continued process verification software · cpv software
Continued Process Verification Software: Build vs Buy
November 17, 2026
30 min read
Analyst report on continued process verification software for pharma in 2026: four architecture patterns, GxP data lineage, statistical control formulas, a pilot scorecard and a five-year TCO model.

- 01CPV sourcing spans four architecture patterns: packaged applications, historian or MES modules, validated analytics stacks, and custom data products. No pattern is a universal winner.
- 02Buying can shorten the path to a governed workflow; building offers control over data models and calculation logic. Each pattern creates a different mix of verification, integration, and ownership work.
- 03Data integration and contextualization are the main constraints. Small, self-selected surveys point to spreadsheet-heavy practice and manual reporting rather than providing market-share estimates.
- 04Statistical rules must be versioned and approved. Supplementary rules can detect shifts sooner while increasing false alerts, and assumptions such as independence need explicit checks.
- 05Evaluate candidates with synthetic batch data, known shifts, missing and late results, and a calculation-version change. Score lineage, detection, false alerts, and report regeneration.
- 06Build total cost of ownership from RFP quotes and local labor rates. The article's labor model is hypothetical and excludes licence, validation, hosting, and revalidation costs.
Executive Summary
Continued process verification (CPV) software is not one product category. It is a sourcing decision across four architecture patterns: a packaged CPV application, a historian or manufacturing execution system (MES) module, a validated analytics and statistics stack, and a custom data product with controlled analytics. No public benchmark ranks these patterns, and this report does not name a universal winner. It gives buyers a way to decide and a way to test.
The regulatory baseline is stable and short. FDA's 2011 guidance defines Stage 3 as "continued process verification" and expects an ongoing program to collect and analyze product and process data, with statistical trending reviewed by trained personnel ([1]) ([2]). EMA uses the term "continuous process verification" and describes it as an alternative to traditional validation, with performance monitored on each batch ([3]) ([4]). EU GMP Annex 15 treats "ongoing process verification" as the same activity as continued process verification ([5]). None of these texts names a software category, and no software creates compliance by itself.
The hard problems are data, not charts. In a 2022 ISPE Iberia survey of 28 companies, most used spreadsheets or standard statistical packages and only 11% used advanced analytics ([6]) ([7]). In a BioPhorum survey of 15 members, 40% still created full CPV reports manually ([8]) ([9]). Those figures come from small, self-selected samples, so they indicate direction rather than market share.
Three findings drive the recommendation. First, every pattern can be made to work, but each moves the validation burden to a different place, and cloud providers state that customers remain accountable for their own GxP systems ([10]) ([11]). Second, statistical defaults matter: a 3-sigma individuals chart has a false alarm every 371 points on average, and adding supplementary run rules raises that to about once in 91.75 points ([12]) ([13]). Third, total cost of ownership (TCO) must come from a request for proposal (RFP), because none of the specialist vendors reviewed publishes list prices. As a hypothetical illustration of method, not a finding, staffing assumptions chosen by the author give a custom build roughly $0.9 million more five-year labor than a bought platform; licence, validation, hosting, and revalidation costs are excluded on both sides, so readers should substitute their own figures.
The practical advice is to run a pilot on synthetic batch data with known injected shifts and gaps, and to score lineage completeness, drift detection, false alerts, and report regeneration rather than dashboard appearance. Statistical settings, rational subgrouping, and limit-recalculation rules should be reviewed by a qualified statistician before they are locked into any pattern.
ISPE Iberia respondents using advanced analytics in a survey of 28 companies
BioPhorum respondents still creating full CPV reports manually in a survey of 15 members
ISPE Iberia respondents able to gather one product's CPV data in under an hour; sample of 28 companies
Introduction and Background
This report examines how pharmaceutical and biotechnology manufacturers can source the software and data architecture behind Stage 3 of the process validation lifecycle. It is written for manufacturing quality, manufacturing science and technology (MSAT), process engineering, site data and automation, statistics, validation, and IT architecture teams. The regulatory history of PPQ and CPV is outside the scope of this report, which focuses on source data, contextualization, metric governance, statistical engines, review workflows, validated reports, and change control.
Scope boundary
CPV is distinct from three neighboring activities. Process performance qualification (PPQ) belongs to Stage 2, whereas Stage 3 is the ongoing assurance that the process remains in control during routine production ([14]). The annual product review is a periodic evaluation: 21 CFR 211.180(e) requires records to be kept so the data can be evaluated at least annually ([15]). Real-time process control acts on the process within a batch, while CPV evaluates performance across batches. One practical consequence is that CPV tools can feed an annual review, but an annual review spreadsheet is not by itself a CPV system. FDA says procedures should guard against overreacting to single events as well as against failing to detect unintended variability ([16]).
