AI that acts for you. Never without your consent.
Steward is a consent-first AI agent for digital government services. It listens, navigates and prepares the work — then stops and asks before anything consequential happens.
digital.
Observed friction
- 38
- fields on one form
- 6
- pages, no save
- 3
- documents to find
- 1
- language offered
Illustrative composite
Fig. 02 — Every field is a decision point. Every error is a reason to give up.
Accessibility didn't always keep up.
Forms written for officials. Workflows that assume a desktop, fluent literacy and an hour of patience. The people who need public services most often meet the most friction.
The intervention
Enter Steward.
Say what you need, in your own words. Steward finds the service, reads the form and prepares every field — then hands the decision back to you.
Where the AI stopsAutomation
with a boundary.
- UnderstandParses your request in plain language
- NavigateFinds the right service and the right page
- PrepareFills fields from what you have shared
- ExplainTells you what each step means
- SubmitSending an application on your behalfPauses → asks you
- ConfirmAgreeing to terms or declarationsPauses → asks you
- CommitAnything payable, binding or irreversiblePauses → asks you
The boundary isn't a setting the model can talk itself past. Consequential actions are gated in code, and only a fresh, explicit human approval opens the gate.
Five moves. One decision.
Ask in your own words and your own language. No menus to learn.
“I need help applying for PM-KISAN.”
The AI doesn't get
the final say. You do.
Every consequential action arrives as a plain-language summary of exactly what will happen. Nothing moves until you choose. Declining is always free, always respected and never argued with.
- A. Specific — one action, one scope
- B. Informed — what, where, and why
- C. Revocable — before it executes
Interactive demo · nothing is submitted
PM-KISAN
Application
Steward prepared
- Applicant details
- Scheme information
- Required fields
Next action
Submit application
To pmkisan.gov.in on your behalf. This cannot be undone once sent.
Consent required
Trust isn't
a prompt.
It's architecture.
A model can be persuaded. A gate can't. Steward's actions pass through deterministic checks outside the model: no valid, current, unmodified consent — no action. Hover a case to inspect it.
Let AI do
the work.
Keep the
decision yours.
Say it the way
you'd say it to a person.
Voice-first, multilingual and forgiving. Steward listens for what you mean, shows you how it understood you, and lets you correct it before anything happens.
“I need help applying for this scheme. ”
- Voice input
- Intent
- Action
- Consent
Eight stations.
One gate you hold.
Reasoning, memory and tools on one side. Verification, execution and audit on the other. Hover a station to inspect how it behaves.
Voice or text, any supported language. The only party who can grant consent.
auth: device + OTP
Plans the task, chooses tools and drafts actions. Has no power to execute consequential steps.
agent · plan → act loop
Session-scoped profile, the open page and prior answers. Minimised and never sold.
ttl: session · PII masked
Read-only navigation, field reading and form drafting. Write tools are wrapped by the gate.
navigate · read · draft
Classifies each action. Anything consequential is paused and summarised for the human.
policy: deny-by-default
Verifies a signed, single-use consent token bound to the exact payload hash.
scope · ttl 120s · hash
Executes exactly what was approved — nothing more. Any change requires new consent.
idempotent · bounded
Append-only log of request, consent, decision and result. Visible to the citizen.
append-only · exportable
Not take the
decision away.
The last 4mm are always yours.
Let AI do the work.
Keep the decision yours.
Steward is a concept for consent-first public-service agents: helpful by default, accountable by design and never the final word.