01 Data engineering
The business finds the failure before the data team does
Stale, duplicated or invalid data breaks no job. It shows up at month-end close, in a board report or in an invoice that never goes out.
On a larger screen, the presentation opens as slides.
Download as PDFProduct presentation · October 2026
BeryX
Data quality, GDPR and the cost of data in one layer: read at the source, read-only, with the calculation behind every number in plain sight.
For the data team, for the DPO and for whoever pays the cloud bill.
BeryxTech · October 2026
1 · The problem
Decisions and AI already run on data. Trust in that data has not kept up.
67%
do not fully trust the data they use to make decisions (55% in 2023)
Precisely and Drexel LeBow, Sep 2024 · 565+ professionals · vendor study
64%
name data quality as their top data integrity challenge (50% in 2023)
71%
worry that incorrect or hallucinated outputs reach decision-makers
dbt Labs, State of Analytics Engineering, Apr 2026 · 363 professionals · vendor study
60%
of AI projects without AI-ready data will be abandoned through 2026, according to Gartner's forecast
1 · The problem
Three costs that appear in no budget
02 GDPR and AI
Accuracy of personal data is a legal obligation
The GDPR requires accurate, up-to-date data, answers to data subjects within a month and breach notification within 72 hours. Infringing the principles carries fines of up to €20M or 4% of worldwide annual turnover. The AI Act adds requirements on training data from December 2027.
03 Data FinOps
The data bill grows with no owner
Tables nobody reads keep being loaded. One badly written query multiplies the day's cost and only shows up in the monthly bill. You pay compute to process data that fails its checks.
1 · Data engineering
Data fails silently: the jobs stay green and the numbers go wrong
Anatomy of a data incident Illustrative example · synthetic data
- Friday, 6 pmThe marketplace connector stops syncing. No job fails: rows simply stop arriving.
- WeekendMarketplace orders never reach the data warehouse. The sales dashboard shows a drop that never happened.
- Monday, 10 amThe weekly close does not add up. The investigation starts: which table, since when, and who else uses this data?
- TuesdayWith the source found, it remains to work out which reports and models have already consumed the wrong data.
Six ways to fail without an error
Freshness
the table stopped receiving data
Volume
half the rows arrived
Schema
a column changed type or name
Duplicates
a retry wrote the same order twice
Validity
an 8-digit tax number, an IBAN with the wrong check digit
Integrity
orders without a customer after a delete without cascade
1 · What it costs
The cost of not knowing is counted with your numbers
| Pillar | What to count | How to estimate | What BeryX measures |
|---|---|---|---|
| Data engineering | Data incidents a month, hours of investigation and days to detection | Incidents × hours × hourly cost, plus days to detection × value at risk per day (invoices not yet issued, for example) | History of issues and incidents with dates, and estimated impact in euros per issue, labelled as a heuristic |
| GDPR | Data subject requests a month and datasets with personal data but no owner or no rules | Requests × hours per request; maximum exposure of €20M or 4% of worldwide turnover | Personal data map with the rule coverage of each column |
| Data FinOps | Tables unread for 90 days, compute spent on data that fails its checks, and daily spikes | Monthly cost of those tables, plus the compute attributed to datasets with open issues | Cost read from the platform's system views, per day (per dataset on BigQuery, per account on Snowflake and Databricks), cost spikes and real usage per table |
1 · GDPR and AI
What the law asks of data, and that can only be proven with measurements
| Obligation | What it requires | Deadline or limit | What it demands of the data |
|---|---|---|---|
| Accuracy · GDPR, Art. 5(1)(d) | Personal data accurate and kept up to date; inaccurate data erased or rectified without delay | Fine of up to €20M or 4% of worldwide turnover (Art. 83(5)) | Knowing which personal data is wrong, where and since when |
| Data subject rights · Art. 12(3) | Answer requests for access, rectification or erasure | One month, extendable by two more | Knowing which systems hold each person's data |
| Personal data breach · Art. 33(1) | Notify the supervisory authority (in Portugal, the CNPD) | 72 hours, where feasible | Knowing which data and which data subjects were affected |
| Impact assessment · Art. 35 | Assess before high-risk processing | Before the processing starts | An inventory of data categories and data flows |
| Data for high-risk AI · AI Act, Art. 10(3) | Training, validation and testing data that is relevant, representative and, to the best extent possible, free of errors and complete | From 2 December 2027 (Annex III) | Measuring errors and completeness in the data sets |
1 · Real cases in Portugal
Portugal's data protection authority has already fined bodies millions over personal data handling
€4.3M
Statistics Portugal (INE)
December 2022
2021 Census. Five infringements: health and religion data processed without a clear lawful basis, insufficient information to respondents, a US processor chosen without due diligence, international transfers without supplementary measures and no impact assessment.
