Bramvia — Business Central Experts

You bought AI. It isn't working. The problem is underneath your ERP

94% of mid-market companies use generative AI. 2% have operationalized it at scale. 86% of CFOs say legacy systems and technical debt — not talent, not budget — are what limit their AI readiness. Here is what actually blocks it, and the order to fix it.

Short answer: if your AI pilots keep stalling, the cause is almost certainly not the model, the vendor or your people. 94% of mid-market companies now use generative AI; 2% have operationalized it at scale (Kaufman Rossin, 100 senior decision-makers). MIT's NANDA study put it bluntly: roughly 95% of enterprise generative AI pilots deliver no measurable return, and an MIT review of 300 deployments concluded that poor integration with legacy workflows — not model malfunction — was the key culprit. The clearest statement of all comes from the finance side: 86% of CFOs say legacy systems and technical debt limit their AI readiness. Not talent gaps. Not budget. Not governance. The foundation. And there is good news buried in the data that nobody tells mid-market companies: you are structurally better placed to fix this than a large enterprise is.

The numbers, in one place

Read the last three rows together. Three quarters of manufacturers are still piloting, nine in ten will spend more next year, and Gartner expects most of what lacks a foundation to be abandoned anyway. That is a lot of money about to be spent on the same outcome as last time.

Why the pilot always looks good and production never arrives

The pattern is consistent enough to be predictable:

The pilot was designed to succeed in isolation. Proofs of concept are scoped to demonstrate capability, not to survive integration. They run on curated data, bypass the legacy systems, and sit outside the authentication, compliance and workflow dependencies that govern real operations. Then production asks them to connect to the system of record they deliberately avoided.

AI gets layered on top of existing processes. McKinsey found nearly 80% of organizations add AI on top of how the work already flows without rethinking it. The capability arrives; the outcome doesn't change. A bad process running faster is still a bad process.

The data isn't ready and nobody said so. Duplicate part numbers, three addresses per customer, costs that never had freight allocated, production feedback that lives on a paper sheet. A model trained or prompted on that produces confident answers that are wrong — which is worse than no answer.

Pilot fatigue arrives before scale does. Deloitte's 2026 report names it: after several stalled pilots, the organization loses the appetite to finish one. By the third failure, executives stop attending the reviews. The fourth pilot launches into a company that has already decided, quietly, that AI doesn't work here.

And it isn't free to stay in the middle. Estimates put sunk cost per abandoned AI initiative at around $7.2 million — before the opportunity cost and the credibility spent.

The part nobody tells mid-market companies

Here is the finding worth acting on:

Large enterprises take an average of nine months to scale a single AI pilot. Mid-market companies, with fewer governance layers, do it in 90 days.

Your size is an advantage here, not a handicap. Fewer committees, shorter decision chains, one system of record instead of eleven. What you lack is not speed — it is the foundation underneath, and that is a shorter project than anyone selling you AI has an incentive to tell you.

What "the foundation" actually means

Not a data lake. Not a platform. Four concrete things:

1. One system of record that holds the truth. If the real stock figure lives in a warehouse spreadsheet and the real margin lives in the controller's file, no agent can answer a question about either. The one-sentence test: if people need a spreadsheet to do their daily work, your ERP has already stopped being your system of record.

2. Master data that is actually clean. Duplicate items, dead SKUs, customers with three codes. This is unglamorous and it is the single highest-return work in the whole sequence, because everything downstream inherits it.

3. Data that can get out — through APIs, not exports. Integration technologies sit at 77% current investment with 75% growth in mid-market IT — the second-highest of any category — precisely because this is where everyone is stuck. A system you can only export from is a system AI cannot work with. What this looks like in practice.

4. A platform that's still supported. On-premises server investment is growing at 0% — nobody is putting money there anymore. And the AI capabilities in modern ERPs are cloud-only by architecture, not by licensing choice. If your system can't be patched, it also can't be the foundation. Which has consequences beyond AI.

The strategic shift worth noticing

Bernstein's mid-year CIO survey found something that should change how you buy:

CIOs do not expect to increase spending on LLM vendors such as OpenAI and Anthropic, "reinforcing the view that enterprises prefer consuming AI through established software platforms rather than building capabilities in-house."

Meanwhile, 40% of enterprise applications are expected to ship with embedded AI agents by the end of 2026, up from under 5% in 2025.

Translation: the winning move is not an AI project. It is AI inside the system you already run. In Business Central that is Copilot included in the licence and agents for payables, sales orders and expenses billed by consumption — switched on over one measured process, with a person approving. No new platform, no new vendor, no new data estate. What that looks like.

The order that works

1. Pick one process with high volume and low judgement. Vendor invoice matching. Order capture from email. Not "AI for the company."
2. Measure it before you touch it. Hours consumed, error rate, cycle time. Without a baseline you will never prove the return, which is how initiative number four gets cancelled.
3. Fix the data that process depends on — only that data. You do not need a clean data estate to start. You need clean data for one process.
4. Switch on what your ERP already ships before buying anything. Most mid-market companies have unused capability sitting in the licence they pay for.
5. Keep a human approving. It is configuration, and it is the difference between a system people trust and one they route around.
6. Measure again at 30 and 90 days. Then, and only then, pick the second process.

That sequence is boring, and it is why it works. Nearly every failed pilot skipped step 3.

How we help

We are not an AI vendor and we are not selling you a platform. We fix the layer underneath: the system of record, the master data, the integrations and the reporting — and then switch on the AI your ERP already includes, one process at a time.

If you want to know what your own data says before deciding anything, that is a fixed-fee engagement that works on any ERP — SAP, Epicor, Infor, NetSuite, Dynamics, QuickBooks or spreadsheets: the Profit Leak Audit. Ten working days, leaks quantified in dollars, and an honest answer on whether your foundation can carry what you want to build on it.

And we use AI in our own delivery work — reading legacy code, cleaning data, generating tests from real documents — which is why the same scope costs 35-40% fewer hours with us. We are not describing something we read about.

FAQ

We already spent a lot on AI. Is it wasted? The tooling usually isn't — the use cases usually weren't the problem. What is missing is the foundation and a single measured process. Most companies can restart with what they already bought.

Do we need a data warehouse first? No, and that assumption is what turns a 90-day project into a two-year one. You need clean data for the one process you are automating. The warehouse conversation comes much later, if ever.

Our ERP vendor says their AI solves this. Embedded AI in your ERP is genuinely the right path — that is the Bernstein finding. But it still reads the data you have. If your stock figure is wrong, Copilot will tell you confidently wrong things.

How do we know whether our foundation is the problem? Ask what questions your system cannot answer today: margin by customer fully loaded, stock that hasn't moved in a year, actual versus standard yield. The list of things it can't tell you is the map of what to fix. Our ERP Health Check does the short version in two minutes.

Is 90 days realistic? For one process, with the foundation in place, yes — that is what the research shows mid-market companies achieving. For "AI across the company", no, and nobody should be selling you that.

Pilots stalling and nobody can say why? Free assessment, no commitment — we'll tell you whether it's the data, the system or the process, and we'll say so if your foundation is already fine.

Bramvia · bramvia.net