Five weeks after a free chat interface made language models a boardroom subject, every enterprise software vendor is rewriting its roadmap presentation. By the time the spring conference season ends, finance and operations leaders will have sat through a dozen slides promising that generative AI inside ERP changes everything. Some of it will be real and modestly useful. A good deal of it will be existing functionality with a new label. Almost none of it will be available on the version you are running. What follows is a filter for the year ahead.
Most of what will be announced this year is either a text box on an old feature, or an old feature with a new label
Both can be worth buying. Neither is a strategy. The problem is that the slide looks identical in either case, so the buyer needs a taxonomy the vendor has no incentive to supply. There are five distinct things currently being presented as one. A language interface over existing data. Ask a question in plain English, get a report you could already have run. Genuinely useful for occasional users who never learned the report builder. Changes nothing about what the system knows. Generation of text artefacts. Product descriptions, job postings, customer correspondence, first drafts of documentation. This is the part that actually became possible in the last year, it is cheap to deliver, and it will be the bulk of what ships first. Prediction relabelled. Demand forecasting, lead scoring, payment-date prediction, churn risk. Much of this is statistical machinery that has been in the product for five years and has just been moved under a new menu heading. Worth having; not new. Autonomous transaction processing. The demo where the system posts the entry, approves the payment or releases the order without a person. This is a slide, not a product, and anyone selling it into a finance function should be asked how the control environment survives it. Implementation and configuration assistance. Code and configuration generated for consultants building on the platform. Real, already working, and worth understanding — because the productivity gain accrues to your implementation partner rather than to you unless somebody negotiates it.
| Claim | Question to test |
|---|---|
| Language interface | Which authorised data and reports can it access? |
| Text generation | Who verifies the draft before use? |
| Prediction | What model, baseline and error limits apply? |
| Transaction action | Which permissions and audit controls govern action? |
| Implementation aid | Who reviews code and how is effort priced? |
Qualitative summary of this article's source text, not a measured outcome or performance estimate.
Four questions that collapse the slide
Ask these of every announced feature and the conversation becomes concrete very quickly. What data does it use? Your transactions only, your transactions plus vendor-held aggregate data, or a general model with no knowledge of your business? Each answer implies a different risk conversation and a different contract. Where does inference run? Inside your tenant, in the vendor's cloud, or at a third-party model provider the vendor has contracted? The third answer means a subprocessor most buyers have not assessed. What happens when it is wrong? Not the accuracy statistic — the workflow. Who sees the suggestion, who accepts it, what is recorded about the acceptance, and what does the audit trail show a year later. How is it priced? Included in your existing subscription, a separately licensed module, or consumption-metered. Ask specifically what happens at renewal, because "included during preview" has a well-established habit of becoming a line item. A vendor who answers all four crisply is shipping something. A vendor who answers the first three by returning to the demo is showing you a research project.
The contractual window is now, and it closes quietly
The agreements signed this year will govern how AI functionality is priced and how your data is used when that functionality actually arrives. Two clauses matter more than the feature list: whether your transaction data may be used to improve the vendor's models or benchmarks, and whether functionality currently described as included can be repriced as a separate module at renewal. Both are negotiable today, while the vendor's own commercial model is unsettled. Neither will be negotiable in three years, when the answer has become standard.
What is actually worth funding this year
Three things, none of which are AI projects. Master data cleanup on the records these features will read — customers, suppliers, items, cost centres. Every prediction and every natural-language answer is only as good as the dimensions it groups by, and the work has independent value if nothing ships. A decided route for enterprise-grade access, with contractual terms on training and retention, so that the teams already using consumer tools on real documents have a sanctioned alternative. One narrow pilot with a baseline measured first. Not to prove AI works, but to build organisational skill in specifying a task precisely and verifying output rigorously.
Practical Guidance for AI Use Case Assessment
- Classify every vendor claim into interface, generation, prediction, autonomy or implementation aid.
- Ask when the feature shipped and what changed in the code, to expose relabelled functionality.
- Establish where inference runs and which third parties are involved.
