Bolt by Sparcle · Sovereign AI for the enterprise

Governance, Data Sovereignty & Compliance

Put AI into every workflow, without sending your data anywhere you don't control and can't see.

View this as a deck Download PDF All roles Empower · Govern · Audit, for the people who have to answer for it.

On this page

Why Sparcle exists The forcing function The market's blind spot Why our mantra is three verbs 01 · Govern · Boundary Govern · Masking · The flagship control Govern · Access, capability & oversight Govern · Agent Gateway · The shadow-AI answer 02 · Comply · Control mapping Comply · Regulation packs Comply · India DPDP Act Comply · Data subject rights 03 · Data sovereignty · Topology Sovereignty · Encryption Sovereignty · The safety net 04 · Audit · The proof Audit · Accountability by design Audit · Transparency The honest posture Every agent path New: the governed browser The ask · Empower · Govern · Audit

Why Sparcle exists

The crux

Your people are already using AI on your data. The only question worth asking is whether that use is one you can control, one you can see, and one you can prove. Sparcle exists to make the answer yes.

The forcing function

AI adoption is outrunning AI control

The productivity is real and unstoppable. The governance to match it is not in place, and the liability still sits with you.

Shadow AI is already inside

45% of employees now use AI on a corporate device, up from 15% a year earlier, and two thirds of that runs through personal accounts no control can see. Shadow AI is the third most common non-malicious insider action in the 2026 DBIR.

Controls are missing where it matters

One in four malicious breaches is now AI-enabled, a 56% rise in a year, costing about $6M against a $4.99M average. More than 1 in 5 organizations reported a breach that targeted an AI model or application.

The accountability is yours

Regulators, boards and customers hold the organization responsible, not the model vendor. 74% of leaders believe they would pass an AI compliance audit; 27% have a mature programme, and 36% of boards discuss third-party AI risk at all.

The market's blind spot

Most AI governance is documented, not enforced

The tools on the market split into two camps, and both leave the same gap open.

Paper, not a control

Policy packs, AI registries, model cards and risk registers describe intent. They sit beside the data flow, not in it. Nothing stops the next prompt.

Enforced, but outside your walls

Cloud proxies and cloud assistants that do enforce inspect your prompts inside their cloud, or only within one vendor's ecosystem. Your data still leaves the boundary you control.

The empty square

No dominant player enforces in the data path, inside your perimeter, in front of any model you choose. That gap is exactly where Bolt sits.

Why our mantra is three verbs

One operating model: Empower, Govern, Audit

Each verb protects the other two. Govern without Empower gets bypassed by shadow AI. Empower without Audit is indefensible. Together they are how responsible AI actually holds.

Empower

Give the whole workforce real AI leverage, across every system, so no one needs an unsanctioned tool. Most everyday work resolves on-device, at zero tokens.

Govern

Every use crosses a boundary you set: PII masked before the model, access scoped to identity, capabilities and models under admin control.

Audit

Agent turns, tool calls, approvals, masking and policy decisions, every search, and the security-relevant identity transitions are recorded in a tamper-evident log you can verify offline. You can answer "what did AI touch?" the moment you are asked.

01 · Govern · Boundary

You define the boundary. Bolt enforces it.

Air-gap, a private model endpoint, or a public model with masking on. You pick the posture per deployment. Sparcle is never in the loop.

Always on

Runs in your infrastructure

Helm chart or Docker in your Kubernetes cluster, VPC, on-prem server, or air-gapped network. Lives entirely inside your perimeter.

Always on · Privacy floor: admin opt-in

Your LLM, your terms

Bring any model: a private endpoint on your GPUs, a local open model, or a public API. Admins can set an org privacy floor that refuses any model whose endpoint sits below it.

Always on

Your keys, no lock-in

Bring your own keys, vaulted with per-tenant encryption. Swap models or providers without re-platforming.

