ERP after AI: where structured discipline meets unstructured intelligence

Introduction: the collision is coming

The question most manufacturers should be asking right now is not whether AI is coming for their business processes. It is how they want to be positioned when it arrives.

Gartner projects that 40% of enterprise applications will feature AI agents by 2026, up from less than 5% in 2025.¹ That is not a five-year horizon. It is the next 12 to 18 months. The structural changes to how businesses manage their operations are already appearing in early enterprise deployments, and the pace is, we believe, faster than most organisations are currently accounting for.

This is not an abstract technology trend. It is a business process question that every manufacturer running an ERP needs to answer. Where does AI fit into the way an ERP environment actually works? What changes inside the system, and what sits outside it? What does the interaction between those two domains look like, and where does the real value get created?

Those are the questions this paper maps. The perspective it draws from is specific: I am writing from both sides of the fence. The Lancea team has delivered more than 100 SYSPRO implementations across the manufacturing sector. Through our work alongside Verbeter, we have practical experience with the AI capability that sits at the other end of that boundary. Working from both sides of that boundary is what gives this paper its angle, and what has shaped the framework at its centre.

This is not a technical deep-dive. It is a practitioner’s map of the territory: what AI is doing inside ERP today, what is coming, where we believe the most consequential value sits, and what manufacturers need to be thinking about, and doing, now.

This paper is written for manufacturers already running SYSPRO who want to understand what AI means for their operation. A practitioner’s map of the territory, not a technical deep-dive.

What ERP was built to do

ERP was designed to do something specific, and it does it exceptionally well.

The brief it was built against: record what happened in the business with precision, enforce business rules consistently, protect the audit trail, and maintain master data integrity. Every order, every transaction, every inventory movement: captured, governed, traceable. Structured, deterministic, guardrailed by design.

That is its strength. And it defines its boundary.

What lives inside an ERP environment reflects that design brief faithfully: financial records, transactional flows, inventory positions, production orders, compliance documentation. The structured execution layer of the business. The system records what the business decided and what it did, and enforces the rules that govern how those decisions translate into action.

What does not live inside the system is different in character, and in volume.

How to handle the exception when the preferred supplier goes short on a Monday. Which customer relationship needs a phone call rather than a system notification. The process adjustment the warehouse team put in place two years ago and never formally documented. The know-how embedded in people, in email threads, in WhatsApp messages, in operating procedures that have not been updated since the last restructure.

Across the 100+ SYSPRO implementations, this is what we have consistently found: ERP delivers what it was designed to deliver, and the structured value it creates is real. But structured systems, by their nature, can only capture structured data. Every business also operates with a substantial layer of unstructured knowledge: the judgement, context, and operational intelligence that exists in people, in conversations, in the accumulated experience of teams. That knowledge has always had real value. It has never had a system that could capture it.

For years, bespoke system customisation partially closed that gap for specific clients. The Lancea team has delivered many such customisations, translating institutional know-how into structured rules the ERP could act on. Effective, but narrow. Each one solved a specific problem for a specific business, and the knowledge remained encoded in that particular implementation.

The gap between what enterprise systems capture and how businesses actually operate has a name in knowledge management literature. Nonaka and Takeuchi identified it in 1995 as the tacit knowledge problem: the competitive advantage embedded in people and processes that resists systematic capture.² ERP encodes the explicit layer with precision. The tacit layer (the knowledge that gives the explicit layer its context) has, until now, had nowhere to go.

AI has a fundamentally different path to this problem. Not by encoding knowledge one rule at a time, but by working with knowledge in its natural, unstructured form: as it already exists in the documents, conversations, and processes of the business.

That is what changes the equation.

Where AI sits in ERP today

The AI capability inside SYSPRO today is real, and it is creating genuine value for manufacturers who are using it.

Document Services within SYSPRO delivers AI-powered invoice processing that reduces processing time by up to 70%, with better than 99% accuracy in automated data extraction (SYSPRO, 2025). The Sidekick AI Knowledge Assistant provides real-time contextual help within workflows, reducing the friction that typically comes with navigating complex ERP environments. Anomaly detection on transactional data surfaces irregularities that would previously have required manual review. These are production capabilities delivering measurable impact inside the SYSPRO boundary today.

Beyond runtime capability, AI is also contributing significantly to the implementation process itself: gathering customer context, supporting specification writing, accelerating development. The Lancea team has integrated AI into implementation workflows in ways that have materially improved both speed and quality. The efficiency gains, in our experience, have been substantial, even where the end-user AI experience is still developing.

