In short. Your ERP can confirm that a purchase order exists. It can't tell you whether the package will land on the date your schedule depends on, because the events that move on that date never reach it. A drawing gets revised; a submittal comes back rejected; an inspection slips two weeks. None of those are transaction. The gap between a package slipping and your organization finding out is measurable; it usually runs weeks. It's the number worth managing.
"On order" is a true answer and a useless one
A project director asks a procurement lead where the compressor skid is. The lead opens the record. PO issued 4 March, status on order, expected on site 18 June. Every field is correct.
The skid is going to be five weeks late. It has been since mid-March.
What happened sits outside the record. The vendor received drawing revision B on 12 March and restarted fabrication planning. Their first material submittal came back rejected on 2 April. A witness inspection booked for 24 April moved to 15 May when the third-party inspector could not travel. On 9 May the vendor re-planned to late July and said so, in an email thread, to one engineer.
Four events, none of which created a transaction. The record wasn't lying; it was answering a different question from the one being asked.

Status is a question about the future
When a project director asks where a procurement stands, the question underneath it is whether the date still holds. If it doesn't, they want to know how much and who caused it.
Procurement systems are built around the transaction skeleton: requisition, order, receipt, and invoice. Those are the moments when something is entered. The date lives in the gaps between them. Those gaps are where engineered packages spend most of their lives.
This is also what buyers ask us for. Across ten procurement conversations this year, bid comparison came up once, from a partner, and never from a buyer. Visibility came up in six of the ten, and the question of who caused the delay in four. The thing people wanted was not a faster award. It was to stop finding out late.
Every slipped date starts in one of four places
In engineered procurement, almost every slipped date traces back to one of four events, and all four happen in systems that sit beside the ERP rather than inside it.
● A drawing revision. The package the vendor is building is no longer the package you ordered. Fabrication planning restarts, and the revision lives in the drawing register.
● A rejected submittal. Material certificates, welding procedures, or test plans come back for rework. The clock restarts in the submittal log.
● A moved inspection. A witness point or factory acceptance test is rebooked, often for reasons that have nothing to do with the vendor. It lives in an inspection calendar.
● A re-plan for the vendor volunteers. The most valuable signal of the four, and the one most likely to arrive in an email thread with two people on it.
Each of these is visible on the day it happens to somebody. The problem is that somebody is a different person each time. Meanwhile the package they are all touching has a required-on-site date none of them is watching.
Name the lag, then measure it
There is a number here worth putting on a page. Call it the detection lag: the days between the first event that moved a package, and the day your organization knew the date had changed.
You can compute it this week without buying anything. Take your last three late packages. Find the earliest event that made the delivery date unachievable, which is usually a revision or a rejection with a date stamp on it. Then find the day the new date was first acknowledged in writing. That distance is your detection lag on that package.
In the operations we have looked at, a procurement lead carrying seventy or more live packages learns that a date has moved when they ask, or when a vendor volunteers for it. Both are events on somebody else's calendar. The lag isn't a function of how hard the team is working. That's why adding a weekly status call moves so little.
There is second place the lag hides, and it is the review meeting. In most engineered procurement, a slipped date is formally discovered when somebody presents a status pack, which means detection is scheduled rather than continuous. The package slipped on 12 March. Its review landed on 20 May. Nothing in between was broken. A calendar decided when you would find out.
Moving detection off that calendar is the change worth buying. In the governed procurement operations we run, supplier non-compliance, technical deviations, missing documentation and contractual risk surface the same day the evidence appears rather than at the next review. The decision still waits for a person to arrive. The discovery no longer does.
Cycle time follows from that. Request through to award runs up to eighty percent faster in our deployments, and it is worth being precise about why. That saving does not come from approving things faster. It comes from finding out early enough that approval was never the thing holding the package up.
What makes the lag expensive is where the package sits. On a critical path, delay converts directly into money at rates; the contract already defines liquidated damages, standstill and escalation. A date discovered in March can be mitigated. The same date discovered in May is a claim.
Half the delays you blame on vendors are yours
Detection is half the answer. The other half decides what you do next.
If the vendor caused the slip, you have a contractual conversation, and the evidence that supports it needs to exist before the dispute rather than after. If your own side caused it, which happens more often than most teams expect, you have a management problem and a remedy you control. A drawing sitting in review for three weeks, a spec clarification that went unanswered, an inspector booked late: none of those appear in a vendor scorecard, every one of them is inside your own organization, and they are the delays you can actually do something about.
Without attribution, every slip becomes a vendor problem by default. Roughly half are not. And the half you misattribute is the half you keep repeating.
An agent closes the lag only if it never guesses a date
This is continuous reading, which is why it has stayed undone. The sources are known, the evidence is dated, and the volume is beyond what a team can watch by hand across seventy packages.
An elsai procurement agent watches the places where the four events appear first, from drawing registers and submittal logs to inspection calendars and the mail threads where vendors say what they are really doing, tying each event back to a package, its purchase order and the date that package is required on site. When the lag opens, it raises it. The evidence comes attached.
Which is what separates this from a status report. A status report says the package is late. What lands here is the date that moved, the event that moved it, the document and revision it came from, the system it was read in, and when.
A project director can act on that the morning it arrives; issue the notice, escalate to the vendor, re-sequence what sits behind it. They can also sign their name to it, because the evidence arrived with the finding instead of being assembled afterwards to justify a decision already taken.
A finding without that evidence waits for the next review meeting to be confirmed, and the lag you just closed reopens between detection and decision.
Three design rules matter more than the detection itself.
● The model never estimates a date or an amount. Dates come from the project's own schedule logic, and amounts from contract rates and purchase order terms. The agent finds and connects evidence. It does not guess at a delivery date; the single question an operator is right to distrust it on.
● Every finding cites its evidence and its as-of-date. Document, revision, clause, system and timestamp, with each step logged against the rule version that produced it. A finding you cannot trace is an opinion with a timestamp.
● Nothing is issued without a named role. No notice, variation or escalation leaves the system unsigned, and every override records who made it, why, and when.
Your quality and project people keep every decision. What changes is that they make those decisions in March.
Where it runs is a decision you make once
In six of ten procurement conversations this year, the first hard question was not accuracy. It was where the data goes. Engineering drawings, contract terms and vendor correspondence are exactly the material an organization is least willing to send somewhere else.
So, elsai runs where you need it to; in your own cloud tenancy, on your infrastructure, or fully air-gapped, and across model providers rather than tied to one. That's a deployment decision you make once, and it is worth making before the pilot rather than after.
Compute the lag on three packages this week
Pick the three packages that hurt most this year and compute the detection lag on each. If the answer is weeks, the constraint on your schedule is not vendor performance and it is not your team's diligence. It is that the events which move dates and the system that reports them were never connected.
Connecting them is an integration problem, and a solvable one. Making the output something a named person will act on the same morning is a governance problem. We wrote about the test it must meet in.
If You Can't Replay the Decision, You Don't Have an Audit Trail.
Recent blogs
Secure your agents
We’d love to chat with you about how your team can secure and govern Ai agents everywhere







