See what your agents did —
and why.

This is the AgenticERP Control Room. It is the console you use to run, watch and govern the agents that work your Oracle JD Edwards ledger.

Getting an agent to do the work is the easy part. The hard part is what a finance organisation asks next. What did it do, why did it do that, and what happens next time. This console answers those three questions on one screen, for a controller, not a developer.

Before anything else, look at the badge next to the logo. It tells you where the numbers came from. Everything in this walkthrough is demonstration data, and the console says so in every frame. When it is reading your live ledger, that badge says live, and names the environment.

The console has eight screens, grouped into three folders down the left. The grouping is not decoration. It is the operating model.

Operate is today's work. Overview, Work queue, and Exceptions. This is where an accounts payable clerk or a controller spends the day. The two numbers in red and blue are live counts, so the folder itself tells you whether anything needs you.

Manage is the agents themselves. Agents, Automation, and Environment. This is where you decide how much rope each agent gets. Nobody edits a configuration file to change an agent's behaviour. It is done here, and it is written to the audit record.

Report is the evidence. Audit, and the CFO view. Audit is what you hand an auditor. The CFO view is what you take into a board meeting. Same underlying data, two different readers.

Two controls sit above all of it. Environment picks which JD Edwards instance you are looking at, so one console can cover a test box and production. Period sets the reporting window, and every figure on every screen is computed from it.

And on the right, an assistant that knows which screen you are on. Ask it about a voucher, a supplier, or any figure in front of you, and it answers from the run log rather than a guess. Every action it takes passes the same two gates the agents pass.

Start on Overview. The large figure is the touchless processing rate. The share of vouchers that went from arrival to posted with no human touching them at all. It is the single number that says whether the agents are actually carrying load.

Underneath it, two ticks. Industry average, and best in class, from the Ardent Partners accounts payable benchmark set. Published figures, not ours. Putting the target on the tile means you never have to look up whether the number is good.

Change the window to ninety days and every figure recomputes, including the comparison, which is always the equivalent period immediately before. Nothing here is painted on.

Below the hero, six operating metrics. First pass match rate. Cost per invoice. Processing time. Exception rate. Value held. And volume. Each carries its own benchmark where a published one exists, and says plainly when one does not.

The chart splits every day into what the agents finished and what they handed back. That split, over time, is the honest picture of how automation is trending. A rising blue band is capacity you did not have to hire.

At the bottom, the agent roster. Named for the process, never the technology. Invoice Capture. Three Way Match. Voucher Posting. That convention runs all the way down into anything the platform builds inside JD Edwards.

The activity feed is newest first, and every line opens the full record behind it. This is the efficiency story in one sentence. You are not reading a report about the work. You are one click from the work itself.

The Work queue is every voucher in the period. Not a top ten. Not a sample. All of them.

The tabs are workflow states rather than object types, so the tab strip is the process. To review. Pending approval. Returned to accounts payable. Held. Posted. A clerk works the tabs left to right and is finished.

Every column sorts. Every filter narrows. Search by voucher, invoice, order or supplier, then filter by exception type, by agent, by business unit, or by confidence band.

Filtering to the low confidence band is the highest value thirty seconds in an accounts payable day. It shows you exactly the work the agents were least sure about, which is exactly where a human should look first.

The tick column follows a convention operators already know. A grey tick means the agent validated it. A green tick means a person did. A red A means automation was blocked. You can read the state of four hundred vouchers by scanning one column.

Click any row and the record opens. This is the centre of the whole product.

The banner at the top states why this voucher did not automate, in one line, before you read anything else. Unit price over tolerance. The Grounding Gate refused the write, and the voucher is recorded and on pay status H.

Below it, the agent's work as a sequence of steps. Resolve the supplier. Locate the purchase order. Verify the receipt. Apply the tolerance policy. Authorise the write. Each step says what it checked and what it found.

Here is the step that stopped it. A variance of nine point six percent against a tolerance of two percent or two hundred and fifty dollars, whichever is lower. Not an error message. A finding, with the policy it was measured against.

And then the rationale. The tolerance applied was this client's configured value, not the shipped default. A looser default would have auto posted this voucher. That sentence is the difference between an agent you can defend and an agent you merely hope is right.

Each step also lists its sources, and this is the part nobody else ships. It shows what the agent consulted and used, and what it consulted and set aside, with the reason. Knowing what an agent looked at and rejected is what makes a record auditable instead of merely plausible.

The record has four views. What the agent did. The extracted fields. Data protection. And a technical trace.

Extracted fields shows every value the agent read off the invoice, and the measured confidence for each one. That is a real number about a real measurement, so it is shown.

What you will never see is a decision confidence percentage. A model's self reported score on a financial write is uncalibrated, and cannot be defended to an auditor. A decision carries a gate outcome and a rationale. A field carries a number. We are deliberate about the difference.