Status INE announced it would challenge the decision in court.
€1.25M
Lisbon City Council
January 2022
Data of protest organisers sent to embassies and to the Ministry of Foreign Affairs: 225 infringements.
Status The courts upheld the infringement. With part of the infringements time-barred, the amount fell to €738k at the Central Administrative Court South (August 2025), in a decision that can still be appealed.
€400k
Barreiro-Montijo Hospital Centre
2018
The first GDPR fine in Portugal: 985 accounts with a doctor profile for 296 doctors, and improper access to clinical records.
Status It became a single fine of €380k. In July 2020, the CNPD agreed to waive it, citing the pandemic and the hospital's finances.
1 · Real cases in Portugal
Fines rise and fall with the CNPD's resources. Notified breaches keep rising.
- 152020
- 602021
- 712022
- 902023
- 232024
- 22025
472
personal data breach cases in 2025, 42% more than in 2024 (357 in private bodies, 115 in public ones)
88
infringement proceedings opened in 2025 which, in the CNPD's words, could have led to 90 fines instead of 2
1 · Data FinOps
The data bill grows with no owner and no link to quality
84%
say managing cloud spend is the top cloud challenge for organisations today
Flexera, State of the Cloud 2025 · 750+ professionals · vendor study
- Tables nobody reads keep being loaded, stored and monitored.
- One badly written query multiplies the day's cost, and it only shows up in the monthly bill.
- You pay compute to process data that fails its checks.
- AI costs grow with no cap and no split by team.
1 · The problem
Where does the truth about your data live today?
Today
- In tests and scripts scattered across teams and repositories
- In the memory of whoever knows the tables
- In a spreadsheet record of processing, reviewed once a year
- In the cloud bill, at the end of the month
- In dashboards nobody links back to where the data comes from
With BeryX
- In versioned rules and contracts that run on every scanExists today
- In a catalogue with owner, domain, sensitivity and usage per datasetExists today
- In a personal data map updated on every scanExists today
- In the daily cost read from the system views of BigQuery, Snowflake and DatabricksExists today
- In incidents with lineage, an owner and a deadlineExists today
1 · Why now
Why now
- 2025Personal data breaches notified to the CNPD rise by 42%, to 472 cases.CNPD, 2025 Activity Report
- Sep 2025The CNPD proposes a faster infringement regime to Parliament, to stop cases becoming time-barred.CNPD, 2025 Activity Report, p. 12
- 2026Gartner forecast: organisations abandon 60% of AI projects that lack AI-ready data.Gartner, 26/02/2025
- 2 Aug 2026The AI Act applies in general (Art. 113).AI Act Service Desk, Art. 113
- 2 Dec 2027High-risk obligations for Annex III, including the quality of training, validation and testing data (Art. 10).AI Act Service Desk, Arts. 10 and 113
- 2 Aug 2028High-risk obligations for Annex I, the regulated products.AI Act Service Desk, Art. 113
2 · The thesis
A trust layer over the data you already have, without copying it
Your sources stay where they are
- Databases
- Data warehouses
- Files and object storage
- Business applications: SAP, Salesforce, Dynamics 365
- Documents
- Streaming
→
BeryX · read-only
- Discover
- Validate
- Monitor
- Explain
- Prove
- Aggregates computed at the source
- Limited samples, with retention
- Proposes fixes, never writes to the sources
- Every number says how it was calculated
→
Data engineering
incidents with lineage, SLOs, rules as code
DPO and legal
personal data map, sensitivity, evidence
FinOps and leadership
cost and quality side by side, real usage, reports per role
Where the team already worksSlackTeamsJiraPower BIEmailWebhooksAPIMCP
2 · The thesis
From connected data to proven data, and back to the start
- 01Connectread-only credentials
- 02Discoverprofile, personal data, usage
- 03Rulesproposed by the profile, accepted by people
- 04Monitorfreshness, volume, schema, anomalies, cost
- 05Incidentlineage, impact, owner, deadline
- 06FixSQL or a pull request, rehearsed first
- 07Provereports, evidence, audit
↺ What is learned flows back into the rules: thresholds, flagged false positives and rules suggested from resolved issues.
