Skip to content
Back to Blog
AI & Governance

An Assistant That Asks Before It Acts

Shi YanJuly 21, 20269 min read

There are two kinds of AI assistant in business software, and both are a poor fit for governance.

The first is a chat box bolted to the side of the product. It answers general questions well and knows nothing about the finding you have open. You end up pasting your own data into it, which defeats the purpose and, in a regulated organisation, may breach a policy.

The second is an agent. It is given a goal and a set of tools and left to get on with it. That is impressive in a demo and alarming in a system of record. If an agent closes a finding, who closed it? If it changes a severity from high to medium at two in the morning, where is the decision?

Architecture governance exists to make decisions visible and attributable. An assistant inside a governance tool has to respect that, or it undermines the thing the tool is for.

This post describes how the assistant in cajeX is built, and why it works the way it does.


Three Rules

We set three rules before writing any of it.

It sees what you see
Never more. Your permissions are its permissions.
It proposes
Every change is shown to you first, in full.
You confirm
Nothing happens until you say so, and the record shows you did.

Figure 1 — The three rules the cajeX assistant is built on.

It sees what you see. The assistant reads your workspace through your own access. If a project is confidential and you are not on its team, the assistant cannot see it either, and will not mention it. Details about members and the plan are for admins, so only an admin's assistant knows them. There is no privileged back door that makes the assistant better informed than the person asking.

It proposes. When you ask for a change, the assistant does not make it. It shows a card describing exactly what would happen: this finding would move from Open to In Progress; this due date would become 30 September; this description would change from the text on the left to the text on the right.

You confirm. The card has a Confirm button. Pressing it runs the same action the ordinary buttons in the product run, with your permissions, recorded in the same audit trail under your name. The assistant is a faster way to reach a control you already had. It is not a second, quieter way to change the record.


What It Knows

An assistant is only useful if it is grounded in something. Ours is grounded in three things.

How cajeX works
The product guide

Answers cite the chapter they came from, so you can read further.

↓
What is on screen
The open record

A finding, a project, a directive, a knowledge base entry, a diagram.

↓
The workspace
Your live data

Frameworks, dashboard totals, applications, sessions, recent documents, your own open findings.

Figure 2 — Three layers of grounding, each limited to what the person asking may see.

The first layer is the product itself. Ask how AI review scope works, or what happens to a finding when its project completes, and the answer comes from the product guide and names the chapter. This sounds modest. In practice it removes most of the questions that used to go to whoever set the workspace up.

The second layer is whatever you have open. Open a finding and the assistant can read it, along with the directive it is meant to satisfy. Open a directive and it knows the findings raised against that directive across the projects you can access. Its suggested questions change with the record: "What is blocking closure?" on a project, "Are the security zones sound?" on a diagram, "Which of our directives implement it?" on a framework.

The third layer is the workspace as a whole. From any screen you can ask which frameworks you should consider for AI governance, and get a ranked answer drawn from the hundreds in the framework catalogue, with a note of which are already enabled. Or open the assistant on nothing in particular and ask what is on your plate: it sees your own open findings across every project you can access and tells you what is overdue.


A Worked Example

Take one finding from an AI review of a payments project.

Nonconformity High severity

The refund service calls the ledger API with a shared static key stored in configuration.

Figure 3 — The finding the example follows.

The architect who owns it opens the finding and the assistant beside it. Four requests follow, and each one shows a different part of the design.

"Explain why this is a nonconformity." The assistant reads the finding and its directive and answers in terms of both: the directive requires credentials that expire, the design uses one that does not, and the directive's rationale is about limiting what a leaked key can do. No change is proposed, because none was asked for.

"The wording is vague. Tighten it." The assistant proposes a revision as a before-and-after card. The architect can accept it, edit the proposed text first, or dismiss it. Accepting updates the finding, and the edit appears in the finding's activity thread like any other.

"The team has started on this. Move it along and give them until the end of the quarter." Two cards: a state change to In Progress, and a due date. The assistant only offers state changes that are valid from the current state, so it cannot propose closing something that has not been through verification.

"Are we seeing this elsewhere?" The assistant switches to the directive and counts the open findings against it across the projects the architect can see. Three projects, all using the same shared-key pattern. That is no longer one team's problem.

You ask
In plain words
→
It proposes
A card, in full
→
You confirm
Or edit, or dismiss
→
The action runs
As you, audited
→
The page updates
No refresh

Figure 4 — Every change the assistant makes follows this path.


What It Can Do Today

The list is deliberately closed. The assistant can propose only actions we have built and reviewed one by one.

It can explain and answerIt can propose, for you to confirm
FindingsWhat a finding means, what is overdue, what to tackle firstA state change, due date, severity, comment, reworded text, or a new finding on a project
DirectivesConformance to a directive, what currently violates it, the impact of changing itA lifecycle move your role allows, a rewrite of a draft, or a new draft directive
ProjectsOpen risks, what blocks closureA sharper name or description, or starting an AI review
Knowledge baseWhat an entry says, answers grounded in itA rewrite, archiving, or restoring an entry
DiagramsWhat a drawing showsAn edit to the open diagram, applied to the editor for you to review and save
WorkspaceFramework recommendations, totals, applications, sessions, recent documentsNothing. Workspace-wide questions are read-only

Figure 5 — What the assistant answers, and what it may propose.

Two details in that table matter more than they look.

Approved directives stay locked. The assistant will help you rewrite a draft, but it will not propose editing an approved directive in place, because an approved directive is something reviews have already been run against. Changing one means creating a new version, with or without an assistant.

And starting an AI review counts as an action. It uses your workspace's AI budget and produces findings people will have to deal with, so it gets a Confirm card like everything else, showing the focus you asked for.


Why Not Let It Do More

We are asked this, usually by people who have seen an agent demo.

The honest answer is that the work an agent would save is not the work that is slow. Moving a finding to In Progress takes one click. What takes time in governance is deciding: whether this is really a nonconformity, whether the deadline is fair, whether a pattern across three projects means the rule is wrong. The assistant makes those decisions better informed and quicker to record. It should not make them.

There is also a plain accountability point. An audit trail that says "changed by the assistant" is an audit trail with a hole in it. Ours says who confirmed the change, because somebody did.

An agent that acts


  • Decides and executes in one step
  • Runs with its own access
  • The record shows the tool
  • You find out afterwards
versus

An assistant that asks


  • Proposes; a person decides
  • Runs with your access
  • The record shows you
  • You see it before it happens

Figure 6 — The difference that matters in a system of record.

This is the same position we take on AI review itself. The AI co-worker drafts findings; an architect accepts, changes, or dismisses each one. How the review works goes through that in detail.


Turning It On

The assistant is off until a workspace owner or admin turns it on in Settings › AI. After that it is the sparkle button in the top bar. It docks to the side of the screen, so the record you are asking about stays in view while you talk about it.

Start with a question about something you have open. The first useful thing it does for most people is not an action at all. It is explaining, in two sentences and with a link to the directive, why a finding they had been ignoring is there.

For the rest of what the AI in cajeX does, see the AI Co-Worker page.

Ready to transform your architecture governance?

cajeX brings AI-powered reviews, knowledge management, and directive lifecycle management to your enterprise architecture team.