Terminology: continued versus continuous
FDA uses "continued", and EMA uses "continuous". EMA's guideline says continuous process verification can be introduced at any time in the product lifecycle ([17]), and it allows a hybrid approach in which the dossier states which approach applies to which steps ([18]). Annex 15 states that ongoing process verification is also known as continued process verification ([5]). For software buyers the practical point is that vendor pages and regulators use the two spellings interchangeably, so an RFP should define the term once and map it to the lifecycle stage and the applicable regional text. ICH Q10 frames the same activity as a process performance and product quality monitoring system that ensures a state of control is maintained ([19]).
Pattern 1: Packaged CPV Applications
Capabilities
Packaged applications bundle connectors, contextualized datasets, charting, alerting, review workflow, and reporting. ValGenesis states that its iCPV application uses multivariate statistical tools and control charts within a statistical process control framework ([20]), and that it is integrated with laboratory information management systems (LIMS), MES, and SAP sources ([21]). A MilliporeSigma-supplied listing says Bio4C ProcessPad monitors critical process parameters against set limits ([22]) and uses principal component analysis to find outlier batches within hundreds of parameters and attributes ([23]). Körber says PAS-X Savvy can pull execution and quality data from distributed control system (DCS) and LIMS sources ([24]), and a Körber case-study summary says it imports, harmonizes, and contextualizes online and quality data automatically ([25]). Aizon describes combining traditional statistical calculations with AI for CPV in real time ([26]). All of these are vendor claims, and the documented connector lists are generic.
Adoption
Survey evidence, rather than market-share data, points to a spreadsheet-heavy base: in a 2022 survey, 11% of the 28 ISPE Iberia respondents used advanced analytics ([7]), and BioPhorum noted that many companies are not yet using a fully validated CPV informatics system ([27]). BGO's overview names data collection and verification as the biggest challenge in CPV system development ([28]) and lists alarm systems among system functions ([29]). It describes categories rather than architecture, which is the gap this report addresses.
Strengths and limitations
The strength is time to a working CPV workflow, since charting, rules, and report templates are pre-built. Limitations are mostly about transparency: public pages do not document calculation versioning, how limits are recalculated, or the deployment model in testable detail, and the vendors in this set do not publish list prices. ValGenesis says iCPV flags out-of-trend conditions with statistical rules and configurable alerts ([30]), but how those rules are versioned and how the audit trail records changes to them should be tested in the pilot rather than assumed. Buyers should treat the product's statistical engine as a component that needs the same prerequisite checks described later in this report.
Pattern 2: Historian and MES Modules
Capabilities
Industrial historians and MES platforms already hold time-series and batch context. AVEVA documents PI buffering as a store-and-forward service that writes stored data back to the archive in chronological order ([31]), and says PI Server 2015 and later can migrate batch data to event frames ([32]). GE Vernova's documentation for regulated environments lists electronic signatures on annotations as a Part 11-supporting feature ([33]). On the MES side, Körber says its electronic batch record audit trail permits no deletion or editing ([34]), Siemens documents review of batch records by exception ([35]), and Emerson says its Quality Review Manager supports review by exception ([36]). Rockwell describes master recipes that follow ISA S88/95 standards ([37]). These are vendor statements about general capabilities, and none of the pages reviewed documents a statistical CPV report as such.
Adoption
Public evidence on how often these systems serve as CPV engines, rather than as data sources, was not found. Körber's PAS-X Savvy page positions statistical trending of long-term performance as a use case ([38]), which suggests vendors are moving in that direction.
Strengths and limitations
The strengths are proximity to raw data and existing validated status of the platform. The limitation is that historians and MES modules are optimized for capture and execution, so cross-batch statistics, limit management, and reviewed reports often need another layer. Licensing can also be opaque: AVEVA's PI software schedule describes an aggregate tag model with a monthly fixed charge set in the order form and no published list price ([39]), whereas Canary Labs publishes a perpetual historian licence of $9,970 plus $1,795 for customer care at its default 3,500-tag configuration, with the price varying by tag count and covering historian licensing only, not a CPV capability ([40]). A neutral limitation is that batch-genealogy context may need to be modeled in the historian's asset framework rather than inherited from the MES.
Pattern 3: Validated Analytics, BI and Statistics Stacks
Capabilities
This pattern composes process analytics, statistics software, and business intelligence (BI) tools. Seeq states that it does not duplicate data, so validated sources remain the single source of truth ([41]), publishes a default quarterly update cycle ([42]), and supplies a validation package, regulatory release notes, and a right to audit ([43]). Its support documentation describes an audit trail for changes to items such as calculated signals ([44]). TrendMiner lists standard integrations with historians such as AVEVA PI ([45]). Sartorius describes SIMCA as software for exploring data and building multivariate models offline, with SIMCA-online for real-time deployment ([46]), and connections to process historians and databases ([47]). For statistics engines, Minitab provides a validation kit so customers can validate the software for intended use ([48]), and JMP's quality white paper notes that FDA-regulated customers might be expected to validate the systems they build with it ([49]).