Every step is recorded: who decided, when and on what evidence.
2 · The thesis
Every dataset carries its own trust record
Owner and domain
who answers for it and which area it belongs to
Exists todaySensitivity per column
public, internal, confidential or personal data, with author and date
Exists todayRules and contract
active rules and a versioned ODCS contract
Exists todayScore and SLO
five dimensions, trend and service objective
Exists todayReal usage
queries per day: hot, active, cold or unused
Exists todayCost
attributed compute, when the platform reports it
Exists todayIncidents and lineage
what failed, upstream and downstream
Exists todayAudit
every change, written in the same transaction
Exists today3 · The cycle · step 1 of 7
Connect
Input
A read-only credential for a database, a warehouse, files or an application.
What BeryX does
Tests the connection, discovers schemas and tables and imports the metadata. The credential is encrypted before it is stored and the address is checked against SSRF before the first request.
Exists todayWho decides
The workspace administrator chooses the source, the scope and how often scans run.
Governance and evidence
Read-only enforced in code: the HTTP client refuses writes before they leave, and Postgres sessions open in read-only mode. In tests, real SQL Server, Oracle and MongoDB servers refused writes.
Exists todayOutput
A catalogue of datasets with schema, schedule and encrypted credential.
Exists todayConnects to
PostgreSQLSQL ServerOracleBigQuerySnowflakeDatabricksS3SAPSalesforceDynamics 365
Handed to the next step
The list of datasets and columns goes on to profiling.
Value
No write credentials, no software installed on the sources and no copied tables.
Exists today3 · The cycle · step 2 of 7
Discover
Input
The connected datasets.
What BeryX does
Column profiling with aggregates computed at the source: nulls, distinct values, patterns and distributions. Recognises tax numbers, IBANs, emails and postcodes, builds the personal data map, classifies sensitivity on four levels and estimates re-identification risk (k-anonymity). On platforms with telemetry, it measures the real usage of each table.
Exists todayWho decides
The DPO confirms or corrects the sensitivity of each column. A scan never overrides a manual decision.
Governance and evidence
Every number says whether it is exact, estimated or sampled, with a 95% confidence interval. Every classification change is recorded with author and date.
Exists todayOutput
Profile, score per dimension, personal data map, sensitivity per column and usage.
Exists todayConnects to
Semantic typesPII mapk-anonymityBigQuerySnowflake
Handed to the next step
The profile feeds the suggested rules.
Value
The DPO sees where personal data lives, read from the sources themselves, without relying on a manual inventory.
Exists today3 · The cycle · step 3 of 7
Rules and contracts
Input
The profile, the suggestions and the tests the team already has.
What BeryX does
Sixteen rule types: formats with check digits (tax numbers, IBANs, cards), uniqueness, ranges, reference lists, referential integrity, freshness, volume and custom read-only SQL. Accuracy against external sources, opt-in: VIES, GeoAPI.pt, numbering plans and MX records. Imports dbt tests, Great Expectations, SodaCL and ODCS contracts.
Exists todayWho decides
The profile and AI propose; a person accepts each rule.