The investment direction confirms where the industry is heading. Deloitte (2025) reports that 43% of enterprises are actively investing in ERP, up from 35% in 2024. Bain & Company (2025) finds that 78% of IT leaders expect at least some ERP functionality to be replaced or augmented by agentic AI within three years.

What is equally clear is that ERP’s AI capability addresses the structured execution layer inside the system. The opportunity that sits at the intersection with the institutional knowledge layer around ERP, and in connecting those two worlds meaningfully, is a different category of problem. One that lives at the boundary between the ERP and the broader operational context that surrounds it.

In my view, that boundary is where the most consequential AI value will be created. It is also where the Lancea team has been building, a pattern we return to later in this paper.

The shift coming: where AI and ERP will collide

The boundary question

The central question for any manufacturer thinking seriously about AI and ERP is this: how much of the business-process intelligence you need sits inside the ERP boundary, and how much sits outside it, and how do those two domains need to work together?

In practice, the answer is both. Most business processes span both worlds. The production order lives in the ERP. The context that explains why a particular supplier was chosen for that order, or why the delivery window was negotiated the way it was, lives somewhere else entirely. Neither domain can be addressed in isolation, because most of what AI needs to be useful spans both.

Research from Deloitte and Bain & Company points toward what has been described as a core-periphery model: a structured view of where different types of intelligence belong.

The core (financial accounting, compliance, audit trail, master data, transactional guardrails) stays inside ERP. These are the processes where structure, determinism, and governance are the point. They should be in ERP, and ERP will increasingly have AI working within that guardrailed layer.

The periphery works differently. AI agents operate in a flexible application layer that connects to ERP data through APIs without disrupting the core. This is not theoretical. It is the architecture already showing up in early enterprise deployments across manufacturing and beyond.

The design question for every manufacturer is therefore specific and practical: for each significant process step, does the intelligence belong inside the ERP, or does it live outside, connecting in? And is your current implementation built to support that connection?

The institutional knowledge domain

Think about how your most experienced team members make decisions. The ERP handles the structured part: the purchase order is placed, the inventory position is updated, the production order is raised. But around every one of those transactions sits a layer of context that no structured system was ever built to hold. Which supplier relationship can absorb a rushed order this week. Whether this customer needs a personal call before the contract renewal lands. How to sequence production when two priority orders arrive simultaneously.

This is not knowledge that went undocumented. It is knowledge that cannot be fully documented, because it is contextual, relational, and different every time. It lives in people. And it has real business value.

Structured systems capture structured data. This knowledge is unstructured by nature. Until now, those two worlds have been incompatible.

This is not an edge case. In South African manufacturing, in our experience, it is the standard state of affairs.

A significant portion of the AI opportunity sits in this domain, and in my view it is consistently underestimated. Because most business processes span both the ERP execution layer and the broader operational context, the two cannot be addressed in isolation. Focusing exclusively on what AI can do inside the ERP boundary misses the context that makes that intelligence meaningful.

What that domain contains is substantial: the customer relationship intelligence that shapes how a contract renewal is handled, the real-time coordination between plant managers and logistics teams, the expertise experienced people have built about how processes actually perform under operating conditions, and the judgement calls that happen dozens of times a day and never make it into a system record.

I believe AI is the first technology with a credible path to making that layer of value accessible at scale. The ability to surface unstructured operational knowledge (from email threads, from voice conversations, from the coordination that happens in real time across teams) and make it accessible, searchable, and actionable alongside the structured data already in the ERP is precisely what this technology was built to do.

The guardrails for what gets executed still live inside the ERP. What lives outside is everything those guardrails need to know in order to run correctly.

At the furthest end of the capability curve, AI agents begin to take on portions of these roles. Not by removing people from the business, but by lifting them to higher-leverage work the organisation could previously never afford: strategy, relationships, exception management that requires judgement rather than execution.

The interaction model

If both domains carry AI capability, the question becomes how they interact. Getting that interaction right is where the real value gets unlocked.

From what we have seen in early enterprise deployments, the write-access problem is largely solved. ERP business rules are enforced at the API layer, so when an external AI agent wants to execute a transaction, the guardrails hold. The rules that govern what the system can and cannot do remain in the ERP, enforced, auditable. An AI agent acting on ERP data from outside the system boundary works through that governance structure, not around it.

The more complex challenges sit on the read side.

Security. Who is allowed to read which ERP data in which context? As AI agents proliferate and begin querying ERP data at high frequency, data governance becomes a live operational question, not just a technical one. Access controls designed for human users need to be rethought for AI consumers.

