AI project assistant

Assistance on every update task. Securely.

Progress, budget, RACI, risks, schedule, workload — whatever a human can update on a project, the assistant can propose too. Securely by construction, not by prompt: it's locked to this one project and to your real rights, in code, before it can ever touch your data — and even then, nothing is applied without your explicit validation.

How it works

Propose, then validate. Every time, no exceptions.

1. You ask

In plain language, in the project's own chat — including pasting raw meeting notes to extract action items from.

2. It proposes

A structured, reviewable proposal — checked against your live data for freshness before it's even shown to you.

3. You validate

Approve or reject, with your real RACI authorization re-checked at that exact moment — not when the chat started.

4. It's logged

Validated or rejected, every proposal lands in a permanent, human-readable changelog for the project.

Locked to this project — the model can never even ask about another one Locked to your real RACI rights — checked in the database, twice Re-verified fresh — stale data is refused, never silently overwritten

From notes to tracked work

Paste your meeting notes. Get back tracked proposals.

Drop in raw notes or a call transcript — the assistant merges the update into the project's tracking note and drafts every concrete action, progress change and risk it can find in it. Nothing invented, nothing applied until you say so.

Pasted into the chat

Kickoff sync — Sept 12

- Design mockups are basically done, client signed off yesterday
- Still need to follow up with the hosting vendor about their SLA — Priya, ideally by Friday
- Security review hasn't started yet, blocked on access — flagging as a real risk now, not just a delay
- Sprint 2 starts next week, same scope as planned

PROGRESS

Update 'Design mockups' progress to 100%

ACTION

New action: follow up with the hosting vendor about their SLA — owner Priya, due 2026-09-18

RISK

New risk on 'Security review': "Blocked on access" — probability: high, impact: medium

NOTE

Updated tracking note on the Kickoff workpackage, merging in today's update

Exact scope

What it can see, and what it can touch — no more, no less.

The full list, not a marketing summary — 9 read tools, 37 propose tools. Full capability parity with a human user: whatever you can do in the app on a project, the assistant can propose too.

Can read (9)

Project overviewFull WBSRecent actionsBudgetRACINotesScheduleCharge / workloadRisks

Can propose (37)

Update deliverable progressCreate a follow-up actionUpdate a follow-up actionToggle an action done/openDelete a follow-up actionCreate a riskUpdate a riskToggle a risk open/closedDelete a riskCreate a budget lineUpdate a budget lineDelete a budget lineCreate a budget paymentUpdate a budget paymentMark a payment paidUn-mark a payment paidDelete a budget paymentAssign RACIUnassign RACIAdd a tracking noteCreate an iterationRename an iterationToggle an iteration activeAdd a catalog workpackageToggle a workpackage activeAdd a catalog deliverableToggle a deliverable activeUpdate the project itselfUpdate the scheduleAdd a predecessorRemove a predecessorUpdate a predecessorSet planned timeLog actual timeUpdate actual timeDelete actual time entryFully unassign someone from an entity

Full risk lifecycle

Create, update, resolve, and remove risks scored on a real probability × impact matrix — not a single guessed severity pick. Same propose-then-validate discipline as everything else. Nothing is silently skipped: if a human can do it on a project, the assistant can propose it too.

Two providers, your choice

Anthropic Claude and OpenAI GPT, independently configurable per environment — each with its own model and its own encrypted API key.

Keys encrypted at rest

AES-256-GCM. The decrypted key is never returned to a view or a JSON response — used only for the outbound call itself.

A monthly spend cap

Set a configurable budget per environment. Once reached, the assistant stops — clearly, not with a generic error.

Freshness-checked proposals

If the underlying data changed since a proposal was created, it's re-verified before being applied — never applied blind.

Authorization re-checked twice

Once when a proposal is created, again when it's validated — roles can change in between, and both moments matter.

A permanent changelog

Every validated change is recorded — a real audit trail of what the assistant did, and who approved it.

See it on your own portfolio

Time savings, without ever losing control of your data.

Get in touch