Governance and evidence
Customer SQL goes through a guard: a single SELECT, in a transaction that is always rolled back. Configuration also lives as code, in a versioned datavault.yaml.
Exists todayOutput
Active rules per dataset and a versioned contract; export to dbt, with SodaCL and Great Expectations still to be switched on.
PartialConnects to
dbtGreat ExpectationsSodaCLODCSVIESGeoAPI.pt
Handed to the next step
Accepted rules run on every scan and feed the monitors.
Value
Rules stop living in one person's scripts.
Exists today3 · The cycle · step 4 of 7
Monitor
Input
Rules, profile history and cost and usage telemetry.
What BeryX does
Automatic freshness, volume and schema monitors, up to 500 per connection on every scan. Anomalies with a modified z-score and false-discovery control (Benjamini-Hochberg), distribution drift, an SLO per dataset and daily cost spikes, with the day's most expensive queries on BigQuery.
Exists todayWho decides
The dataset owner sets the SLO and the severity. A false positive marked «ignored» silences the series.
Governance and evidence
Published methods, tested against the code. No quality, anomaly or cost figure comes from an LLM.
Exists todayOutput
Alerts in the app, by email, in Slack, in Teams and by webhook, with on-call escalation.
Exists todayConnects to
SlackTeamsEmailSigned webhooks
Handed to the next step
An alert opens or updates an issue, which can join an incident.
Value
The failure reaches the data team first, not the month-end close.
Exists today3 · The cycle · step 5 of 7
Incidents and root cause
Input
An alert or an open issue.
What BeryX does
Groups issues into incidents through lineage (foreign keys, view SQL, dbt manifest, OpenLineage) and shows the downstream impact. AI proposes the likely cause from metadata and limited samples. Response deadlines by severity, with escalation.
Exists todayWho decides
The domain owner accepts, assigns and closes the incident.
Governance and evidence
Audit trail written in the same transaction as the change and exportable to the SIEM: Splunk, Sentinel or Google SecOps.
Exists todayOutput
An incident with lineage, impact, owner, deadline and history, and a ticket in Jira, Linear, ServiceNow, Azure DevOps, GitHub or GitLab.
Exists todayConnects to
JiraLinearServiceNowAzure DevOpsGitHubGitLabOpenLineage
Handed to the next step
The likely cause and the impact go on to the proposed fix.
Value
Knowing which reports and models were affected before anyone asks.
Exists today3 · The cycle · step 6 of 7
Fix
Input
An incident with a likely cause.
What BeryX does
Proposes the fix in SQL, in the source's dialect, or in PySpark, and rehearses it on a sample before it is accepted. From the Business plan, the approved fix opens a pull request in the customer's GitHub or GitLab. Standardisation and deduplication also come out as proposals.
Exists todayWho decides
A person approves and the customer applies it through their own CI. BeryX never writes to the sources.
Governance and evidence
A proposed, accepted, applied and verified flow: the next scan confirms whether the issue is gone.
Exists todayOutput
A pull request or approved SQL, linked to the dataset and the ticket.
Exists todayConnects to
GitHubGitLabSQLPySparkdbt
Handed to the next step
The next scan's verification closes the loop and goes into the history.
Value
The fix is reviewed and versioned like any other code.
Exists today3 · The cycle · step 7 of 7
Prove
Input
The history of scans, incidents and decisions.
What BeryX does
Reports per role (engineering, management, leadership) with estimated impact in euros, as PDF and on a schedule; a CDO panel with targets and traffic lights; the AI Act Art. 10 dossier (Business); feeds for Power BI; audit export; a public badge per data product.
Exists todayWho decides
Leadership reads; the DPO and internal audit use the evidence.
Governance and evidence
The impact in euros is a declared heuristic that tends to underestimate. Every number says how it was calculated.
Exists todayOutput
Reports, dossiers and exports with date and source.
Exists todayConnects to
Power BITableauLookerPDFCSV
Handed to the next step
What is learned flows back to the rules and thresholds of step 3.
Value
Answering an auditor or the regulator with history, not with a hurried reconstruction.
