Head-to-head · updated 13 September 2026

DataShield vs Ataccama: golden records an agent can query, and proof it was allowed to

Ataccama has been building this since 2007, and it shows. Ataccama ONE puts data quality, MDM, catalog, lineage, control and watch tools on one metadata layer. The MDM module does the real work: match, merge, survivorship, stewardship. Plus two-way sync back to the source systems. Their own MDM page keeps the promise plain: "Consolidate critical data and build trusted assets." Quoted in full. They have been a Gartner Magic Quadrant Leader, for boosted data quality. Five years running, at that. Their logo wall has Allianz, Prudential and T-Mobile on it. We would not argue with any of that.

We sell a different shape of the same job. DataShield Ontology is a master data management platform, built agent-first. Matching is Fellegi-Sunter probabilistic record linkage. It uses Jaro-Winkler, Levenshtein, Soundex and Double Metaphone comparators, LSH blocking and EM-trained parameters. So you can read the method, instead of trusting the adjective. The surface is the part that differs. Mastering alone is 143 MCP commands. An agent can ask a business user questions. It can derive a spec. It can work a decision gate you sign off, one step at a time. It then builds entity types and trees. Plus draft match configs, link fields and check rules. All in a single change. Nothing it makes goes live on its own. A golden record also has to earn its way to an agent. It clears a six-axis trust score. And a fail-closed consent check. Both, before any agent read returns it.

We tokenize datasets at ingest. Agents then query tokenized data over MCP. Turning a token back into real data is a gated step, always audited. Each checked tool call gets checked against the agent's current access. Each ruling gets sealed into a hash chain you can verify, without trusting us. Below is the honest split. It includes the rows Ataccama wins.

Get free Demo Center access Get an instant quote

The short version

Pick DataShield when

You want the model stood up in a conversation, not a services engagement. An agent asks you, proposes the model, and builds it as drafts you approve. How the ontology works.
You want to read the matching method before you buy it. Fellegi-Sunter weights. Five survivorship strategies. A promotion gate that refuses any config without a green battery run.
Agents are already reading your golden records. Someone will ask you to prove which agent read what. Our chain answers with math, not a policy PDF. Run the verifier.
You need MDM, PII classification, tokenization and agent authorization in one self-hosted footprint. On keys you hold. At a price you can see, before the call.

Pick Ataccama when

Data quality is the actual project. Profiling, rule authoring, cleansing and now observability are their core. They are five-time Gartner Leaders in that category, for a reason.
You need multidomain MDM at enterprise scale. Packaged customer, product and supplier models. Two-way sync into SAP and Salesforce. A steward team of thirty people. We are not that.
Procurement runs on analyst reports and reference calls. They have both. Plus nineteen years of insurance and banking deployments.
You want one vendor for quality, catalog, lineage, rules and MDM. You would rather merge tools than build a stack piece by piece.

Bottom line: Ataccama tells an agent whether the data is good. We let an agent help build the model. Then we decide whether it may read a given record. We keep proof of the call, too. A thousand source systems and a steward bench of thirty? Buy theirs. A few domains, no setup budget, agents already live? Start here.

Feature by feature

Competitor cells describe what Ataccama's public site and news posts say. We use the date above. If we got something wrong, email support@myorg.ai. We will correct it and credit you.

