Grain
Start a workflow

Grain · Governed AI execution

AI operations your team can trust to run.

Grain turns a request, files, and rules into a tracked workflow: it routes the right agent, runs the steps, pauses for decisions, and returns outputs with evidence.

Start a workflow See how it runs

Durable tasks · Governed agents · Human approvals · Evidence with every output

How it runs

From request to reviewed output.

Every job becomes a tracked workflow. Task state, approvals, and evidence stay attached from intake to delivery — nothing lives in a one-off chat.

01

Intake

The request, files, rules, and history become one durable task that outlives the session.

02

Route

A versioned agent profile brings the right skills, model settings, and tool grants to the job.

03

Execute

Steps run with checkpoints and retries. Failures recover from known-good state — they don't restart.

04

Approve

The run pauses for the decisions a person should make, and records each one.

05

Inspect

Evidence, safe previews, and output lineage — reviewed before anything ships.

What Grain hands back

A finished packet your team can review, send, or archive.

Each artifact carries its source refs, origin, content hash, checkpoint, and run metadata — the output is a packet, not a chat response.

Capabilities

Built for work your business repeats.

Each capability pairs the business outcome with the mechanism that makes it dependable.

Capture the request, files, and rules in one place.

Every job starts with what the business already knows — the ask, the attachments, the constraints, and the history around it.

MechanismDurable task intake stores files, context, and operator thread history on a task that persists — not a one-off chat that evaporates.
ExampleAn RFP lands with 48 pages, a pricing matrix, and one rule: pricing needs sign-off.

The right playbook for the job.

Repeatable work runs the way your team decided it should — the same skills, the same limits, every time.

MechanismRegistry-backed agent profiles define skills, tool grants, and model settings — versioned by revision and content hash, so what runs is exactly what was approved to run.
Examplerfp-response@12 can read the drive and CRM and write documents — and can never send email.

The steps get done — research, drafts, checks, packaging.

Grain moves the work across the stages a person would, without a person pushing each one.

MechanismThe execution harness runs agents, tools, and sub-agents with checkpoints, retries, leases, and recovery — long-running work stays alive for hours or days.
ExampleA CRM timeout retries and recovers; the run resumes from its last checkpoint, not from zero.

People approve the consequential decisions.

Grain does the legwork; your team makes the calls that matter — and every call lands on the record.

MechanismNeeds-input, approval, override, stop, resume, and reactivate are explicit workflow states — not chat messages that scroll away.
ExampleThe pricing section waits for Sarah. Nothing ships until she says so.

See what happened before you trust the result.

Any output traces back to the run that produced it — what it read, what it did, who approved what.

MechanismRun inspection exposes task state, journal evidence, safe previews, output metadata, and event provenance — end to end.
ExampleOpen the evidence drawer on any deliverable and walk the run from intake to delivery.

Review-ready deliverables, not chat responses.

Work comes back assembled: the documents, the numbers, and the record of how they were made.

MechanismOutput plugins produce packets with attribution, lineage, and artifact references — every file traceable to its sources and runs.
ExampleOne packet: the response, the approved pricing, the questionnaire answers, and the decision record.

Each correction improves the next run.

Fix something once and the fix sticks. The workflow gets more reliable the more you run it.

MechanismAccepted drafts, correction records, deterministic reruns, and lineage-aware routing reduce rework run over run.
ExampleMay's approved proposal language is reused in July — automatically, with attribution.

Audit and control around AI work.

Adopt AI execution in regulated work without giving up the controls your auditors expect.

MechanismPHI-safe runtime logs carry safe metadata only; detailed content lives in private journals, never in log lines.
ExampleA benefits case review runs end to end without member details ever landing in a log.

Use cases

Start with a workflow your team already runs.

Eight agent fleets, one anatomy: a ledger that cross-checks the sources, a deterministic gate against a written rulebook, a reviewer, and human-approved outbound. Click any fleet to see how it's implemented.

Healthcare

Patient healthcare-bill advocacy

Months-long cases from a shoebox of bills, EOBs, and collection letters — with appeal and charity-care clocks ticking.

Fleetadvocacy-case-agent · see the implementation →
 Franchising

Franchise agreement assembly

Deal documents become a state-compliant agreement packet — where one wrong rider or missed registration rule voids the deal.

Fleetfranchise-agreement-agent · see the implementation →
 Logistics

Detention & demurrage invoice disputes

Rebuild the free-time math and void defective container charges under the FMC billing rule — inside the 30-day window.

Fleetdnd-dispute-agent · see the implementation →
 Construction

Lien waiver exchange & deadline tracking

Statutory waivers chased every pay cycle, matched to the pay app — while per-state lien deadlines tick.

Fleetlien-rights-sentinel · see the implementation →
 Legal & privacy

Incident notification matrix

One discovery date against GDPR's 72 hours, HIPAA, 50 state statutes, and contractual notice clocks.

Fleetbreach-notice-agent · see the implementation →
 Food service

Supplier contract-price audit

300 invoice lines checked against the contracted price list, every week — inside the dispute window.

Fleetcase-price-audit · see the implementation →
 Equipment dealers

Warranty claim recovery

Every warranty repair filed correctly — right serial, covered parts, allowed hours — before the window closes.

Fleetwarranty-claim-desk · see the implementation →
 Consumer brands

Retailer fine recovery

A third of retailer compliance fines are wrong. Prove it — before the 30-to-90-day dispute deadline.

Fleetshipping-fine-recovery · see the implementation →

Trust & governance

Governed execution, not a black box.

The controls a CTO asks about, in terms a CEO can sign off on: every capability granted, every run evidenced, every output traceable. Safer than ad hoc AI use — by design.

Governed agents

Profiles, skills, tools, and models are versioned. What runs is exactly what was approved to run.

Durable execution

Checkpoints, retries, and recovery keep long-running work moving. Failures resume — they don't restart.

Human approval

Grain pauses for decisions before consequential actions. Every approval is on the record.

Inspectable output

Every deliverable connects back to its runs, artifacts, and lineage — reviewable before you rely on it.

PHI-safe operations

Runtime logs avoid prompts, raw content, stack traces, and sensitive text. Detail lives in private journals.

Get started

Run the work your team repeats every week.

Bring one recurring job. Grain will intake it, route it, run it, and hand back an output you can inspect.

Start a workflow Talk to us