Exists today3 · One story, end to end
Monday at the online shop: from a stalled connector to the weekly report
A demonstration with BeryX's «Loja online» pack: synthetic data from a Portuguese online shop, with eight planted issues. Illustrative example · synthetic data
- 01ConnectThe orders database connects read-only. The first scan profiles customers, products, orders and order lines in seconds.
- 02DiscoverThe profile flags seven columns with personal data (name, email, tax number and address) and the format rule reports 4% invalid emails.
- 03RulesThe e-commerce pack installs rules for email, ISO country, postcode, order uniqueness and freshness. The profile proposes more rules, which a person accepts or rejects.
- 04MonitorThe freshness rule fails: the most recent order is 72 hours old, against a limit of 24. The marketplace connector stopped on Friday.
- 05IncidentLineage groups related issues into incidents, each with its origin, severity and the datasets affected.
- 06FixFor the 14 repeated order_id values, BeryX proposes the diagnostic SQL and the fix, also in PySpark. The fix is flagged to run only after the diagnosis is validated.
- 07ProveThe executive report sums up the business risk, the estimated financial exposure, the trend and the critical incidents, with the recommended action.
4 · Inside
What the team sees
4 · Inside
Reference architecture
People and tools
Web appSlack and TeamsPower BI, Tableau and LookerAI agents over MCPVersioned public API
BeryX · Google Cloud, europe-west1
APIAsynchronous workers for scans and reports, with queues and a DLQPostgres holding the metadata, never the tablesAI facade with a per-task routerMCP server with 21 read-only tools
Your sources
Read-only connections through each vendor's official methods
What leaves the sources: aggregates, up to 20 example values per column and, for row rules, a sample of up to 5,000 rows per table.
- Shared SaaS in the EU (europe-west1)Exists today
- Dedicated instance per customer, with its own egress IP and optional CMEKTo be switched on
- Running inside the customer's perimeterVision
4 · Inside · AI
AI with rules: where it comes in, where it does not, and who decides
Where AI comes in
- Rule suggestions, likely cause of incidents and reports per roleExists today
- A natural-language assistant over the catalogue and the issuesExists today
- Glossary and dataset documentation, always for human approvalExists today
Where it does not
- No quality, anomaly or cost figure comes from an LLM: the methods are deterministic and publishedExists today
- Proposing never executes: every action asks for human approval and is auditedExists today
Customer controls
- AI switched off on any plan, with reports in offline modeExists today
- Raw or synthetic samples, as the workspace choosesExists today
- Your own Anthropic key (Business) or Claude on Amazon Bedrock or Vertex AI, in European regions (Enterprise)Exists today
Guards
- Daily AI cost cap and a monthly token quota per planExists today
- Per-task router: by default, only Anthropic receives customer dataExists today
- Output guard that masks secrets, emails, IBANs, cards and tax numbers before storing (it does not yet cover the streaming assistant)Partial
- Defence against indirect prompt injection, with adversarial tests in CIExists today
- Marking of AI-generated content, as AI Act Art. 50 requiresExists today
To be switched on By default, limited samples go to Anthropic, in the US, under standard contractual clauses. The anonymiser (Presidio) is built but not yet in the production image.
4 · Inside · GDPR
GDPR: from a manual inventory to a map that updates on every scan
Knowing where personal data lives
- Personal data map: which datasets hold which types of data, how each was detected and whether it has active rulesExists today
- Sensitivity per column on four levels, with inheritance through lineage (suggested) and curation with author and dateExists today
- Re-identification risk from quasi-identifiers (k-anonymity); examples sent to AI are generalisedExists today
Keeping it accurate
- Check digits for tax numbers, IBANs and cards, and reference listsExists today
- Accuracy against VIES, GeoAPI.pt, numbering plans and MX records, opt-inExists today
- Accuracy audit by human sampling, with a confidence intervalExists today
Answering and proving
- A register of data subject requests with each jurisdiction's legal deadline: EU, United Kingdom, Switzerland, Brazil, California and AngolaTo be switched on
- Searching for and erasing a data subject's value in the samples BeryX storesTo be switched on
- Sample retention configurable per workspace and an audit trail exportable to the SIEMExists today
- Scanning for forgotten credentials in code and notebooksExists today
The CNPD cases, as questions the data answers
- Where is the health, religious or financial data?
