You can­not lay­er intel­li­gence on top of chaos

A few years ago, the key ques­tion in every ERP trans­for­ma­tion was: when do we go live? Today our cus­tomers are ask­ing some­thing else: what do we do with all this data once we are live? ERP is still the foun­da­tion, but the real advan­tage comes from what you build on top of it – and increas­ing­ly, that means AI.

At Islet, we have now deliv­ered close to 20 AI projects. Here is what we have learned.

The con­ver­sa­tion often stops at the first question

A com­pa­ny comes to us excit­ed about AI and wants to deploy agents. The first thing I ask is: what does your data look like? Where does it reside, who owns it, is it clean? That is often where the con­ver­sa­tion paus­es. Mas­ter data is incon­sis­tent, sys­tems are siloed and nobody has clear own­er­ship. You sim­ply can­not lay­er intel­li­gence on top of chaos.

Four things an AI-ready data foun­da­tion needs

Mas­ter data qual­i­ty. If your mate­r­i­al mas­ter has dupli­cates or your cus­tomer records are incon­sis­tent, your AI will reflect all of that, just faster. Garbage in, garbage out does not dis­ap­pear when you add a lan­guage mod­el on top.
Fresh data. A data ware­house updat­ed night­ly works for month­ly report­ing, not for oper­a­tional deci­sions. If some­one asks for the sta­tus of a pur­chase order and the answer is 12 hours old, trust dis­ap­pears fast.
Seman­tic lay­er. “Show me rev­enue by region” sounds sim­ple, but the sys­tem must know what rev­enue means in your com­pa­ny and which SAP tables map to which busi­ness con­cepts. With­out that, you get answers that are tech­ni­cal­ly cor­rect but oper­a­tional­ly mean­ing­less.
Data own­er­ship. Tech­nol­o­gy sur­faces prob­lems, humans fix them. Some­one in the busi­ness must be account­able for each data domain, or trust erodes and projects stall.

Bet­ter deci­sions, faster, every day

Think of a Project Man­ag­er who needs to know a project’s com­mit­ted costs. Today they ask finance, wait for hours and often walk into a client meet­ing with­out the answer. When they can ask the ques­tion in Teams and get an answer in 30 sec­onds, cost over­runs are caught before they esca­late. These are not dra­mat­ic AI moments. They are bet­ter deci­sions made faster – and the effect com­pounds over time.

I expe­ri­enced this myself when we recent­ly moved our own ERP to SAP Cloud ERP (Pub­lic Cloud). I expect­ed to lose my usu­al vis­i­bil­i­ty for a cou­ple of months. Instead, three or four days after go-live, I walked into a steer­ing group meet­ing with a clear pic­ture of the busi­ness, thanks to an AI solu­tion our own data team had already built.

AI is not always the answer

Last year we received close to 50 requests ask­ing whether some­thing could be solved with AI. More than half were solved anoth­er way: bet­ter user train­ing, small user inter­face changes or adjust­ments to inte­gra­tions. Our job is to help our cus­tomers suc­ceed, not to sell a par­tic­u­lar technology.

From pilot to production

Com­pa­nies that scale AI start with a real busi­ness prob­lem, not a tech­nol­o­gy ambi­tion. They do not say “we want to use AI”. They say “we lose two days every month rec­on­cil­ing invoic­es, and we want that solved.” That speci­fici­ty gives clear suc­cess met­rics, a moti­vat­ed Busi­ness Own­er and a bound­ed scope.

They also take change man­age­ment seri­ous­ly. AI that nobody trusts or uses deliv­ers zero val­ue, no mat­ter how ele­gant it is. Com­pa­nies stuck in pilots are often chas­ing a demo for the board.

When AI starts tak­ing action

An agent that answers ques­tions is low risk. An agent that cre­ates pur­chase orders or approves invoic­es needs guardrails, audit trails and human check­points. The tech­nol­o­gy is ready, but the prac­tices are not yet stan­dard. Design for account­abil­i­ty from day one, because retro­fitting it lat­er is expen­sive and messy. And rather than just automat­ing today’s process­es, use agents as a rea­son to ask whether those process­es still make sense.

The EU AI Act is less of a block­er than many fear. Most enter­prise use cas­es are not high risk, and the core prin­ci­ples – trans­paren­cy, human over­sight, good data gov­er­nance – are the same ones that make AI trust­wor­thy in the first place.

My three pieces of advice

1. Get your data foun­da­tion right first. Clean mas­ter data and clear own­er­ship are pre­req­ui­sites, not after­thoughts. Every short­cut slows you down later.
2. Take one high-val­ue use case all the way to pro­duc­tion. Mea­sure it, build con­fi­dence and then scale.
3. Invest in your peo­ple as seri­ous­ly as in tech­nol­o­gy. Change man­age­ment and train­ing make the dif­fer­ence between adop­tion and aban­don­ment. Tech­nol­o­gy is often the eas­i­est part.

Lis­ten to the full con­ver­sa­tion on Suc­ceed­cast pod­cast: S1E7 Jan­i­na Luo­to: You can­not lay­er intel­li­gence on top of chaos!

More infor­ma­tion:

#arti­fi­cial­in­tel­li­gence #data #AI #blog #Suc­ceed­cast