Evaluations fail in a predictable way. Everybody attends the same demonstration, it goes well because the vendor chose the data, and the questions that would have exposed the gaps were never asked because each role assumed somebody else was covering them.

These are organised by role so that between four people, the whole surface is covered.

CFO
CIO
Controller
Program Leader
Shared Test
Evidence
Exit

Questions for the CFO

The CFO's concern is whether this produces a defensible benefit and what the exposure is if it does not work as described.

What is the measured baseline: A good vendor asks for your numbers before quoting a benefit. One that leads with a benchmark is guessing.
What happens when it is wrong: Specifically: how is an incorrect proposal detected, corrected and prevented from recurring.
What does internal audit see: The control design should be demonstrable rather than described, ideally before you sign.
What is the exit: If this is removed in two years, what is left behind and what does the team go back to.
A specific question answered in seconds, sourced

Ask whether it needs AI at all

The sharpest question a CFO can put to a JD Edwards AI vendor is which of your processes they would not use AI for. A vendor who answers everything is selling, not advising.

AI is not a solution to every problem. There are other solutions out there. AI could be a solution, but it may not be the best, most efficient or effective one.

Adam Swoverland, on the AgentCFO podcast

Use the questions below to find out where the vendor draws that line. The honest ones will name work a scheduled rule already does better.

Questions for the CIO

The CIO carries the integration risk and the security posture, and both have concrete answers that separate serious products quickly.

What is the integration surface: Which endpoints and orchestrations are used. A short documented list is good; a vague answer is not.
Whose permissions apply: If everything runs under one shared account, your role design has been flattened.
What is installed inside JD Edwards: The right answer is nothing. Custom objects are a permanent retrofit obligation.
Where does data go: Which content leaves the boundary, to which provider, under what retention and training terms.

Questions for the Controller

The controller asks the questions that expose whether a product understands finance, and these are the ones a demonstration is designed to avoid.

Each has a specific correct answer, and hearing a general one tells you the product is a chat interface rather than a finance system.

How are voids excluded: Reversed and void records left in totals is the most common source of wrong numbers.
How are ledger types separated: If actual and budget are not distinguished unprompted, they will get mixed.
What happens to an unmatched invoice: A specific answer indicates process knowledge; a general one indicates a canvas.
What does the approver see: Full proposal with source rows, or a summary. The difference is the strength of your control.
Checklist split between automated steps and human judgement

Questions for the JD Edwards Program Leader

This role carries the long term operability and is best placed to spot the answers that will create work in two years.

Which orchestrations does it call: A named list you can review and control, not a general description of write back.
What happens at a Tools Release: Whether revalidation is a documented short exercise or an unknown.
How are environments separated: Distinct credentials and whitelists for test and production.
What does it do to performance: Query patterns, aggregation behaviour and whether heavy analysis can be pointed at a replica.

The One Question Everybody Should Ask Together

If only one thing gets tested, make it this one, because it separates the field faster than any other and it is entirely objective.

Ask for a transaction created in a test environment, through a supported orchestration, with a named approval record and the resulting document and batch number returned. Everything else in the evaluation is context for this.

Test environment: Real structures and real data shapes, not demonstration data.
Supported method: Through an orchestration, not a direct table write or screen automation.
Approval record: A named person confirmed, and the record shows what they saw.
Reference returned: Document and batch number, so the action is traceable afterwards.

"Evaluations fail because four people watched the same demonstration and each assumed somebody else was asking the hard question."

From question to routed, approved transaction

Answers That Should Concern You

Certain responses are worth noticing not because they are disqualifying on their own but because together they describe the shape of what you are buying.

Discovery phase first: If the product cannot touch a test environment until after a scoping engagement, the engagement is the product.
You author the rules: If matching and coding rules must be written before anything works, the learning is being done by people.
Custom objects required: An upgrade obligation you will carry indefinitely.
Approval is optional: If posting without named confirmation is configurable, approval is a setting rather than a control.

How to Run It

Two weeks, one painful process, your worst real cases, and a measured before and after. Four roles each asking their own questions, and one shared test that ends with a posted transaction.

That produces a decision based on evidence rather than on impressions, and it is considerably cheaper than three months of briefings.

Finance results after ChatFin deployment

Ask by Role, Test Together

ChatFin expects these questions and answers them concretely: a documented orchestration list, nothing installed inside EnterpriseOne, sessions inheriting your role design, enforced approval, and a posted transaction in a test environment.

If only one thing gets tested, make it the posted transaction with an approval record.

Book a Demo

Frequently Asked Questions

What is the single most important evaluation question?

Ask for a transaction created in a test environment through a supported orchestration, with a named approval record and the resulting document and batch number returned. It is entirely objective, it takes an afternoon, and it separates the field faster than any other question because most of the market can read but not write. Everything else in the evaluation is context for this.

What should a controller ask that a demonstration will not cover?

Four specifics. How reversed and void records are excluded from totals, which is the most common source of wrong numbers. How actual is separated from budget by ledger type. What specifically happens to an invoice that matches no purchase order line. And what the approver sees, meaning the full proposal with source rows or a summary, because that difference is the strength of your control.

Which vendor answers should raise concern?

Four, and they matter as a pattern rather than individually. A discovery phase required before the product can touch a test environment, which means the engagement is the product. Rules that you must author before anything works, meaning the learning is being done by people. Custom objects required inside EnterpriseOne, which is a permanent retrofit obligation. And approval being configurable rather than enforced.

Related Articles

ChatFin Compared With Other JD Edwards AI Vendors
The Business Case for AI Agents in JD Edwards Finance
How to Sequence Your JD Edwards AI Agents
The CFO AI Readiness Checklist