What matters	DataShield	Ataccama	Edge
Standing up the model	An agent does it with you. The mdm_project tool asks a business user. It derives a spec from the answers, and from live platform state. Then it presents a dependency-ordered decision gate. Each plan carries what it is, why, the impact and the choices. You confirm or fix it. A fix flows to each dependent call. There is deliberately no accept-all button. One atomic change then builds the entity types, trees, draft match configs. Also link fields and check rules. Each gets labelled with an honest gate posture. The generator never promotes a config and never runs matching. Fifteen commands, no setup services.	Model-driven MDM, configured inside Ataccama ONE by a data architect or a setup partner. Their 2026 news adds an AI Agent that automates data-quality work. It gets data AI-ready faster, they say. We found nothing that stands up a mastering model from a work-user interview.	◆
Entity resolution and matching	Fellegi-Sunter probabilistic record linkage. Per-field comparison levels, with m and u probabilities. Additive log-odds weights. LSH blocking, with a bucket-size cap so common names do not blow up. EM-trained parameters, with a held-out sample. Comparators are exact, Jaro-Winkler, normalized Levenshtein, Soundex, Double Metaphone, n-gram, plus date and numeric proximity. A config is drafted, trialled read-only, and explained pair by pair. It is trained, then run against a frozen corpus of known-true and known-distinct pairs. No config goes live without a green battery. An absolute-ID tier routes ID conflicts to review, instead of merging blind.	AI-powered record matching, with customizable rules. Plans a steward accepts or rejects. A system that learns from each match. Proven at large scale. The method itself is not open.	◈
Multidomain modelling	An open entity-type word list, not a template you bend. Each type declares one of three wired matching modes. Text matched. Id matched. Or not masterable at all. The People, Places and Things taxonomy is a table row, not an enum. So a fourth category is an insert. Hierarchies let a child config resolve inside its parent's scope first. They cross parents only at a stricter threshold, with measured rollups. Relationships carry validity dates. Check rules can gate publication outright. What we do not ship: a packaged Client 360 or Supplier 360.	Cross-domain MDM across client, product, supplier and more. On the same metadata layer as their quality rules, catalog and lineage. With two-way sync back into source systems. Nineteen years of models and proven builds behind it.	◇
Getting data in, and keeping it current	Six on-ramps: inline ingest. URL and batch ingest. Presigned object storage. Storage watch sources. A resumable file repo. A catalog binding. Bind a live PostgreSQL object to an entity type once. Mastering then stays current, on a declared cadence. Under an auto, agent-review or steward-review remaster rule. Differential loads mark removed rows stale and changed rows for re-check. A weekly full pass reconciles deletes. That is PostgreSQL today, not your whole estate. We are not an ETL product, and will not sell you one.	Connectors across the enterprise estate. Snowflake, Databricks and ServiceNow integrations. Marketplace listings. Two-way sync that writes mastered data back into the source systems. This is their home ground.	◇
Survivorship and golden records	Five declarative strategies, as pure fold functions: most trusted, most recent, longest, most frequent, aggregate. The default chain is most frequent, then longest, then deterministic lexicographic. Most trusted reads a source's known track record. That means how often a source's values get overturned by stewards. Not a hand-ranked list someone wrote in a workshop. Each value carries an origin envelope. golden_explain names the winner. It names the strategy, the winning source, and how many values it beat. Version log is append-only, and survives delete. A value-free CloudEvents outbox streams golden changes in commit order.	Golden record creation for single and cross-domain models. Side-by-side record comparison. Duplicate elimination. Two-way sync back into source systems, lakes and warehouses.	◈
Stewardship and earned automation	Stewards decide patterns, not items. The plan ships with a scan for proof against itself. One human OK can clear thousands of like cases. It still counts as one OK. Trust is a three-rung ladder. A pattern earns auto-clear after 50 OKs at a 98% rate, with zero pushback. One pushback demotes it. At the top rung, a set sampler acts. It sends 10% of auto-cleared items back to a human. So you can prove why any item was picked. Standing rules are born staged, and need a second OK. A blast-radius dry run calls the writer's own gate code. So a preview cannot promise what the writer would refuse.	A stewardship workbench refined over nineteen years. Match proposals a steward accepts or rejects. A system that learns from each match. Their steward tooling and their screens are ahead of ours.	◈
Data quality depth	A 20-section dataset profile. Completeness, field stats, patterns, column meaning and trust metrics. Plus a six-axis trust score, per entity type, domain and estate. With open weights, a floor per axis and an hourly refresh. No rule-authoring studio, and no data-watch product. Accuracy scores only where you have written check rules. We say so, rather than imply estate-wide coverage.	The core of the company. Profiling, rule authoring, cleansing, matching, and an agent observability layer, launched in February 2026. Five straight years as a Gartner Magic Quadrant Leader for boosted data quality.	◇