Govern · Masking · The flagship control

Sensitive data is masked before it reaches the model

One chokepoint on the LLM boundary, always on for structured identifiers, independent of which model you picked. Personal and health identifiers are detected, masked and pseudonymized before any prompt leaves your boundary, across streaming responses and text inside images, then restored on-device in the answer. Enforced in the path, not written in a policy.

1 · Identify

Detect personal and health identifiers in the prompt, contacts, financial and national IDs, validated with real check-digit algorithms so look-alikes are not flagged. Text inside attached images is scanned by OCR through the same detectors.

2 · Mask and pseudonymize

Replace each identifier with a stable, reversible pseudonym before the prompt leaves your boundary. The model reasons over the request, and on an unknown checksum the pipeline fails closed.

3 · Restore on device

Map the pseudonyms back to the real values inside your perimeter when the answer returns. The real values stay inside the boundary you control.
Always on Same safety floor on any model, public or private Free-text name detection: admin-enabled NER

Govern · Access, capability & oversight

Policy and oversight, not trust

Who can use which systems, models and capabilities is set by your directory, and the consequential actions wait for a human. Nothing is left to the honor system.

Admin opt-in

Directory-driven RBAC

Access scoped by role, department and group from your IdP over SCIM. Every action reaches each system as the user's own identity, never a shared bot.

Admin opt-in

Capability lockdown

Enable, disable or restrict capabilities by role or group, connectors, models, clipboard, screen capture, with a kill-switch on each.

Admin opt-in

Versioned policy

Capture, simulate against your user population, activate, and roll back by pointer. Change the rules without fear.

Always on

Human approvals

Send, delete, schedule and pay wait for explicit human sign-off, a durable request that survives a restart, so nothing consequential auto-fires.

Govern · Agent Gateway · The shadow-AI answer

Govern the AI your people already use

The shadow-AI problem needs an answer, not just a policy. Bolt puts the traffic from the AI tools your people already reach for through the same masking, policy and audit boundary, so their convenience stops being your exposure.

Admin opt-in · Claude today

A governed relay

Third-party AI clients route through Bolt. Prompts, tool calls and image content are masked, checked against policy, and logged before anything leaves your boundary.

Warn by default · block by policy

On-device AI-DLP

On paste, or as your people type into public AI sites, sensitive data is detected and redacted in place, and can be blocked outright under a managed policy.

Always on

The same evidence

Every mask, allow and deny on that traffic seals into the same audit chain. Shadow AI becomes governed AI you can prove.

02 · Comply · Control mapping

The same controls, mapped to every framework

Buyers now expect this mapping as table stakes. The differentiator is that on Bolt these are running controls, not documented intentions.

Comply · Regulation packs

One engine, any jurisdiction

Regulations ship as swappable packs. Packs with identifier recognizers feed the same masking chokepoint. A new country is one pack, not a new product.

Always on

Built-in coverage

29 packs: 22 jurisdictions (GDPR, UK GDPR, DPDP, CCPA, LGPD, PIPL, APPI, PIPA and more) plus 7 cores. Each declares its identifiers, rights and residency; the dashboard shows the rights, residency and safeguards.

Always on

Real identifier validation

National IDs are checked with their true check-digit algorithms (Aadhaar, PAN, GSTIN, NHS, Singapore NRIC, Australia TFN, Brazil CPF and CNPJ, Japan MyNumber, Korea RRN and more), so look-alike strings do not trigger false masking.

Always on

Signed and additive

Built-in packs ship inside the signed Bolt app. Custom packs install as signed.boltpack bundles, reviewed on install, activated per organization. Author a new jurisdiction without touching the engine.

Comply · India DPDP Act

Govern the point of use, the leg no consent manager can see

Consent is captured at your front door by your consent manager. Bolt sits downstream, governing the internal leg where your people and your AI actually touch that data.

