Inside my new research paper: 60 AI capabilities, 24 reference implementations and nine experiments on building an AI-powered ERP
Cars already sense the road, plan a route, steer and brake on their own, and hand control back to the driver when it matters. Business software has not caught up.
Surveys report that 78 percent of organisations used AI in 2024, and nearly nine in ten use it today. Yet only about two in ten have scaled AI agents. In most companies, AI still sits beside the ERP as separate chatbots and tools. Those tools cannot see the company's data. They cannot act through its transactions. And they are hard to govern.
I see this every week. In Doha I build and customise Odoo, ERPNext and Zoho, and as a freelance engineer I design ERP and AI systems for businesses across Qatar, the UAE, Europe and India. People still re-type supplier invoices, chase stock figures by phone and wait days for reports that answer simple questions. The gap is not intelligence. It is connection.
So I wrote a research paper about closing that gap properly. This article walks through what I designed, how I tested it and what I learned. Links to the full paper are at the end.
What Is a Self-Driving ERP?
The Self-Driving ERP is an autonomous enterprise system. It is an ERP in which artificial intelligence is part of the product, not an add-on. It keeps everything an organisation expects from an ERP: one database, integrated modules, approval workflows, access rights and an audit trail. On top of that, it builds 60 AI capabilities into ten modules, all connected through one shared AI driving system.
The ten modules are Sales and CRM, Purchase, Inventory, Accounting, HR and Payroll, Projects, Manufacturing, Helpdesk, Reporting and BI, and a Developer Studio for the engineers who build and maintain the system.
Inside each module, AI shows up in five consistent ways. Smart capture fills a form from a PDF, photo or voice note. Inline suggestions show a forecast or recommendation next to a quantity, price or date, with a short reason. Ask the ERP answers questions about any record, policy or report, with citations. An approval inbox holds the actions agents propose, each with its risk score and evidence. And a daily briefing tells each user what changed, what needs attention and what the system did automatically.
What Self-Driving Cars Teach Us About ERP
A self-driving car is not one clever algorithm. It is a set of components that work together. I mapped eleven of them to an ERP, and each pairing became a design requirement.
Cameras, radar and lidar become perception: OCR, vision and speech models that read invoices, site photos and voice notes. The vehicle data bus becomes connectors and event streams. HD maps become the knowledge layer, built on retrieval-augmented generation and knowledge graphs. The driving computer becomes the cognitive core, a shared pool of large language models, small language models and classical machine learning.
The route planner becomes the agent orchestrator. Steering and brakes become tool calls through the Model Context Protocol into the ERP's own transactions. Collision avoidance becomes a risk gate. Driver takeover becomes human approval with an explanation. The dashboard becomes chat and voice in Arabic and English. Safety systems and the black box become governance, and telemetry becomes operations.
The car industry also gave me a maturity model. Its six levels of driving automation translate directly to business processes. At Level 0, people do all the work. At Level 2, the ERP drafts complete transactions and people approve each one. At Level 3, it executes low-risk actions within set limits and hands over otherwise. At Level 4, it runs a whole process and people handle exceptions. Level 5, full automation, is not recommended for financial processes.
The level is set per process, not per company. Accounts payable intake might run at Level 4 while hiring stays at Level 1.
Under the Hood
The architecture has an ERP core and five AI layers, plus two planes that span everything.
The ERP core holds business objects, workflows and access rights. Perception turns documents, images and audio into structured facts. Knowledge and memory store and retrieve company knowledge with hybrid search. The cognitive core reasons and predicts. The agent runtime plans and acts through tools. The interaction layer is how people reach the system. The governance plane handles guardrails, prompt-injection defence, personal data redaction and a tamper-evident audit ledger. The operations plane handles evaluation, drift monitoring and deployment.
These layers are connected by a loop: sense, understand, reason, decide, act and learn. This loop is the main difference between an AI-powered ERP and an ERP with an assistant attached. An assistant waits for questions. A Self-Driving ERP senses events and closes the loop from data to action and back to learning.
Everything can be built with open-source tools on a single server, on top of the Odoo framework. The paper includes 24 reference implementations, from the ERP connector and hybrid search to the MCP tool server, the risk gate, the audit ledger and a Docker Compose deployment.
Putting It to the Test: Nine Experiments
An architecture paper is easy to write and hard to trust. So I ran nine reproducible experiments on synthetic ERP and software operations data. Every number can be regenerated from one script with fixed random seeds.
Demand forecasting
A gradient-boosted model cut forecast error to 15.0 percent WAPE, compared with 17.3 percent for a four-week moving average and 21.3 percent for using last week's sales. For a product that sells 100 units a week, the typical miss falls from about 21 units to about 15.
Anomaly detection in the general ledger
Across 20,000 journal lines with 200 planted anomalies, an isolation forest reached an AUROC of 0.997, against 0.814 for a common three-sigma audit rule. It placed 70 percent of anomalies in a 200-item daily review list, against 55 percent for the rule.
