Skip to content

Resources

Taiga vs. Factory vs. Cognition

By Mikko Laakkonen · Taiga co-founder and CEO

Published · Updated

Read against three questions an auditor asks about a single change, the three vendors sort quickly. Cognition sells an autonomous software engineer whose output your reviewers evaluate. Factory sells a software factory your engineers or partners assemble. Taiga sells a monthly service that requires eight documents before the build and keeps a seven-year audit log under Finnish law. The same questions show where Taiga loses today: no certificate or penetration test, no SAML or SCIM, GitHub only, one model provider, no air-gapped option, and Amazon Bedrock Guardrails switched off.

DORA requires a financial entity to make sure that every change to its ICT systems is "recorded, tested, assessed, approved, implemented and verified in a controlled manner." The same regulation makes it keep a register of its contracts with ICT providers, and an exit strategy for the ones that support critical or important functions. When an AI agent wrote the change, part of that evidence can only come from the vendor: the record of what was built, from which documents, and why.

The quickest test of a vendor is a single change picked at random from last quarter, and three questions an auditor asks about it: which document existed before it was built, who approved it against which written policy, and under which contract and law the record sits. We put those questions to Cognition, Factory and Taiga by reading what each one publishes in its product pages, documentation, trust centre, terms and data processing agreement, as they stood on 4 October 2026.

Cognition reported a $1 billion annualized revenue run rate on 25 September 2026, seventeen days after raising more than $2 billion at a $48 billion valuation. Factory raised $200 million at a $5 billion valuation on 15 September and has not published a revenue figure. Taiga's last published figure is EUR 100,000 of annualized revenue in June 2026, unaudited, and we are four founders in Helsinki. None of those numbers answers the auditor's questions.

Every quote below is verbatim from a page that was live that day, and the sources list names each one. Where a page says nothing, we write not stated, which is not the same as absent, because both companies sign enterprise contracts that are not public. If either shows us a page that corrects a row, the correction goes on our receipts page with its date, as our own corrections do.

Where the three of us agree

Factory wrote, in the post announcing its partner network on 19 August 2026, that "even with frontier models, writing code is about 20% of the software development lifecycle." Cognition wrote, in Devin's annual performance review in November 2025, that "human review is still necessary, because code quality is not straightforwardly verifiable." We would sign both sentences.

All three of us hand the merge decision to your Git host. Factory's compliance documentation says to "enable branch protection to require PR reviews before merging Droid-assisted changes." Devin's GitHub integration page recommends "enabling branch protection rules on your main branch to ensure all required checks pass before Devin can merge changes." Our own walkthrough says "required repository checks and reviews are never bypassed."

All three of us also give you what is built. Cognition's terms say the customer "owns all right, title, and interest, including all intellectual property rights, in and to Customer Data, including Outputs, to the fullest extent permitted by applicable law." Factory's pricing page answers "Who owns the code generated by Factory?" with "You do." Our master terms say "all right, title and interest in the Delivered Software is the Customer's from delivery."

None of the three takes over your on-call duty, either. Factory's Release stage, in Private Preview, will "execute approved release steps or hand off to the deployment system," and its Incident Response begins when "an alert bot posts a top-level message in the configured Slack channel." Devin's PagerDuty integration, added on 11 September, lets it "act as an oncall responder that triages PagerDuty incidents and posts its investigation as incident notes." At Taiga, your pipeline executes the deployment, and our product page says the monitoring signals "do not imply complete infrastructure observability, incident response or autonomous production remediation."

The disagreement is about who carries the other 80 percent, and under which law.

What each one says it sells

Cognition's homepage calls Devin "the first autonomous software engineer," one that "plans, writes, tests, and ships production code on its own, working inside your codebase and the tools your team already uses." On devin.ai, Devin "won't stop until the PR is ready to merge." Cognition's enterprise terms put the review on the customer: "It is Customer's responsibility to evaluate whether Output is appropriate for its use case (including where human review is appropriate) before any use or sharing of any Output." Cognizant's press release of 23 September 2026 shows that arrangement in production. A seven-person Cognizant team used Devin to rebuild Odyssey Logistics' legacy forms as cloud-native applications. "Every change was reviewed and signed off by a Cognizant technical lead or architect before merging," the release says, and it reports a 37 percent net cost saving.

Factory's homepage heading is "Build your software factory," and the post that defined Factory 2.0 on 15 June 2026 says who does the building: "This system will be built by engineers, and in turn will build their software." Partners are the second route. They will "embed with customer teams to design, deploy, and run the software factory." The factory around the Droids is yours to assemble, inside your tools, under an autonomy ceiling your organization sets between Off and High. In Factory's documentation the Software Factory overview, Triage, Release and Incident Response carry a Private Preview tag, which Factory defines as "limited or invite-only access that is already running in production." Automations became generally available on 30 September.

