Agilent teaches AI to reproduce a laboratory’s peak integration decisions

Agilent applies machine learning to a small but consequential analytical task: choosing how chromatographic peaks are integrated. Its MassHunter add-on learns from a laboratory’s existing corrections, aiming to make repeated analysis more consistent while keeping the measured signal available for scientific review.

01 / ProductA learned integration method inside a wider instrument business

Agilent supplies instruments, consumables and laboratory software across analytical science. Its MassHunter portfolio covers acquisition, qualitative analysis, quantitative analysis, library management and reporting for GC/MS and LC/MS systems. AI Peak Integration occupies one part of that larger workflow. It does not replace the instrument, sample preparation, calibration, or the laboratory’s responsibility for deciding whether a reported concentration is defensible.

The current product page specifies phthalates and TDCPP using single-quadrupole gas chromatography/mass spectrometry. A model observes manual integration events and adapts to the laboratory’s analysis patterns. The scientific job is to locate and integrate relevant peaks, including difficult shapes that otherwise require repeated analyst intervention. A successful integration still needs the surrounding analytical method to be valid.

Agilent’s technical overview describes a cloud-based learning system with model versioning and a registry. That creates two connected objects to manage: the original analytical evidence and the version of the learned method applied to it. Treating the output as merely a new number in a spreadsheet would discard much of what makes the approach inspectable.

Agilent’s 13 October 2026 agreement to acquire Proscia also points towards AI-enabled digital pathology. The announcement expected closing early in fiscal 2027, subject to conditions. That is a pending portfolio expansion, separate from the available MassHunter application examined here; it does not make Proscia’s software part of an AI Peak Integration licence.

02 / AudienceUseful where a repeated analytical task has a stable definition

A consumer-products testing laboratory running supported chemical analyses is a plausible fit. Its analysts already recognise the compounds, understand the expected chromatograms and correct integration when necessary. Capturing those decisions may be more useful than asking a general AI assistant to interpret scientific files without the same instrument and method context.

A research laboratory working across unrelated compounds should be more cautious. The scope documented for this particular add-on does not establish support for every LC/MS or GC/MS application in the broader MassHunter family. Agilent’s 2023 research poster discusses additional substances as future possibilities; that historical prospect is not a current entitlement or a supported production method.

Bruker offers a useful adjacent comparison: its AI processing work is tied to NMR signals and a different analytical discipline. Thermo Fisher Scientific spans another instrument and software ecosystem. The relevant comparison is the task, supported instrument and review process, rather than whether each company uses the phrase artificial intelligence.

03 / WorkflowProposed workflow: preserve the analyst’s judgement as training evidence

Start with a defined consumer-product sample family and one existing quantitative method. Assemble representative historical chromatograms that the laboratory is authorised to use, including ordinary peaks, broad or overlapping regions and cases requiring difficult baseline decisions. Preserve the original files and document why each reference integration was accepted. Disagreement between qualified analysts should be resolved before it becomes contradictory training material.

Next, establish the supported MassHunter and instrument configuration with Agilent. In the proposed trial, analysts continue their normal processing while the learning system gathers relevant integration examples. The technical overview explains how the plug-in captures chromatographic information and corrections for model development. Keep the trial separate from any decision to change the laboratory’s released reporting method.

Reserve a different sample batch for evaluation. Compare the learned integration with the existing method and an independently reviewed reference. A useful result records the difference in integrated area, the cases requiring intervention and whether the final interpretation changes. Counting fewer mouse clicks alone would miss a model that is convenient but systematically handles one troublesome compound differently.

When a candidate model is accepted, record its version with the result set. Re-run selected historical samples with that saved version to check reproducibility. Agilent describes version-controlled models on its product page; the laboratory should demonstrate that its installed workflow actually retains the identifiers needed for an investigation, rather than relying on a general statement about reproducibility.

Finally, decide how new corrections enter future training. A changed sample matrix, a different preparation protocol or a new analyst can alter the examples supplied to the model. In this proposed evaluation, a model update receives a defined review before replacing the accepted version. Sequenced has not run MassHunter, processed samples or measured the product’s claimed productivity gains.

04 / PricingBuy the supported application and the cloud arrangement together

The published material describes a cloud-based software solution and a commercial enquiry route, but this review did not verify a current public monetary tariff. Request the price for the named AI Peak Integration capability, the required MassHunter installation, support and any relevant instrument configuration. A quote for an existing workstation does not establish that the AI service is included.

Regional access is consequential. The product page limits current availability and support to the United States and specified European countries: Austria, Belgium, Denmark, Finland, France, Germany, Ireland, Italy, the Netherlands, Sweden, Spain, Switzerland and the United Kingdom. A laboratory elsewhere should obtain explicit eligibility confirmation before planning an implementation around the add-on.

The technical overview identifies the software-as-a-service model. Ask which data leave the workstation, where training and prediction occur, how service interruption is handled and what happens to model access at contract end. These are questions for the actual agreement; public descriptions do not establish a particular customer’s retention period, processing region, fee unit or renewal terms.

05 / DistinctionsThe valuable distinction is learning a defined analytical convention

Agilent’s approach learns from decisions already made within an established analysis environment. That can make the project easier to explain than a broad effort to introduce AI across an entire laboratory. The training examples correspond to a recognisable action: correcting how a measured signal is integrated. Experts can examine those actions and decide whether the model has captured the intended convention.

The quality-control discussion presents AI as a way to reduce repetitive work. Those vendor examples are useful for identifying a potential bottleneck, but their outcomes should not become the laboratory’s forecast without local measurement. Saving time on routine peaks may produce little overall benefit if sample preparation or final scientific approval remains the dominant constraint.

Another distinction is that a saved model can be associated with a historical analytical result. This makes version discipline central to the operating model. An investigation should be able to distinguish a change in the sample, a change in conventional processing and a change in learned integration, instead of treating every difference as unexplained AI variability.

06 / QuestionsCheck scope, teaching quality and the consequences of a correction

The product’s narrow supported analytes are an advantage for a clearly matched laboratory and a boundary for everyone else. Avoid extrapolating from a successful phthalate demonstration to pesticide screening or a different mass-spectrometry platform. Those may be worthwhile research questions, but they need their own supported configuration and evidence.

The public FAQ link returned navigation without substantive answers during this research. The technical overview and current product page were readable, so the article uses them for mechanism and access boundaries; it does not infer missing FAQ commitments. Authenticated licence terms, customer data-processing agreements and a live installation were not inspected.

A model trained on inconsistent manual decisions may reproduce inconsistency more efficiently. Agree how analysts handle ambiguous peaks and retain examples rejected from the reference set. Also distinguish an integration correction from a conclusion about sample safety: the software assists a measurement step, while the laboratory remains responsible for interpreting the complete method and reporting context.

07 / DecisionBegin with a supported method that the laboratory already understands

Agilent is relevant to AI adoption because its software puts learning next to real analytical measurements. The strongest starting case is a repeated, supported analysis whose manual corrections can be explained and checked. Establish regional eligibility, preserve raw evidence and model versions, and judge the trial by reviewed analytical results as well as the work it saves.