Write the rules into the worker.
Hard-code the limits, the approval paths, the exceptions. It works, precisely and reliably, for exactly as long as none of them change.
Every policy change becomes an engineering ticket
MeetAlex.
The operating intelligence layer
Work moves forward, with the record attached.
ALEX builds a governed model of how your company actually operates.
The gap
AI workers do not stall because they are not clever enough.
They stall because nothing tells them what actually governs the action in front of them. Which policy is current. Who holds the authority for this amount, on this account, today. Whether two sources disagree. Whether it is safe to proceed alone.
That knowledge exists in your company. It is just spread across contracts, approval matrices, a handbook nobody has revised since 2023, and the four people who have been here long enough to know which of those to ignore.
A capable model with no access to that is not an operator. It is a very confident guess.
The choice today
Two ways to give an AI worker the rules. Both of them fail.
Let the model work them out.
Point it at the documents and let it infer. It will answer fluently every time, and it has no way to tell you which of those answers was a guess.
The failures are silent, and they are about money
Govern them once, outside the worker.
One current model of what governs, who may decide and what is unresolved, maintained as the company changes and queried by every worker before it acts.
A policy change updates the model, not the code
Rules written into a worker are a snapshot. Rules held outside it can stay true.
Before it acts
Four questions, answered every time, before anything runs.
A Worker does not ask ALEX for information. It asks for permission, and gets a decision it can act on, with the reasoning attached.
Pick an action a Worker wants to take.
The Worker wants to
What governs this
Who may decide
What ALEX could not establish
May it proceed
3 of 3 checks answered
Every request is answered before anything runs, against the states and refusal conditions the Order-to-Cash Worker uses today. Pick any one to see the reasoning behind its decision.
Where it sits
Between the workers and the systems they act on.
ALEX does not replace your ERP and it is not another agent. It is the layer that lets one speak to the other without either of them guessing.
Workers
1
One worker, five connections. It works. That is exactly why nobody looks past it. Each connection is an integration someone maintains and a rule someone had to decide. At five, that is a project you can hold in your head.
Add a worker and watch the count. Nothing unreasonable happens at any single step, which is why the shape of it is only ever obvious in hindsight.
Your first AI worker will work. The cost arrives at the second one, and adding a worker should add a workflow, not another version of your company to keep in step.
What it holds
Six distinctions your documents never had to make.
Standing
Which version actually applies.
A document existing is not the same as a document governing. ALEX tracks which statements currently carry weight, and refuses to treat “most recent” as a synonym for “the one that counts.”
Reality
Documented vs. observed.
What the policy says and what people actually do are tracked as separate states, not merged into one comfortable narrative.
Conflict
Contradictions held open.
When two sources disagree, ALEX doesn't pick a winner by guessing. It surfaces the contradiction and holds both until someone with the authority to resolve it does.
Authority
Who may actually decide.
Approve, escalate, override. Modelled explicitly, not inferred from who happens to be in the room or what a title implies.
Completion
What “done” means here.
Agreed before the work starts, so a Worker is not guessing at a moving target. What the case was aiming for and what actually happened are recorded separately, and they are allowed to differ.
Decay
What breaks when something changes.
When a policy or a limit moves, everything that was decided under the old version is marked, not left to quietly keep running on it.
The hard part
Reconcile.
Reading a company's documents is the easy part. The hard part arrives three months later, when the approval limits change and nothing tells you which decisions were made under the old ones.
Register a change to a policy or an authority in ALEX and it supersedes the old version, works out exactly which downstream fields depended on it, and marks them stale by name. Anything compiled from the old version is blocked from being presented as current until it is rebuilt and re-approved.
The change is registered by whoever holds the authority for it. That is the person who would have signed the new limit anyway, not a role you have to staff. What they do is publish the change once. What ALEX does is the tracing, the naming and the blocking: the part that otherwise falls to whoever happens to remember.
A model that doesn't survive a change is a report.
This is the step that makes it something else, and the reason the model is worth keeping. A model that quietly decays is not an asset, however good it looked on the day it was delivered.
What you actually use
Six screens, and no dashboard among them.
Everything here reads from the recorded event log. Nothing is reconstructed at the moment you look at it.
Today
What needs you, derived from recorded waits, owners, deadlines and open questions, with the source event behind each item.
Cases
Live work: yours, waiting, needing authority, overdue, completed. Business outcome and operator state stay separately labelled.
Decisions
What is waiting on you specifically: the question, the authority, the alternatives, the evidence, and what remains unknown.
Waits
What the system is waiting for, and from whom, with delivery receipts, and failures shown as failures rather than as sent.
Changes
Contradictions, publication checks, and what a change to a policy actually invalidated downstream.
Ruling controls are disabled for anyone without the authority. A forged request returns 403 and leaves the case file byte-identical.
Proven
The same four numbers, every time.
The guarantee comes from the checks, not from the model, so it holds whichever model is behind it.
0
Fabricated claims
100%
Traced to source
0
Conflicts collapsed
42
Graded runs
Measured across 42 graded runs on three independent document sets, using two language models roughly thirty times apart in capability, and scored against answer keys written before the runs. The stronger model produced more unsupportable material, and the checks caught all of it.
The ALEX offer
Not ready to automate the work yet? Start by understanding it.
We'll build a governed model of how one consequential operation actually runs: where documentation and practice disagree, who has real authority, what's contradicted, and what's missing.
Deploy a Worker afterwards and it runs on the model we already built with you. nothing gets rebuilt. Either way you own an explicit, current account of a part of your company that until now existed only in people's heads.