Adoption
The ECA and Concept Heidelberg SPC survey had 34 participants, and Nelson rules were the most used rule set at 55% ([50]) ([51]). That is a small sample, but it shows that rule sets are a configuration decision users already make. R is a further option: the R Foundation reproduces FDA's statement that FDA does not require use of any specific software for statistical analyses ([52]), and notes that validation is more than a verification or testing exercise ([53]). An industry working group concluded there is minimal risk in using base and recommended R packages in validated systems, with a risk-based approach for contributed packages ([54]).
Strengths and limitations
The strength is flexibility: teams can choose the statistical engine and keep source systems authoritative. The limitation is integration ownership. Workbooks, calculated signals, and dashboards are configuration that must be versioned and reviewed, and a tool's validation kit does not validate the customer's workbooks. Published BI prices are modest compared with specialist platforms: Tableau Cloud Standard lists Creator at $75 ([55]), Explorer at $42 ([56]), and Viewer at $15 ([57]) per user per month, billed annually. These are list prices for the BI layer only, and they exclude the data pipeline, validation effort, and statistics governance.
“No pattern dominates: packaged and analytics options shift effort from build to verification, whereas a custom build shifts it from verification to ownership.
Pattern 4: Custom Data Product with Controlled Analytics
Capabilities
A custom data product builds ingestion, a canonical model, a calculation library, and a reporting layer on a cloud or on-premises data platform. The platform vendors document shared responsibility. AWS states that the regulated company is still accountable while the cloud provider handles its side ([10]), and that GxP applications deployed in the cloud still need validation ([58]). Microsoft states that there is no GxP or FDA 21 CFR Part 11 certification for cloud service providers ([11]). Google Cloud says responsibility for GxP compliance ultimately lies with its life science customers ([59]), and Databricks documents built-in audit logging that the customer must configure ([60]). FDA's data integrity guidance explicitly lists cloud infrastructure within the computer systems covered by the CGMP computer-system rules ([61]).
Adoption
No public data quantifies how many sites run custom CPV data products. In the ISPE Iberia survey, 71% of respondents claimed personnel trained in statistics or data science ([62]), a precondition for owning calculation code.
Strengths and limitations
The strengths are control over the data model, calculation logic, and change cadence. The limitations are ownership and cost. The 2025 draft revision of EU GMP Annex 11 states that the regulated user remains fully responsible for validation-related activities even when relying on vendors, service providers, or internal IT ([63]). A custom build also needs a documented calculation library, because the 2011 Annex 11 text requires that the application be validated and the IT infrastructure qualified ([64]). Infrastructure usage prices are published and small per unit: AWS lists S3 Tables standard storage at $0.0265 per GB-month for the first 50 TB ([65]), and Databricks data engineering starts at $0.15 per DBU ([66]). Total cost depends on team size, licences, and validation effort, which the buyer must quantify in the RFP.
Feature Comparison
Table 1 below compares the four patterns on the dimensions that tend to decide the choice. It is a qualitative synthesis built from the vendor and regulator statements above, not a scored ranking.
| Dimension | Packaged CPV application | Historian or MES module | Validated analytics stack | Custom data product |
|---|---|---|---|---|
| Primary strength | Pre-built CPV workflow and report templates ([67]) | Proximity to raw and batch data | Flexible engine over authoritative sources | Full control of model and calculations |
| Source data lineage | Vendor connectors, generic claims | Strong inside the vendor domain | Depends on connectors and workbook discipline | Designed to the data contract |
| Statistical engine | Built in; versioning must be tested | Often limited cross-batch statistics | Tool-specific; validation kit or package | Own code; needs a calculation library and tests |
| Review and e-signature | Workflow claimed by vendors | MES review by exception documented | Often external to the tool | Must be built or integrated |
| Validation burden | Shared with vendor | Shared with vendor, large scope | Split across tool and workbooks | Highest, fully the user's |
| Published pricing | None found | Mixed (AVEVA none, Canary published) | BI and statistics prices published, specialist not | Infrastructure prices published; labor and validation costs must be estimated locally |
| Typical risk | Black-box limit logic | Capture-optimized, analysis-light | Configuration sprawl | Key-person and maintenance dependency |
No pattern dominates: packaged and analytics options shift effort from build to verification, whereas a custom build shifts it from verification to ownership. Hybrids are plausible: a historian or MES as source, an analytics layer for exploration, and a controlled reporting layer for the formal CPV record.
