet. Start with the task, data type and decision owner. Arrange the exchange method before supplying recipes or customer information. You can also explore the development direction of Pulsar GRC and CrewShift.

One synthetic nonconformity through CAPA

Use a fictional food-production record describing a missed verification step. Keep it explicitly synthetic and avoid real batch or customer data. Record the requirement, observed deviation and immediate action taken in the exercise. Containment addresses the immediate situation; analysis of the cause and corrective action need their own evidence and review.

Assign the person investigating the cause and define the information needed. An AI draft can help organize the description, but it does not establish the cause. Compare the proposed analysis with the fictional records and reject unsupported conclusions. An authorized person decides which corrective action fits the evidence.

Define an effectiveness criterion before closing CAPA. For example, the exercise can require a reviewer to inspect the revised verification instruction and a specified follow-up record. That is an illustrative method, not a universal food-safety acceptance rule. A real facility determines the criterion from its process and applicable requirements.

In Pulsar, connect the action, owner, deadline, source version and evidence. Read the record after saving and preserve the review state. An uploaded document is not automatically proof that the action prevented recurrence. Keep an unresolved effectiveness check visible rather than closing it to improve the completion percentage.

The 14-day Pulsar trial lets you evaluate this recorded workflow. It does not replace HACCP responsibilities, laboratory checks or a release decision for real product. Review the payment-method, post-trial charge and cancellation terms before starting, then reproduce the synthetic case to judge whether the evidence is easy for your team to find.

Sources

Verified on 2 October 2026. The data structure, automation workflow, measures and fictional example are Brillnet recommendations. This article is not a complete legal inventory for every food business or a validation of a HACCP plan.

European Union, Regulation (EC) No 852/2004, particularly Article 5: https://eur-lex.europa.eu/eli/reg/2004/852. Accessed 2026-10-02. HACCP based procedure obligations, within the scope of the actual activity.
European Commission, Food law general requirements: https://food.ec.europa.eu/horizontal-topics/general-food-law/food-law-general-requirements_en. Accessed 2026-10-02. Official explanation of operator responsibility, traceability and responses to unsafe food under Regulation 178/2002.
European Commission, Guidance platform, Notice 2022/C 355/01: https://food.ec.europa.eu/food-safety/biological-safety/food-hygiene/guidance-platform_en; notice text: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ%3AJOC_2022_355_R_0001. Accessed 2026-10-02. Guidance on GHP and HACCP based procedures.
FAO/WHO Codex Alimentarius, Codes of Practice, CXC 1-1969 General Principles of Food Hygiene: https://www.fao.org/fao-who-codexalimentarius/codex-texts/codes-of-practice/en/. Accessed 2026-10-02. Official public code listing.
NIST, AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework. Accessed 2026-10-02. Voluntary AI risk management framework.
OWASP Gen AI Security Project, LLM01:2025 Prompt Injection: https://genai.owasp.org/llmrisk/llm01-prompt-injection/. Accessed 2026-10-02. Instructions embedded in model input. Sources checked: 2026-10-10.