← All insights AI and ERP: a field read · 1 of 3

AI will not fix a broken ERP

An agent does not run on the demo tenant. It runs on your data. In most ERP estates the data is barely clean enough for a person to trust, let alone a machine that acts on it without asking first.

A robotic hand pressing a red approve button on a cracked, failing surface

Every buyer I talk to this year is hearing the same pitch. Agents that reconcile your accounts, chase your approvals, answer your employees, and close your books while you sleep. The demo is clean. The slide says autonomous. The count on the slide keeps climbing: hundreds of agents, thousands of skills, live now.

Here is what the pitch leaves out. An agent does not run on the demo tenant. It runs on your data. And most estates I walk into are nowhere close. The data is not clean enough for a person to trust, let alone a machine that acts on it without asking.

What a new analyst does with bad data
  • Notices the numbers look wrong within a week
  • Flags it, walks down the hall, asks somebody
  • Holds the payment until it is confirmed
  • Every one of those is an unpaid control you have been relying on
What an agent does with the same data
  • Reads the field, believes the field, acts
  • Pays the same invoice against two vendor IDs
  • Processes a change for someone who left in March
  • Fast, consistent, and completely confident

An agent is a worker who never questions the record

Give a new analyst a report full of duplicate vendors and stale cost centers, and within a week they start asking why the numbers look wrong. They flag it. They walk down the hall. They hold the payment until someone confirms.

An agent does none of that. It reads the field, believes the field, and acts. Point it at a supplier master with duplicate records and it will happily pay the same invoice against two vendor IDs. Point it at worker data where termination dates contradict payroll history and it will process a change for someone who left in March. The agent is fast, consistent, and completely confident. On bad data, that is not an assistant. That is an error machine with a nicer interface.

The vendors are right that the technology works. What they stay quiet about is the precondition. An agent that acts needs three things your estate probably does not have yet: master data clean enough to trust, workflows disciplined enough to follow, and approval logic explicit enough to encode. Miss any one and the pilot produces confident nonsense.

1Master data clean enough to trustNot a dashboard showing green. A measured error rate on the objects that move money and touch people.Test: what is your duplicate rate on vendors and workers, right now?
2Workflows disciplined enough to followIf the real process routes around the system half the time, the agent will automate the half nobody uses.Test: does the documented path match the logs?
3Approval logic explicit enough to encodeJudgment that lives in someone's head cannot be handed to a machine. It has to be written as a rule first.Test: can you write the exception rule without saying "it depends"?
Infographic: a geological core sample of ERP data layers, from clean surface data down to the field-error layer agents act on
What your agent actually reads. The demo shows the surface layer.

On a recent program, 10% of active worker records were wrong

I am not speaking in theory. On a recent program, 10% of the "active" worker records were duplicates or carried termination dates that contradicted payroll history. Call that a data-quality issue and you have mislabelled it. That is payroll-tax and compliance exposure, and it was sitting inside a system everyone described as clean.

On another program, we found error rates north of 20% in the legacy data that discovery had signed off as ready. And the classic: 400 "critical" legacy reports collapsed to 90 once we checked them against the actual run logs. Nobody asked for the other 310 again. Every one of those reports would have been a surface an agent could read and act on. Most of them were clutter, and some were wrong.

10%Of "active" worker records were duplicates or carried termination dates contradicting payroll history. That is compliance exposure, not a data-quality issue.Field, anonymized
20%+Error rates in legacy data that discovery had already signed off as ready, on a separate program.Field, anonymized
400 → 90"Critical" legacy reports once checked against actual run logs. Every one of the 310 would have been a surface an agent could read and act on.Field, anonymized

The wider research says the same thing in bigger letters. Gartner found that 63% of organizations either lack AI-ready data-management practices or are not sure they have them, and has gone on record that through 2026 a large share of AI projects will be abandoned specifically because the data underneath was not ready. MIT's 2025 field study of enterprise AI put a harder edge on it: about 95% of generative-AI pilots delivered no measurable return to the P&L. The technology was rarely the reason. The inputs were.

