01 · Where work happens
Decisions happen out loud.
The launch moved on a call. The scope changed in a thread. By the time any tool hears about it, it has already happened — in conversation.
Capabilities · Provectus' company AI · In production
An AI for the company itself. Not another copilot.
Barley is the intelligence hub Provectus runs on. It joins our meetings and reads our tickets, docs, deals, and Slack. Records stay updated and questions come back with sources — and workflows fire without waiting for anyone to type it in.
The ceiling
Copilots make people faster. The company still runs on data entry.
02 · What tools know
Systems of record are second-hand.
Jira, the CRM, the status deck: none of them witness anything. They know what someone typed, after someone typed it. Every record is a copy of a conversation.
03 · Why automation waits
So "automation" waits on a human.
A workflow built on those tools fires only after the data entry is done. The human is the sensor — and the bottleneck. Automation that waits on a person is just enablement.
Your tools don't know what happened. They know what someone typed.
The mechanism
One sentence. Every system downstream.
A worked example: one decision on a client call, and everything that has to change because of it. Today, every arrow below is a person remembering to type.
# call · Acme Corp · quarterly review · Meet
client: "…then let's move the launch to March 30 —
and pull the CSV export out of this release. Ship it in the next one."
client: "…then let's move the launch to March 30 —
and pull the CSV export out of this release. Ship it in the next one."
Barley is on the call
It resolves who
Acme Corp — the account in the CRM, the deal attached to it.
It resolves what
Acme Launch — the project, its epic and stories in Jira.
It extracts the decisions
Two of them: a date moved, and a scope cut.
and the representations follow
Jira · the plan
So the board says: launch epic due Mar 30; CSV-export story moved to the next release — with a comment linking back to the recap.
before: whoever remembers, after the call
CRM · the deal
So the deal record says: timeline shifted at the customer's request — noted on the Acme account, visible to whoever picks it up next.
before: the account exec's Friday admin
Status report · the narrative
So this week's report says: launch re-planned to Mar 30, scope trimmed by one item — assembled, not authored.
before: Sunday-night doc archaeology
The team · the broadcast
So the project channel hears: what moved, and what it means for QA — before anyone asks "wait, when did we decide that?"
before: hallway rumor, secondhand
And the obvious question — who gets to hear all this? — is answered at the source: Barley inherits permissions from the systems it connects. A private channel stays private; a restricted deal stays restricted. Everyone sees exactly what they could already see, just without the hunting.
The proof
An ordinary day on Barley.
The loop Provectus itself runs on, shown as one day in five moments.
- Morning a client call
Barley joins the Meet alongside the team and comes back with more than a transcript: the decisions, their owners, and a recap anyone can forward.
# recap — Acme quarterly review
decisions launch → Mar 30 · CSV export → next release
owners re-plan: Dana · client comms: Arif
next recap → project channel · tickets suggested ↓ - After the call tickets, suggested
From the recap, Barley drafts the Jira changes. A person approves; Barley files them: dates, assignees, comments linking back to where it was decided.
# suggested from recap
→ move launch epic due date to Mar 30 [approve]
→ new story: ship CSV export in next release [approve] - Midday a question in Slack
"What's happening with Acme?" One answer, assembled from the recent calls, open tickets, latest docs, and the deal, with every claim linked back to its source and tagged by how fresh it is.
- Release day the announcement writes itself
A release goes out; the release-notes scenario posts what shipped and who it matters to. Nobody compiles it.
- Friday status, assembled
The weekly status report fires on schedule, built from the week's calls, tickets, and docs instead of memory. The RAID log updates the same way.
We run our own company on this loop. It's the same one we build for clients.
The keystone
Nobody types what was already said.
Smarter chat is a side benefit. The point is systems of record that stay current as a side effect of the work itself.
Records keep themselves
Tickets, status, logs stay current because the conversation is the trigger. No one has to remember.
Answers carry receipts
Barley navigates to the sources rather than replacing them: every answer links back to the call, ticket, or doc it came from.
Workflows don't wait
Scenarios fire on events: a release, a decision, a Friday. Humans approve wherever it writes.
Enablement makes people faster. Automation makes work happen.
Under the hood
The machinery, for the technical eye.
What a platform team asks before trusting a company AI — answered the way we built ours.
01 · Retrieval
Vector search plus a knowledge graph
Hybrid RAG: semantic search with a knowledge graph beside it, so relationship questions like "who worked on projects using Bedrock?" resolve by walking real edges instead of model guesswork.
02 · Connectors
One brain over every system
Meetings across Zoom, Meet, and Teams; Slack; Jira read and write; Confluence and Drive; the CRM; the people directory; GitHub and GitLab. Read everywhere; write where the workflow needs it.
03 · Permissions
Permissions come from your systems
Every query runs as you: what's restricted in the source system stays restricted in the answer. Responsible AI enforced by architecture instead of a policy document.
04 · Entity resolution
It knows which "Acme" you mean
Typo-tolerant, LLM-ranked resolution scopes every question to the right project, company, or deal before retrieval even starts.
05 · Scenarios
Workflows fire on events
The automation layer: workflows fire on events in the stream (a release, a decision, a date), with human approval gates wherever they write to your systems.
06 · Surfaces
Wherever you already work
A web UI built calls-first, and an MCP server that puts the same brain inside Claude Code, Cursor, Claude Desktop, or ChatGPT. Model routing matches cost to task.
Build yours
Yours won't be Barley.
Barley isn't a product in a box; it's our reference architecture for a company AI. The retrieval, permission, and workflow machinery transfer; your stack, your systems, and your access rules stay yours. Blueprints are the entry point: one business process, automated end to end.
Blueprint 01 · CRM Flow
The record keeps itself.
Sales teams don't hate the CRM; they hate feeding it. CRM Flow turns every call, thread, and meeting a deal generates into the record itself: fields current, notes filed, next steps tracked.
Blueprint 02+ · Next in the series
The next ones are already running.
Status reporting, meeting-to-ticket, RAID logs: the next blueprints are drawn from workflows Barley already runs internally, packaged one business problem at a time.
Beyond the series
Bring the process nobody wants to own.
A record-keeping ritual your team hates is a candidate. We scope a custom build on the same core — your conversations in, your systems kept current.
The first blueprint automates a process. Every next one inherits the platform.
Give the company an AI.
Start with CRM Flow: one process, end to end, on your stack. Or bring the workflow your team hates most, and we'll scope it against the same core.