The company brain that remembers decisions

Ravi, IT admin. Ten minutes.

Verity reads the company's sign-in system and finds the 38 apps Northwind actually uses, ranked by how sure it is. It connects to none of them. Each one becomes a proposal Ravi approves, with the exact permissions listed, and every connection starts read-only.

Behind the list is a catalogue with three trust tiers: vetted managed connectors, official vendor servers, and community ones that stay off unless an admin allows a specific one with a reason.

Priya. One click on a consent screen, then nothing to do.

Verity pulls from the five tools, works out that the Acme in Salesforce, the Acme in Zendesk and the Acme paying invoices in Stripe are the same customer, and builds one record. Each fact carries where it came from, when the tool last confirmed it, and a grade: a number from Salesforce is authoritative, an email is observed, a Slack opinion is asserted, a guess by an enrichment vendor is only a hypothesis.

Permissions come with the data. Verity mirrors what each tool lets Priya see, so she sees her accounts and her colleague sees his.

Priya. Thirty seconds, coffee in hand.

The brief is written only from facts in the ledger; every number has a footnote you can click. The signals underneath are produced by plain rules — an overdue invoice plus open P1 tickets plus a pricing question is a renewal at risk — and the score shows its own arithmetic.

Priya. One question instead of five tabs.

The answer is written from the ledger and every sentence carries a footnote to the exact fact, with its grade and how fresh it is. When the ledger does not have something, Verity says so and says why — here, that call recordings are not connected — rather than guessing.

Her colleague Sam asks the same question and gets nothing about Acme, because Salesforce does not share that account with him. The answer is filtered before it is written, not after.

Priya. One sentence in plain English.

Verity turns the sentence into a plan of three steps and files each as a proposal — a preview of exactly what would change, checked against the company's rules. The date change passes. The fee waiver needs Dev because policy says so. The email is allowed because Jordan wrote first, but one check comes back unknown: nobody knows Jordan's timezone, so Verity will not quietly assume.

Unknown is treated like a stop, not a pass. That single design choice is where most agent products go wrong.

Dev, from his phone. Then something happens that no one planned.

Dev approves the waiver in Slack. Three minutes later Acme pays the invoice. When Verity goes to execute, it checks the facts one last time, sees the invoice is no longer open, and refuses: the approval was for a world that no longer exists. Priya rebases the step, sees the waiver is now pointless, and discards it. The other two steps go through, and each is read back from the tool to confirm the write landed.

The email step needed a reason for the unknown timezone; Priya gave one and it is on the record with her name.

Priya. One click and one reason.

Undo is not a delete. It is a new proposal that restores the "before" picture from the receipt, and it goes through the same rules and the same freshness check as any other change. If someone had moved the date again since Monday, the undo would have stopped and shown the difference, exactly like scene 6.

Priya, twenty minutes. No engineer.

She describes what the agent may and may not do, how sure the facts must be before it acts, how much it may spend, and which model version it uses. Then she rehearses it against a copy of the last month with fake versions of the real tools. Only after she has seen the 14 proposals it would have made, and the 2 it would have been blocked on, does she publish.

From then on the Watcher proposes; it never executes. Its proposals land in the same Approvals inbox, checked by the same rules, receipted the same way.

Meera. Before anyone at Acme noticed anything.

The Doctor gives every agent a health score from its own receipts. When the Watcher's score falls, it is quarantined automatically: it can still read, it can no longer propose, and its open proposals are put on hold. Then the Doctor walks the parts lists of its recent receipts and finds the one thing that changed — someone moved the model pin — and writes the fixes as proposals for Meera to approve. Nothing is fixed behind her back.

Meera. The screen the auditor asks for.

Every agent that can propose a change is registered with an owner and its permissions — including the ones Northwind built in Salesforce Agentforce and the assistant it runs in Glean, because both connect to the company's systems through Verity. Rules and contracts have version histories with effective dates. The coverage number is generated from configuration and names the write paths that still bypass Verity, so nobody claims control they do not have.

One button produces the evidence pack for the quarter; another freezes the whole company into read-only during an incident. Privacy is built in for India's data law: every field has a class, consent is a fact, and erasure destroys the key rather than the history.

Meera and finance. An afternoon instead of a quarter.

Because every receipt kept its parts list, Verity can look up every decision that depended on discount policy v4, replay each one under the corrected rule using the facts as they were at the time, and sort them: unaffected, correctable, needs a human. The correctable ones become compensating proposals, approved in one batch by their owners, executed through the same checks as everything else. The pattern becomes a defect that other companies can screen for without seeing Northwind's data.