Whether an agent may read the record at all	A golden record has to earn the right to be read by an agent. It needs a combined trust score of 0.70 or better. Each axis above its floor. A live consent basis for the purpose, too. Consent is a hard veto. The absence of a recorded lawful basis is not permission. The check fails closed. One shared predicate enforces this, on the SQL path and the application path. So a read cannot route around it. Golden reads come from views that mask personal data unconditionally. The reader role holds no grant on the base table.	The Data Trust Index scores data. It hands the score to an agent through their MCP server. Their September 2026 Ossie work gives agents readable quality warnings before they act. That is a signal an agent may ignore. We found no method that withholds a record from an agent.	◆
Audit evidence	SHA-256 hash chain, with Ed25519-signed checkpoints. Those checkpoints are themselves chained. A check returns clean, old damage, or new tamper. It names the failure: tamper, insert, delete, or cut-short. Each merge, unmerge, pin and control override lands on it. Each has a before and after picture, in the same step. Try the checker.	An immutable-sounding control trail is implied by their stewardship workflows. We found no open crypto proof of tamper on their site.	◆
Agent authorization	Each checked tool call passes a scope ceiling, an access tier, a consented-tool allowlist. Plus a pull-back re-check, before dispatch. The call fails closed.	Their MCP server hands agents ruled data plus a Data Trust Index signal. That is a quality gate, not an authorization gate. We found no per-call choice point.	◆
Break-glass	Scoped, time-boxed emergency access for agents. It auto-revokes. It is step-up and IP gated. It cannot be quietly removed from the log. Mastering has its own stop switch, as well. A global control flag halts matching across the fleet. A kill is final.	Not described anywhere on their site.	◆
GDPR erasure	Crypto-shred of per-subject key material. Plus ISO 27560 consent receipts. Plus a consent substrate. It covers all six GDPR Article 6(1) lawful bases, across nine purposes. One live fact per subject and purpose. Golden version history keeps a tombstone with the rationale. The audit chain still verifies afterwards. The change stream carries no attribute values. So there is no event backlog to purge.	Stewardship and control workflows cover retention calls. The erasure method on a golden record is not described. Soft-delete is the usual MDM answer. Worth asking.	◆
Tokenization and PII handling	Deterministic, join-preserving, vault-reversible tokens, applied at ingest. Plus quasi-ID generalization, as a rule. Dates become year, decade, or age band. ZIPs become 3 or 4 digits. Phones, SSNs and emails get partial masking. Each column gets a measured score. Matching runs on real source values. A matching field that is tokenized at rest makes the run refuse. It throws a typed error, rather than mega-merge each colliding token. Derived phonetic and n-gram fields inherit the base field's PII verdict. So a name cannot be rebuilt from its Soundex. Masking and generalization are switches you turn on, not defaults.	Automatic classification detects business terms and sensitive data. Tokenization is not vocabulary they use for MDM.	◆
MCP and agents	More than 200 MCP tools across Ontology, Auth, Corpus and Lighthouse. Mastering alone is 143 commands, across the mdm, stewardship and mdm_project tools. About a third of the whole agent surface. An agent can propose a merge. It can work a stewardship queue. It can enact a standing rule. Or read a golden record under rule. Auth issues MCP tool tokens with scope ceilings. It meters each call against the agent.	A real MCP server, shipped with the November 2025 agent relaunch. It lets Claude and ChatGPT reach ruled data, with a Data Trust Index attached. In September 2026 they open-sourced an Apache Ossie converter. So agents can read quality warnings before acting.	◆
Deployment	Self-hosted, in your own cloud or data center. Or a dedicated single-tenant server we operate. Docker images cover Auth, Ontology, Corpus and Lighthouse. Each has a signed deploy manifest, that Guardian verifies. Ed25519 keys stay on your instance, in a key directory you control. Kid-keyed rotation fails closed on an unknown key. KMS or HSM custody is a built seam in Auth, not a shipped link yet. HMAC tokenization keys sit on your system today. Not in a KMS.	SaaS, self-hosted and hybrid. With AWS and Azure marketplace listings, and Databricks Marketplace since June 2026. Built for mixed estates.	◈
Maturity signals	Live and in use. Guardian and Lighthouse since April 2026. Entity matching is validated against roughly 370,000 real public records, across three City of Chicago datasets. Not only synthetic ones. We publish no record-count ceiling and no accuracy figure. We do not have one we would defend. SOC 2 not yet certified, and we say so.	Founded 2007. Roughly $150M from Bain Capital Tech Opportunities. Five straight years as a Gartner Leader. A Forrester Wave Leader placing for data quality, in Q1 2026. Their own news claims record Q4 2025 growth. An outside profile puts them at about 250 clients. Either way, their banking and insurance client list is long, and ours is not.	◇
Pricing	Open model, scoped instant quote, no sales wall.	Quote only, priced by part and volume. An outside estimate puts Ataccama ONE at roughly $90,000 a year to start. Ataccama posts no figures.	◆

