Head-to-head · updated 13 September 2026

DataShield vs Tamr: whose AI-native MDM can show its work?

Tamr has mastered records since 2013. It came out of Michael Stonebraker's lab. The match engine is the real thing. Point it at eleven source systems. It hands back one view. They sell it as boxed data products. They price it by the golden record, not the seat. Their FY26 release claims SaaS revenue more than doubled. Net retention hit 109%. Want someone else to run the machine? That's a fair answer.

We run the same job in your building. We let an agent build the model with you. Guided onboarding interviews a business user. It writes a spec. Then it walks a choice gate. You confirm or override. One step builds the whole model, all as drafts. Matching is Fellegi-Sunter fuzzy linkage. Each merge has per-field weights. You can read them. You can argue with them. Data gets tokenized at ingest. Agents query tokenized data over MCP. Detokenization is a privileged, audited step. A golden record has to earn its way to an agent. Six quality axes. A floor on each. A consent check that fails closed. Below is the honest split. It includes the rows Tamr wins.

Get free Demo Center access Get an instant quote

The short version

Pick DataShield when

An agent should build the model with you. Guided onboarding interviews a user. It writes a spec. Then it works a choice gate. Proposals come in order. Each proposal says what it does. It says why. It says the impact and the options. You confirm or override. An override flips each dependent choice back for re-derivation. There's no accept-all button. One step then builds the whole model. Everything lands as a draft. Nothing promotes itself.
A steward has to defend a merge. Ours is Fellegi-Sunter fuzzy linkage. Per-field agreement weights. EM-trained m and u parameters. Plus an explain call. It prints why two records met. Not "the model was confident".
Source records cannot leave your building. We ship Docker images. They cover Ontology, Auth, Corpus and Lighthouse. Self-hosted, in your own cloud or data center. Or use a dedicated single-tenant server we run for you. You can see a price first. You never pay by the record for your own master data.
Agents will read the golden records. A golden has to earn that first. Six quality axes with public weights, a floor on each, and a lawful-basis check that fails closed. Then it arrives through a masked view. That's under a tool token with a ceiling. Re-checked against current access. How Auth does it.

Pick Tamr when

You are mastering hundreds of millions of records. Across dozens of systems. You want a vendor whose whole budget goes there. Ours is deep, but it's one part of a platform.
You want enrichment built in. Tamr resolves against outside data as it runs. We normalize enrichment. We register it. We don't sell you the data.
A boxed data product beats a project. Fair enough. Healthcare Providers, Suppliers, Customer 360: their models shorten the first ninety days. Curator Hub gives stewards a polished place to work. We ship no boxed domain. We won't pretend otherwise.
You'd rather buy SaaS than operate anything. Multi-tenant hosting in one of five regions meets your residency rule.

Bottom line: Tamr is the better bet if scale is the whole problem. Or a fast first golden record. We are the better bet if you want an agent to build the model with you. Also if you need to gate which records agents may read. Or if you must answer an examiner who asks why two people became one person.

Feature by feature

Competitor cells describe what Tamr's site and press releases say. As of the date above. If we got it wrong, email support@myorg.ai. We will fix it, credited.