And the technical trace is the same record in the vocabulary every agent observability tool uses. One data model, two readers. The controller reads the step map. The engineer reads the span tree. Neither is a different truth.

Data protection answers the question your security team will ask in the first ten minutes. This is what the model actually saw. The supplier name, the taxpayer identifier, the bank account, the routing number, the remit to address. All surrogates.

The real values sit in a local vault and are put back immediately before the JD Edwards call. They never travel in a payload. Row level security governs which records an agent may retrieve. It does nothing about the tax identifier inside a record it is entitled to read. So the question worth asking any vendor is not whether data leaves your perimeter. It is whether the regulated value appears in the prompt at all.

Exceptions is a purpose built workbench. At the top, the full taxonomy. Fifteen types, enumerated, each with live volume, average age and value exposed. You can see which category costs you the most time, and which costs the most money. They are rarely the same one.

Load one and the screen splits. On the left, the agent's evidence, read only. The invoice as read. Why it stopped. And the voucher's state in JD Edwards right now. That last box matters. The liability is visible to the close, the payment is blocked, and nothing is reversed while it waits for you.

On the right, your decision, and only your decision is editable. A resolution, a corrected value if one is needed, a reason code, and a note that goes into the audit record.

And each outcome tells you what happens next. Approve, and the agent continues. Reject, and the agent stops. That is the management experience in three buttons. You are not chatting with a bot about a voucher. You are making a decision, and routing a process.

Your decision is written against the voucher with your identity, the reason code and the time, and fed back as precedent. The next comparable exception arrives with your last answer already in the record. That is how the queue gets shorter over a quarter instead of staying the same size forever.

The Agents screen is the fleet. One card per agent, and each card answers the questions a risk committee asks.

Every agent has a named human sponsor accountable for what it can reach, a scope listing which companies it may touch, and an explicit write list. Most of this fleet writes nothing at all. It reads. The ones that do write name the exact tables, and nothing else.

Each card also carries its Grounding Gate refusals. A refusal is not a failure. It is the agent declining to write because a governing fact was missing, which is exactly the behaviour you are paying for. Counting them tells you whether the guardrails are doing anything.

And at the bottom, the lifecycle. New, assess, build and test, deploy, operate, retire. An agent moves through these in order and nothing reaches production without passing the state before it. This is what turns a pile of automations into a managed fleet.

Automation is split by agent, because the agent is the unit you actually manage. Each panel controls one agent. Whether it runs at all, how much autonomy it has, and every threshold it owns.

Autonomy has three settings. Confident acts above threshold, and is the default. Always waives the thresholds, which quietly degrades learning, because only checked work becomes precedent. Never reviews everything first, and is right for a prove it period on a new client.

Now the detail that changes how people actually tune this. The threshold input sits directly beside the histogram of the confidence it is being set against. You are not typing a number into a void. You can see the shape of your own data while you set the line through it.

And the moment you move it, the strip at the bottom tells you what the change costs. How many more vouchers you just routed to review. How many more afternoons of somebody's time you just bought. Every automation product lets you set a threshold. This one tells you the price of the setting before you commit to it.

Some stages are hard gates and cannot be relaxed from this screen at all. Receipt verification. The tolerance policy. Both gates. If a policy genuinely changes, it changes in configuration, deliberately, with a change record. Not something a user quietly widens on a Friday afternoon.

Environment is where the console meets JD Edwards. The environment identifier, the authentication mode, whether credentials are configured, and whether access is read only. Probe the connection and it reports which layer is answering, without changing anything.

Every collection is logged with its row count and outcome, so when a number looks wrong the first question, is this data current, is answered on the same screen. In practice this is the screen that ends the argument about whether the dashboard or the ledger is right.

Audit is one row per action, across the whole fleet. Its Grounding Record. Both gate outcomes. The vault reference for any regulated value it touched. And who authorised it. One governed execution path, whether the action came from an agent, the assistant, or a person.

Export the period as a spreadsheet and you have handed your auditor the evidence pack. Not a screenshot of a dashboard. The records themselves. Most organisations spend more time assembling that pack than running the process it describes.

The CFO view is deliberately short. Days payable outstanding, which belongs here and not on an operator dashboard, because it is a treasury metric. Then the three figures a controller asks for first. Value actioned. Value awaiting approval. Value reversed.

Value reversed is the one to watch. It answers, in a single number, whether anything the agents posted has had to be undone. No competitor in this market publishes that figure, and it is the first thing a finance leader should demand before letting software write to the ledger.

So where does the efficiency come from. Three places. The agents remove the touches. The workbench removes the hunt, because the evidence is assembled before you arrive rather than after you ask. And the audit trail removes the reconstruction, which is the work nobody budgets for and everybody does.

And the management experience is the same story from the other side. You can see what the fleet did, tune what it is allowed to do next, and prove both. All of it in one container, on your own infrastructure, reading your JD Edwards directly.

AgenticERP Control Room. See what your agents did, and why.