The decision logic can be reduced to four questions:
- Is source data already contextualized? If a historian or MES already carries batch context, favor a module or analytics layer over a new connector build.
- Does the site need to own calculation logic? If metrics are proprietary or change often, favor a custom data product or analytics stack.
- Can the organization staff data engineering and statistics? If not, a build is a risk, not a saving.
- Is vendor audit access available? Without it, packaged and analytics options carry unverified assumptions.
A workable RACI (responsible, accountable, consulted, informed) split for any pattern is as follows:
- Quality assurance: accountable for the CPV plan, signal responses, and approval of reports.
- Statistics or MSAT: responsible for metric definitions, subgrouping, limits, and rule sets.
- Site data and automation: responsible for source interfaces, buffering, and time synchronization.
- IT and validation: responsible for qualification, change control, and periodic evaluation.
- Vendor or platform team: responsible for release notes, defect handling, and audit access.
- Fits when time to a governed workflow matters and vendor validation material can be reused.
- Pre-built charting, rules, and report templates shorten the path to a working CPV workflow.
- Test rule versioning and the audit trail for rule changes in the pilot.
- Fits when data models and calculations are strategic and the organization can sustain the team.
- Build ingestion, a canonical model, a calculation library, and a reporting layer.
- Control the data model, calculation logic, and change cadence; take on ownership and cost.
Neither is universally better. The article also considers historian and MES modules and validated analytics stacks.
Reference Architecture, Data Contracts and the Canonical Model
A CPV data stack has six layers: sources, ingestion, a contextualized canonical store, a calculation layer, a review workflow, and a reporting archive. FDA defines metadata as the contextual information required to understand data ([68]), and that definition is the basis of the canonical model: product, process, process version, site, line, batch, critical process parameter (CPP), critical quality attribute (CQA), analytical method, limit, event, exclusion, and approval. Each exclusion and limit change needs a reason, a reviewer, and an effective date.
Table 2 below sketches a source-to-metric lineage register. The refresh intervals, owners, and retention values are design examples to be set by each site, not regulatory requirements, apart from the retention rule that the QMS and applicable regulations define.
| Source system | Metric examples | Refresh | Owner | Quality rule at ingestion | Retention |
|---|---|---|---|---|---|
| Process historian | Mean hold temperature, pH, tag-derived phase values | Per batch close, with late-data reprocessing | Site automation | Buffered data accepted; gaps flagged, not filled | Per record retention SOP |
| MES or electronic batch record | Batch genealogy, phase times, manual entries | Per batch release | Manufacturing | Batch ID and recipe version required | Per record retention SOP |
| LIMS or chromatography data system | Assay, impurities, bioburden, results below limit of quantitation | Per result approval | QC laboratory | Method and version captured; censored values flagged | Per record retention SOP |
| ERP and materials | Raw material lot attributes, supplier lot | Daily | Supply chain | Lot-to-batch link required | Per record retention SOP |
| QMS | Deviations, CAPA, change records | Event-driven | Quality assurance | Link events to batches and to exclusions | Per record retention SOP |
| Environmental monitoring | Counts, excursions | Per sampling event | Microbiology | Zero-inflated counts handled by an approved method | Per record retention SOP |
The lineage register matters because the evidence expectations are explicit. MHRA says raw data must permit full reconstruction of the activities ([69]), and PIC/S defines data governance as the arrangements that provide assurance of data integrity across the data lifecycle ([70]). Where data move between systems, the 2011 Annex 11 requires that validation include checks that data are not altered when transferred ([71]). Store-and-forward buffering of the kind AVEVA documents ([31]) protects against loss during connection interruptions, but the pipeline must still record when late data arrived and whether any reported metric changed.
Statistical Control: Methods, Prerequisites and Warnings
FDA recommends that a statistician or a person with adequate training in statistical process control develop the data collection plan and the methods used ([72]), and it recommends quantitative statistical methods whenever appropriate and feasible ([73]). EMA expects a continuous process verification scheme to describe the statistical models or tools used ([74]), and Annex 15 says statistical tools should be used where appropriate to support conclusions on variability and capability ([75]). The formulas below are standard and should be confirmed by a qualified statistician before they are configured.