Taiga's service description, version 1.0 of 6 September 2026, defines a Build: "Each Build delivers the implementation package: code, infrastructure definitions, CI/CD configuration and the audit trail for that Build, with the human contributions recorded against it." Before the first Build on a new project, eight documents have to exist: specification, user flows, architecture, technology decisions, data flow, threat model, DPIA and risk register. After that, each change is an initiative that Taiga plans and builds "using its documents, policies and repository context," and our product page adds: "Plans can wait for approval, or build on their own." Our shared-responsibility page keeps the review where the other two keep it: "Reviewing generated output is yours, and it is the responsibility easiest to let slide."

The test

Which document existed before the build?

Both competitors can start from a written spec. Devin accepts a "Build from a spec page" input and generates architecture documentation through DeepWiki when a repository is onboarded, and Factory's Spec Mode plans before code changes. Neither one's public pages describe a threat model or a data protection impact assessment that the product writes before the build. Factory's Security Review applies STRIDE threat modelling to the pull request.

Taiga requires the eight documents above, threat model and DPIA among them, before the first Build on a new project, and plans every later change against them. For imported software the documents are derived from the code. The worked example on our site is labelled illustrative, and it says "passing these sample checks is not a production-readiness claim."

Who approved the change, against which written policy?

Cognition's audit log APIs "record user actions and feed the monitoring and compliance tools a customer already operates," and its security exhibit keeps audit logs for at least one year. Devin's security profiles became generally available on 7 August 2026, and an enterprise can now enforce them, with network and MCP allowlists and read-only git, as a floor nobody below can loosen.

In Factory, organization-level hard controls are authoritative, and lower levels "can never weaken or remove" them. Droid Shield scans staged diffs for secrets and "fails closed when a diff is too large to scan safely." Audit-log export is arranged "during your enterprise engagement," and the retention period of Factory-hosted logs is not stated.

Taiga records the human contributions to each delivery in the audit trail, and our master terms say Taiga "keeps that record for seven years and produces it on request." Control-plane audit logs sit in an isolated immutable archive for the same seven years, and the tenant audit trail is on the list of data you can export yourself. The same terms tie our warranty to what they call a Governed Run, "a run for which a named competent reviewer is recorded in the Platform audit trail before its output is promoted to production." Our pages also say what the record is not: the 22 readiness questions "are not one universal automated release gate", and "policy evidence does not certify compliance." The one item in our public run record is a build that failed, was fixed by the agent and passed on re-run.

Under which law, and where is the data?

Cognition's enterprise contract is with Cognition AI, Inc. of San Francisco, under New York law with arbitration in New York. Devin's brain, which the documentation describes as "a stateless, cloud-based service that powers Devin's intelligence," is "always residing in Cognition's Cloud" in every deployment model. A dedicated customer can "choose the region where it runs" for its single-tenant AWS environment, which Cognition says helps "meet internal policies and local data residency requirements." Devin's deployment documentation, enterprise and trust pages name no EU region; a May 2025 post on devin.ai about the self-hosted offering of Windsurf, the editor now called Devin Desktop, says "we launched a cluster in Germany to satisfy European regulations around data residency." Breach notice is "without undue delay following any confirmed breach."

Factory's contract is with The San Francisco AI Factory Inc., under California law, and its data processing agreement states that "Factory's primary Processing operations take place in the United States." Factory's article on its EU deployment, an enterprise option, describes the split: "Factory's EU deployment keeps session content and inference in Europe, while organization pointer records, user profiles, and billing data remain in the global US deployment." Its documentation adds that those records "contain no user-generated content." Its airgapped build runs with no outbound internet on models you host, and Factory Private, which moves the control plane into your VPC or on premises, has been in Private Preview since 18 September. The same article states the limit: "Neither regional placement nor customer ownership alone establishes compliance with a particular law or contract." Breach notice is "without undue delay."

Taiga's contract is with Taiga AI Oy of Helsinki, under Finnish law with arbitration seated in Helsinki. Customer data at rest is in AWS eu-central-1, Frankfurt, and non-EU regions are denied by organization-wide guardrails at the infrastructure boundary; our trust page adds that "inference may run in more than one EU region for capacity and resilience" and that "the CDN's TLS certificate is managed in a US region because the provider requires it." Inference runs in Amazon Bedrock, where the model provider "has no access to that account, to the prompts and responses passing through it, or to the service logs." Both companies under that stack, AWS and Anthropic, are US-parented. Breach notice comes within 24 hours, and erasure completes within 35 days with a signed deletion certificate. The terms include a DORA add-on, extra terms that apply automatically, with no opt-in, to any customer that is a financial entity under DORA, and a clause that makes Taiga the provider of the platform under the EU AI Act and the customer its deployer. They do not mention NIS2.

