INTRODUCING STEWARD the enterprise platform
with embeddable
AI product experts
and verifiable outcomes.

STEWARD: The enterprise platform that provisions platform-independent, surface-agnostic AI product experts with verifiable outcomes. STEWARD gives a company that operates software an expert inside that software. It reads the company's documentation and its live systems under each user's own login, answers with sources, and leaves a record for every answer.

STEWARD is built in three parts: the Data Plane, the Control Plane, and the widget. STEWARD is not a widget. It ships with one. The Data Plane is where the expert works: it retrieves from your knowledge base, reads your API on the user's behalf, runs analysis in a sealed sandbox and streams a cited answer. The Control Plane is where you govern it: products, sources, keys, members, billing, traces and the audit log, in the console or through its API. The widget is the reference surface.

STEWARD is for any company that operates software with its own users. Operators who run their work on their own software. Developers and agencies who build for clients. Software companies and institutions with their own platforms. Live reads need a host application with signed-in users, because STEWARD borrows their token. A public site without login gets knowledge-base answers only.

No. STEWARD is a platform that ships with a widget, and the widget holds no knowledge, no agency, no judgment and no record. Everything that makes the expert lives behind the platform API. Call the chat API directly and you get the same streamed answer with the same trace.

A STEWARD expert does four things, held at once. Knowledge: it has read every page you designated and cites the section. Agency: it reads your own API on the user's behalf, under the user's own login, and runs analysis over what it finds. Judgment: below the evidence threshold it says it has no material, and does not improvise. Record: every answer leaves a trace and every change leaves an audit entry.

STEWARD reads. The person decides. It never books, sends, moves, approves or changes a record. It reads the REST API you configure, at a fixed base URL, with an endpoint allowlist if you set one, under the user's own token. It does not answer from material it was not given. It does not record who asked.

Every STEWARD answer carries its record: the sources it read, the live reads and analysis runs it made, the full prompt and completion, the cost and why it stopped. You open that record in the console at the same fidelity we see it. The audit log holds every change to your configuration. Nothing edits it, and only deleting the account removes it. An evaluation that scores whether the answer matched the question is specified and staged.

Yes, STEWARD runs inside the console too. Ask it about a trace, its sources or the credits behind an insight, and it answers from the record.

STEWARD says so, names what it did not find, and hands the question back to the person. Below the evidence threshold it states it has no relevant material and does not improvise. The refusal is recorded like any other answer.

95% of enterprise AI pilots show no measurable ROI.

Source: MIT NANDA, The GenAI Divide, 2025.

We think the cause is verification. If nobody can show what the AI read and why it answered, the institution cannot put it in front of a customer, an auditor or a regulator. The record is what lets it.

Documents you designate are ingested into your STEWARD account's knowledge base. Your live systems are read at the moment of the question, under the user's own login, with no index and no copy. AWS services alone process institutional content. Inference runs on Amazon Bedrock through a private endpoint, and AWS commits that Bedrock stores no inputs or outputs and that model providers never train on them. Personal data enters a prompt only when the question requires it, and personal data patterns are redacted before any text reaches the user. Conversations and telemetry are deleted at your account's retention window or on demand. Those deletions, and deleting the account itself, are how a data-subject request is fulfilled.

AWS Bedrock is STEWARD's sole model provider. Inference runs through a private endpoint signed by the infrastructure's own identity. No model-provider key exists anywhere in the platform. You pick the model per product in the console, with every change versioned and reversible. The model predicts text. STEWARD does the retrieval, the live reads, the guardrails and the record.

Every answer STEWARD delivers is an insight. It costs one credit, plus one for each live read or analysis run on the way, up to four more. An answer that never arrived costs nothing. A PDF costs half a credit a page, once per version. Every other file costs nothing to process. Every charge links to the trace that earned it. Plans and bundles are in the pricing section above.

No. STEWARD has no overage charges and no auto-refill. You are billed for your plan and for the bundles you choose to buy, nothing else. Subscribe mid-month and you pay only for the rest of that month, with the full allowance from day one. Cancel any time in the console and keep your plan to the end of the period you have paid for.

Yes. By default, STEWARD emails every member of your account when you pass 80% of the allowance, when it runs out, when a bundle balance falls below 100 credits, when a payment fails, and when your card will expire before the next renewal. Every alert also shows in the console, whether or not email is on. Users do not see these alerts and these settings can be modified by the account members.

On every STEWARD plan below Enterprise, cards and invoices are managed in Stripe's customer portal, opened from Payment & invoices in the console. Payments run on Stripe's hosted checkout, so card and bank details never touch STEWARD. Plan changes are made in the console.

Because STEWARD did the work. The expert made the call to your API or ran the analysis. What failed was downstream. A call STEWARD itself refused or could not run costs nothing.

No. Every STEWARD plan can embed the expert in your product with a publishable, origin-bound key. Enterprise adds the option of secret keys for direct API access, so your own software can call the expert without the widget.

No. STEWARD bills per account. Every member shares one allowance and one record.

Yes. The STEWARD chat API returns the same streamed answer and trace the widget shows, on every tier, with a publishable key bound to one product and its origins. The control-plane API automates everything the console does, with a secret key, on Enterprise. Secret keys are minted once in the STEWARD Console, scoped to a product, and revoked from the same place, with revocation reaching every surface in 60 seconds or less. The docs are at docs.steward.link.

An insight is a STEWARD answer in which the expert read your live system or ran an analysis, or both. A live read calls your API on the user's behalf and returns the record as it is now. An analysis run computes a figure or a chart over that data in a sealed sandbox. The chart arrives inside the answer, with its sources.

You install STEWARD. Your client pays. You never chase the invoice. Your account is free. Each client is an account on this pricing page, with its own expert, its own sources and its own record. Provisioning experts from your own code uses the control-plane API, on Enterprise.

STEWARD setup is configuration. Create a product in the console. Connect a repository or a shared folder. Give it your API's base URL and, if you want, an endpoint allowlist. Mint a key. Take the snippet from the playground in HTML, React or Angular, or hand the embed skill to your coding tool. Configure and embed. Nothing to rip out.

No. The STEWARD record holds the source, the time, the input and the output. Reading under the user's login is not recording the user. The token is forwarded, never stored.

STEWARD keeps conversations and telemetry 90 days by default, 6 months on Enterprise, up to a year. Billing records outlive every window. Export what you keep longer.

A SOC 2 Type II report on STEWARD is in progress. Production and staging run in separate AWS accounts. Production deploys only by manual dispatch of a commit with green CI. More than 8,000 automated tests gate every merge. Security materials are available to buyers in diligence.

The STEWARD expert is called an AI product expert on every surface, and the word AI is in its definition. AI-interaction disclosure at first contact and machine-readable output marking under EU AI Act Article 50 are specified and staged.

Yes. Ask STEWARD about STEWARD on this page, open the trace of the answer you received, then start for free at steward.link.