◆ DataShield leads◇ Ataccama leads◈ comparable

Ataccama claims come from ataccama.com and Ataccama's own press posts. 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 data trust platform

A model an agent helps you build

Most MDM projects start with a workshop and a partner statement of work. Ours starts with a conversation. An agent asks what your shop calls a client. Which systems are authoritative. Who stewards what. It proposes the model back as a set of calls. You confirm or fix it, one at a time, in order. Then it builds the whole thing as drafts. You still hold each promotion. How the ontology works.

Proof that survives an audit

A stewardship log that can be silently edited proves nothing. Ours is a hash chain, with signed checkpoints. The checker tells you what broke, not just that something did. Re-tamper an already-attested break un-attests it. So the trick of "we knew about that one" does not work twice. Try it in your browser, no signup.

Authority that can change mid-flight

An analyst resigns on a Friday. Their agent is 20 minutes into a 40-minute merge run, against your client master. With DataShield, the next checked tool call gets re-checked against current access. It fails closed. A trust score does not have an opinion about this. How Auth does it.

Where Ataccama is genuinely stronger

We would rather you heard this from us. Ataccama has run MDM projects in insurance and banking since before the phrase "data trust" existed. The part lives on the same metadata layer as their quality rules, catalog and lineage. That matters in practice. The steward who fixes a match also sees why the source field was flagged. All in one place. Their connector estate. Steward screens are part of it, too. Their cross-domain models. Their two-way sync back into source systems. All built out, in ways ours are not. We ship no packaged Client 360. We bind live sources for PostgreSQL only. They are a five-time Gartner Leader for boosted data quality. Also a Forrester Wave Leader for data quality, in Q1 2026. Their November 2025 relaunch shipped a working MCP server. That's more than most old MDM names can say.

If your shortlist is Informatica, Reltio and Ataccama, they deserve to be on it.

Here is the push-back, and their new CEO wrote it for us. In August 2026, Martin Zahumensky said clients used to ask one thing. Was their data clean, written, ruled and mastered? Now they ask "whether that same data can safely fuel an AI agent." Right question. Their answer is a signal. A Data Trust Index. An Apache Ossie converter that gives an agent "AI-readable quality warnings" before it acts. Read that September 2026 update. Count the mentions of authorization, audit or rule. There are none. A warning an agent may ignore is not a control. Our answer is a gate. The record does not come back until it clears the trust floor. A live consent basis, too. The call gets authorized against current access. The ruling gets sealed. That layer sits underneath a platform like theirs, without complaint.

Questions worth asking both of us

These are the questions we would want answered if we were the ones buying. 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. A check tells delete apart from cut-short and from tamper. Each merge, unmerge and field pin is on that chain. Each carries a before and after picture, saved in the same step as the edit. Run it against a sample chain at /verify. Ataccama: their site describes control and steward trails. We found no open tamper-check method. Ask them to show one. An MDM merge log is exactly the thing an examiner will want to trust.

What happens to a revoked agent mid-session?