The blind spot today

The internal AI leg

  • Consent managers and CRMs record that consent exists
  • They are not in the path when an employee runs AI on the data
  • No CMP, DLP or CRM can prove that leg honored the purpose

What Bolt adds

Enforcement plus proof, in your perimeter

  • Purpose enforced at the point of use, derived from context and policy
  • PII masked before any model, on-prem or BYO-LLM
  • Signed record that your people and AI honored consent
Masking, audit, residency: Always on Consent-linkage purpose enforcement: Admin opt-in Live CMP vendor wiring: in progress

Comply · Data subject rights

Rights work, on a clock you keep

Erasure, retention and key destruction are wired as running jobs. The dashboard shows the rights each active regulation declares.

Always on

Erasure and DSAR

A typed per-person cascade erases a subject's account, sessions, credentials and pseudonym bindings (and chat history on the grace-period path), tenant-scoped so one erasure never crosses into another tenant, and revokes the user's upstream OAuth grants so a deleted user keeps no live access.

Always on

Retention sweepers

Messages, sessions, runs and audit entries age out on the retention window you set, enforced daily, not left to manual cleanup.

Per organization · admin action

Crypto-shred

Destroy a tenant's key and its data becomes unrecoverable, the strongest form of erasure. The shred is irreversible and guarded by an explicit org-id confirmation. A disposal receipt is sealed into the audit chain on every attempt, and the shred is refused outright if that receipt could not be written.

03 · Data sovereignty · Topology

Where your data lives

The workforce, the engine, the masking, and the store all sit inside a boundary you own. Only masked prompts cross to the model you selected. Sparcle has no server in the path.

  1. Your workforce

    on their desktops

  2. Bolt engine

    Helm / Docker

  3. Masking

    PII removed

  4. Encrypted store

    your keys

  5. LLM you chose

    masked only
Your infrastructure VPC / on-prem / air-gap

Sovereignty · Encryption

Encryption and key control, held by you

Stored secrets and indexed content are encrypted under per-tenant keys you control, so possession of the disk is not possession of that data.

On when a KMS is configured · warns if it fails

Per-tenant keys, your KMS

Envelope encryption with a customer master key per tenant, backed by a local KMS, AWS KMS or HashiCorp Vault. Bring your own keys.

Always on

Token vault

OAuth and model credentials sealed with authenticated encryption, bound so a token cannot be lifted and replayed for another user.

Per organization · admin action

Crypto-shred and rotation

Rotate keys on schedule; shred a tenant key to make its data unrecoverable. Rotation is non-destructive: keys already wrapped keep decrypting under the old key id while new writes use the new one. Both are sealed into the audit chain, and a shred is refused outright if its disposal receipt could not be written.

Sovereignty · The safety net

Even the most permissive choice stays safe

The question every CISO asks: "if my analyst picks a public model, does sensitive data leak?" The answer is no, because masking happens before the call, regardless of the model.

Consumer AI

Convenience, no floor

  • Raw PII typed straight into the model
  • No record of the transfer, no way to mask
  • Data retained and processed outside your reach

Bolt, any model

Convenience with a floor

  • Identifiers pseudonymized before the prompt leaves
  • The model reasons over tokens, never the real values
  • Real values restored on-device, and the event is logged
A privacy floor no consumer AI or copilot can match

04 · Audit · The proof

A record you can verify, without trusting us

Most tools give you a log you have to take on faith. Bolt gives you a cryptographic proof of integrity that a regulator can check with a standalone verifier, needing neither our database nor our keys.

Always on

Tamper-evident chain

Recorded events are hashed into a Merkle tree and the root is signed by your key, so altering any past entry breaks the proof. The chain covers agent and tool activity, approvals, masking and authorization decisions, search, admin policy changes, the security-relevant identity transitions, and the privileged key and identity operations. The audit-seal operation itself is not in it yet.

Always on

Offline-verifiable export

Export an evidence bundle that verifies on its own. No live connection to Sparcle, no dependence on our infrastructure to prove it.

Always on

The differentiator

Competitors hand you a log. Bolt hands you proof. For a compliance buyer, that is the difference between an assertion and evidence.

Audit · Accountability by design

Accountability at every layer

Governance is not one gate. It is a stack of controls, each enforcing something and each emitting evidence, so no path can skip a rule and every question has an answer.

