Secure by design. Compliant by process.
Millwright builds every AI workflow under the controls below. Each control is mapped to the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act, and the whole set is built to the United States regulations that apply to AI today.
Controls that apply to every system
Nine controls, applied to every client and every system. Under each one is where it sits in the three frameworks we build to.
-
Least privilege
Every automation runs under its own identity with only the access it needs. Credentials live in a secrets manager, never in a prompt or a code repository.
NIST AI RMF: Govern, Manage. ISO/IEC 42001: A.3 internal organization, A.6 life cycle. EU AI Act: Article 15, cybersecurity.
-
Your approval
Nothing goes live, and no high impact action runs, without a person approving it. An approval that is not given expires. It never passes by default.
NIST AI RMF: Govern. ISO/IEC 42001: A.9 use of AI systems. EU AI Act: Article 14, human oversight.
-
Tested before launch
Every system is tested for prompt injection, data leakage, and misuse against the OWASP Top 10 lists for language model and agentic applications before it reaches your customers.
NIST AI RMF: Measure. ISO/IEC 42001: A.6 verification and validation. EU AI Act: Article 15, accuracy and robustness.
-
A stop switch
Every system can be stopped at once, by workflow, by client, or across the board, and the stop is tested on a schedule.
NIST AI RMF: Manage. ISO/IEC 42001: A.6 operation and monitoring. EU AI Act: Article 14, the ability to interrupt the system.
-
Your data stays yours
Collection is kept to the minimum necessary. Client data is isolated per client and is never used to train a model. Your data, recordings, and records remain yours.
NIST AI RMF: Map, Govern. ISO/IEC 42001: A.7 data for AI systems, A.10 suppliers. EU AI Act: Article 10, data governance.
-
Compliance by process
Before a workflow goes live we identify the rules that apply to it and, where a rule calls for it, have counsel review the design. We do not launch first and check later.
NIST AI RMF: Govern, Map. ISO/IEC 42001: A.2 AI policy, A.5 impact assessment. EU AI Act: Article 9, risk management.
-
Audit trails
Every action is attributable to a person or a named system identity, with structured decision records you can review.
NIST AI RMF: Measure, Manage. ISO/IEC 42001: A.6 event logging. EU AI Act: Article 12, record keeping.
-
Ongoing checks
Every live client receives automated monthly security checks and a concise summary of findings, open risks, and anything that needs your attention.
NIST AI RMF: Manage. ISO/IEC 42001: A.6 operation and monitoring, clause 9 performance evaluation. EU AI Act: Article 9, continuous risk management.
-
People know it is AI
Every agent we deploy identifies itself as an AI assistant at the start of each interaction, whether or not a law requires it in that place.
NIST AI RMF: Govern, transparency. ISO/IEC 42001: A.8 information for interested parties. EU AI Act: Article 50, transparency.
The frameworks we build to
Three references shape how every system is designed and run. Building to them is how we work. It is not a certification, and we say so.
-
NIST AI Risk Management Framework
The United States government's voluntary framework for trustworthy AI, version 1.0, organized around four functions: Govern, Map, Measure, and Manage. We also apply its Generative AI Profile. NIST is revising the framework under the federal AI Action Plan, and we track that revision because at least one state law treats alignment with the current version as a defense.
-
ISO/IEC 42001
The international standard for an AI management system. Our policies, roles, impact assessments, life cycle controls, data handling, logging, and supplier reviews are structured to its Annex A areas, so that certification is a scoping decision rather than a rebuild. Millwright is not certified today.
-
EU AI Act
The European Union's AI regulation applies to systems placed on the EU market or whose output is used in the EU. That does not describe our clients today, and we build to its requirements anyway: risk management, data governance, record keeping, transparency, human oversight, and accuracy, robustness, and cybersecurity, plus the duty to tell people they are interacting with AI, in force since August 2, 2026.
United States regulations we build to
There is no comprehensive federal AI statute as of October 2026. What applies to an AI system in the United States is a set of existing federal laws, a growing number of state laws, and an unresolved federal effort to preempt the states. This is how we read it today, and it is a summary, not legal advice.
Federal law that applies to every client facing system
- FTC Act, Section 5. Unfair or deceptive practices are unlawful whether or not AI is involved. Claims about what a system can do must be truthful and substantiated, and a system must not deceive the people it talks to.
- Telephone Consumer Protection Act. The FCC ruled on February 8, 2024 that AI generated voices are artificial voices under the Act, so automated calls need prior express consent, and its consent revocation rules have applied since April 11, 2025.
- CAN-SPAM Act. Commercial email must carry accurate headers, a truthful subject line, a physical postal address, and a working opt out honored within ten business days.
- Federal AI policy. Executive Order 14179 of January 23, 2025 and the AI Action Plan of July 23, 2025 set the direction. Executive Order 14365 of December 11, 2025 directed the Justice Department to challenge state AI laws, and the White House published a legislative framework for federal preemption on March 20, 2026. Congress has not enacted it. State laws remain in effect until a court or Congress says otherwise.
State AI laws that shape how we build
- State AI governance acts, the first of which took effect January 1, 2026. These prohibit AI built to manipulate people toward harm, to discriminate unlawfully, or to produce abusive sexual content, and at least one recognizes alignment with the NIST AI Risk Management Framework as an affirmative defense.
- Colorado SB 26-189, signed May 14, 2026 and effective January 1, 2027. Replaces the 2024 Colorado AI Act. Where automated decision technology is used in a consequential decision about a person, it requires notice at the point of interaction, a plain language explanation after an adverse outcome, a right to correct data, and human review.
- California's automated decision making regulations under the California Consumer Privacy Act, effective January 1, 2027. Pre use notice, opt out rights, and access to the logic behind a significant decision.
- Other states have passed narrower rules, most of them requiring that people be told when they are dealing with AI. Disclosure is our default everywhere, so the common thread of state law is already built in.
Current as of October 17, 2026. Where a workflow would make or support a consequential decision about a person, or reach a client, answering location, or customer base in a new state, counsel reviews the rules for that state before we build.
Compliance by process
Compliance is not a sticker on a product. It is a step that happens before launch, every time. For each workflow we identify the rules that apply to it, including commercial email, telephone and text consent, call recording, state privacy and AI laws, and breach notification duties, and where a rule calls for it, counsel reviews the design before the system goes live.
We do not make compliance statements about your business without counsel review, and we do not describe framework alignment as certification.
Where we draw the line
Some work we do not take on, because the risk to a client outweighs anything a system could save them.
-
No protected health information
We do not build systems that create, receive, store, or transmit protected health information, so we do not work with health care practices.
-
No payment card data
No system we build collects, stores, or repeats a card number. Payments go through the processor's own hosted page, never through an agent.
-
No recording by default
Calls are not recorded unless a written recording policy covering every jurisdiction on the call is in place first.
-
No outreach to your customers without consent
A system that calls or texts your customers does not go live until the consent design has been reviewed by counsel.
If something goes wrong
Every system fails safe. If an AI provider is unavailable, the workflow stops rather than improvising a different behavior.
If an incident occurs, the process is fixed in advance: contain the affected system, preserve the evidence, assess the impact, notify you and anyone else the contract or the law requires within the required time, fix the cause, and review what happened so it does not repeat.
Every agent we deploy identifies itself as an AI assistant at the start of each interaction. That is our standard for every client.