ERP or AI Agent: Where the Boundary Runs Through Mittelstand Operations in 2026

At 09:14 on a Wednesday morning, an incoming customer email lands in the shared mailbox of a German Mittelstand industrial supplier. The customer wants to bring forward a delivery by two weeks, mentions a new ship-to address in passing, and asks for an updated price for an extended quantity. The order desk reads the email, opens SAP, looks up the customer, checks the credit limit, opens the contract in SharePoint, walks down the hallway to production planning, comes back, drafts a reply, copies the new ship-to into the master data, and books a follow-up call with sales. Elapsed time: 47 minutes for a request the ERP could have answered in seconds.

That is not an ERP failure. It is an ERP being used for what it was never built to do. The ERP recorded everything that finally got into it - the new order, the new ship-to, the credit check, the price. What it cannot do is read an email, weigh it against a contract, ask production planning, and draft a reply. That is the work of an AI agent on top of the ERP - and confusing the two is what wastes the next round of Mittelstand IT budget.

This article is not about whether ERP is dead (it is not). It is about understanding precisely what each tool is built for, where the boundary runs, and what the hybrid architecture actually looks like in a real Mittelstand operations function in 2026.

Why This Question Hits Hardest in 2026

The ERP is the most expensive piece of software the typical Mittelstand company runs. SAP accounts for about 48 percent of the German ERP market¹², Microsoft Dynamics for roughly 10 percent, with Sage, Infor, abas, and proALPHA filling much of the rest. The investment runs in millions and the implementation in years. So when AI vendors pitch “the AI-native ERP that replaces SAP”, every operations leader has to make a real call - and the answer matters more than for any other software category.

The forces shaping the question right now

• The SAP 2027 deadline - SAP confirmed mainstream ECC support ends 31 December 2027¹¹. Only about 39 percent of ECC customers have so far licensed S/4HANA. Migrations take 18 to 36 months and 69 percent of DACH enterprises report insufficient internal capacity. AI agents that remove manual work during the migration window are now business-critical for many Mittelstand IT functions.
• SAP Joule and Microsoft Dynamics Copilot - The big two ERP vendors have wired agentic AI directly into their platforms. SAP rolled out the Cash Management Agent, Production Planning Agent, Order Reliability Agent, and Bid Analysis Agent across 2026². Microsoft Dynamics 2026 release wave 1 ships agentic AI in sales, service, finance, supply chain, commerce, HR, and ERP³. These features are useful but bounded to the data the vendor sees.
• Cross-system reality of Mittelstand operations - Most exceptions in finance, sales, procurement, and service cross system boundaries. The order is in SAP, the email is in Outlook, the contract is in SharePoint, the supplier confirmation is a PDF, the receivable history is in DATEV, the customer asked a follow-up in WhatsApp. Embedded ERP agents reason on a fraction of that context.
• The data gap is the real bottleneck - Gartner forecasts that 60 percent of AI projects will be cancelled by end of 2026 due to inadequate data foundations¹⁰, and Gartner predicts over 40 percent of agentic AI projects will be cancelled by end of 2027 due to escalating costs and unclear value⁵. The ERP is usually where the critical data lives - or fails to live cleanly.
• The Mittelstand is under-invested - Bitkom finds 41 percent of German firms now use AI actively, up from 17 percent the year before⁷, but the spend per firm remains far below the global average. The instinct is to consolidate, not to multiply tools - which makes “ERP vs agent” the dominant framing for most decision makers.

The German Mittelstand context

German operations carry specific weight that makes the ERP versus AI agent question different from a US or Asian conversation. The decision has to fit five concurrent realities, not just one technical preference.

• SAP dominance plus a long tail of specialists - SAP S/4HANA, ECC, and Business One sit at the centre of most mid-market plants. Around them, German-built specialists like abas, proALPHA, IFS, and Sage handle vertical needs. Any AI agent strategy has to handle this heterogeneity.
• DATEV and Lexware on the finance side - DATEV alone handles the bookkeeping for the majority of German SMEs, often parallel to the ERP. Cross-system agents are essential to integrate these realities.
• GoBD and audit burden - German tax law (GoBD) requires immutable, auditable records of every business transaction. The ERP is the audit trail. AI agents must respect it - logging decisions, never bypassing posting rules, deferring to the ERP for the official record.
• Betriebsrat co-determination - German works councils have rights over any technical system that monitors employees. Operations agents (finance, sales, procurement) stay clear of this. HR-adjacent agents need explicit Betriebsrat consultation.
• Skills shortage and demographics - Fewer than 40 percent of German firms can find AI-qualified workforce¹⁰. The architecture has to fit a small in-house IT team, not a 200-person enterprise IT function. That favours agents on top of existing ERP over multi-year transformations.

What an ERP Actually Is

An ERP (Enterprise Resource Planning system) is the software layer that records and orchestrates the business: finance, procurement, sales, inventory, production planning, HR, payroll. It encodes how the company books, buys, sells, ships, hires, and pays. Its job is to enforce defined business rules, hold the canonical master data, and produce the audit trail.

