What AI still cannot do in ERP

We sell an AI-native ERP. So read this as a confession rather than analysis, and discount accordingly.

But the fastest way to waste money on AI is to buy it for a problem it does not solve, and the industry is currently very bad at saying which problems those are. Here is our list. None of these are "not yet" — most are structural.

1. It does not know your business

It knows how to build what you describe. That is not the same thing, and the gap catches people out constantly.

"Build me an inventory system" produces something generic and disappointing. "We hold stock in two outlets and a warehouse, transfers need receiving confirmation at the other end, and we count monthly" produces something useful. Same tool, wildly different outcome, and the variable is you.

This will not be fixed by better models. The information is not in the model — it is in your building, in the heads of people who have never written it down. A model cannot infer that your best customer gets 45-day terms because of something that happened in 2011.

What this means practically: the bottleneck moved from consultant availability to your own clarity. If two people in the same department describe the same process differently — normal, in our experience nearly universal — no amount of AI resolves that. You have an organisational disagreement, and it has to be settled by humans before software helps.

2. It does not clean your data

Your records are messy in ways specific to you. Three customer records for the same company. A notes field where someone stored payment terms for a decade. Product codes that changed meaning in 2018 and nobody flagged.

AI genuinely helps with mapping and with spotting probable duplicates. It cannot tell you which of the three duplicate customers is the real one. Only your team knows, and only by looking, one at a time.

This remains the single most reliable way to blow a timeline, and it is the most under-quoted line in every proposal you will receive — ours included. Anyone quoting you a six-week ERP is quoting you the build. Ask them about the migration.

3. It does not fix your politics

The finance manager who does not trust the new numbers. The sales team keeping a shadow spreadsheet. The branch that quietly never migrated.

This has always been the real reason ERP projects fail, and it is a human problem. No model touches it. If anything, AI makes it worse: you can now out-build your organisation's capacity to absorb change. Shipping a module in an afternoon is easy. Getting fourteen people to change how they work on Monday is exactly as hard as it was in 1995.

4. It does not make third-party APIs stable

When a marketplace changes its settlement format, generated code does not save you. Somebody has to notice, understand the change, and fix the integration. That is an ongoing relationship with someone else's roadmap, and it is work forever.

This is why we are cautious about promising integrations that do not exist yet. An integration is not a feature you ship once; it is a commitment to keep something working while another company changes it underneath you.

5. It does not remove the need for human judgement on anything that matters

AI can prepare a reconciliation, a tax submission, a payment run. A human should still sign it.

This is not conservatism. Generative systems produce plausible output, and plausible is not correct. A model that misreads a figure does not flag uncertainty the way a person hesitating does — it presents the wrong number with the same confidence as the right one. That failure mode is specifically dangerous in finance, and it is the subject of AI hallucinations in financial data.

The mitigation is not a better model. It is structural: thresholds, approvals, audit trails, and a named human above a certain number.

6. It does not scale down the hard problems, only the tedious ones

A useful way to sort work: is this tedious or is it hard?

Tedious means well-defined but laborious — typing invoices, matching POs, building the same form for the fortieth time. AI ate this. Genuinely, completely, and that is a big deal.

Hard means the difficulty is in the deciding — what should our approval threshold be, is this customer worth the credit risk, should we carry this SKU. AI can inform these. It does not decide them, and the ones where it appears to decide them are usually the ones where you have quietly accepted a default you never examined.

Most software marketing conflates these two categories deliberately. Automating the tedious is enormously valuable. It is not the same as automating the hard, and buying the first while believing you bought the second is how disappointment happens.

7. It does not stop you building a mess — it accelerates it

This is our own failure mode and we would rather name it than have you find it.

When anyone can create a module in an afternoon, you can end up with forty half-finished modules, three competing versions of the same workflow, and nobody who owns any of it. The old world's change-request queue was slow and infuriating and it did do something: it forced a moment of thought and gave one person a view of the whole.

Remove the queue and you need to replace what it provided — deliberately, with permissions, ownership and review. That is the subject of the new shadow IT. It is not free, and a vendor telling you governance comes for free is telling you they have not run this at scale.

What it genuinely does

Balance, since the above is unrelenting.

It removed the translation layer between how you work and how software must be configured. That was the single most expensive thing in business software for thirty years, and it was the reason ERP never reached small businesses. Removing it is not a marginal improvement. It changes who is allowed to have software that fits them.

It is very good at reading, matching and explaining — documents, exceptions, anomalies, "why is this number what it is". These are real hours in real finance teams.

And it made being wrong cheap. That sounds minor. It is the biggest one. Almost every pathology of traditional ERP — the blueprint phase, the change-request queue, the big-bang go-live — existed because changing your mind later was expensive. Make that cheap and the pathologies dissolve.

How to use this list

As a filter. When a vendor demos something, ask which category it is in.

Is it removing tedium? Probably real, probably valuable, buy it.

Is it claiming to remove judgement? Look much harder. Ask what happens when it is wrong, who finds out, and how fast.

Is it claiming to know your business? It does not. It knows what you told it, and the quality of what you get back is the quality of what you said.

The honest pitch for this whole category is narrower than the marketing: AI will not run your business, but it will stop your business being run by the limitations of your software. That is a smaller claim. It is also, we think, the true one — and it is worth more than the oversell.

Common questions

Can AI work out how my business runs on its own?

No. It knows how to build what you describe, which is not the same thing. "Build me an inventory system" produces something generic and disappointing; "we hold stock in two outlets and a warehouse, transfers need receiving confirmation at the other end, and we count monthly" produces something useful. Better models will not close that gap, because the information is not in the model — it is in the heads of people who never wrote it down.

Will AI clean up our old, messy records?

Partly. It genuinely helps with mapping and with spotting probable duplicates, but it cannot tell you which of three duplicate customer records is the real one. Only your team knows, and only by looking, one at a time. This remains the single most reliable way to blow a timeline and the most under-quoted line in any proposal. Anyone quoting a six-week ERP is quoting the build, so ask them about the migration.

Does a human still need to sign off if AI prepared the work?

Yes, on anything that matters. AI can prepare a reconciliation, a tax submission or a payment run, but generative systems produce plausible output, and plausible is not correct. A model that misreads a figure presents the wrong number with the same confidence as the right one, rather than hesitating the way a person does. The mitigation is structural: thresholds, approvals, audit trails, and a named human above a certain number.

How do I tell a genuine AI claim from marketing?

Sort what you are shown into two categories. Is it removing tedium — well-defined but laborious work like typing invoices or matching POs? That is probably real and worth buying. Is it claiming to remove judgement, the deciding part? Look much harder, and ask what happens when it is wrong, who finds out, and how fast. Is it claiming to know your business? It does not. It knows what you told it.

Related: what is an AI-native ERP and AI forecasting for business inventory, which is the same argument aimed at one feature.