See an AI role at work

Watch an AI role do the work.

Not a chatbot answering a question. One defined job, the business context and systems to finish it, limits on what it can decide, and a person who signs off where it matters. Here are three of them, start to finish.

Start the run Build your AI Workforce

About 90 seconds. No form, no video, nothing to install.

The run

Three roles. Three pieces of work. Start to finish.

Pick a role and watch it run. Each one is a different shape of recurring work, and each one shows the same four things: what it checks, what it is allowed to do on its own, where it stops, and what it hands to a person.

Invoice Review AI Role

06:00/Step 1 of 10

On duty, nothing to do

The role sits on one queue with one job. No prompt, no session, no person at a keyboard. It waits for work.

Activity log

Nothing yet. The role has not been given work.

Illustrative scenarios, built from the shape of work we see in production. Not a client system, and not client data.

Work that checks

Invoice Review AI Role

06:00 / Step 1
On duty, nothing to do
The role sits on one queue with one job. No prompt, no session, no person at a keyboard. It waits for work.
06:14 / Step 2
Invoice received
A supplier invoice lands in the shared accounts inbox, hours before anyone is at a desk.
Work arrives on its own schedule. The role starts from the work, not from somebody asking it a question.
06:14 / Step 3
Read and extracted
Supplier, purchase order reference, eleven line items and totals, pulled straight off the document.
The supplier changed their layout since last month. It reads what is actually there rather than a template somebody set up.
06:15 / Step 4
Checked against the purchase order
Line by line against what was actually agreed. Ten lines match. One is 12% over, and the delivery missed the agreed window by six days.
This is the part a chatbot cannot do. It is not recalling what invoices usually look like, it is opening what your business agreed.
06:15 / Step 5
Your rule applied, not its own judgment
Variances under 5% and under $1,000 post automatically. This one clears both limits, so it does not post.
The threshold is yours. The role does not decide what counts as acceptable, it applies what you decided.
06:16 / Step 6
Query drafted, payment held
It writes the supplier query with the line comparison attached, files the exception against PO 4418, and holds the invoice out of Friday's payment run.
Work performed, not advice offered. Three things changed in your systems while nobody was awake.
06:16 / Step 7
Stopped at its boundary
It does not approve the variance and it does not release the payment. Above the threshold this role has no authority, so it stops.
A prompt tells AI what it should do. A control decides what it can do. This limit sits in the infrastructure, not in a sentence.
06:16 / Step 8
Sent for one decision
The Accounts Payable Manager gets one question, with the evidence already attached.
Escalation is part of the job, not a failure of it. The decision takes seconds because the work is already done.
09:02 / Step 9
Decision returned, job finished
The manager sends the query. The role releases the other ten lines for payment, annotates PO 4418, and closes the item.
The person decided the one thing only a person should decide. The role did everything on either side of it.
09:02 / Step 10
Recorded and measured
Every step logged and attributable, and a result this role can be judged on next month.
A role you cannot measure is a role you cannot trust with more.

Work that prepares

Proposal Prep AI Role

08:00 / Step 1
On duty, nothing to do
The role watches one queue: incoming proposal requests. No prompt, no session, no person at a keyboard.
09:41 / Step 2
Request received
A 34 page request for proposal arrives from a prospect already in the CRM. The response is due in four working days.
Work arrives on its own schedule. Nobody had to notice it and forward it to the right person.
09:41 / Step 3
Read and broken down
It pulls out the scope, the required sections, the submission format and the deadline, then opens the prospect's record.
It reads the document this buyer actually sent, not a template of what these requests usually ask for.
09:52 / Step 4
Drafted from approved content
It assembles the response from the approved proposal library, using passages that have already been through legal, and pulls in what was proposed to two similar buyers.
It is reusing work your business already signed off, not inventing new claims about you.
10:04 / Step 5
Priced from your rate card
Standard scope lines are priced straight from the rate card. Two lines are not on it, so they are left blank and marked.
The rate card is yours. The role prices what it is allowed to price and marks what it cannot.
10:06 / Step 6
Stopped at its boundary
It does not price the two uncovered lines, does not accept the liability clause, and does not send anything to the prospect.
A prompt tells AI what it should do. A control decides what it can do. This limit sits in the infrastructure, not in a sentence.
10:07 / Step 7
Sent for a decision
The account lead gets the draft with three open items marked and everything else already written.
A Monday morning of work is waiting when they sit down, and what is left is only what needs their judgment.
14:20 / Step 8
Decisions returned, job finished
The lead prices the two lines and declines the clause. The role updates the draft, rebuilds the pricing table, formats it to the required submission template and logs it against the opportunity.
The person made three decisions. The role did the twenty hours on either side of them.
14:20 / Step 9
Recorded and measured
Every reused passage traceable to the approved source it came from, and a result the work can be judged on.
When a claim in a proposal is questioned later, you can see where it came from.

