Last updated: August 31, 2026
AI in ERP is real today in three narrow places, which are document capture, drafted text, and natural-language reporting. Everything wearing the word autonomous is a roadmap slide with a delivery window on it. The sorting rule is whether you can switch a thing on this afternoon without starting a project.
A CFO forwarded me a PDF in June. Fourteen pages, nice cover, produced by a firm neither of us had heard of. His company’s AI Readiness Score was 61 out of 100.
He wanted to know what to do about the 39.
So I asked him what the 61 measured. He didn’t know. Neither did his account manager. I read the fourteen pages twice and I still don’t know. Buried in there was a maturity curve with five stages, and a recommendation that he engage a partner for a Phase 2 roadmap, which is a sentence that costs about forty thousand dollars.
Nobody has ever failed an AI readiness assessment. Not once. Not anywhere. There is no version of that report that comes back and says stop, go home, you are not ready. It comes back low enough to justify the next engagement and high enough that you feel like you’re nearly there. That is the product.
Quick disclosure so you can discount me properly. I run an ERP and business systems consulting group, so a reader who decides their stack needs a hand is a reader who might call my office. You’re also reading this on a staffing firm’s site, and KORE1’s ERP recruiters place the people who end up owning whatever you switch on. Neither fact changes the feature list below. That part is public. Go check it.
Sort Everything Into Three Piles
AI in ERP means machine learning and language models running inside your ERP against your own transaction and master data, doing narrow jobs like reading a vendor bill, drafting a description, or answering a reporting question in plain English. The scope is a task on a record. Not a department.
Every claim a vendor makes goes in one of three piles, and the test takes about a minute per feature.
Pile one is shipping. Generally available, in your edition, in your region, and your administrator can enable it without a statement of work. Pile two is scheduled, which means it appears in a published release plan with a quarter attached, and it is not in your account yet. Pile three is slideware. Announced on a keynote stage, demonstrated on somebody else’s data, no date, no documentation page. That is the whole taxonomy.
Most of what you have been shown lives in piles two and three. That isn’t a scandal. It’s how enterprise software has always been sold, and the vendors are not lying, exactly. They’re describing a future they intend to build, in the present tense, because the present tense sells better. Sales works.
| Pile | What it looks like | The question that sorts it |
|---|---|---|
| Shipping | A help documentation page, a settings toggle, a known license tier | Can my administrator turn it on today? |
| Scheduled | A published release plan with a quarter and a preview flag | What is the general availability date, in writing? |
| Slideware | A keynote, a press release, a demo on a clean sandbox | Show me the documentation page. Any documentation page. |
Run your last vendor deck through that table. It takes ten minutes and it will annoy somebody in sales, which is a reasonable price. Do it anyway.
The Pile You Can Turn On Before Lunch
Here is what is genuinely live, using NetSuite because that’s the platform my team works in every day and I can speak to it without hedging.
Bill Capture reads vendor invoices. It uses document object detection and optical character recognition against Oracle Cloud Infrastructure’s document understanding model, it drops the extracted values onto a bill record, and a human approves the result. Text Enhance drafts and rewrites content inside records, running on OCI generative AI, so item descriptions and customer emails start from something rather than from nothing. SuiteAnalytics Assistant takes a reporting question in plain English and returns a dataset. All three sit on Oracle’s own features that use AI page, which is the single most useful bookmark in this entire conversation.
Notice the size of those. Not one changes an operating model. Bill Capture takes a task that consumed a person’s morning and turns it into a review queue.
That’s the win. Small is what works. Everything bigger is still a slide.
The document one pays first almost every time. We watched an accounting team switch from keying vendor bills to reviewing extracted ones. Nobody argued about accuracy afterwards. They argued about who owned the exception queue now, because that sat in no job description, and volunteering for an unnamed queue is how a person ends up doing it forever while everyone else quietly gets on with their quarter. Which is the actual pattern in most of these. The software works and the org chart doesn’t. Every time.
The Microsoft side reads similarly if you look in the right place. Dynamics 365 agents for purchase and sales scenarios in Business Central are real, and Microsoft publishes a release wave plan stating in public which capabilities land between April and September 2026. That document is doing your due diligence for you. Read the plan.
SAP announced more than fifty Joule assistants and over two hundred specialized agents at Sapphire this year. Some of that is generally available, including the agentic ABAP developer tooling. Some of it is a number on a stage. Big difference. The two categories are not distinguished in a keynote, and they are distinguished in the documentation, which tells you where to spend your reading time.
Adoption Is a Number About People, Not Processes
Now the buzzwords. AI readiness and AI adoption are the two things that get measured, and they get measured because they are easy, not because they matter.
