Example
The live map
Your operation as a living map: every department, agent and workflow, with work moving between them and the places that need you lit up.
The operating layer for AI work.
One platform runs everything we build, wherever it is deployed: the agents that judge, the workflows that execute, the rules that compute, the gates where you decide and the record of all of it. You follow it live in the Dephrast app.
Six kinds of piece. One grammar.
Every system we build is made of the same six kinds of piece, drawn the same way everywhere: in the app, on the map and on this page. Once you can read one department, you can read them all.
- Judges
Agent
Reads and interprets: documents, emails, requests. Calls a model through one gateway, with only the tools it is given.
- Executes
Workflow
Durable steps that survive a failure and resume where they stopped, instead of starting again.
- Computes
Rules
Code for anything with a right answer: sums, dates, tax, policy. The same input always gives the same answer.
- You decide
Gate
The place a person approves or rejects, with the evidence beside the decision. The only thing drawn in crimson.
- Reaches out
Port
The one way anything leaves: a send, a post, a posting. A port refuses, by name, until its conditions are met.
- Remembers
Record
Append-only history of what happened, who decided and what it cost. Nothing is edited after the fact.
Everything a production AI system needs, built in.
Build
- AgentsTyped modules, each with one job, a declared set of tools and a budget.
- OrchestrationDurable workflows and schedules, with retries on the steps that are safe to repeat.
- IntegrationsYour CRM, accounting, inboxes and documents, reached through their own APIs and through ports.
- Model routingAgents ask for a capability; one gateway chooses the model and records every call.
Control
- PermissionsEach agent can use only the tools it is given, enforced when it runs, not written in a document.
- RulesPolicy and arithmetic in code, so a check can be read, tested and trusted.
- ApprovalsOne inbox of decisions from every department. No button decides anything unread.
- BudgetsA spending limit per agent, checked every time it runs.
See
- ObservabilityEach run traced step by step. Secrets are redacted before a trace is written.
- Audit trailRuns, steps, decisions and model calls on an append-only record.
- ReplayPlay back the last 24 hours of work across the map, or reproduce a decision from what it read.
- Cost trackingEvery model call priced, and totalled per run, per agent and per department.
What you see in the Dephrast app.
Dephrast configures and runs the system; you watch it work and make the decisions. These are the screens you do it in.
Example
One inbox of decisions
Every approval from every department, each opening onto its evidence. No button decides anything unread.
Example
Every run on the record
Each run recorded step by step: what it did, how long each step took and what it cost.
Example
The morning briefing
What happened since yesterday, what needs you today, and what it cost, on one page.
Example
Budgets and permissions, enforced
Each agent may use only the tools it is given, within a budget checked every time it runs.
Example
Replay the day
Scrub back through the last 24 hours of work across the whole map.
Product screens show example data.
One view of the whole operation.
Above the departments sits an orchestrator: it tracks what each one is doing, sequences work between them, and briefs you each morning on what happened, what needs you and what it cost.
- The map
- Every department, agent and workflow, with its live state.
- The inbox
- Every decision waiting for you, from every department.
- The briefing
- Since yesterday: what ran, what needs you, what it cost.
- Replay
- The last 24 hours of work, played back across the map.
An example of the Dephrast app: an activity feed of agents working across departments, and a decision waiting for a person to approve it.
Your stack stays. Dephrast connects the work.
Your CRM, accounting system, inboxes, calendars and databases stay where they are. We connect agents to them through their own APIs, so there is nothing to migrate and no new seats to buy.
- Slack
- Google Workspace
- Notion
- Asana
- Jira
- Salesforce
- HubSpot
- Zoho
- Zendesk
- Intercom
- Airtable
- Dropbox
- Stripe
- Xero
- QuickBooks
- SAP
- Shopify
- Snowflake
- Twilio
- Zapier
And anything else with an API, a database or an inbox.
Names and logos belong to their owners. They show the kinds of tools we connect to, not partnerships.
Not a chatbot. Not a zap. Not a project that ends.
Each of these is good at what it is for. Dephrast is for a different job: running real business processes, with controls, on your tools, and keeping them running.
AI assistants
- Good for
- Helping one person think, write and search.
- Where Dephrast is different
- Dephrast runs a process end to end, inside your tools, with approvals and a record, not a conversation someone has to have every time.
Workflow tools
- Good for
- Moving data between apps when the rules are simple.
- Where Dephrast is different
- Dephrast handles the judgement in between, such as reading documents, deciding and drafting, with rules in code and a person at the gate.
Automation agencies
- Good for
- Building a one-off automation.
- Where Dephrast is different
- Dephrast combines agents, deterministic rules, approvals and observability on one platform, and operates it after launch day.
The questions operators ask first.
Do we have to change our tools?
No. That is the point. We build against the CRM, accounting system, inboxes and spreadsheets you already run, through their own APIs. If something genuinely cannot be automated where it stands, we tell you before you pay for it.
What is an AI operating system?
It is what an AI department becomes when every department works together. One orchestrator sequences work between them, you get one morning briefing and one inbox of decisions, and there is one record of everything that ran. Most businesses reach it in stages, starting with one automation.
Why not just use ChatGPT, Zapier or Make?
For a single task, you might. What we add is everything those tools leave to you: finding the work worth automating, rules in code for anything with a right answer, approvals with the evidence beside them, one record of everything that ran, and a team that keeps it running.
Will the AI act on its own?
Only where you have allowed it, and only once its record shows it can. Every agent starts supervised; anything that leaves your business starts as a draft or a proposal a person approves. Running unattended is earned task by task, on your own records.
Where does it run?
Wherever suits you. We can host and run it for you, deploy it into your own cloud account, or run it on your own infrastructure. Where it runs, what it may reach and which services process your data are agreed in writing before we build. Each client is its own tenant, and Dephrast does not train models on your data.
See it on your own process.
Bring us one workflow. We'll show you how it would run on the platform: where the agents sit, where the rules compute and where you decide.
Free, 30 minutes, and a reply from a person within one working day.