Table 3 below is a formula sheet with prerequisite checks. Cp and Cpk follow the definitions in the NIST/SEMATECH e-Handbook, and Pp and Ppk use the same arithmetic with the overall (long-term) standard deviation instead of the within-subgroup estimate, a convention that the reviewing statistician should confirm.
| Tool | Formula (plain text) | Prerequisites and warnings |
|---|---|---|
| Individuals chart (I-MR) | Limits = mean plus or minus 3 x (average moving range / 1.128) ([76]) | Independence and stable mean; autocorrelated data narrow limits and inflate false alarms ([77]) |
| Xbar-R chart | Sigma-hat = average range / d2 | Range method suits small subgroups ([78]); subgroups must be rational (same conditions within, differences between) |
| p and c charts | p-bar plus or minus 3 x sqrt(p-bar(1 - p-bar)/n); c plus or minus 3 x sqrt(c) | p-chart needs independent units ([79]); c-chart normal approximation needs a Poisson mean of at least 5 ([80]) |
| EWMA and CUSUM | Weighted or cumulative sums of deviations | More sensitive to small drift than a single-point rule ([81]) ([82]) |
| Cp, Cpk | Cp = (USL - LSL) / (6 sigma); Cpk = min[(USL - mean)/(3 sigma), (mean - LSL)/(3 sigma)] | Stable, approximately normal data ([83]); at least 25 values for confidence-interval approximations ([84]) |
| Pp, Ppk | As Cp and Cpk, using overall standard deviation | Do not report from an unstable process; label the sigma used |
| Hotelling T2 (multivariate) | Distance of a batch from the multivariate center | Limits come from distribution theory, not 3 sigma ([85]) |
Several warnings follow from the cited sources. A 3-sigma limit corresponds to a 0.27% point probability, and for an in-control process the average run length to a false alarm is approximately 371 ([12]). Adding Western Electric supplementary rules changes the false alarm frequency to about once in every 91.75 points ([13]). Rule sets also differ between sources: NIST lists six in a row trending ([86]), whereas a BioProcess International article defines Nelson test 2 as nine points in a row on one side of the center line ([87]). The same article reports a false alarm rate of roughly 0.87% per lot for independent data under Nelson tests 1 to 4 ([88]). A rule set should therefore be a versioned, approved configuration, not a vendor default.
Small and unusual datasets need explicit handling. NIST warns that when limits are computed from a limited amount of data, actual chart properties might differ considerably from the theoretical ones ([89]), and cites a Shewhart rule of thumb of at least 25 samples ([90]). BioPhorum's CPV guidance says about 30 batches might reflect all sources of variation, which it does not present as a hard rule ([91]), and prefers long-term sigma ([92]). Bioprocess data are often non-normal, so a quantile limit can suit count data better than three standard deviations ([93]), and a PDA Journal paper examined how non-stationarity, autocorrelation, and outliers affect individuals-chart limits ([94]). NIST also advises recomputing limits only when at least 30 more data points are available and no process change is known ([95]).
Alert governance completes the picture. BioPhorum says responses to CPV signals should not depend on a set of default responses ([96]), and a PDA Letter article notes that no regulatory guidance exists on how industry should react to such signals ([97]). For limited early-commercial data, the same article calls for a deep dive into the available data ([98]).
GxP Controls, Validation and Change Control
The control set is the same for every pattern. An audit trail is a secure, computer-generated, time-stamped record, per FDA's definition ([99]), and 21 CFR 11.10(e) requires such audit trails to record the date and time of operator actions, with record changes not obscuring previously recorded information ([100]) ([101]). Electronic signatures must be linked to their records so that they cannot be transferred to falsify a record by ordinary means ([102]). The Annex 11 consultation draft would require audit trail functionality that logs all manual user interactions ([103]) and is enabled and locked at all times ([104]). As of November 2026, the European Commission consultation page states that the response period has ended ([105]), and no final adopted text has been published in the sources cited here, so the 2011 Annex 11 is treated as the operative text in this analysis.
Validation effort should scale with risk. FDA says the extent of validation studies should be commensurate with risk ([106]), MHRA says validation effort increases with complexity and risk ([107]), and ICH Q9(R1) says the effort, formality, and documentation of quality risk management should be commensurate with risk ([108]). GAMP 5 Second Edition emphasizes critical thinking by knowledgeable subject matter experts ([109]). FDA's computer software assurance guidance was reissued on February 3, 2026 ([110]), but its stated scope is software used as part of medical device production, so it should not be presented as a drug CGMP rule ([111]). PIC/S adds that validation alone does not necessarily guarantee that records are protected ([112]), so administrative controls and training are part of the design.
Change control is where CPV stacks most often drift in practice, so an explicit trigger list helps. The list below is a suggested starting point, not a legal rule; each organization should justify it under its own risk assessment.
- Reassess formally: a change to a calculation formula, limit-setting method, rule set, subgrouping, or the handling of censored values.
- Reassess formally: a new or modified connector, a change in time-stamp or unit conversion, or migration of data to a new platform.
- Verify and document: a vendor release, because the 2011 Annex 11 requires controlled changes including system configurations ([113]); Seeq's default is quarterly releases ([42]).