Kaufman Rossin surveyed a hundred senior mid-market decision-makers with NewtonX in December 2025, all of them holding real budget authority, at companies with revenue between five million and a billion. 94 percent were using generative AI. Two percent had operationalized it at scale. Eighty-three percent said they had moved past dabbling into deliberate trials or embedded processes, and ninety-three percent planned to spend more over the following twelve months.
Ninety-two points apart.
Other people found the same cliff from other angles. RSM’s 2026 middle market survey of 1,030 executives across the US and Canada put 86 percent at partially or fully integrated, with 17 percent pursuing anything transformational. Netrio asked 401 IT leaders at companies with between 200 and 5,000 employees and got 82 percent with AI in production somewhere against 26 percent who could call it scaled and governed. Different samples, different questions, same shape. We lined all five of those datasets up against each other in the 2026 mid-market ERP and AI adoption report if you want the full reconciliation.
An adoption percentage counts how many humans have typed into a chat box at least once. It does not count a single process that changed. Not one. Which is how a company can be 94 percent adopted and 2 percent operational without anybody in the building noticing the contradiction.
One fair distinction before I get letters. Readiness as a word is not the problem. A genuine ERP readiness assessment checks specific things, like whether your item master holds one row per item and whether your close takes nine days or nineteen. Those are facts about your operation and they have answers. An AI readiness score is a number generated by the firm that would like to sell you the remediation. Same word. Entirely different object.
The Word to Look For Is Generally Available
Gartner named the thing that is happening. They call it agent washing, the rebranding of assistants, robotic process automation, and chatbots you already owned as agentic AI, and they estimate that only around 130 of the thousands of vendors claiming the label are building anything meaningfully different. Their prediction, published in June 2025, is that more than 40 percent of agentic AI projects get canceled by the end of 2027, on cost, unclear value, or absent risk controls.
Forty percent. Canceled. Not underperforming, not delayed.
The tell is tense. Listen for “will be able to” and “is designed to” and “our roadmap includes.” Honest phrases doing dishonest work in a room where nobody is writing the distinction down. A shipped feature gets described in the past tense by somebody who has already used it.
My rule is boring and it has never been wrong. Ask for the documentation page URL. Just the URL. Not a datasheet, not a one-pager, not a recorded demo. The help article a support agent would send a customer who filed a ticket about that feature. If it exists, the feature exists. If it doesn’t, it doesn’t. And if the answer is that it’s coming in the next release, you have just moved that item from pile one to pile two, and now you can plan honestly instead of budgeting against a video.
Panorama’s 2026 ERP Report surveyed 170 organizations over a year of data collection, at a median revenue of $200.5 million, and found AI adoption inside ERP projects had barely moved year over year. Their read is normalization, which I think is right. The excitement curve and the deployment curve stopped being the same curve about eighteen months ago. That gap is the story.
Your API Allowance Does Not Care About the Demo
Here’s the part nobody puts on a slide, and it’s the part that decides whether your AI feature ships at all.
We priced an integration this summer for a regulated manufacturer, and the binding constraint on the entire architecture turned out to be arithmetic. Their licensed API call allowance was 130,000 a year. A design that touched the API once per unit, at roughly 70,000 units a month entering serialized inventory, would have burned about 1.35 million calls. Over by ten times.
The same flows batched at the document level come to something like 141 documents a month at four to six calls each. Under a quarter of the allowance, with headroom left over for verification traffic and growth.
Same requirement. Same platform. Same data. One design is impossible and one is comfortable, and the difference is a decision somebody makes in week two that never appears in a demo, never appears in a license discussion, and surfaces about four months later as a support ticket nobody can close.
Anything agentic multiplies calls rather than reducing them, because the model checks, retries, reads back, and confirms. That is what makes it useful. It is also what makes the allowance math get away from you when nobody has run it. Ask any vendor pitching an agent into your ERP what their expected call volume is against your licensed ceiling. Watch what happens.
A second one in the same family, because it costs teams two weeks every time. A NetSuite RESTlet will not accept a plain API key. External callers need OAuth 1.0 token-based authentication or signed OAuth 2.0, so any vendor whose callback documentation says to supply your API key needs an authenticated relay standing in front of it. Nobody mentions this until integration week. It bites everyone. I’ve written more about how the connector and agent layer actually behaves in the piece on the NetSuite MCP connector and SuiteAgents, permissions included, which is its own adventure.
What the Real Version Costs, in Hours
People ask for a number and then get annoyed at the number. Fair enough. Here are ours, and they are benchmarks off real delivery rather than a list price.
We run a blended rate of $215 an hour across delivery roles. The only differential in our entire pricing model is $275 for an independent validation lead, and it’s independent for a reason I’ll get to. Discovery, risk assessment, and a validation plan come to 182 hours, which we fix at $44,110. Build, validation, and release on a substantial regulated integration land around 1,060 hours and somewhere between $241,577 and $316,735, firming to a fixed fee once the requirements specification is approved. Holding that validated state across the two ERP releases a year that arrive whether you’re ready or not costs another 120 hours and about $31,046 annually.