Who did it

Identity

Every action runs as the user's own identity, via your IdP and SCIM.

Where it went

Boundary

Data moves only to endpoints you designate. Air-gap capable.

What the model saw

Masking

PII and PHI pseudonymized before any model call, on any model.

What was allowed

Policy

RBAC and per-capability rules from your directory attributes.

Who authorized it

Approval

Durable human sign-off on destructive or outward-facing actions.

Proof you can hand over

Audit

A Merkle-sealed, offline-verifiable record of the boundary, masking, policy and approval decisions above.

Audit · Transparency

The evidence you hand an auditor or a board

Transparency is not a dashboard you hope is right. It is a record you can produce, filtered and signed, the moment you are asked.

Always on

Every request logged

Method, path, user, status, duration and correlation IDs on every API call, with credential paths redacted.

Always on

Evidence export

Export the tamper-evident audit history for the time range you pick, verifiable offline, ready for an assessment or inquiry. It is not split by regulation.

Always on

Answer the hard question

"What did our AI touch, for whom, and where did it go?" is a query, not a research project.

Always on

Know your coverage

An org-wide read of whether masking and policy cover every surface, and whether any config has gone stale, so assurance is measured, not assumed.

The honest posture

What you can attest to today

We don't sell you a certification. We give you the running controls that get you certified, and we're precise about what's always on, what your admin turns on, and what's still roadmap.

Always on

  • PII / PHI masking, text and images
  • Merkle-sealed, offline-verifiable audit
  • Per-tenant encryption, BYOK, shred
  • Sealed deletion receipts, OAuth revoke
  • Regulation packs with real ID checksums
  • Durable human approvals
  • BYO-LLM, in-perimeter deployment

Admin opt-in

  • Directory-driven RBAC
  • Capability lockdown by role and group
  • Versioned policy: simulate and rollback
  • Command and terminal governance
  • Agent Gateway for third-party AI (Claude)
  • Model privacy-floor enforcement
  • In-app egress allow-listing
  • DPDP consent-linkage enforcement

Design partner roadmap

  • DSR intake console with SLA clocks
  • Live consent-manager vendor wiring
  • Independent read-only auditor role

Every agent path

One policy, including the agent CLIs.

The same governance that covers clipboard and browser now covers Claude Code, Gemini CLI and Codex, so a coding agent is not a hole in the policy.

New: the governed browser

The browser Bolt ships, and the check on the way out

Most AI exposure happens in a browser tab. Bolt ships its own, so what a person types or pastes into a page is screened on the machine before it is allowed to leave it.

A normal browser

Sees the paste, says nothing

  • A customer list reaches the AI site intact
  • The extension is the only control, and only where it is installed
  • No record of what was sent, or why
  • Every AI-enabled extension holds its own permissions

Bolt's browser

Names it, holds it, records it

  • What was found, grouped by type with a count, without repeating the values back
  • The page is held while the person decides, and a written reason is required to proceed
  • The reason lands in the audit record as a durable incident, not a counter
  • The tab says Guarded only when the guard reported, and Not screened when it cannot

The ask · Empower · Govern · Audit

AI you can put your name behind.

Enforced, in your perimeter, in front of any model. Empower the whole workforce, govern every use at a boundary you control, and audit it with proof you can hand over, without surrendering control or visibility to a cloud provider.

3

verbs, one operating model

0

data paths back to Sparcle

bolt@sparcle.app

start the conversation

Every other seat

Same platform, argued from where you sit

Each role page makes the case in that role's own language, and links back here.

Overview · All roles Decision Makers · Board CISO · Security CIO · IT CFO · Finance Compliance · DPO General Counsel CHRO · People VP Sales · CRO Platform · Integrations Developer Product Manager Finance · Ops Security · SOC Account Executive Support · CS Everyday

See it on your own data

Bolt runs inside your perimeter, on the LLM you choose.

Download Bolt Book a demo