Data ownership. As AI reads ERP data at scale, questions about whether that data belongs to the customer or the ERP vendor become commercially significant. This is a live discussion in enterprise software, and I expect it will shape the architecture decisions SYSPRO customers make over the next 18 months.

Pricing and governance. How AI read-access is priced, and by whom, will determine what architectures are financially viable. SYSPRO’s current positioning, which allows access to all SYSPRO APIs without additional module licensing, is a meaningful advantage here. As AI adoption scales, API usage volumes will become a consideration, but SYSPRO’s current posture provides meaningful room to grow.

For South African manufacturers specifically, data sovereignty is worth naming directly. Where AI processes and stores your ERP data is not an abstract question. The offshore data concern, for businesses that need their operational data to remain in South Africa, is a material consideration.

The interaction model between the ERP and the AI layer around it is not only a technical architecture decision. It is a business decision, and it deserves to be made deliberately.

The openness imperative

The ERP implementations I believe will deliver the most AI value in the next five years share one characteristic: they are built for openness.

Easy to read from, easy to integrate with, easy to execute against from external agents. Closed architectures limit what AI can do with your data, and limit your ability to connect ERP intelligence to the broader operational context where a significant portion of the value sits. The two limitations compound each other.

The strategic orientation that makes sense is also the one that preserves ERP’s core value: treat the structured layer (financial accounting, compliance, audit trail, master data) as the authoritative domain for what it does best, and ensure the surrounding architecture allows AI to operate fluidly across both sides of the boundary. Fight for the core where guardrails genuinely matter. Partner for everything outside it.

SYSPRO’s direction is worth noting here. The current positioning (API access without additional module licensing) removes a friction point that constrains AI value in some competing ERP environments. That is a meaningful step toward the kind of openness that will serve customers well as AI matures.

The architectural choices made now (how open the ERP layer is, how well-governed the data access is, how deliberately the interface between the ERP and the surrounding AI layer has been designed) will determine how much AI value can be unlocked over the next five years. These are choices worth making deliberately, before the AI conversation arrives in your business without you.

What this means for your SYSPRO operation

Understanding the landscape matters. But the more pressing question, for a manufacturer running production today, is what to do with it.

Three questions are worth asking now, before the AI conversation arrives uninvited.

Where does our institutional knowledge live?

This is a practical audit, not a rhetorical question. Start with the processes your ERP records, then ask how those processes actually run. Where does execution diverge from what the system shows? Where is the exception-handling know-how concentrated? Which supplier relationships are managed by a single person, and what happens if that person is unavailable?

If three key people left tomorrow, what would SYSPRO not be able to tell you? The answer to that question maps the institutional knowledge domain that AI now has a credible path to reaching, and gives you a starting point for where to focus.

If three key people left tomorrow, what would SYSPRO not be able to tell you?

What AI could operate against our SYSPRO data today?

The conversational interface to ERP data is not a future concept. The capability exists now for your team to query live SYSPRO data in natural language: without building a report, without navigating dashboards, without calling someone who knows where to look (see the AGX section below).

Questions your team currently answers by spending 20 minutes in the system, or by pulling a monthly report that is already two weeks out of date, can be answered in seconds from live data. The starting point is not replacing your ERP or restructuring your processes. It is adding an intelligent layer that makes your SYSPRO data more accessible to the people who need it, and connects it to the operational context surrounding it.

Is our SYSPRO implementation AI-ready?

Data quality matters more when AI is reading it at high frequency. The anomalies and inconsistencies in your master data that your team has learned to work around become significant when an AI agent is operating against that data at scale.

API access and governance are worth reviewing: is your SYSPRO instance configured for external integration? This is not an expensive question to answer, but it is better answered before an AI initiative is underway rather than during it.

These three questions do not require an AI strategy to address. They require a clear-eyed look at where your business stands today, and they position you to make deliberate choices about where to go next.

The South African context

The numbers are worth naming. Forty-nine percent of South African enterprises are now creating dedicated AI budgets (Dell Technologies, Q2 2024). AI and IoT investment in South African manufacturing is projected to contribute R380 billion to manufacturing GDP by 2030 (Tech in Africa, 2025).

The manufacturers who capture that value will not necessarily be those who moved fastest. They will be those who moved most deliberately: those who understood where the real value sat, and built toward it with the right expertise on both sides of the ERP boundary.

The bridge in action: AGX MCP Server

The architecture described in this paper (a conversational AI layer that operates at the boundary between SYSPRO’s structured data and the broader operational context) is not theoretical. The Lancea team has built an early, working example of exactly this pattern.

