Research & Publications
Applied research on agentic AI, done inside a real business system.
Red Digital Factory runs a small research practice alongside its products. We study how autonomous AI agents can operate business software safely for small and medium companies, what it takes to move from narrow assistants towards more general capability, and how ERP and digitalisation actually play out for SMEs in the Gulf and North Africa. RedERP is our laboratory: 30+ modules, real transactions, three languages.
- Focus
- Agentic AI for business operations, AGI-oriented capabilities, ERP and SMEs, Arabic-first interfaces
- Testbed
- RedERP in production — 30+ modules, live tenants in several countries, English, Arabic and French
- Output
- Field notes and documentation today; preprints, papers and datasets as work matures
- Open to
- Universities and labs, graduate students, industry partners for pilots
Research themes
Five questions we keep coming back to.
Each theme grows out of something we ship. The questions below are the ones we are actively working on, not a wish list.
- 01
Agentic AI for business operations
Agents that do work inside an ERP — create an invoice, move stock, book leave, reconcile a statement — under the same permissions, validations and audit trail as a human user. We study how to decompose business tasks for specialist agents, how a routing agent hands work to them, when to ask for confirmation, and how to recover when an action was wrong.
Questions we are working on
- Which ERP actions can be delegated end to end, and which must stay preview-then-confirm?
- How should specialist agents (sales, inventory, finance, HR) share context without sharing permissions?
- What does a useful audit trail of agent actions look like for an accountant?
- How do we measure agent success on real tasks rather than on benchmarks?
- 02
Towards general capability, grounded in operations
We treat business operations as a demanding, measurable environment for increasingly general agents: long-horizon goals, incomplete information, many tools, hard constraints and real consequences. Rather than speculating about artificial general intelligence, we test which general capabilities — planning, memory, self-correction, coordination between agents — actually help, and how to keep them bounded.
Questions we are working on
- How far can planning and memory carry an agent through a month-end close without human steps?
- Which failure modes appear only when several agents coordinate on one dataset?
- How do we bound a general agent's authority inside a company: budgets, scopes, approvals?
- What evaluation protocol makes progress on operational tasks comparable across models?
- 03
ERP and SMEs in the Gulf and North Africa
Most research on enterprise software studies large companies. We study the twenty-person firm in Doha or Tunis: what it costs to adopt an ERP, where implementations stall, which modules matter first, how multi-branch and multi-currency realities shape data quality, and what 'unlimited users from $5 a month' changes in adoption behaviour.
Questions we are working on
- What predicts a successful ERP rollout in an SME with no IT department?
- How does pricing structure change which modules SMEs actually use?
- Which data-quality problems recur in SME ledgers and how much can software prevent?
- How do regional fiscal rules (Qatar, Tunisia) shape product architecture?
- 04
Arabic-first natural-language interfaces
Operating software by asking, in Arabic — including dialects such as Tunisian Arabic — is different from doing it in English. We work on understanding operational requests in mixed-language, right-to-left contexts, on evaluating language models against real business instructions, and on interfaces that read naturally to Arabic speakers rather than translated from English.
Questions we are working on
- How well do current models resolve business intents in Gulf and Tunisian Arabic?
- What is lost when Arabic operational vocabulary is mapped through English?
- How should numbers, dates and currencies behave in mixed-direction agent conversations?
- 05
Trust, safety and evaluation
An agent that can post journal entries needs stronger guarantees than a chatbot. We build evaluation suites from anonymised real scenarios, study guardrails that hold up under adversarial or ambiguous input, and look at cost per action, success rate and human override rate as the metrics that decide whether a feature ships.
Questions we are working on
- Which guardrails reduce harmful actions without making agents useless?
- How can SMEs verify what an agent did without reading logs?
- What is an acceptable cost per successful action for everyday ERP tasks?
How we work
Research that has to survive production.
- 01
Grounded in a live system
Every hypothesis is tested against RedERP and anonymised operational data, with the constraints a real company has: permissions, fiscal rules, latency, cost.
- 02
Measured on real tasks
We report success rates, override rates and cost per action on scenarios drawn from actual businesses, not on synthetic benchmarks alone.
- 03
Published, carefully
Findings are written up as field notes, technical reports and, where they hold, preprints and papers. Customer data never leaves the company; only aggregated, anonymised results do.
Publications
What we have published so far.
Our written output currently takes the form of field notes and documentation from Agentic ERP in production. Peer-reviewed work follows as the results stabilise.
Papers, preprints and technical reports
In preparation.
Our first technical reports — on evaluating specialist agents on ERP tasks and on Arabic operational intent understanding — are being written up. Preprints will be listed here as soon as they are public; write to us if you would like an early read or want to collaborate on a submission.
Field notes and documentation
Published on the RedERP sites. They describe how the agentic layer behaves in production and are the basis for the research above.
Field note · 2026
Agentic ERP for Sales and Inventory
How a routing agent assigns invoices, delivery notes and warehouse quantities to specialist agents, with preview before anything posts.
Field note · 2026
Agentic ERP for eCommerce
Storefront orders routed to specialist agents that update catalogue, stock and invoices in live RedERP, with human confirmation.
Field note · 2026
Agentic ERP for Accounting and Audit
Books, statements and document capture handled by an accounting agent — every posting confirmed by a person.
Field note · 2026
Agentic ERP — the overview
The product page for the agentic layer of RedERP: specialist agents, permissions, audit trail and the preview-then-confirm model.
Documentation · 2026
Agentic ERP documentation
Reference documentation in RedERP Academy: how to enable agents, what they can do, and how their actions are recorded.
Open problems
Questions we would like help with.
If one of these is your topic — as a researcher, a student or a company willing to run a pilot — we would like to talk.
- 01
Confirmation fatigue: how should an agent ask for approval often enough to be safe and rarely enough to be used?
- 02
Delegation boundaries: a principled way to decide which ERP actions are reversible enough for full autonomy.
- 03
Dialect robustness: benchmarks for operational Arabic — Gulf, Tunisian and mixed with French or English — on business tasks.
- 04
SME data quality: how much of the messiness in small-company ledgers can software prevent at entry time?
- 05
Cost-aware agents: choosing models and tool calls per task so that the cost per successful action stays below its value.
- 06
Explainable audit: presenting an agent's chain of actions so a non-technical owner can verify it in under a minute.
Collaborate
Work on these questions with us.
We partner with universities, research labs and companies, and we host graduate students for theses and internships in Doha and Tunis. Partners get a production system to test against and anonymised operational scenarios; we get sharper questions and better methods.
- Joint projects and grant applications
- Master's and PhD theses on agentic AI, NLP or information systems
- Pilots with companies that want to test agents on their own operations
- Datasets and evaluation suites built together, published where possible