- The personal data map and sensitivity per column. Exists today
- Who has access, and does it match who should?
- A referential integrity rule between the account directory and the staff register. Exists today
- Where does the data go?
- Lineage between the connected sources. Data sent by email or to third parties stays out of scope. Partial
4 · Inside · FinOps
Data FinOps: cost read from the platform itself and crossed with quality
Cost and quality side by side (Business)
- Native telemetry: BigQuery (INFORMATION_SCHEMA.JOBS), Snowflake (WAREHOUSE_METERING_HISTORY) and Databricks (system.billing.usage)Exists today
- How much of the month's compute ran on data that fails its checksExists today
- Cost per table on BigQuery; on Snowflake and Databricks the cost belongs to the account, and the split is labelled as an attribution. Without telemetry, no number appears at allExists today
Cost spikes
- A daily spike opens an issue and alerts the channels; on BigQuery, with the day's 5 most expensive queries and their literals maskedExists today
- Published criterion: z ≥ 3, at least 50% and 1 USD above the median, with 7 days of historyExists today
Real usage
- Hot tables (20 or more queries a day), active, cold and unused (0 queries in 90 days, only with telemetry covering at least 80 days)Exists today
- Unused tables become candidates for decommissioningExists today
- Usage-driven scans: hot tables deeply on every scan, unused ones only in the monthly profile, opt-in and approved by the stewardExists today
BeryX's own cost
- A bytes cap before reading on BigQueryExists today
- AI cost per report, per task and per provider; scheduled reports through the batch API at half priceExists today
- Chargeback of usage per workspace, domain and team, with a monthly CSV and a Power BI feedExists today
Partial Cost telemetry is built and tested, but not yet proven on a real BigQuery, Snowflake or Databricks account. It measures compute cost, not how efficient the data format is.
4 · Inside · Security
Security: read-only enforced in code, and the gaps put in writing
Access to the sources
- Read-only in the code, tested against real serversExists today
- Each vendor's official methods, and credentials encrypted (Fernet) with key rotationExists today
- Anti-SSRF with pinned DNS, for webhooks and tickets tooExists today
Isolation and identity
- A test walks every API route looking for leaks between customersExists today
- SSO over OIDC and SAML 2.0 on every plan; MFA mandatory for organisation administratorsExists today
- An IP allowlist per organisation, for API tokens tooExists today
Trail and control
- Audit written in the same transaction as the change, with export to the SIEMExists today
- BeryxTech support access only with consent, read-only and for 72 hours at mostExists today
- Export of the workspace's data and cascading deletion with a signed certificateExists today
4 · Inside · Ecosystem
Ecosystem: 29 source types and the tools the team already works in
Databases 6
PostgreSQLMySQL and MariaDBSQL Server, Azure SQL and SynapseOracleMongoDBSQLite
Warehouses 4
BigQuerySnowflakeDatabricksMicrosoft Fabric*
Files and object storage 6
CSVGoogle Sheets*Google Cloud StorageAmazon S3Azure Blob and ADLS Gen2SFTP
Business applications 5
SalesforceSAP S/4HANA and SAP GatewayOData v2 and v4Dynamics 365REST and JSON
Documents 4
ConfluenceJiraSharePoint*Google Drive*
Streaming 2
Kafka*Pub/Sub*
Collaboration 2
SlackTeams*
* Code done; the driver is not yet in the SaaS image.
