ERP AI budget line items: what to put on the sheet before finance asks
ERP AI budget line items
Also searched as
- what to put in the budget for an ERP AI project
- AI project budget template manufacturing CIO
- hidden costs of ERP AI
- capex vs opex for on-prem AI
Short answer
Budget ERP AI in eight groups: infrastructure, software and model licences, integration, data preparation, security and compliance, people and change, run cost, and contingency plus exit. The costs that surprise teams are data cleanup, internal time and ongoing evaluation, not the model itself.
Applies to: Any ERP with a pilot or first production use case. Figures are structures and ratios to fill with your own quotes, not prices.
Build the budget in this order
- 1Fix the scope first: use cases, number of users, read-only or write-back, on-prem or private cloud. A budget without scope is a wish list.
- 2List infrastructure: GPU servers or reserved cloud GPU, storage for vector indexes and logs, network and backup. Decide capex versus opex with finance before quoting.
- 3List software: model serving stack, retrieval or vector database, orchestration, observability, and any commercial assistant licences. Open-weight models have no per-token fee but have serving and evaluation cost.
- 4List integration: ERP connectors, API or read replica setup, single sign-on, role mapping, and document source connections. Add partner or vendor time at their rate card.
- 5List data preparation: master data cleanup, document ingestion, metadata, and test question sets. This is usually the most underestimated group.
- 6List security and compliance: threat modelling, penetration test, data protection assessment, audit evidence work, and legal review of contracts.
- 7List people: internal project owner, business subject experts, training, support, and ongoing evaluation. Cost internal time even if it does not hit a cost centre.
- 8Add run cost, contingency and an exit line, then split the sheet into one-off and recurring before submission.
Infrastructure: capex, opex and sizing honestly
For private deployments the biggest sizing driver is concurrent users and context length, not total employees. A pilot for a few dozen planners and buyers can run on a modest GPU server; an entire plant asking questions all day needs more. Ask any supplier to size against your expected concurrent peak and your longest realistic prompts, which include retrieved ERP rows and documents.
Capex versus opex is a finance policy question, not a technical one. Owned GPUs are capex with depreciation and power and cooling in opex. Reserved cloud GPU or a sovereign private cloud is opex. If you are not sure the use case will survive the pilot, opex or short reserved terms reduce stranded asset risk. If you are certain of a multi-year steady load, ownership can be cheaper per query.
Do not forget non-production environments. You will want a test instance of the assistant pointed at an ERP test company or copy, otherwise every change is tested against production data.
Software, integration and the costs that hide
Software lines are usually smaller than expected for open-weight stacks, but check licence terms of each model: some open-weight licences restrict certain uses or large-scale deployment. Have legal read them. If you also license a vendor copilot, list it separately so the comparison is visible.
Integration is mostly time. Typical work includes a read-only database user or API client per ERP, mapping ERP security groups to assistant permissions, SSO, and connecting document stores such as quality procedures or supplier contracts. Partner-led delivery is common; get fixed-scope quotes for the pilot and time-and-materials caps beyond it.
Hidden costs: ongoing evaluation (a test set of real questions rerun when models or ERP versions change), prompt and retrieval tuning, user support, and ERP upgrades that change table or field names. Budget a recurring line for these rather than hoping they vanish.
Budget sheet skeleton (fill with quotes) A. Infrastructure one-off | recurring B. Software + models one-off | recurring C. Integration one-off | recurring D. Data preparation one-off | recurring E. Security/compliance one-off | recurring F. People + change one-off | recurring G. Run cost (power, support, evaluation) recurring H. Contingency + exit one-off
Contingency, exit and how to present it to finance
Carry contingency as an explicit line, not hidden padding. For a first deployment a range of 15 to 30 percent is a reasonable conversation starter, adjusted for how much is unknown. Finance respects an honest line more than inflated unit prices.
Include an exit line: the cost to switch off, export logs and indexes, and return or decommission hardware. Designs on open-weight models and standard retrieval stores have low exit cost compared with a platform where prompts, fine-tunes and workflows live in a proprietary format. That is a real budget argument for portability.
Present the sheet with three columns: pilot, first production, and steady state. Boards and finance leaders are more comfortable approving the first column while seeing that the others exist and are bounded.
Common pitfalls
- !Budgeting only GPUs and licences and forgetting data cleanup, which often equals or exceeds integration.
- !Counting internal subject expert time as free. If it is not on the sheet, it is not protected on the calendar.