The breakdown mattered more than the average. The rule caught every inflated amount but missed every off-hours posting. The model caught most round-sum payments to new partners but only a quarter of off-hours postings. The practical lesson is to combine both: rules for what auditors already know, and machine learning for patterns nobody wrote a rule for.
How much autonomy is safe?
This was the experiment I cared about most. I simulated 10,000 agent actions. The most common real-world rule, approving anything under a value limit, automated 85 percent of actions but let 48.9 percent of the error cost through unreviewed. Many costly mistakes are irreversible or low-confidence rather than large.
A risk gate that also weighs reversibility, data sensitivity and model confidence automated 67 percent of actions while leaving only 12.2 percent of error cost unreviewed. That is about four times less exposure for one-fifth less automation.
Safety and audit
A guard for model-generated SQL blocked 23 of 24 unsafe queries. The one it missed used a file-reading server function, which showed why the guard must be one layer among several. A hash-chained audit ledger detected all 1,000 tampering attempts and wrote about 150,000 entries per second on one CPU core.
Production software
Over 30 days of per-minute service metrics, a seasonality-aware isolation forest detected 23 of 24 planted incidents with about 2.9 false-alarm episodes per day. Static thresholds caught 19 and raised about 15.1 false alarms per day. For an on-call engineer, that is the difference between a pager that is trusted and one that is ignored.
The experiments also caught my own mistakes
Two results surprised me. Naive hybrid search made exact record look-ups worse, dropping from 100 percent to 78.7 percent, until simple query routing brought it back to 93.3 percent. And my first invoice check used a 1 percent relative tolerance, which let about half of swapped-digit errors through. Switching to an absolute tolerance of 0.05 QAR caught every error with no false alarms.
Finding these flaws before going live is exactly why experiments matter.
What It Means for a Business
Independent research already shows large gains from AI on routine knowledge work. In published studies, customer support agents resolved 14 percent more issues per hour, professionals finished writing tasks 40 percent faster, and developers with an AI pair programmer finished a task 55.8 percent faster.
In the paper I worked through an example for a trading company with 40 office staff. With conservative assumptions for supplier invoices, quotations, management questions and inventory, the system saves about 170 hours a month, roughly one full-time employee, worth about 14,673 QAR. With a server costing about 2,000 QAR a month to run and a one-off implementation of 60,000 QAR, the investment pays back in about 4.7 months.
Every company should recalculate with its own numbers. But the shape of the result holds: when AI removes routine work from processes that run hundreds of times a month, a modest investment can pay back within a year.
How It Compares
SAP, Oracle and Microsoft are all adding AI agents to their ERPs, mostly for large companies and mostly in their own clouds. Odoo 19 and Zoho have added AI features too.
The Self-Driving ERP aims at the many smaller companies those products do not serve well. It can run privately on one server with open-weight models. It works in Arabic and English. It uses one risk gate across every module. And it shows its evidence openly.
It also has clear limits. Large vendors have decades of accounting, tax and localisation rules, certified compliance programmes and thousands of production deployments. My experiments use synthetic data and test the deterministic parts of the system. End-to-end runs with language models on a live Odoo instance are the next stage of this work.
Where to Start
If you want to try this in your own company, start with one process that is frequent, document-heavy, measurable and low-risk when wrong. Accounts payable invoice intake and internal policy questions usually fit. Autonomous payments and hiring decisions do not.
Measure your baseline first. Start the agent in suggestion mode. Move up one level of automation at a time, and only when the measured error rate at the current level is acceptable. Autonomy should be earned, not granted.
Final Thoughts
The data that matters to a business lives in its ERP, while the AI people use lives in a browser tab that knows nothing about their orders, stock or customers. The Self-Driving ERP is my attempt to connect the two safely.
My first paper measured how AI helps people build software. This one asks what an ERP looks like when AI helps people run the business and the software behind it. Wherever I could, I tested the answer rather than just argued it.
Read the Full Paper
- Zenodo (DOI): 10.5281/zenodo.23880861
- Academia.edu: Self-Driving ERP on Academia
- Medium: Read this article on Medium
- My earlier study on AI app builder platforms: 10.5281/zenodo.19280643
Let's Connect
If you work with ERP systems or AI in business, I would love to hear from you. Which processes in your company would you automate first? Where would you never let an AI act alone? What would you like to see tested next?
You can reach me at thowfeeksalim002@gmail.com or visit my website at thowfeek.in.
This article is based on my preprint "Self-Driving ERP: An Autonomous Enterprise System Integrating Sixty AI Capabilities, from LLMs and RAG to AI Agents" (November 2026), which has not been peer-reviewed. The experiments use synthetic data, so absolute numbers will differ on real ERP data. Vendor features and market figures were current at the time of writing.
Top comments (0)