The AGX MCP Server exposes a conversational AI interface directly against SYSPRO ERP data. What it replaces is the traditional path to ERP insight: navigating dashboards, building reports, writing SQL queries, or calling someone who knows where to look. The AGX layer makes SYSPRO data queryable in natural language, in real time, by the people who need it, without requiring technical intermediaries.

A query like “What were our five slowest production orders last month, and what caused the delays?” used to require a report built over hours, or a conversation with someone who already knew the data. It is now answered in seconds from live SYSPRO data.

This is aligned with where the broader industry is heading. Tier 1 ERP vendors have deployed the same architecture pattern in recent engagements.³ Critical Manufacturing has documented conversational analytics replacing traditional dashboards in manufacturing environments. For Lancea clients, this capability is available now, not on a roadmap.

The next step for any SYSPRO manufacturer who wants to understand what this looks like in practice is not a product decision. It is a conversation about where your institutional knowledge lives, what your SYSPRO data looks like today, and what it would take to put an intelligent layer on top of it.

Lancea’s position is the result of deliberate investment on both sides of the boundary: deep SYSPRO process expertise built across 100+ implementations, and the AI development capability the team works alongside through Verbeter. Practitioner knowledge of how SYSPRO environments actually operate, combined with the technical capability to build the bridge layer. That combination is what makes the AGX MCP Server what it is, and what makes the broader AI opportunity accessible to our clients in a way that is grounded rather than speculative.

Conclusion: position now, not later

The intersection of AI and ERP is not a story about what comes next. It is about what is coming now.

Gartner’s projection (40% of enterprise applications featuring AI agents by 2026) is an 18-month horizon, not a decade-long trend. I expect the manufacturers who navigate this well will not be those who moved fastest. They will be those who understood the territory clearly, and made deliberate choices about where to build.

Three things are true at once.

ERP remains essential. The guardrails, the audit trail, the master data layer: these do not go away. They become more important as the operational context around them grows in complexity.

A significant portion of the AI opportunity sits outside ERP, in the institutional knowledge domain that ERP was never built to capture. Because most business processes span both domains, bridging the two is not optional. It is where the real value gets unlocked.

And the manufacturers who think about this now (across both domains, with the right expertise on each side) will not be caught by this transition. They will be the ones shaping it.

The ERP implementations that unlock the most value will be those built for openness: easy to read from, easy to integrate with, easy to execute against, and connected to the operational context that gives ERP data its meaning.

The Lancea team’s position (deep SYSPRO expertise on one side, Verbeter AI capability on the other) is not accidental. It is where the next five years gets interesting.

AI and ERP will collide. The only question is whether your business is positioned to shape that collision or absorb it.

References

1. Gartner (August 2025). Enterprise AI adoption forecast: 40% of enterprise applications will embed AI agents by 2026, up from less than 5% in 2025. Gartner. ↩
2. Nonaka, I., & Takeuchi, H. (1995). The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation. Oxford University Press. The tacit knowledge concept distinguishes between knowledge that can be codified and transferred (explicit) and knowledge embedded in practice, skill, and experience (tacit). ↩
3. Microsoft Dynamics 365 Blog (November 2025). Dynamics 365 ERP MCP Server: Adaptive and Analytics-Ready. Microsoft. ↩

Further sources

• Altron (October 2025). Altron launches South Africa’s first operational AI Factory delivering enterprise-grade AI infrastructure, platform and services. Altron.
• Bain & Company (2025). Is Agentic AI the Inflection Point for Scaling ERP Transformations? Bain & Company. Source: Bain Technology Maturity Assessment Benchmarking Survey 2025 (n=480).
• Deloitte (October 2025). How ERP is evolving in the agentic AI era. Deloitte. Source report: “AI is capturing the digital dollar. What’s left for the rest of the tech estate?” (Deloitte, October 2025).
• SYSPRO (2025). AI capabilities: Document Services, Sidekick AI Knowledge Assistant, anomaly detection. SYSPRO.
• SYSPRO and Versori (July 2025). Partnership: 25 pre-built agentic integrations for SYSPRO environments. SYSPRO.
• Tech in Africa (2025). AI Adoption in Africa 2025: South Africa Leads, Others Catch Up. Tech in Africa.
• Dell Technologies (Q2 2024). Innovation Catalyst Research: 49% of South African enterprises creating dedicated AI budgets. As reported by: Intelligent CIO Africa (January 2025).