The 2026 AI ERP Transparency Index
Scoring 50 ERPs Against the 5-Trait Test for AI-Native Architecture
Status: In progress
Methodology locked May 2026. Per-vendor scoring in progress. Full ranking publishes Q3 2026 with all 50 vendors scored against all 12 criteria, every score sourced, and the public GitHub repo live alongside this page.
What this is
A public, reproducible scorecard of the 50 ERPs that claim AI capabilities in 2026, measured against 12 transparency criteria. Each ERP gets a score from 0 to 100. Every score is backed by a public source. Methodology and data are open so any analyst, journalist, or buyer can verify a score or rerun the entire scoring against new evidence. The Index publishes annually; the next edition lands May 2027.
The 12 criteria
Each ERP is scored 0 to 10 on each of these 12 dimensions, then the 12 raw scores are summed and normalized to a 0 to 100 score. Criteria extend the established 5-trait test for AI-native architecture (see the methodology page) by adding pricing, sovereignty, model disclosure, and architectural recency.
AI writes to the GL directly
Can the AI agent post journal entries autonomously, or does every action need a human to approve and submit?
Pre-AI data model
Was the database schema designed before AI existed, or was it built with AI agents in mind from day one?
Native action layer
Is there a programmatic action surface the AI calls directly, or does the AI translate intent into form-fills?
Pre-write invariant enforcement
Are GL safety checks (debits equal credits, period open, accounts exist) enforced before the write, not patched up after?
Single AI tier (no gating)
Is the AI feature set available to every customer, or gated behind the enterprise pricing tier?
Open documentation of AI behavior
Is the prompt, behavior model, and decision logic documented publicly, or treated as a black box?
Reproducibility of AI outputs
Can a customer replay an AI action and inspect the exact inputs, outputs, and GL entries produced?
Foundation model transparency
Does the vendor name the LLM that powers their AI, and pin a version, or keep the model selection opaque?
Customer-owned data and model isolation
Does customer data ever leave the vendor tenant for AI training? Opt-in, opt-out, never, or shared by default?
Pricing transparency for AI
Is AI feature pricing published on the website, or only available through a custom sales quote?
Open-source posture
Is the AI layer open source so customers can audit it, or is it closed code only the vendor can read?
Year of last meaningful AI architecture change
When did the vendor last revise their AI architecture in a substantive way (not a marketing rebrand)?
How scoring works
- · 0 to 10 per criterion. Each of the 12 criteria gets an integer score from 0 (the dimension is absent or undisclosed) to 10 (the dimension is fully present and publicly documented).
- · Sources required. Every score must be backed by a public source: vendor documentation, press release, RFP response, marketplace listing, or product walkthrough. The source URL is recorded in the dataset.
- · "Unknown" is treated as 0. If a dimension is undisclosed by the vendor and we cannot find a public source, the score is 0 with a footnote. Vendors can challenge this by submitting a public source.
- · Normalization to 0 to 100. Raw scores (out of 120) are normalized to a 0 to 100 final score for readability.
- · Vendors can challenge. Any score can be disputed by filing a public issue on the GitHub repo with a new public source. See the FAQ for the challenge process.
- · Snapshot dated. Each annual edition locks a snapshot date. Architectural changes after the snapshot roll into the next edition, not the current one.
The 5 tiers
The 50 vendors split across 5 tiers by segment and license model. Tier composition is published; the per-vendor roster within each tier is locked at methodology lock and published with the full ranking in Q3 2026.
| Tier | Segment | Count | Description |
|---|---|---|---|
| Tier 1 | Enterprise | 10 | Large multi-entity ERPs serving Fortune 1000 buyers. |
| Tier 2 | Mid-market | 15 | Mid-market ERPs and finance specialists serving 100 to 5,000 person companies. |
| Tier 3 | SMB | 10 | SMB bookkeeping and small-business accounting platforms. |
| Tier 4 | AI-native startups | 10 | The 2024 to 2026 cohort of finance and ERP startups built AI-first. |
| Tier 5 | Open source | 5 | Open-source ERP and accounting projects with public source repositories. |
Total: 50 vendors. Tier roster published with the full ranking in Q3 2026.