What matters	DataShield	Tamr	Edge
Standing up the model	An agent does it with you, through 15 MCP commands. It talks with a user. It writes a spec. Then it presents a choice gate, one proposal at a time. Each proposal carries a what, a why, an impact. Plus the options. You confirm or override. An override propagates. Each dependent choice flips back for re-derivation. There's no global accept-all, on purpose. One step then builds the model: entity types, hierarchies, configs, bindings, rules. Each piece gets labelled with an honest gate posture. The generator never promotes a config. It never runs matching. It never skips the check. No setup services involved.	Boxed data products shorten the first ninety days. Setup comes with the price. That's a good deal. It's still their project, not yours. We found no path where your own business user drives the modelling.	◆
Matching method, and whether you can read it	Fellegi-Sunter fuzzy record linkage. Per-field comparison levels with m and u probabilities. Match weight is log2(m/u). Added up across fields. Comparators: Jaro-Winkler, normalized Levenshtein, Soundex, Double Metaphone, n-gram. Plus date and numeric proximity. LSH blocking caps bucket size. A common surname can't blow up the run. Parameters are EM-trained against a frozen holdout. A trained polarity inversion is a hard refusal. An absolute-identifier tier routes a contradicted identifier pair to review. Not a blind merge. No config reaches production without a green paired-gate check. It runs the real scoring path. Confidence is typed honestly: a text match carries a probability. A steward or identifier match carries none. Not a fake 1.0.	ML-based entity resolution. Eighteen claimed patents. Trained with human feedback. Their docs describe accuracy and automation. Not the scoring model. We could not find a public explanation of how a pair scores.	◆
Multidomain modelling and packaged domains	An open entity-type vocabulary. You model your domain: three wired matching modes per type. A People, Places and Things taxonomy. It's a table row, not an enum. Hierarchies, where a child config resolves inside its parent's scope. At a base threshold. Stricter across parents. Relationships with bitemporal validity. Domains that carry an owner and a compliance framework. Validation rules gate publication. What we don't ship: a boxed Customer 360. You'd bend that yourself.	Boxed data products cover B2B and B2C customers. Plus contacts. Also healthcare providers, companies, suppliers, products and locations. Plus a multi-domain product. Does one of those fit your shape? Then you start ahead of us.	◇
Getting data in, and keeping it current	Six on-ramps. Inline, URL and batch ingest. Storage watch sources. A file repo. A catalog bind, the mastering-specific one. Point an entity type at a live PostgreSQL object once. Mastering then stays current, on a declared cadence. Pick auto, agent-review, or steward-review remastering. Differential loads mark deleted source records stale. Updated ones get flagged for re-evaluation. A weekly full pass reconciles deletes. That's PostgreSQL today, not your whole estate. We are not an ETL product. We won't sell you one.	Native connectors for the leading lakes and warehouses. AWS, Azure, Databricks, Google Cloud and Snowflake are named on their platform page. Plus a public BigQuery case study. Connector breadth is theirs.	◈
Mastering at very large scale	Validated against roughly 370,000 real public records. Across three City of Chicago datasets, not just synthetic ones. Zero false merges on each trap class in our public benchmark. That includes forty distinct people sharing one phone number. Reference deployments are mid-market sized. Matching runs batch and on demand. Not a sub-second stream. We publish no record-count ceiling. We won't invent one.	The core competence and a decade of tuning. Documented warehouse-scale runs and a public BigQuery case study. Enterprise logos include Toyota, Danaher and Société Générale.	◇
Enrichment and reference data	A normalization and enrichment registry. Plus curated ISO reference sets. Pinned, write-once versions, with staleness rollups. Bring your own. We don't resell external data.	Entity resolution with built-in external data. Sold as part of the platform. A genuine advantage if your records are thin.	◇
Stewardship, and how automation is earned	Stewards review patterns, not items. One human call disposes a whole pattern. The proposal runs a counterevidence scan first. It argues against itself before you see it. Trust is a three-rung ladder. 50 approvals at a 98% rate, no dissents. That promotes a pattern. One dissent demotes it. At the top rung, a deterministic sampler picks 10% of auto-disposed items. Those go to human audit. That way you can prove why any one item was sampled. Standing disposition rules are born staged. They need a second explicit release. They auto-suspend on an overturn. A blast-radius dry run uses the writer's own gate code. A preview can't say yes and then refuse. Each decision lands on an immutable before-and-after trail.	Curator Hub launched in November 2025. It pairs agents with human reviewers for the hard cases. It has a no-code interface stewards seem to like. It's available to Tamr Cloud clients. The screens are nicer than ours. What we couldn't find: a sampled audit. Of the decisions their agents took on their own.	◆