- Periodic review: Annex 11 requires periodic evaluation to confirm that systems remain in a valid state ([114]), and PIC/S expects evaluation of deviations, changes, and performance ([115]).
- Limit changes: treat as controlled changes with a rationale, since ICH Q10 expects monitoring outputs to drive change management ([116]) and management review across the lifecycle ([117]).
Report reproduction is the acid test. A validated report should regenerate from archived inputs and the versioned calculation configuration, and should match the approved report to a stated tolerance. ICH Q10 also frames knowledge management as a systematic approach to acquiring, analysing, storing, and disseminating information ([118]), which supports keeping signal rationales and limit histories queryable.
“A validated report should regenerate from archived inputs and the versioned calculation configuration, and should match the approved report to a stated tolerance.
Pilot Design, RFP Scorecard and Performance Benchmarks
No public benchmark compares CPV platforms on detection performance, so the buyer has to create one. The recommended pilot uses synthetic batch data generated by the buyer, with a documented random seed, so every candidate sees the same data. The data set should contain:
- Known injected shifts: mean shifts of 1.0, 1.5, and 2.0 sigma at documented batch numbers.
- Missing and late data: whole batches missing and results that arrive after the batch is reported.
- A censored series: results below the limit of quantitation.
- An autocorrelated series: to test whether the engine warns about independence.
- A calculation-version change: introduced midway to test change control and report reproduction.
The expected detection delay for a mean shift can be computed in advance. For an individuals chart with a single-point 3-sigma rule and known parameters, a calculation from the normal distribution gives an expected detection delay of about 43.9 batches for a 1.0 sigma shift, 15.0 for 1.5 sigma, and 6.3 for 2.0 sigma. Run rules and EWMA should shorten these delays at the cost of more false alerts, so both are scored. These figures are theoretical benchmarks for independent normal data, not vendor results.
Table 4 below is a pilot scorecard. Thresholds are examples for the buyer to set before the demonstration.
| Test | Injected condition | Measure | Example acceptance criterion |
|---|---|---|---|
| Drift detection | 1.5 sigma shift at batch 40 of 120 | Batches to first alert | Within the configured rule set's expected delay, documented |
| False alerts | 100 in-control batches, seed fixed | Alerts raised | Matches the theoretical rate for the declared rules (0.27 expected for single-point 3-sigma) |
| Missing and late batches | 5% missing, 3 late-arriving results | Handling and flagging | Flagged, not silently filled; metrics restated with audit trail |
| Censored data | 30% results below limit of quantitation | Documented method | Method named, versioned, and approved |
| Reproducibility | Re-run with the same inputs | Output difference | Identical within stated tolerance |
| Lineage completeness | Trace 20 random values to source | Percent traced | 100% to source record, user, and time |
| Report regeneration | Regenerate a signed report | Match to approved copy | Match, with a recorded configuration version |
| Change control | Alter a calculation | Audit trail and re-approval | Change logged, impact assessed, report flagged |
An RFP built from this scorecard should also require:
- Calculation documentation: the sigma estimator, subgrouping rule, and rule-set version for every chart.
- Versioned limits: an effective date, reason, and approver for every limit change.
- Export of raw and derived data: with audit trail and metadata, in a documented format.
- Release and validation material: release notes, test evidence, and audit access.
- Exit terms: data return and report regeneration after contract end.
The scorecard shifts the evaluation from appearance to evidence. Audit access and defined responsibilities, which MHRA expects for cloud providers ([119]), are contractual items to confirm in the RFP.
Define the canonical model and metric catalog before running the pilot.
Use buyer-generated synthetic batch data with a documented random seed so every candidate sees the same data.
Choose a pattern against a lineage and statistics scorecard, and build a TCO from RFP quotes and local labor rates.
Data Analysis and Evidence
Published quantitative evidence is limited and mostly survey-based, so each figure below carries its sample size and sponsor.
- ISPE Iberia (2022): 28 companies ([6]); 46% may not have implemented real, non-annual CPV ([120]); only 18% could gather one product's CPV data in under an hour ([121]).
- BioPhorum (2022 report): 15 member companies ([8]); 40% create full CPV reports manually ([9]); most data acquisition uses manual entry ([122]).
- ISPE Pharma 4.0 survey: 418 respondents in 2023 ([123]); respondents who had not started fell from 31.2% in 2021 to 15.1% in 2023 ([124]).
The evidence suggests that data integration, not statistical engines, remains the bottleneck: ISPE's authors note that data come from disparate sources in different formats and with different contextualization ([125]), and BioPhorum's earlier report says CPV IT systems must be validated and have sufficient data interfacing ([126]). The surveys are small and from different years, so they cannot be pooled.