- !Sizing by headcount rather than concurrent users and context length.
- !Ignoring ERP upgrade effects. A SyteLine, SAP or NetSuite release can change fields and break retrieval mappings.
- !Assuming per-token pricing is the whole cost of a hosted model. Egress, logging, evaluation and security review are separate.
- !No test environment line, so changes are validated against production.
How an ERP-grounded AI assistant handles this
A grounded assistant such as ERPray can help with the data preparation estimate itself: it can query the ERP for counts of items with missing descriptions, vendors without payment terms, or open orders with zero-cost lines, giving you evidence-based scope for the cleanup line rather than a guess. It does not set your budget, but it replaces some of the guesswork in the riskiest line.
Frequently asked questions
What share of the budget is the model itself?
For open-weight deployments, usually a small share. Most spend goes to integration, data preparation, security work and people time. That is why comparing vendors on model price alone misleads. Compare total cost over three years for the same scope, including run cost and evaluation.
Should ERP AI be capex or opex?
Follow your finance policy and your certainty. Owned GPU hardware is capex plus running cost. Reserved or sovereign cloud capacity is opex. If the use case is unproven, shorter opex terms limit stranded assets. If load is steady and long-lived, ownership can be economical. Decide with your controller before requesting quotes.
How much contingency is reasonable?
For a first project many teams carry 15 to 30 percent as a visible line, higher when data quality is unknown. The right number depends on how much of the scope is still undiscovered. Show it explicitly with what it covers, such as data fixes and ERP version changes.
Do we budget for ongoing evaluation?
Yes. Keep a test set of real questions with known good answers and rerun it after model changes, ERP upgrades or retrieval changes. It is a modest recurring effort, but without it quality drifts silently and trust drops. Treat it like regression testing for any integration.
Related
How Much Does ERP AI Actually Cost?
A scoped pilot against one ERP module typically runs 15,000 to 50,000 USD; a single-department production deployment runs 50,000 to 150,000; a multi-module enterprise rollout runs 150,000 to 500,000 or more, plus ongoing hosting and usage costs. Native vendor copilots are usually priced per active user per month on top of existing licensing, not included free.
CIO briefingHow to present ERP AI to the board: what to include and what to leave out
A board does not need a model tour. It needs one decision, a bounded first step, the risks and who owns them, a cost envelope, and the evidence you will bring back. Structure the session as decision, exposure, pilot, money, and proof, in about eight slides.
CIO briefingStaffing an ERP AI Team: What Roles You Actually Need
Most mid-market organizations do not need a dedicated ERP AI team at pilot stage - one ERP-literate technical owner working part-time with an implementation partner is enough. A dedicated team becomes worthwhile once you have more than two or three live AI use cases in production, at which point you need a named owner for prompt/data quality, one for integration and access, and a business-side owner for adoption.
CIO briefingIs Your ERP Data Ready for AI? A Readiness Checklist
Most ERP data is ready enough to start a scoped AI pilot immediately; full data-quality remediation is not a prerequisite. A handful of specific gaps - duplicate item or customer masters, inconsistent units of measure, missing descriptions, and orphaned records - will visibly degrade AI answers and are worth checking before the pilot, not after.
CIO briefingNegotiating AI add-ons with your ERP vendor at renewal
Treat the AI add-on as a separate decision from the renewal, and ask for four things in writing: unbundled pricing with consumption caps, no training on your data, full API and data access regardless of the AI SKU, and clean exit terms. Use a credible alternative as leverage, not as a threat.
CIO briefingERP Upgrade or AI First? A CIO Decision Framework
In most cases, add a grounded AI layer on top of your current ERP first, because it proves value in weeks and shows exactly what an upgrade would need to fix. Reserve a full ERP replacement for cases where the platform itself is end of support, unsupported, or structurally blocking the business, not simply because it feels dated.
AI for ERPAI for Infor SyteLine and CloudSuite Industrial
Add grounded AI to Infor SyteLine or CloudSuite Industrial: natural-language answers, agents over IDOs and ION, on-prem or CloudSuite deployment.
AI for ERPOn-Prem AI for CloudSuite Industrial, Without the Multi-Tenant Cloud Move
Add generative AI to CloudSuite Industrial without moving to Infor's multi-tenant cloud. On-prem private LLM options for CSI, with governance and audit built in.
Stuck on ERP AI (any ERP)?
Talk to engineers who work inside ERP AI (any ERP) every week, and who build private AI that answers these questions from your own ERP data.