Reproducibility
The Index is built to be reproducible by any third party. When the full ranking publishes in Q3 2026, the following will be public:
- ·All scoring data on GitHub at
avansaber/ai-erp-transparency-indexin JSON and CSV. - ·Per-cell sources cited with public URLs.
- ·Methodology document in Markdown, version-controlled in the same repo.
- ·Reproducibility script that re-runs the scoring against any vendor's published documentation, so a third party can verify or contest a score.
- ·Public challenge process: file an issue with a counter-source, get a public response within 14 days, see corrections in the changelog.
Licensed CC BY 4.0. Free to cite, free to fork, free to build on. Attribution required.
Annual cadence
The Index publishes once per year. The 2026 edition methodology locked May 2026; the full ranking publishes Q3 2026. The 2027 edition will publish May 2027 on the same scoring rubric, with the snapshot date and any methodology revisions documented in the changelog.
Between editions, vendor architectural changes get noted but do not retroactively change the current edition's scores. Major mid-year revisions (a vendor ships an architecture change that meaningfully alters a score) may be published as a notable update outside the annual cadence; these are clearly marked and dated.
How to cite
Citation
AvanSaber Inc. 2026. The 2026 AI ERP Transparency Index. Available at https://www.erpclaw.ai/research/ai-erp-transparency-index/
For BibTeX, IEEE, or APA-formatted citation entries, email [email protected].
Frequently asked questions
Why these 12 criteria specifically?
Each criterion has a single answer the vendor can either prove with public evidence or cannot. We avoided subjective dimensions (ease of use, support quality, brand sentiment) because those are not auditable. The 12 dimensions extend the established 5-trait test for AI-native architecture by adding pricing transparency, customer data sovereignty, foundation-model disclosure, and architectural recency. Each adds a separable axis that vendors either publish or do not.
Why does ERPClaw publish this?
Two reasons. First, the category is genuinely under-served. Gartner Magic Quadrant ranks ERP on overall capability, not on AI-architecture honesty. IDC reports cost $5,000 and sit behind a paywall. A free, reproducible, open scorecard is a hole in the market. Second, ERPClaw is a vendor in this category. We disclose that openly, we self-score using the same rubric, and we publish our own score even when it is unflattering. Independent journalists and academics can rerun the methodology and verify or dispute any score.
How can a vendor challenge a score?
File a public issue on the GitHub repo for the artifact (avansaber/ai-erp-transparency-index, published at launch). Include the criterion number, the proposed new score, and a public source (vendor documentation, press release, RFP response, or product walkthrough) that supports the change. We respond within 14 days. Disputed scores get a public correction and a footnote in the next annual edition. Every revision is documented in the changelog.
Will ERPClaw self-score?
Yes. Self-scoring is required for the methodology to be credible. Where ERPClaw scores high (open source, single AI tier, pre-write invariant enforcement) we publish the score. Where ERPClaw scores low (criterion 12 if applicable, since we are a younger product than NetSuite or SAP S/4HANA) we also publish the score and a note about what we plan to do. Honest self-scoring is the trust signal that makes the whole artifact credible.
What if a vendor changes architecture mid-year?
We lock a snapshot date for each annual edition. The 2026 edition will note its scoring snapshot date (likely Q2 2026). Architectural changes published after the snapshot are noted in the next annual edition, not retroactively backfilled into the current one. This keeps the dataset reproducible. Vendors who ship major changes can request an interim re-score, which we may publish as a notable update outside the annual cadence.
Why is the data not on this page yet?
Methodology was locked in May 2026. Per-vendor scoring is in progress and takes 40 to 60 hours across 50 vendors. We could publish partial data and update incrementally, but a half-scored ranking is worse than a fully-scored one because readers and journalists treat it as final. The full ranking publishes Q3 2026 with all 50 vendors scored against all 12 criteria, all sources cited, and the public GitHub repo live alongside the page.
How does this relate to the 5-trait test?
The 5-trait test is the architectural foundation. The 12 transparency criteria extend it by adding dimensions that matter to a buyer beyond pure architecture: pricing transparency, model disclosure, data sovereignty, and architectural recency. Read the methodology page at /research/ai-native-vs-decorated-five-trait-test/ for how the 5 traits map to the 12 criteria.
Subscribe for the full release
Full ranking publishes Q3 2026. Subscribe and we will send the link the day it goes live.