Newton Research

The vendor makes action a headline feature. Under end-to-end execution: "Newton's agents take action on your behalf: writing code, executing analyses, and pushing data to any of your platforms & tools". The marketer use case states the target: "push audiences and campaign changes directly to the relevant platform". Writing an audience segment or a budget change into a live ad platform is authenticated action in an outside system, not retrieval.

The targets are named on the vendor's partners and integrations strip: Meta, Google Ads, Google Display, Google Analytics, LiveRamp, BeesWax, HubSpot, Salesforce, Snowflake, Databricks, Looker, Tableau, Google Sheets, Excel, PowerPoint, OneDrive, Microsoft 365 and Google Cloud. The published Snowflake Cortex launch and LiveRamp Clean Rooms and XMI execution corroborate action inside third-party environments.

The custom route is also open, independently: "Newton also seamlessly integrates with nearly any business", with agents writing and executing their own code and the vendor describing fully flexible integrations, so the customer is not confined to a fixed connector list. Pulling data from those systems is retrieval rather than action.

Sourcenewtonresearch.ai/solutions/for-brands and newtonresearch.ai/aboutread 2026-09-12

Multi-agent and multi-step execution are both documented, and the multi-agent half is the product's own framing. Newton sells "a team of specially-trained AI agents" and describes itself as agent-native, built on a multi-agent framework since 2023, with agents that "take your direction and work hand-in-hand with you and your team to execute complex analytics tasks".

The execution chain is named end to end: "Newton's agents take action on your behalf: writing code, executing analyses, and pushing data to any of your platforms & tools". Those are three distinct stages with state carried between them, author the code, run it against the customer's data, then write the result outward, which is orchestration rather than a chained prompt.

The blueprints are themselves multi-step pipelines: Data Exploration and Preparation chains exploratory analysis, cleaning and standardization before anything downstream runs, and Planning and Activation chains budget allocation, forecasting and predictive optimization. Customers can also "build and train your own custom AI agent in days, not months", so the runtime accepts new participants. No named runtime is published.

Sourcenewtonresearch.ai/solutions/for-brands and newtonresearch.ai/aboutread 2026-09-12

Grounding is assembled per run from the customer's live data environment, and no maintained retrieval structure is documented. Agents work over the customer's own data in place: "pull information and insights across datasets in real time, before things become stale", and "integrate your proprietary data and models".

Newton can pull data from multiple sources and environments and transform or clean it before executing analytics, and the Data Exploration blueprints exist to join, analyze and investigate data sets from various sources. Answers are grounded in the customer's corpus, but every documented act of grounding happens inside a run: pull, transform, analyze, return.

No index, graph or embeddings layer over the customer's corpus is described, nothing is said to persist between analyses, and there is no queryable structure the buyer could point at. Insight Automation comes closest, generating insights by pulling from past reports, presentations and analyses, but a store of past deliverables is the application's data model rather than a maintained retrieval layer. The architecture explains why: with zero data movement, there is deliberately no Newton side index.

Sourcenewtonresearch.ai/solutions/for-brands and newtonresearch.ai/contactread 2026-09-12

No oversight mechanism is documented: no approval gate, review-and-approve step, pre-execution policy engine, confirmation before an action lands, per-agent scoping the buyer configures, escalation path, or administrator surface where any of it would be set. The collaboration language is not a gate: "Newton's team of specialized AI agents take your direction and work hand-in-hand with you and your team" describes who instructs the agent, not who approves what it does, and a customer testimonial about leaving room for human judgment and creativity is a sentiment about working style rather than a shipped mechanism.

The absence matters because of where the agents write. They "push audiences and campaign changes directly to the relevant platform" and "push data to any of your platforms & tools", writes into live advertising systems where a wrong budget allocation spends real money, and nothing published sits between the agent's decision and that write. The code transparency lets an analyst check work after the fact, but being able to read what happened is not being able to stop it.

Sourcenewtonresearch.ai/solutions/for-brands and newtonresearch.ai/aboutread 2026-09-12

Neither attestations nor access controls are published. No attestation of any kind: no SOC 2, ISO 27001, HIPAA or GDPR certification, no trust center, no compliance badge, no auditor named, no security kit on request. No customer-facing access surface: no SSO, SAML, SCIM, RBAC, roles, permissions or admin console. No security or trust page is published, and the only legal link is the Privacy Policy.

Newton's security pitch is architectural: "Access state of the art AI without putting sensitive data at risk", "keep all your sensitive data safe and secure", "no movement of data, full compliance". Zero data movement is a deployment property, not a control surface or an attestation, and "full compliance" names no framework. A buyer asking who inside their organization can run which agent, or for an auditor's report, has nothing to read.

Sourcenewtonresearch.ai/about navigation and footer, newtonresearch.ai/solutions pagesread 2026-09-12

Run-time transparency, no retained record. The brands page offers full transparency into the code and processes the platform executes, and a named customer on the same page reports seeing the code it generates. Nothing documents an audit log, retention, export, a SIEM path or attribution of which agent pushed which audience to which platform, so a buyer cannot reconstruct an action afterwards.

Sourcenewtonresearch.ai/solutions/for-brandsread 2026-10-01

State carried across turns or sessions is not described anywhere, and there is no per user or per tenant memory scope, no published lifetime, no expiry or purge path and no read or write surface. The one statement that sounds like memory describes absorbed learning: "We will never use your proprietary data to train Newton or make it smarter for general use... anything you show or teach to Newton will only be used to make your specific instance of Newton better for your needs".