Five-year TCO model
Specialist CPV vendors publish no list prices, so a TCO must be assembled from cost categories and verified through the RFP. Table 5 below lists the categories and the public anchors found.
| Cost category | Driver | Public anchor |
|---|---|---|
| Platform licence | Users, tags, sites, modules | Specialist: none published; Canary historian perpetual $9,970 plus $1,795 care at the default 3,500-tag configuration, varying with tag count and not a CPV application ([40]); Tableau Cloud Standard Creator $75 per user per month (Standard Edition list price; Enterprise is higher) ([55]) |
| Connectors | Number and type of sources | Vendor-quoted |
| Validation | Risk-based scope; vendor package reuse | Effort, not price list |
| Data engineering | Pipelines, lineage, monitoring | Software developer median wage $135,980 (May 2025) ([127]) |
| Statistics governance | Rule sets, limit reviews | Statistician median wage $105,650 (May 2025) ([128]) |
| Support and hosting | Cloud storage and compute | S3 Tables $0.0265 per GB-month ([65]); Databricks from $0.15 per DBU ([66]) |
| Change and revalidation | Release cadence, vendor updates | Vendor release cadence |
BLS reports median annual wages of $135,980 for software developers ([127]), $120,230 for data scientists ([129]), and $105,650 for statisticians ([128]). BLS also reports that benefits were 30.0 percent of private-industry employer compensation costs in June 2026 ([130]). Treating wages as 70% of compensation, an illustrative loaded cost is $194,257 for a developer, $171,757 for a data scientist, and $150,929 for a statistician. This is a rough loading from an all-industry average, not a pharma-specific rate.
Illustrative labor model (Hypothetical Example)
Assume a custom build needing 2.0 developer, 0.5 data scientist, and 0.5 statistician full-time equivalents (FTEs) in year 1, then 1.0, 0.25, and 0.25 FTEs in years 2 to 5. Using the loaded rates above, five-year labor is about $1.65 million. Assume a bought platform needs 1.0 developer and 0.25 statistician FTEs in year 1 (connectors, configuration) and 0.5 and 0.25 afterwards, which is about $0.77 million. The labor difference is about $0.88 million, or roughly $176,000 per year. If a platform's licence, hosting, and vendor services cost less than that per year under these assumptions, buying is cheaper, and if they cost more, building can be competitive. The assumptions are the buyer's to replace, and validation, infrastructure, and revalidation costs are excluded.
Implications and Future Directions
The Annex 11 revision and the new Annex 22 on artificial intelligence are 2025 consultation drafts, so CPV software embedding machine learning should be assessed against them as they evolve. The Annex 22 draft applies to static models that do not adapt their performance in use ([131]) and states that it does not apply to generative AI and large language models ([132]). The Annex 11 draft cites increased use of cloud services as a driver of the update ([133]). A conservative design keeps any adaptive or generative component outside the validated calculation path, using it for drafting commentary that qualified reviewers must verify.
Pattern convergence is likely. Vendors in every pattern already claim connectors to the others: packaged applications cite LIMS and MES integration ([21]), analytics platforms cite historian integration ([45]), and a cloud AI vendor describes an open, flexible architecture ([134]). The durable differentiator is therefore governance: versioned metrics, tested reports, and explicit change triggers. A CDMO's own description of its CPV program, which says it often uses control charts for trend analysis ([135]), shows how little of this governance is visible publicly, which is why a pilot is needed.
The sensible sequence is to define the canonical model and metric catalog first, run the pilot on at least two patterns, and choose the pattern whose evidence is strongest rather than whose interface looks best. Nothing in this report is legal or regulatory advice, and statistical and validation decisions should be confirmed by qualified professionals at each organization.
Frequently Asked Questions (FAQs)
What does FDA Stage 3 require, and is CPV software mandatory?
FDA's guidance calls for an ongoing program to collect and analyze product and process data ([1]), with a system for detecting unplanned departures from the process as designed ([136]). It does not mandate software, and the Stage 3 goal is continual assurance that the process remains in a state of control ([137]).
Should we buy or build CPV software?
Neither is universally better. Buy when time to a governed workflow matters and vendor validation material can be reused. Build when data models and calculations are strategic and the organization can sustain the team. The illustrative model above is a hypothetical that demonstrates the method; its labor gap follows from the assumed staffing levels and excludes licence, validation, hosting, and revalidation costs, so it does not show what drives real-world cost differences.
Who are the best CPV software vendors?
No independent ranking exists. Vendors with public CPV claims include ValGenesis, MilliporeSigma (Bio4C ProcessPad), Körber, Sartorius, and Aizon, while Seeq and TrendMiner are analytics platforms often layered over historians. Evaluate them with the pilot scorecard, not a list.