Whether a golden record is fit for an agent to read	A golden has to earn the read. A composite score covers six quality axes. Public weights, a floor on each. Refreshed hourly. Plus a consent veto that fails closed. It covers all six GDPR Article 6(1) lawful bases. Across nine purposes. One live fact per subject and purpose. No recorded basis means no. One predicate enforces the gate. The SQL and the application paths share it. They can't drift apart. Only the derived consent score crosses the masking boundary. The consent facts themselves stay behind it.	Human oversight on review is the control they describe. We found nothing that gates agent access on the quality score. Or on the lawful basis of a record. Worth asking: what stops an agent from reading a record nobody trusts yet?	◆
Tokenization and masking	Deterministic, join-preserving, vault-reversible tokens applied at ingest. Plus quasi-identifier generalization. Dates go to year, decade, or age band. ZIPs go to 3 or 4 digits. Phones, SSNs and emails get partial masking. Each column gets a measured cardinality-reduction score. These are features you switch on, not defaults. Golden consumption views mask in the view. And in the reader role. That role has no grant on the base table. The mask can't be bypassed from SQL. View-to-application parity is a release gate. One catch: matching runs on real source values. A matching field tokenized at rest makes the run refuse. It throws a typed error. Not a blind merge.	Not vocabulary Tamr uses. Data sits in per-tenant storage inside their SaaS. It gets mastered there.	◆
Audit evidence	SHA-256 hash chain with Ed25519-signed checkpoints, themselves chained. Verification returns one of three things: clean, attested damage, or tampered. It names the failure: tampering, insertion, deletion or truncation. Ed25519 signing keys stay on your instance. You control the key directory. Rotation is kid-keyed. KMS or HSM custody is a built seam in Auth. Not a shipped integration. Governance mutations get their own immutable before-and-after trail. In the same step. Try the verifier.	Lineage on the 360 pages. The review flow implies a record of choices. We found no public proof of tamper evidence.	◆
MCP and agents	More than 200 MCP tools, across Ontology, Auth, Corpus and Lighthouse. Master data alone is 143 commands. About a third of the whole agent surface. That's 99 for mastering, 29 for stewardship, 15 for guided onboarding. Goldens reach agents through a masked consumption relation. It reports what it excluded, and why. A value-free CloudEvents stream. 21 event types. Delivered in commit order. RAG grounding documents projected only from agent-ready records. A DuckDB dimension the reader role can't escape. Auth issues MCP tool tokens with a scope ceiling. It meters each call. It attributes each call to the agent.	Their platform page now lists LLM connectivity over MCP. Plus agentic curation and a Bring Your Own Agent architecture. Their FY26 release claims API traffic grew roughly threefold. So agents are arriving. What we couldn't find: any description of per-call authorization. Or tool-level scopes over that surface.	◆
Agent authorization and revocation	Each governed tool call passes a scope ceiling. An access tier. A consented-tool narrowing step. Plus a mid-session revoke check. All of it before dispatch. It fails closed. Delegation is RFC 8693 token exchange. It carries an enforced scope ceiling.	Human-in-the-loop review is the control. We found no agent identity, scope or revoke model in their docs.	◆
Break-glass	Scoped, time-boxed emergency access. Admin and IP-allowlist gated, step-up authenticated, fully audited. It auto-revokes.	Not described in their docs.	◆
GDPR erasure and lawful basis	Crypto-shred of per-subject key material, plus ISO 27560 consent receipts. This runs over a consent substrate. It records a lawful basis per subject and purpose. Append-then-supersede. One live fact at a time. The audit chain still checks out after the subject is gone. Actor identities are committed with an HMAC. The change stream carries no attribute values. It needs no purge cascade of its own.	Residency in five regions, with independent tenants. The erasure method isn't public. Not for a mastered record, nor its source links.	◆
Deployment	Self-hosted, in your own cloud or data center. Or a dedicated single-tenant server we operate. Docker images for Ontology, Auth, Corpus and Lighthouse. A signed deploy manifest, Guardian verifies it.	True SaaS, multi-tenant, five regions with per-tenant separation of interim data and metadata. No self-hosted option is described.	◆
Maturity signals	Auth, Guardian and Lighthouse are live now. Guardian and Lighthouse since April 2026. SOC 2 not yet certified. We say so. Small team, hands-on, source escrow for design partners.	Founded 2013. Turing Award pedigree. Roughly $69M raised, by public estimate. Their FY26 release claims 102% SaaS revenue growth with 97% gross retention. Plus a named client list. The growth numbers are theirs. We haven't audited them.	◇
Pricing	Published model and a scoped instant quote. No per-record meter on your own master data.	A clear model, no numbers. Pay for golden records, not duplicates. Output-based on Tamr IDs, volume discounts, no per-seat charge, setup comes free. Dollar figures come only by quote.	◆