Work that chases

Document Chase AI Role

Mon 07:00 / Step 1
On duty, nothing to do
The role reads every job on the schedule each morning. There is nothing to chase until something is missing.
Mon 07:02 / Step 2
Gap found
A job starts in nine days. Three of the eleven documents it requires are not on file, and one of them expired four days ago.
Nobody asked it to look. It checks every job every morning, which is exactly the work nobody gets to until it is late.
Mon 07:05 / Step 3
Matched to whoever owns it
Each missing document is tied to the one person responsible for it, rather than broadcast to a group.
Chasing only works when one named person is asked for one specific thing.
Mon 07:06 / Step 4
First chase sent
Three separate messages, each naming one document, why the job needs it and the date it is needed by.
Work performed, not a reminder to someone else to perform it.
Wed 09:00 / Step 5
One arrived, checked, two still open
The safety acknowledgement comes back. The role opens it, confirms it is the current version and signed, files it, and leaves the other two open.
It checks what arrived rather than ticking a box because an email came in.
Fri 09:00 / Step 6
Second chase, and a risk raised
Two follow-ups go out. The insurance certificate is now the item putting the start date at risk, so the job is marked at risk in the operations system.
The escalation schedule is yours. It follows the one you set rather than nagging on its own rhythm.
Fri 10:20 / Step 7
Stopped at its boundary
The supplier offers last year's certificate instead. The role will not accept it, and it will not waive the requirement to keep the job moving.
A prompt tells AI what it should do. A control decides what it can do. This limit sits in the infrastructure, not in a sentence.
Fri 10:21 / Step 8
Sent for a decision
The Operations Manager gets one question, with the chase history and the start date attached.
The judgment call reaches a person early, while there is still a week to act on it.
Fri 16:45 / Step 9
Decision returned, job finished
The manager holds the start date and calls the supplier. The current certificate arrives that afternoon. The role verifies it, files it, clears the at-risk flag and confirms the job is ready to start.
The person made one call. The role did the five days of checking and chasing around it.
Fri 16:46 / Step 10
Recorded and measured
Every request, every response and every check kept against the job, so the file stands up later.
The reason a requirement was met, and when, is the part auditors ask for.

Illustrative scenarios, built from the shape of work we see in production. Not a client system, and not client data.

Anatomy

Six things that made that a role, not a tool

Every role you just watched has all six, whichever run you picked.

A defined job

One piece of recurring work, written down, with a standard it can be measured against. Not help with finance. Check these invoices against what was agreed.

Business context

Your payment terms. Your rate card. The requirements this job type carries. Context is what makes AI useful inside your company rather than in general.

System access

The finance system, the CRM, the document store, the shared inbox. Scoped system by system, reading where reading is enough, writing only where it has been granted.

Authority with a limit

A prompt tells AI what it should do. A control decides what it can do. Every run above ends up somewhere the role has no authority, and stops there, because the limit sits in the infrastructure rather than in a sentence.

Human oversight

At the limit it stops, prepares the decision, and hands it to a named person with the evidence attached. Escalation is part of the job, not a failure of it.

A record and a measure

Every step logged, every decision attributable, and a number at the end that the work can be judged on.

Change the job and the systems and you have a different role. The shape never changes.

The difference

Three things that can look similar from the outside

All four are useful. Only one of them takes work off your team.

An AI tool

You open it, you ask, it answers. The work still sits with you, and nothing happens until you act on what it said.

A copilot

It helps the person doing the job move faster inside one application. Useful, and still one person doing the work.

An automation

It runs the same steps every time. It does not read an exception, it breaks on one.

An AI role

It holds a job, with the context and systems to finish it and limits on what it can decide. It handles the exception, stops at its boundary, and hands a person the decision with the work already done.

Most companies end up with all four. Only the fourth one changes where the work sits.

Three tiers of trust

Autonomy is earned, not switched on.

Authority is granted task type by task type against thresholds you set. Here is the invoice role from the runs above, and the same role nine months later. One rung moved, on the evidence of what it had already done. Any grant can be moved back down.

Per task typeActs on its ownHuman approvalNot permitted

Invoice Review AI RoleThe role you just watched · month three

Read the invoice and match it against the purchase orderActs on its own
Post a clean match with no varianceActs on its own
Draft the supplier query with the evidence attachedActs on its own
Send that query to the supplier without reviewHuman approval
Release payment where a variance is above your thresholdHuman approval
Approve a price variance, or change supplier bank detailsNot permitted

Invoice Review AI RoleThe same role · month nine

Read the invoice and match it against the purchase orderActs on its own
Post a clean match with no varianceActs on its own
Draft the supplier query with the evidence attachedActs on its own
Send that query to the supplier without reviewActs on its own
Release payment where a variance is above your thresholdHuman approval
Approve a price variance, or change supplier bank detailsNot permitted

The control questions that matter

Which task types are allowed to act without a human?
What evidence moves a task type up a rung?
Who owns the threshold, and who can move it back down?

Six months of clean sends moved one rung. The bottom row never moves.

Your turn

Where would a role fit in your business?

We do not sell a catalogue of AI employees. We start from the work your business needs done. Most of the recurring work that survives a growing company falls into one of six kinds, and each one is somewhere a role can take a job off your team.

Work that prepares

Proposals, reports, briefs, estimates, documentation.

Work that checks

Quality reviews, compliance checks, reconciliations, audits.

Work that chases

Approvals, missing information, follow-ups, exceptions.

Work that processes

Documents, requests, tickets, transactions.

Work that monitors

Changes, thresholds, status, operational conditions.

Work that coordinates

Multi-step processes across people and systems.

The three runs above are work that checks, prepares and chases. Not sure which one is worth starting with in your business?Score one piece of your own work →

Ownership

Stop the service, keep the workforce.

The role runs inside your own cloud account, and it gets more valuable the longer it works with how your company actually operates.

Yours, permanently

The deployed role, the business context it has built up, the operating rules and the audit history. All of it sits inside your own security boundary and keeps running on its last deployed version.

What Metacto does

Keeps the role accountable, measurable and improving. We monitor it in production, check how it is performing against its standard, and tune it as your business changes.

This page shows one role. The offer around it is the Managed AI Workforce.See how it works →

You have just watched somebody else's work. What would yours look like?

45 minutes, nothing to prepare. Tell us about one repetitive or capacity-constrained piece of work and we will map the first one or two AI roles around it.

45 minutes

No prep required

Leave with one or two AI opportunities mapped