- Get pricing at renewal in writing, not pricing during preview.
- Refuse blanket consent to your transaction data improving vendor models.
- Fix master data first, because every feature reads it.
- Keep a human preparer for anything that posts to the ledger.
- Judge the roadmap on your version, not on the vendor's latest release.
The Regional Angle
Three factors change the calculation for groups operating in this region. The first is the localisation queue. Regional buyers already live at the back of it, waiting for value-added tax handling, withholding, wage protection files, end-of-service calculations and now electronic invoicing. The Saudi integration phase began five days ago, and finance teams are consumed by it. Artificial intelligence features will reach this market late, in English first, and configured for accounting conventions that are not ours. That is not an argument against them; it is an argument against letting a persuasive AI roadmap influence a platform decision that must be made on localisation readiness, which is verifiable today. Buy the system that files correctly this year. Treat the AI roadmap as an option with no delivery date attached. The second is who actually decides. In most regional implementations the platform choice is heavily shaped by the implementation partner, and the first genuinely working AI in enterprise software is the code and configuration assistance those partners are beginning to use. That productivity gain is real and it lands entirely on the supply side. If your partner's consultants are producing configuration faster, your statement of work should reflect it — through fixed-price deliverables rather than day rates, or through an explicit review of estimates mid-project. Very few clients here will think to ask, and the ones who do will pay materially less for the same implementation. The third is the state of group master data, which is worse here than the general discussion assumes and for structural reasons. Regional groups hold multiple entities across free zones and mainland jurisdictions, created for ownership, licensing and tax reasons rather than operational ones, each with its own chart of accounts and its own supplier and customer records. The same supplier commonly exists five times with five codes. Any feature that answers questions at group level will read that and produce confident nonsense, and the executive asking the question will have no way to tell. Deduplicating supplier and customer masters across entities is the single highest-return preparation available, and it pays for itself in procurement leverage and credit control whether or not a single AI feature ever ships.
The objection worth taking seriously
The strongest objection is that the correct response to all of this is to wait. Enterprise software AI has over-promised in every previous cycle. Whatever eventually proves useful will arrive in a future release you are already entitled to, at a price you will negotiate then rather than now. Spending this year on "AI readiness" is a consultant's invoice in search of a business case, and the organisations that ignore the entire subject for eighteen months will lose nothing. That is largely correct about the features. Early adopters of enterprise AI modules this year will pay beta prices for preview functionality and spend real effort on things the vendor will fix for free in two releases' time. But the two items with genuine urgency are not features. One is contractual, because the data-use and repricing terms you accept this year will still be operating when the functionality matters. The other is master data, which takes longer than anyone estimates, has value independent of AI, and is the reason most of these projects will underperform their demos. Doing those two things is not an AI programme. It is housekeeping that happens to have a deadline.
Common Questions
Should we delay an ERP decision until AI features are clearer?
No. Decide on localisation, functional fit, total cost and implementation capability. Every serious vendor will ship comparable AI functionality within two years, so it is unlikely to be the differentiator anyone expects.
Are the natural-language query features worth paying extra for?
For organisations with many occasional users who never adopted reporting tools, often yes. For teams with a competent analyst, the payback is smaller than the demo suggests.
Can these systems post transactions automatically?
They can be configured to propose transactions. Allowing them to post without human approval breaks segregation of duties and should not survive an internal audit conversation.
What should we expect over the next twelve months?
Expect every major vendor to make an announcement by mid-year, with general availability arriving late in the year or slipping into the next, and confined to narrative, search and drafting rather than to transactions. Expect separate pricing to become the norm after an initial period of inclusion. Expect procurement questionnaires to acquire an AI section asking where inference runs and whether customer data trains models. And expect at least one prominent implementation to be described as an AI failure when the actual cause was master data nobody cleaned — which is the most predictable event of the coming year, and the cheapest to avoid.
AI Use Case Assessment — we sort the vendor roadmap into what ships, what is relabelled and what is fiction, then tell you which two preparations are worth funding this year.