TL;DR
AI on top of the ERP has become four different things sold to four different buyers. Here's how to tell them apart — and which layer actually moves your margin — on Microsoft Dynamics 365 and Business Central.
Enterprise AI has quietly changed direction.
For a few years the pitch was "replace your systems with something smarter." The bet that is actually winning — and the one drawing serious venture funding through 2026 — is the opposite: leave the ERP exactly where it is, and put AI on top of it.
That is the right instinct. An ERP is the system of record a business already trusts, already trains its people on, and has already customised over years. Ripping it out to chase intelligence is how good projects die in month four. Adding a thin layer of intelligence on top of it is how they ship.
But "AI on top of the ERP" has become a phrase that means four genuinely different things — built by four different kinds of company, for four different buyers. If you are evaluating this for a team on Microsoft Dynamics 365 or Business Central, the single most useful thing you can do first is separate them. A vendor who cannot tell you which of these four they do is the tell.
The four layers
1. Maintenance and operations
Agents that keep the ERP running: triaging the daily incidents, handling change requests, absorbing the work that large outsourcing teams have historically billed by the hour. This is the layer attracting the biggest enterprise funding right now, and it makes sense — global spend on maintaining ERPs runs into the hundreds of billions a year, and most of it is people-time.
Who it is for: very large estates, usually SAP and Oracle, with heavy system-integration contracts. The win is cost taken out of maintenance.
2. Decision optimization
The ERP records a long list of decisions it never actually computes: the safety-stock level, the reorder point, the discount floor, the price on the quote, the way work is routed through the warehouse. Each of those is a judgement call, and in most companies it is a static field set once at go-live and never revisited — or rebuilt by hand in a spreadsheet every week.
This layer does the arithmetic the ERP was never built to do, on the company's own history, and writes the recommendation back for a person to approve.
Who it is for: any operation where a pricing, inventory or fulfilment decision moves real money. The win is margin, not cost.
3. Document and contract intelligence
Reading the paper the business actually runs on — contracts, invoices, agreements, forms — and turning it into structured data, with the exact source behind every value. A renewal date becomes a field on the record. An obligation sits against the vendor it binds. A figure pulled from a scanned document arrives with the page it came from.
Who it is for: legal, procurement and finance teams drowning in PDFs. The win is risk caught before it costs you, and hours given back.
4. Conversational access
Ask the system a question in plain language — "what did we commit to this supplier, and when does it renew?" — and get an answer, with its source attached, without exporting anything.
Who it is for: everyone who currently waits on a report. The win is speed of access.
Which layer actually moves your margin
Here is the distinction worth internalising: maintenance takes cost out; the other three make or protect money.
That is not a knock on maintenance — for a Fortune 500 running a vast SAP estate, taking cost out of a hundred-million-dollar maintenance line is enormous. But for most mid-market teams on Dynamics 365 and Business Central, the recorded-but-uncomputed money is in layer 2 and layer 3. The price that was approved by rule of thumb. The safety stock that never matched real demand variability. The auto-renewal nobody read until it had already rolled. That is where the margin quietly leaks, and it leaks in the exact place the ERP has all the data and none of the math.
If you have to choose where to start, start where a decision you currently make by hand, in a spreadsheet, is worth real money when it is right.
How this gets built without touching the core
The reason "AI on top of the ERP" is winning over "AI instead of the ERP" is that the first one is safe. Done properly, it should never put your system of record at risk.
Everything we build at Cognilium runs beside Dynamics 365 — in the customer's own environment, on the platform's own supported integration surface. We read the system of record, compute the decision or read the document, and write the result back behind a human approval step. The ERP core is never modified. The partner who implemented it keeps running it. And the engineering discipline underneath — routing model calls instead of firing them at everything (on one contract-intelligence build, that cut model calls by roughly 75%), returning every extracted value with the quote it came from — is what makes the output trustworthy enough to act on.
Add the missing layer. Never rip and replace.
Where to start
You do not buy all four layers at once, and you should be suspicious of anyone selling them that way. Pick the one decision — the price, the stock level, the contract review, the renewal you keep missing — that is currently made by hand and matters when it is right. Prove the layer there. Then widen it.
If your business runs on Dynamics 365 or Business Central and the real decisions still happen in a spreadsheet, tell us the workflow that is actually costing you — you will get a straight read on which layer helps, what it would take, and where it would not be worth doing. Or see the two we build most: Dynamics 365 decision optimizers and legal AI ops built into the ERP.
Share this article
Agentic ERP
Topic Cluster