◆ DataShield leads◇ Tamr leads◈ comparable

Tamr claims are drawn from tamr.com and Tamr's own press releases, last checked 13 September 2026. We link them below rather than work from memory.

Three things you get here that you won't get from a mastering SaaS

A model you built in a conversation

A business user answers questions about their business. The platform writes a spec. Then it hands back a choice gate. One proposal at a time. Each with its reasoning and options. Override one. Each choice that depended on it comes back for another look. When you're done, one step writes the whole model. It arrives as drafts, gated. An agent can drive the whole path over MCP. See how the ontology works.

Golden records that earn their way to an agent

Six quality axes. Public weights. A floor on each. A consent veto that fails closed. Miss any of it. The agent never sees the record. Simple as that. The records that pass arrive through views. Those mask on the way out. The reader role has no grant here. Not on the table underneath. Raw values need a separate, privileged, logged step. What Auth checks on each call.

Proof the record's history wasn't edited

Merges, splits, steward decisions and detokenization all land in a hash chain. Checkpoints are signed. The verifier tells you what broke, not just that it did. That's the property EU AI Act Article 12 and HIPAA §164.312(b) reviewers ask about. Try it in your browser, no signup.

Where Tamr is genuinely stronger

Tamr has spent thirteen years on one problem. And it shows. The engine handles estates we haven't been asked to handle. It has a public BigQuery case study. It has logos that would survive any reference call. Their enrichment is built in. That matters more than people expect. Thin records match badly. Buying outside data separately is its own project. Their boxed data products are real. We have nothing like them. Fair enough. We publish no data-product object at all. Does a Healthcare Providers or Supplier model fit your shape? Then they start ahead. We'd say so on a call. Curator Hub gave stewards a pleasant place to work. Pleasant matters. Especially when someone has four hundred cases to clear by Friday. Their pricing idea beats most of the category. Pay for golden records, not the duplicates you started with.

Their FY26 release claims 102% SaaS revenue growth. Plus 97% gross retention. That reads like people staying.

Here's the push-back. It's narrow. Tamr's answer to a hard match is a model. Plus a human reviewer. Ours is a score with named parts. Plus a model your own business user helped build. Choice by choice, with an agent doing the typing. Both produce a golden record. Only one gives you an argument. Hand it to a regulator. Ask why Maria Santos and Maria Santos-Oliveira became the same patient. The same gap shows up downstream. They speak MCP now. Their API traffic grew threefold by their own count. That tells you where the work is going. What Tamr doesn't say: which records an agent may read. Or what limits that agent on a call. Or how you'd prove afterward what it read. We gate all three.

Questions worth asking both of us

These are the questions we'd want answered as the buyer. Ask them on each call, ours included.

Can you cryptographically prove an audit log entry wasn't deleted?

DataShield: yes. Each record commits to the one before it. Checkpoints are signed and chained. The verdict tells deletion apart from truncation. And from tampering. Re-tampering a marked chain un-marks it. Nobody can hide behind an old attestation. Run it at /verify. Tamr: their 360 pages show lineage. Their review flow keeps a record of choices. We found no tamper-evidence method in their docs. Ask to see one.