Learning folded into a per-customer instance is a model property, not a store the agent writes and reads back, and the buyer cannot see, scope or purge it. The architecture points the same way: data never leaves the customer's environment, so there is deliberately no Newton-side store. Insight Automation pulling from past reports, presentations and analyses is retrieval over stored deliverables, not memory.

Sourcenewtonresearch.ai/contact and newtonresearch.ai/solutions/for-brandsread 2026-09-12

Running inside the customer's environment is the product's central design claim. The vendor states "Newton is a containerized application architected for deployment to your cloud platform with zero data movement", and repeats it across every segment page: "Newton runs natively in your cloud with no movement of data" and "Newton's agents work alongside you, natively in your enterprise data environment, with no movement or leakage of sensitive data".

That is not a region choice within the vendor's cloud but no vendor cloud at all. The target environments are named on the vendor's own blog, a contained system that lives inside a customer's own infrastructure on AWS, GCP, Azure, Snowflake or Databricks, corroborated by the partners strip on the About page carrying Databricks, Snowflake and Google Cloud, a first-party Databricks Marketplace listing and a published Snowflake Cortex launch.

Sourcenewtonresearch.ai/about and newtonresearch.ai/solutions/for-brandsread 2026-09-12

Newton calls them blueprints: "Newton's team of AI-powered agents come pre-loaded with a curated set of blueprints that make up its marketing science playbook. These blueprints give non-technical users the ability to execute common marketing analytics use cases with the push of a button, all on your own data".

The catalog is browsable and paginated, with seven named categories each holding three named blueprints and a Load More control: Data Exploration and Preparation (Exploratory Data Analysis, Data Cleaning, Data Standardization); Audience Management (Audience Composition Analysis, Audience Segmentation, Lookalike Modeling); Customer 360 (LTV Analysis, Customer Journey Analysis, Predictive Modeling); Planning and Activation (Budget Allocation, Forecasting & Benchmarking, Predictive Optimization); Measurement (Efficiency & Effectiveness, Test Design, Incrementality Analysis); Reporting and Insight Generation (Dashboard Reporting, Insight Automation, Whitelabeled Reporting); Anomaly Detection (Anomaly Detection, Outlier Identification, Issue Prevention).

That is twenty-one named units on the first page alone, and each does its own job: Lookalike Modeling and Anomaly Detection work on different data and neither depends on the other. The Test Design and Incrementality blueprints measure advertising, not the agent.

Sourcenewtonresearch.ai/solutions/for-brands Blueprints sectionread 2026-09-12

Every documented route by which work reaches the agent is a person asking. Newton is invoked by a natural-language question ("ask analytics questions in plain English and get answers in seconds"), by pushing a blueprint button, or by an analyst directing the agent team. No event, webhook, schedule, cron, inbound queue or monitoring trigger is published, and no channel exists beyond the product's own interface and the white-label chat: no email address, messaging platform, mobile app or extension.

The closest thing is the Predictive Optimization blueprint, which reads "optimize ongoing planning and audience development based on recurring automated analysis", with Anomaly Detection framed as catching issues early. Recurring automated analysis sounds like a schedule, but it is one phrase inside a use-case description: no scheduling surface, cadence, configuration or statement that a run starts without a person is published. Dashboards that refresh in minutes are a data pipeline cadence, not an agent trigger.

Sourcenewtonresearch.ai/solutions/for-brandsread 2026-09-12

No model statement and no choice. The About page's Partners & Integrations strip shows Anthropic and OpenAI logos among Meta, Looker, Databricks, Google Ads, PowerPoint, OneDrive and others, which does not say whether Newton runs on those models or integrates with them; no page states which model serves any feature, and no selection, bring your own model or admin policy is offered.

Sourcenewtonresearch.ai/about and newtonresearch.ai/solutions/for-brandsread 2026-10-01

Nothing is documented for outside callers: no REST or GraphQL API, no SDK in any language, no developer portal open or gated, no MCP server, no A2A agent card, no webhook, and no endpoint or auth scheme. The site has no Developers or docs entry and no documentation or developer subdomain.

A Databricks Marketplace listing shows Newton can be installed into a Databricks environment, which is distribution and a deployment fact, not a way to call the platform from outside.

The one candidate is the white-label offering, "Whitelabel an AI analytics chat interface for use directly with your customers" and "incorporate the intelligence and capabilities of Newton's agents into your own UI", which implies an interface exists, but nothing technical is published: no reference, no auth, no shape.

Sourcenewtonresearch.ai/about navigation and footer, newtonresearch.ai/solutionsread 2026-09-12

The product measures advertising, not itself. Test Design and Incrementality Analysis blueprints evaluate campaigns, the About page's regular testing of the agents on more than 110 marketing analytics questions is the vendor's internal regression suite, which the customer neither invokes nor receives, and the benchmark against general LLMs is a vendor claim about its own product. No sandbox, dry run, harness, scored test case or regression check on a change the customer makes is documented.

Sourcenewtonresearch.ai/about and newtonresearch.ai/solutions/for-brandsread 2026-10-01

Newton is built around programmatic interfaces, the opposite of an agent driving screens: it runs as a containerized application inside the customer's own cloud, queries their warehouses directly, writes its own code, and pushes results through platform APIs to Meta, Google Ads, LiveRamp, Salesforce and the rest.

No hosted or local browser, desktop session, or remote or local computer control is documented. Writing code and executing analyses is code execution inside the customer's environment, not interface control, and the white label chat interface is Newton's UI embedded in someone else's product, the vendor being displayed rather than the agent operating a display.

Sourcenewtonresearch.ai/solutions/for-brands and newtonresearch.ai/aboutread 2026-09-12