In a GAMP 5 category 5 environment, validation is roughly 30 percent of base hours. Thirty percent. That number is why a materially cheaper proposal usually turns out to be cheaper by exclusion rather than by efficiency. Read the exclusions.
We also priced the same scope on a low-code integration platform, got $265,900 before platform licensing, then argued against our own cheaper option. A realistic build still contains scripted transformation steps that stay category 5 no matter what category the platform claims. The tool doesn’t reduce the validation burden. It relocates it, and somebody discovers that during qualification. Bad week.
Most of you aren’t in a regulated environment and your numbers will be smaller by a lot. The structure survives the change in scale, though. Someone has to specify the thing, someone has to test the thing, and the testing is not a rounding error.
Provenance Is an Audit Question Now
We wrote our AI policy down, partly because a client asked and partly because I wanted to know what my own answer was.
A named qualified person is author of record on every deliverable, and model output stays draft material until that person can explain any line of it to an auditor without opening the tool that produced it. Generation happens against an approved specification rather than a prose description, because category 5 validation needs traceability from requirement through design and code to test, and a prompt written in a hurry breaks that chain in a way nobody notices for six months. Review before merge is mandatory, documented, and done by a qualified person who is not the author. Test protocols are written by a human, always, because independence of the test from the implementation is a control and generating both from the same source quietly defeats it.
None of that is exotic. It’s four paragraphs. Nobody publishes theirs, which is odd given how many firms are selling AI services this year.
You’ll want somebody in-house who can read a release note and tell you which pile it belongs in. That skill is rarer than the job title implies. Much rarer. When we need that seat covered temporarily during a build, contract staffing is usually the honest answer rather than a permanent hire, and KORE1 fills those in about 17 days with 92 percent of placements still in seat a year later. Their MCP and agentic AI consultant practice exists because the demand showed up faster than the supply did.
Questions I Get After the Demo
Our vendor says the AI is included in our license. Is it?
Three separate bills hide inside that sentence. The feature entitlement, the consumption cost of the model calls, and the edition tier that gates the feature in the first place. All three bill. Ask which of them your contract covers, in writing, and ask what happens at volume. Included usually means included up to a threshold nobody quoted you.
How do I tell a shipped feature from a release-plan feature?
Look for the phrase generally available. Then find the help documentation page. Vendors publish release plans with quarters attached, and those plans are public and honest. Check the quarter. A capability listed under a future one is not something to build a business case on this year.
We are two versions behind. Are we locked out of all of this?
Partly. Less than you fear. Most of the document and text features ride the platform rather than your customizations, so an upgrade turns them on. What actually blocks companies is not the version number. It’s the pile of unmanaged customizations that makes upgrading feel dangerous, and that’s a change-control problem wearing a technical costume.
The board wants an AI number for the ERP. What is a defensible one?
Cycle time on one named process, measured before and after. Days to close, hours to code a batch of vendor bills, or touches per sales order. Two dates. Adoption percentages make the slide look better and will not survive a follow-up question from anybody who runs operations.
Our auditors are asking about AI-generated code. What do we need in writing?
Four things, and none of them are long. Who is the named author of record, what approved specification the generation ran against, who reviewed it before merge, and who wrote the tests. Write them down. Answer those four in a paragraph each, and you are ahead of most of the vendors bidding on your work.
My peers are all waiting a year. Should I?
Wrong axis. Nothing is settling. The pile-one features are boring and stable and they were boring and stable last year too, so waiting buys you nothing there, and waiting on pile three costs you nothing either since it doesn’t exist yet. There’s no version of this where sitting out the document-capture stuff pays.
Hype Is Just a Feature With a Date on It
I don’t think the vendors are running a con. Roadmaps are real work, and most of what’s on those stages will eventually ship, some of it well.
The problem is that the deck and the documentation get read in the same meeting by the same people, and only one of them is a commitment. That’s a decision-hygiene failure inside your building, not a vendor failure. That one is yours.
So go do the ten-minute version. Open your ERP’s AI documentation page. Find the two or three features live in your edition right now. Turn one on next Tuesday against a process somebody in your building complains about out loud, and measure the cycle time before and after. Ten minutes.
Then, when the next fourteen-page readiness assessment lands in your inbox with a score on the cover, you’ll have something better than a score. You’ll have a process that changed.
If you get into it and want a second read on where the arithmetic breaks, or you’ve worked out that the person who should own this doesn’t work for you yet, talk to a KORE1 recruiter about the seat, or hit me up on LinkedIn. Bring the deck. I like reading them.