- Catalogues · OpenMetadata and DataHub; Collibra and Purview validated only against a simulated transportPartial
- BI · Power BI, Tableau and LookerExists today
- Tickets · Jira, Linear, ServiceNow, Azure DevOps, GitHub and GitLabExists today
- Orchestration · dbt Cloud, dbt Core, Airflow, Fivetran and AirbyteExists today
- AI agents · Published MCP server, with 21 read-only toolsExists today
- Developer packages · Python SDK, Terraform provider, VS Code extension, Airflow and Claude Code plugin, not yet publishedTo be switched on
- SIEM · Splunk, Microsoft Sentinel and Google SecOpsExists today
4 · Inside · Methods
Every number says how it was calculated
Three kinds of number
exact, estimated or sampled, marked in the app
Exists todayConfidence intervals
at 95%, Wilson with finite population correction; Fisher's z for correlations
Exists todayScore with fixed weights
completeness 25%, uniqueness 25%, validity 20%, timeliness 15% and consistency 15%
Exists todayAnomalies
modified z-score (median and MAD) with Benjamini-Hochberg false-discovery control
Exists todayImpact in euros
a declared heuristic, always estimated, that tends to underestimate
Exists todayPublic methods
the document is published, and a test fails if a documented constant changes in the code
Exists today5 · Who it is for
Six roles, each with its own guide and its own tour
Workspace administrator
Connects sources, manages the team, the alert channels and the AI policy.
Exists todayRole guide →
Member
Works on issues, rules, suggestions and incidents.
Exists todayRole guide →
Viewer
Reads scores, reports and the catalogue without changing anything. Cannot yet be assigned on the Team page.
PartialRole guide →
Domain owner
Decides on the datasets, incidents and rules of their domain.
Exists todayRole guide →
Organisation administrator
Governance of every workspace: SSO, IP allowlist, data policy and chargeback.
Exists todayRole guide →
Billing
Plan, invoices and the cost report per team.
Exists todayRole guide →
5 · Who it is for
What changes for each function
| Function | Today | With BeryX |
|---|---|---|
| Data engineering | Finds failures through the business; tests scattered across repositories | Automatic monitors, incidents with lineage, versioned rules and a CI gate |
| DPO and legal | A manual inventory, reviewed once a year | A personal data map on every scan, sensitivity with an author and exportable evidence |
| FinOps and finance | The bill at the end of the month, with no link to quality | Daily cost per dataset or per account, cost spikes with the most expensive queries, and unused tables |
| Leadership and CDO | Reports with no source and no trend | A panel with targets and traffic lights, and reports per role |
| AI teams | Training data nobody measures | Readiness per purpose, the Art. 10 dossier and synthetic data with a privacy assessment |
6 · Value
The landscape: where each approach is strongest
Enterprise suites · Collibra, Informatica
- Catalogue and glossary as the company's system of record
- ISO 27001 and SOC 2, with a public trust centre
- Cleansing and MDM that write to your systems, with write credentials
Cost Public price on UK G-Cloud: £74,999 to £122,919 a year, plus implementation
UK G-Cloud — Collibra Data Quality (service 136629392504870) · UK G-Cloud — Informatica IDMC (service 727452156234281) · Collibra — Trust Center
Open source · Great Expectations, Soda
- Free and running inside your perimeter
- Unlimited SQL and Python in the tests
- Every test is written and maintained by your team
Cost Licence 0; Soda Team at $750 a month, plus compute and engineering hours
Great Expectations — plans and pricing · Soda — plans and pricing
Observability · Monte Carlo, Sifflet
- Mature metadata monitors on cloud warehouses
- SOC 2 Type II and ISO 27001
- They detect and alert; the fix stays with the team
Cost Annual contracts; Sifflet at about $48k a year on its published price
Monte Carlo — docs: security and compliance overview · Monte Carlo — Trust Center · Sifflet — plans and pricing
BeryX
- Quality, personal data and cost measured in the same catalogue
- Proposes the fix in SQL or as a pull request, without write credentials
- No certifications or contractual SLA yet
Cost Free · from €490/month
6 · Value
Where the difference is, and what we do not claim
What can be tested today
- Quality, personal data and cost measured over the same catalogue and the same history
