ERP vendors are not asleep on AI. Serious platforms have been weaving AI into planning, forecasting, control and user experience. AI will not replace a real ERP. It will replace small tools that masquerade as ERP, while making true ERPs faster, smarter and safer.
The myth: "AI will kill ERP"
The claim pops up every few days. A shiny demo shows a chatbot answering a question about stock or margin, then someone concludes, “Who needs ERP?” What those demos skip is everything that makes ERP valuable in the first place: trusted data models, auditable workflows, financial discipline, compliance, and end-to-end process control across sales, purchasing, inventory, production, service and finance.
AI can summarise, predict and recommend. ERP executes, records and proves. One is an accelerant, the other is the engine.
Why ERP remains the system of record
-
Authoritative data: ERP maintains a single source of truth that can be audited end to end
-
Governed processes: Approvals, segregation of duties and compliance are built into ERP flows
-
Financial rigour: Real-time postings, reconciliations and statutory outputs are native to ERP
-
Scale and reliability: High volumes, multi-company, multi-currency and cross-border operations
AI sitting outside this core cannot guarantee integrity. AI inside ERP can enhance it without breaking the chain of trust.
What AI actually changes inside ERP
Practical, high-impact areas where AI adds real leverage:
-
Forecasts that learn
Demand, supply and cash-flow forecasts that adapt to seasonality, promotions and lead times rather than static averages. -
Anomaly and risk detection
Spotting price leaks, suspicious discounts, duplicate vendors, erroneous costs and outlier transactions before they become losses. -
Natural language access
Ask “Why did service response time slip in July?” and get an answer tied to live data, drillable to the document level. -
Document automation
Extracting structured data from invoices, POs, delivery notes and certificates, then validating it against masters and rules. -
Adaptive workflows
Routing approvals by risk score, past vendor behaviour or deal size instead of one-size-fits-all chains. -
Recommendations in the flow
Next-best-offer, parts compatibility, substitute items, dynamic safety stocks, technician scheduling and route optimisation. -
Security intelligence
Unusual login patterns, geo-fencing breaches and permission drifts flagged in real time.
Architecture that works: AI inside the ERP fabric
-
Event driven: Stream key ERP events to AI services, then feed decisions back as native actions or recommendations
-
Feature store: Keep curated, explainable features built from ERP data, not raw, messy logs
-
LLM guardrails: Constrain prompts to ERP schemas, user rights and customer data boundaries
-
MLOps: Version datasets and models, monitor drift, roll back safely, document lineage
Anything less becomes a fragile sidecar that cannot be trusted in audits.
How to evaluate an ERP's AI, in 10 questions
-
Does the AI respect user permissions and branch or company boundaries
-
Can every AI action be traced to documents and postings
-
Is training done on your data silo with explicit consent and isolation
-
Are models versioned with changelogs and rollbacks
-
Can you override, approve or reject AI suggestions with reason codes
-
Are metrics like forecast MAPE or anomaly precision visible
-
Does AI write back to ERP through supported APIs and workflows
-
Is latency acceptable in real operations, not just in demos
-
Can you cap or schedule compute to control cost
-
Are security events from AI logged alongside ERP security logs
Measuring ROI that actually matters
-
Working capital: Fewer stock-outs and overstock, lower ageing, improved turns
-
Conversion: Higher quote-to-order rate from guided pricing and cross-sell
-
Cycle time: Faster procure-to-pay and order-to-cash with automated document flows
-
Service outcomes: First-time-fix rate, response and resolution times
-
Financial hygiene: Reduced write-offs, fewer manual journals, cleaner reconciliations
-
Risk: Lower fraud or leak incidents per period
If a vendor cannot tie AI to these levers, it is theatre.
Risks and how to avoid them
-
Shadow AI: Unconnected tools that scrape exports and invent their own truths. Fix with embedded, governed AI.
-
Data bleed: Mixing customer-specific data with general models. Fix with encryption, tenancy isolation and strict scopes.
-
Prompt drift: Unstable outputs from vague prompts. Fix with templated prompts and constrained retrieval.
-
Cost creep: Models left running without caps. Fix with scheduling, budgets and usage policies.
-
Over-automation: Letting AI book entries blindly. Fix with thresholds and human-in-the-loop.
The Tuhund point of view
We do not position AI as a showpiece. In Tuhund, AI has long powered predictions, system-driven planning, security and communication. The Ruaa bot has been sending smart notifications and reports since 2012. Our current work strengthens three pillars:
-
Trust
Customer data is isolated by design with custom encryption between servers. Common knowledge and customer-specific knowledge never mix. -
Explainability
Every AI suggestion links to the exact ERP records you can drill into. No black boxes. -
Action in the flow
AI outputs enter the same approval layers, audit trails and security groups that already govern ERP activity.
The result is simple. AI does not sit beside Tuhund. It runs through Tuhund. Most of it might be invisible to the end user and that is the real beauty of AI.
Bottom line
AI will not replace ERP. It will expose shallow tools, and it will elevate the platforms that already run businesses with discipline. The winners will be ERPs that embed AI where it moves money, time and risk, without ever compromising the ledger or the law.