We found no mention of DORA, NIS2 or the EU AI Act in the public pages we read from either US company: Factory's complete documentation export and its security, enterprise, legal and EU deployment pages, and Cognition's documentation export and its product, trust and legal pages. Their enterprise contracts may say more.

Side by side

Every cell is drawn from a public page checked on 4 October 2026. Text in quotation marks is verbatim; not stated means the page is silent.

Cognition, Factory and Taiga compared row by row, from public pages checked on 4 October 2026
QuestionCognition (Devin)FactoryTaiga
Sold as"the first autonomous software engineer""Build your software factory"A monthly service: documents, build and record
Who carries review and operationYour reviewers; output evaluation is the Customer's (enterprise terms 2.4)Your engineers, or partners who "design, deploy, and run the software factory"Taiga produces the output, documents and record; you review and accept; you or Taiga run it, by hosting mode
Documents before the buildSpec input; no product-generated threat model or DPIA statedSpec Mode plan; no threat model or DPIA before the build statedEight required, including threat model and DPIA; a plan before each change
Approval before mergeYour branch protectionYour branch protectionYour branch protection; a named reviewer recorded for a Governed Run
Audit recordAudit log APIs; kept at least one year (security exhibit)Audit events; export arranged in the enterprise engagement; retention not statedHuman contributions per delivery kept seven years, produced on request; tenant audit trail you export yourself
Data locationBrain "always residing in Cognition's Cloud"; single-tenant AWS in a customer-chosen region; EU region not named in deployment docsEU deployment: "session content and inference in Europe"; "organization pointer records, user profiles, and billing data" in the US; airgapped buildAWS eu-central-1; non-EU regions denied by organization-wide guardrails; inference within EU regions; CDN TLS certificate managed in a US region
Contracting entity and lawCognition AI, Inc.; New YorkThe San Francisco AI Factory Inc.; CaliforniaTaiga AI Oy; Finland
Breach notice"without undue delay following any confirmed breach""without undue delay"Within 24 hours
DORA, NIS2, AI Act in public documentsNot statedNot statedDORA add-on; AI Act provider and deployer clause; NIS2 not in the terms
Certifications and auditsSOC 2 Type II since September 2024; ISO/IEC 27001:2022SOC 2 Type II, ISO 27001, ISO 42001 (security attachment, trust centre)None held; ISO 27001 in progress, SOC 2 planned
Third-party penetration testAt least annually, plus a bug bountyAt least annuallyNone yet
Source controlGitHub, GitLab, Bitbucket, Azure DevOpsGitHub, GitLabGitHub.com and GitHub Enterprise only
IdentitySAML/OIDC SSO, SCIMSAML/OIDC SSO, SCIMCognito with Entra ID or Google; no SAML, no SCIM
Model providers"does not currently support third-party LLM API keys"Multiple providers, BYOK, your own endpoints in airgapAnthropic Claude via Amazon Bedrock only
PricingFrom $20 a month on individual plans; enterprise priced per contract, billed in agent compute unitsFrom $20 a month on individual plans; enterprise tier with unlimited members, no published priceFrom EUR 2,000 a month for Learn; from EUR 6,000 with Deliver + Operate; unlimited users; always-on cap
Training on your dataEnterprise: never without written consent; paid self-serve plans: opt-out"in no event" without the customer's prior consentNever; organization-wide opt-out

Pricing does not compare like for like, and finance will ask anyway. Taiga starts from EUR 2,000 a month for Learn and from EUR 6,000 once Deliver + Operate is added, with unlimited users and an account cap that stops consumption-driven work instead of invoicing past it. Both competitors start at $20 a month on individual plans and price enterprise contracts one by one.

What the test says about Taiga

The size gap from the opening raises the question DORA asks in Article 28(8): what happens if the provider fails. Our master terms answer it by hosting mode. In customer-hosted mode, where you use your own repositories and your own cloud account, "there is nothing to escrow: the Customer loses the Platform, not the product." In Taiga-hosted mode, "source and deployment configuration are mirrored to the Customer continuously through the term." Leaving while we still exist is covered as well: export is self-service in machine-readable formats, the tenant audit trail and the Discovery documents are on the export list, and there is no switching charge.