What happens to a revoked agent mid-session?

DataShield re-checks access on each governed tool call. Revocation lands on the next call, not the next token refresh. An analyst leaves on a Friday. Their agent is twenty minutes into a forty-minute job. The next call fails closed. Tamr: we couldn't find an agent revoke model in their docs. Ask how long an agent keeps working after you pull its access.

How does GDPR erasure interact with the audit trail?

DataShield crypto-shreds the per-subject key material. It issues an ISO 27560 consent receipt. Actor identities in the chain carry an HMAC commitment. The proof still checks out once the subject is gone. Under the golden views sits a lawful-basis record. Per subject and purpose. No recorded basis counts as no. The change stream carries no values at all. Nothing to purge there. Tamr publishes residency and tenant separation. What happens to a mastered record on an erasure request? Not spelled out. Nor its source links or review history. Ask for the mechanism.

Can our own people stand up an MDM model here, or do we need consultants?

Your people, with an agent doing the work. It talks with a user. It writes a spec. It walks a choice gate. Each proposal states its reasoning and impact. Plus the options you turned down. Override one. Its dependents come back for re-derivation. Then one step builds the whole model, as drafts. Nothing is promoted. No config goes live until it passes a full check. Tamr's answer is a boxed data product with setup thrown in. That's faster on day one. If your domain matches one of theirs. Ours is faster on the day your domain doesn't.

Tamr does MCP now too. What's actually different?

Their platform page lists LLM connectivity over MCP. They report API traffic roughly tripled last year. This isn't a checkbox. The difference is what sits under the call. A record has to clear a quality composite. And a consent veto. Only then can an agent see it. The call carries a tool token with a scope ceiling. Narrowed to consented tools. It's re-checked against current access. Metered against the agent. Then sealed into the chain. One caution about our own stack. We'd rather say it than be caught. Auth's MCP surface checks its own signed tokens. The Ontology surface authenticates by API key. With per-tool tier gating. Different methods. Same fail-closed answer. We're not going to blur them.

Does DataShield have SOC 2?

Not yet. We will not imply otherwise. Auth is live with a public threat model. A verifier anyone can run backs it. Guardian and Lighthouse have run live since April 2026. Design-partner terms include source escrow. A small vendor is not a single point of failure. Details on the security page.

Tamr's positioning: "AI-Native Master Data Management" with data products for B2B and B2C customers, contacts, healthcare providers and companies, suppliers, products and locations. - tamr.com, 13 Sep 2026
Platform features now include "LLM Connectivity with MCP" and "Agentic Data Curation" combining AI agents with human-in-the-loop oversight. - tamr.com/platform, 13 Sep 2026
FY26: 102% YoY direct SaaS revenue growth, 97% gross revenue retention, 109% net revenue retention, and nearly 3x growth in API web requests. Introduces a "Bring Your Own Agent" architecture. - PR Newswire, 5 Mar 2026
Curator Hub pairs AI agents with human stewards for the hardest curation cases. It's part of the AI-native MDM platform, available to all Tamr Cloud customers. - PR Newswire, 19 Nov 2025
Tamr RealTime is delivered on "real-time processing on our SaaS architecture." It's underpinned by 18 claimed patents in AI, machine learning and data curation. - PR Newswire, 30 Jul 2024
Pricing model: "Pay for golden records, not duplicates." Output-based on Tamr IDs, volume discounts, no per-seat pricing, setup included. No dollar figures published. - tamr.com/pricing, 13 Sep 2026
Deployment is multi-tenant SaaS across five regions (US, Canada, EU, UK, APAC). Customer data, including interim datasets and metadata, is held in separate tenants. - tamr.com/security, 13 Sep 2026

Other head-to-heads

MDM

Get free Demo Center access

You've seen the proof

Ready for a number? Scope your deployment and we'll price it against your own economics.

Get your quote →