DataShield re-checks access on each checked tool call. So a pulled grant lands on the next call, not the next token refresh. There is also a fleet-level stop for mastering itself. One flag halts automated matching everywhere, and a kill is terminal. Ataccama's MCP server gives agents ruled data, with a trust score attached. We could not find a mid-session pulled-access method in their public docs. Ask how long a compromised agent keeps reading your client master. After you pull its access.

How does GDPR erasure interact with the audit trail?

DataShield crypto-shreds per-subject key material and issues an ISO 27560 consent receipt. Actor identities in the chain are HMAC-committed. So the proof still checks out once the subject is gone. The golden record leaves a tombstone version with the rationale. The change stream carries no attribute values at all. So there is no event backlog to chase. Ataccama's public docs cover retention and stewardship workflow. What happens to a golden record and its merge log, under a wipe request? Not spelled out. Ask for the method, not the workflow.

We only need MDM. Do we have to license the rest of Ataccama ONE, and who builds the model?

The licensing half is a question for their sales team. We cannot answer it for them. Packaging is modular and quote-only. An outside estimate starts Ataccama ONE around $90,000 a year. Ask which modules are in the quote. Which ones are add-ons? Ask what the number does when your data volume doubles. Then ask who builds the model, and what that engagement costs. On our side, the MDM work is part of Ontology. Self-hosted, with a open price and an instant quote. The model gets built by an agent that asks you and builds drafts you approve. No setup partner in the middle.

Ataccama ships an MCP server too. What's different?

Theirs serves ruled data plus a Data Trust Index. So an agent learns whether the data is reliable. Good design for a quality platform. Ours is the whole mastering surface. 143 commands for master data alone. An agent can build the model. It can train and promote a match config. It can work the stewardship queue, and read the result. And the read is gated. A golden record comes back only if it clears the trust floor. And a live consent basis. Both, through a view that masks personal data. Want an agent to ask "is this data fresh?" Theirs answers. Want to prove which agent read which golden record last March? Ours does.

Does DataShield have SOC 2?

Not yet, and we will not imply otherwise. Auth is live now, with a public threat model and a checker anyone can run. Guardian and Lighthouse have been live since April 2026. Our mastering engine is validated against about 370,000 real public records. We publish no accuracy score, because the honest one is not flattering in each leg. Design-partner terms include source escrow. That way a small vendor is not a single point of failure. Details on the security page.

Ataccama's MDM promise, verbatim: "Consolidate critical data and build trusted assets… turn fragmented, duplicate data into a single, reliable source of truth." Quoted in full. Short version: one trusted record, from many messy ones. - ataccama.com, 13 Sep 2026
Homepage positioning: "Trust your data, accelerate your AI." With a 5-time Gartner Magic Quadrant Leader badge. And a Forrester Wave Leader badge for Data Quality, Q1 2026. - ataccama.com, 13 Sep 2026
Ataccama's own relaunch release claims an autonomous ONE AI Agent, delivering AI-ready data 83% faster. Plus an MCP server. It lets Claude and ChatGPT reach ruled data, with a Data Trust Index attached. - Ataccama newsroom, 6 Nov 2025
New CEO Martin Zahumensky: customers "wanted to know whether their data was clean, documented, governed and mastered. Now they want to know whether that same data can safely fuel an AI agent." Quoted in full. - Ataccama newsroom, 4 Aug 2026
Apache Ossie converter gives agents "AI-readable quality warnings" before they act. The release describes no authorization, audit or policy control, at all. - Ataccama newsroom, 9 Sep 2026
Ataccama's own release reports record Q4 2025: a 30% three-year CAGR. More than 75% of new fourth-quarter customers ran its generative-AI features in production. Customer counts are the company's or third-party profiles', not audited figures. - Ataccama newsroom, Feb 2026
Named a Leader in the Gartner Magic Quadrant for Augmented Data Quality Solutions. First of five straight years. - PR Newswire, 11 Mar 2024
Third-party estimate: Ataccama ONE pricing starts at roughly $90,000 annually. It varies by deployment, modules and data volume. Ataccama itself publishes no price list. - SelectHub, 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 →