- Read-only enforced in code, with fixes as proposals and pull requests
- Every number with a published method and a confidence interval
- AI with output rules, an offline mode and the option to switch it off
- A published price and a free account with no sales call
What we do not claim
- That BeryX would have prevented a particular fine
- Certifications: we do not have ISO 27001 or SOC 2 yet
- A contractual SLA
- That every connector has been proven on a real account
- That it replaces an MDM or an enterprise catalogue
- That it fixes the data on its own
6 · Value
What becomes measurable
| Metric | How BeryX measures it | Status |
|---|---|---|
| Quality score per dataset and per domain, with trend | Five dimensions with fixed weights, on every scan | Exists today |
| SLO compliance per dataset | A service objective per dataset and its compliance history | Exists today |
| Open and resolved issues by severity | Issues and incidents with dates, owner and response deadline | Exists today |
| Columns with personal data but no rules or no owner | The personal data map crossed with rules and owners | Exists today |
| Deadline of each data subject request | A request register with the jurisdiction's deadline | To be switched on |
| Compute spent on data that fails its checks | The platform's cost telemetry crossed with the score | Exists today |
| Tables unused for 90 days | Query counts per table from the system views | Exists today |
| AI cost per report and per team | Cost per call, per task and chargeback per team | Exists today |
6 · Value
The honest ledger
Exists today
- Profiling, 16 rule types, automatic monitors, anomalies and SLOs
- Incidents with lineage, likely cause and tickets
- Proposed, rehearsed fixes and pull requests (Business)
- Personal data map, sensitivity and re-identification risk
- Cost and quality side by side, cost spikes, real usage and chargeback
- Reports per role, the assistant and the MCP server
- SSO over OIDC and SAML, MFA, IP allowlist, audit and SIEM
- A tour and a written guide per role
Partial
- Connectors not yet proven on a real account, BigQuery, Snowflake, Databricks and Oracle among them
- Collibra and Purview validated only against a simulated transport
- Metadata pages and the critical data element dossier (BCBS 239) only through the API
- Entity matching on samples of up to 50,000 rows
- The AI output guard does not yet cover the streaming assistant
- The Viewer role cannot yet be assigned on the Team page
To be switched on
- Data subject requests and export to SodaCL and Great Expectations: built, in the next deploy
- Seven connectors not yet in the SaaS image
- The anonymiser in the production image
- Dedicated instance, high availability, WAF and a fixed egress IP
- SDK, Terraform, VS Code and Airflow packages not yet published
- Online payment
Vision
- ISO 27001, SOC 2, ISO 27701 and ISO 42001 certifications: we do not have any yet
- External penetration test
- SCIM and security keys (WebAuthn)
- Private connectivity and running inside the customer's perimeter
- Two-way sync with ServiceNow
6 · Value
Plans
Free
€0/month
- 1 source
- 25 datasets
- 5 AI reports a month
- 30 days of history
Team
€490/month
- 5 sources
- 500 datasets
- 200 AI reports a month
- 365 days of history
- API tokens
- Alert webhooks
- Document analysis for RAG
- Scheduled email reports
- Synthetic data from the profile
Business
€1,190/month
- 15 sources
- 2,000 datasets
- 1,000 AI reports a month
- 730 days of history
- AI Act compliance dossier
- Platform cost telemetry
- Remediation as a Pull Request (GitHub/GitLab)
- Your own Anthropic key (BYOK)
- Choice of AI models per task
Enterprise
On request
- unlimited
- Claude in your organisation's cloud account (Amazon Bedrock or Google Vertex AI)
7 · Next step
How to start
01
Try it on your own
A free account with one source and the one-click demo, with four synthetic data packs: Loja online, ERP, Neobanco and Hospital.
Create a free account →02
A conversation with BeryxTech.
Thirty minutes to fill in the cost table with your numbers and see the demo on your case.
Talk to us →03
A pilot on one domain
One data domain, the metrics from the previous slide measured at the start and at the end, and a decision made with numbers.
Propose a pilot →Data that engineering, the DPO and finance trust, with the same evidence.
BeryX, a BeryxTech product.
beryx.lbostuff.tech/en/apresentacao
BeryxTech · October 2026 · 39/39