The research says it in bigger lettersGartner found 63% of organizations either lack AI-ready data-management practices or are unsure whether they have them, and has gone on record that through 2026 a large share of AI projects will be abandoned specifically because the data underneath was not ready. MIT's 2025 field study put a harder edge on it: about 95% of generative AI pilots delivered no measurable return to the P&L. The technology was rarely the reason.

What "ready" actually means

Readiness has nothing to do with a data-quality dashboard showing green. It is four questions you can answer with a name and a date.

Who owns each master. Not the system that stores it, the person accountable for it. On most programs no single person owns worker, supplier, or customer data end to end, so every team assumes another has it. When the loads fail, and they do, there is nobody to call.

How dirty is it? Measure it, do not assume it. Run the duplicate check. Run the completeness check. Put a real percentage on it before anyone promises an agent will "handle" it.

Do the workflows exist the same way twice. An agent automates a process. If your process lives in three people's heads and changes by region, there is nothing consistent to automate.

Is the approval logic written down. An agent needs to know when it may act alone and when it must stop and ask. If your controls are tribal knowledge, the agent has no rules to inherit, so it either does too much or nothing useful.

The 90 days before you spend a dollar on agents

Before the agent conversation, run a short, unglamorous cleanup. It is the cheapest insurance in the program.

First 30 days: name an owner for every master data domain and give each one a single measured error rate. No owner, no agent. This costs nothing but a decision, and it is the decision most programs skip.

Days 30 to 60: rationalize before you migrate. Kill the reports nobody runs, dedupe the vendors and workers, reconcile the control totals to the dollar. On the programs that go well, finance signs off on data that clears under a 2% error rate on the final mock, and we run at least three mocks to get there. That is the bar an agent should inherit, not a hope.

Days 60 to 90: write the approval logic down. For each process you would hand an agent, state the threshold where it acts alone, the threshold where it routes to a human, and the human who owns the exception. If you cannot write that rule, you are not ready to automate that process, and no vendor can write it for you.

Infographic: a six-rung ladder of data readiness work across three 30-day zones, from naming owners to writing approval logic down
None of this is AI work. All of it decides whether AI works.

None of this is AI work. All of it is the work that decides whether AI does anything but embarrass you in month two.

The cleanup is your job, and no vendor can take it

The reason this piece is hard to hear is that the cleanup is your job, not the vendor's and not the integrator's. The platform ships the agent. The SI configures it. Neither one owns your master data, your process discipline, or your controls. Those were your problem before AI and they are more expensive now, because a person who inherits bad data slows down and a machine that inherits bad data speeds up.

Agents will change ERP work. I am not in the camp that says this is hype. I am in the camp that says the value lands only on estates that did the boring part first, and that is a small club right now.

Do this week: pick your highest-volume transaction, the one a vendor would love to demo an agent against, and pull the real error rate on the data it runs on. One real number, measured this week. If it is above a few percent, you have your Phase Zero, and it has nothing to do with AI.

Sources

Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk," Feb 26, 2025 (63% lack or are unsure of AI-ready data practices; prediction that 60% of AI projects will be abandoned through 2026 without AI-ready data). gartner.com

MIT NANDA, "The State of AI in Business 2025" (~95% of enterprise generative-AI pilots showed no measurable P&L return; 300+ deployments studied). Report PDF; corroboration: Fortune, Aug 18, 2025.

Field numbers (10% duplicate or contradictory worker records, 20%+ legacy error rates, 400 reports to 90, sub-2% final-mock error over at least three mocks) are from 9Nation program experience and the published Workday Lessons Learned material. Clients anonymized to program level.

Let's talk

Find your real error rate before the vendor demos against it.

A short, independent data-readiness read is the cheapest insurance an AI roadmap can buy. Client side, no platform to sell.

Field notes

Get the next lesson in your inbox.

One hard-won program lesson at a time. No cadence promises, no spam.