The other follow-up questions go against us. We hold no certificate. ISO 27001 is "In progress. Not held. The management system is running; the certificate is not ours yet and you will not see the logo here until it is." SOC 2 Type II is planned, and no third-party penetration test has been carried out yet. Cognition has held SOC 2 Type II since September 2024 and lists ISO/IEC 27001:2022, Factory lists SOC 2 Type II, ISO 27001 and ISO 42001, and both test externally at least once a year.

"SAML and SCIM provisioning are not built yet," and both competitors ship them. Our platform connects to GitHub.com and GitHub Enterprise only, while Devin also works with GitLab, Bitbucket and Azure DevOps, and Factory with GitLab. "Every agent runs an Anthropic Claude model reached exclusively through Amazon Bedrock," with "no second provider," where Factory routes across several providers and your own endpoints. And the platform needs the internet: "The platform is a managed EU SaaS, so no, and we will not pretend otherwise." Factory ships an airgapped build.

Our two Amazon Bedrock Guardrails profiles are built and versioned but switched off in every environment while support cases with AWS are open. The correction went up on our trust updates page on 5 August 2026, with what covers the gap until they return: "retrieval is scoped to your tenant, agents run least-privilege and are invoked only by the API, and agent output is validated at the API layer." No newer entry has replaced it.

Which one to buy

If you have a platform team that wants to assemble its own factory inside the tools it already runs, possibly with no outbound internet at all, Factory sells that, including an airgapped build we do not offer. If you want an autonomous engineer on the team, with your reviewers or an integrator carrying the review, or your code lives on Bitbucket or Azure DevOps, Cognition sells that. Both of them also pass a vendor-risk gate that asks for a certificate, SAML or SCIM before onboarding; we do not yet.

Taiga is for the regulated EU enterprise that would rather not staff a platform team to build its own factory, and whose purchase turns on the auditor's questions.

Compare the internal work of assembling and operating a factory with the published Taiga service plans. The evidence hub shows what our research and public reference implementation establish, with their limitations.

Ask all three of us the three questions from the top of this page on the first call, and do not let any vendor pick the change, us included.

Sources

All pages fetched and re-checked on 4 October 2026. Taiga's own claims carry receipts that link to the page they come from.

Frequently asked questions

What is the difference between Taiga, Factory AI and Cognition's Devin?+

They are sold as three different things. Cognition sells Devin, "the first autonomous software engineer," which works inside your repositories and tools until a pull request is ready to merge; your engineers or an integrator review and evaluate the result. Factory sells a software factory that, in its own words, "will be built by engineers, and in turn will build their software," with partners available to design, deploy and run it. Taiga sells a monthly service that requires eight documents before the build, plans every later change against them and keeps a seven-year audit log, under a Finnish contract.

Do Factory or Devin keep data in the EU?+

Factory documents a dedicated EU deployment for enterprise customers that keeps session content and model inference in Europe while organization pointer records, user profiles and billing data stay in its US deployment; it also offers an airgapped build. Devin's deployment documentation and trust pages name no EU region, though a May 2025 post on devin.ai about Windsurf, the editor now called Devin Desktop, mentions a cluster in Germany for European data residency; dedicated customers choose the AWS region of their single-tenant environment, while Devin's brain stays in Cognition's cloud. Taiga runs in AWS eu-central-1 with non-EU regions denied by organization-wide guardrails. All three checked 4 October 2026.

Which of the three is certified?+

As of 4 October 2026, Cognition has held SOC 2 Type II since September 2024 and lists ISO/IEC 27001:2022. Factory's security attachment and trust centre list SOC 2 Type II, ISO 27001 and ISO 42001. Taiga holds no certificate: ISO 27001 is in progress and SOC 2 Type II is planned, and no third-party penetration test has been carried out yet.

Which one should a regulated EU enterprise choose?+

It depends on who will carry the review and the paperwork. For a regulated EU enterprise that would rather not staff a platform team to build its own factory, and whose purchase turns on an auditor's questions, Taiga's public terms answer them: eight documents required before the build, a plan before every change, a seven-year audit log, a 24-hour breach notice and a DORA add-on under Finnish law. An organization that needs an air-gapped platform, SAML and SCIM, GitLab, Bitbucket or Azure DevOps, a choice of model providers, or a vendor with certificates in hand today is better served by Factory or Cognition.

What should I ask all three vendors?+

Pick a change from last quarter at random and ask which document existed before it was built, who approved it against which written policy, and under which contract and law the record sits. Ask Taiga the same three questions.

Our July list is in ten questions to ask an AI development vendor. Run the six-question self-test

See how a requirement or a repository finding becomes a reviewed change.