How do CPV tools compare, and what is a CPV data platform?
Compare tools on lineage, calculation versioning, rule-set configuration, review workflow, and report regeneration. A CPV data platform is the set of sources, canonical store, calculation layer, and reporting archive described above.
How is statistical process control software validated?
Risk-based and for intended use. Minitab offers a validation kit ([48]), and R guidance emphasizes that validation covers the whole lifecycle ([53]). The customer still validates its own workbooks and configurations.
Can CPV be automated?
Yes for collection, calculation, alerting, and report drafting, but review and signal responses need trained people ([2]). Automation changes the validation scope and does not remove accountability.
Conclusion
CPV software sourcing is best treated as an architecture decision with four patterns, each with a different place where validation effort lands. Regulators set the outcome, which is ongoing assurance of a state of control, and leave the tooling to the manufacturer. Older, small surveys suggest spreadsheet-heavy practice, manual reporting, and data integration as the main constraints.
The recommended path is to define metrics and the canonical model, choose a pattern against a lineage and statistics scorecard, and build a TCO from RFP quotes and local labor rates. Statistical settings, limit recalculation rules, and change triggers should be reviewed by a qualified statistician and the validation function before go-live. A pilot with injected shifts, gaps, and a mid-study calculation change will reveal more about a platform than any interface.
About IntuitionLabs
Build practical AI for pharma and biotech with IntuitionLabs. We help life-science teams turn complex information and workflows into useful software, governed knowledge systems and AI tools.
IntuitionLabs is an AI consulting, custom software development and data engineering firm serving pharmaceutical, biotechnology, medical-device and other life-science organizations. We work with clinical, regulatory, medical-affairs, commercial, quality and IT teams to connect technology decisions with the work people need to accomplish.
AI consulting and adoption
Our AI enablement services cover readiness assessments, use-case selection, governance and policies, team workshops, adoption measurement and ongoing advisory support. We help organizations structure the information layer behind AI: source material, context, permissions and maintained knowledge that make generated answers useful and reviewable. Private LLM inference and hosted AI options support teams evaluating how to operate AI with appropriate control over their data and infrastructure.
Software, data and life-science workflows
IntuitionLabs develops custom software for pharma and biotech, integrates enterprise systems, and builds data engineering and business intelligence solutions. Areas of focus include AI agents, regulatory research, medical writing, medical affairs, CMC information, competitive intelligence and clinical-document workflows. Our eTMF intelligence work includes cross-system reconciliation and inspection-readiness support.
Enterprise platforms and regulated delivery
We provide Veeva services, application support, managed services, integrations and custom applications, alongside enterprise content work involving platforms such as Egnyte. For regulated workflows, our services include GxP enablement, computer-system validation and software development addressing 21 CFR Part 11 requirements. The applicable controls, validation responsibilities and acceptance criteria are defined for each engagement.
Work with IntuitionLabs
Explore AI enablement, pharma and biotech software development, data engineering and BI, and Veeva services. Contact IntuitionLabs to discuss your workflow, information sources and implementation needs.
IntuitionLabs publishes educational research to help life-science teams make informed technology decisions. Coverage of a product or organization does not imply a client relationship, endorsement or partnership.
Sources / 137

Need Expert Guidance on This Topic?
Let's discuss how IntuitionLabs can help you navigate the challenges covered in this article.
I'm Adrien Laurent, Founder & CEO of IntuitionLabs. With 25+ years of experience in enterprise software development, I specialize in creating custom AI solutions for the pharmaceutical and life science industries.
The information contained in this document is provided for educational and informational purposes only. We make no representations or warranties of any kind, express or implied, about the completeness, accuracy, reliability, suitability, or availability of the information contained herein. Any reliance you place on such information is strictly at your own risk. In no event will IntuitionLabs.ai or its representatives be liable for any loss or damage including without limitation, indirect or consequential loss or damage, or any loss or damage whatsoever arising from the use of information presented in this document. This document may contain content generated with the assistance of artificial intelligence technologies. AI-generated content may contain errors, omissions, or inaccuracies. Readers are advised to independently verify any critical information before acting upon it. All product names, logos, brands, trademarks, and registered trademarks mentioned in this document are the property of their respective owners. All company, product, and service names used in this document are for identification purposes only. Use of these names, logos, trademarks, and brands does not imply endorsement by the respective trademark holders. IntuitionLabs.ai is an AI software development company specializing in helping life-science companies implement and leverage artificial intelligence solutions. Founded in 2023 by Adrien Laurent and based in San Jose, California. This document does not constitute professional or legal advice. For specific guidance related to your business needs, please consult with appropriate qualified professionals.