The defining characteristics

• Transactional core - Posts every business event according to chart of accounts, document types, and approval workflows. The ERP is the source of truth for finance, inventory, and order status.
• Master data of record - Customer master, vendor master, material master, employee master, chart of accounts. Every operations decision eventually references the master data the ERP holds.
• Rule-based workflow - Three-way match, credit hold, release strategies, approval thresholds, payment terms. The rules are predictable, deterministic, auditable.
• Module integration - Finance, sales, procurement, inventory, production, HR all live in the same data model. A goods receipt updates inventory, financial accounts, and material valuation in one transaction.
• External integration - DATEV, banks, payment platforms, e-invoicing portals, customs systems, tax authorities. ERPs are the system that integrates with the outside world in a regulated, audit-aligned way.
• Audit and compliance trail - GoBD, IFRS, HGB, e-invoicing mandates. Every document, every change, every reversal is tracked. Compliance-critical for German operations.
• Reporting and consolidation - Closing the books, profit-and-loss, balance sheet, working capital, cash flow. The ERP holds the numbers and the dimensions to slice them.

Where the ERP is genuinely brilliant

• Audit-grade transaction posting - GoBD-compliant, IFRS-compliant, immutable trail of every business event. Replicating this with an AI agent would be reinventing financial audit infrastructure.
• Master data as single source of truth - Customer 100023 is the same customer in finance, sales, service, and reporting. Without a master-data system of record, nothing else works.
• Three-way match and procurement workflow - Purchase order, goods receipt, invoice. The ERP enforces the rule. Done right, it prevents most accounts-payable fraud and process drift.
• Period close and consolidation - The boring, structured work of producing financial closing in a controlled, auditable, repeatable way. The ERP is built for it.
• Regulatory integration - E-Rechnung (mandatory in Germany), DSGVO data subject requests, tax filings, customs declarations. ERPs hold the integrations the regulator accepts.
• The 24/7 backbone - Orders post overnight, payroll runs on a schedule, ASN files flow in and out automatically. Deterministic, structured, no human in the loop per transaction. That is exactly right for the work it handles.

Where the ERP hits its design limits

• Cross-system reasoning - The ERP sees ERP data. It does not see customer emails, supplier PDFs, contract clauses, Slack messages, Excel sheets the controller maintains, or the WhatsApp message the sales rep got at 22:00.
• Unstructured input - An incoming order arrives as a PDF in shared mailbox. A supplier announces a delay in free text. A customer complaint reaches the service inbox as a screenshot. The ERP cannot read any of it.
• Exception handling beyond defined rules - The order needs a discount the price list does not cover. The vendor is duplicated under two numbers. The customer asks for a credit term not in the master data. The ERP routes the case to a human queue. There is no “just figure it out” option.
• Document-heavy back-office work - Customer complaint responses, supplier dispute letters, dunning letters, internal audit packs, ad-hoc management reports. The data is in the ERP. The drafting is not the ERP’s job.
• Cross-departmental orchestration - When sales, finance, logistics, and procurement all touch the same issue, no single ERP transaction coordinates them. The orchestration falls to whoever happens to be on shift.
• Inbound communication - Email, WhatsApp, customer portals, partner portals, shared mailboxes. The volume is large, the work is repetitive, and the ERP is structurally blind to it.

What an AI Agent on Top of ERP Actually Is

An AI agent in an operations context is not a chatbot, not a dashboard, not an ERP replacement. It is a goal-directed reasoning layer that reads from the systems already in place (ERP, CRM, DMS, email, supplier portals, customer portals, spreadsheets), reasons about what to do toward a defined operational goal, and either acts within bounded authority or escalates to a human with full context.

The defining characteristics

• Goal-directed, not script-directed - You define the outcome (process the inbound order, resolve the credit hold, draft the supplier complaint, reconcile the open item, propose the master-data fix). The agent figures out how to get there using available tools.
• Reads across systems - ERP, CRM, DMS, email, supplier portal, customer portal, contracts, spreadsheets, DATEV. The agent reasons about the full context, not just ERP data.
• Handles unstructured input - PDFs, emails, free-text notes, contract clauses, screenshots, shared mailboxes. About 80 percent of enterprise data is unstructured, and most of it is invisible to the ERP.
• Sits above the ERP, not beside it - The ERP stays as system of record. The agent reads from it and writes back through proper APIs (BAPI, RFC, OData, IDoc, REST, MCP). No bypassing, no replacing.
• Respects audit and posting rules - Agents call the same posting interfaces a human user would. The ERP enforces three-way match, release strategies, GoBD-compliant logging. The audit trail stays intact.
• Escalates with context - When the agent encounters a decision outside its authority or confidence, it escalates to a human with what it found, what it tried, and what it recommends. Reviews are 5 to 10 times faster than handling the raw exception alone.
• Pricing by use case, not by user - Agents are typically priced per active use case (per agent, per process), not per ERP seat. The economics fit a small team handling high volume.
• Portable across ERP editions - A well-built custom agent reads through stable interfaces (BAPI, OData, IDoc, REST). It works on ECC, S/4HANA on-prem, S/4HANA Cloud, Business One, Dynamics 365 BC, abas, and proALPHA. Migration to a new ERP does not require rebuilding the agent.

What an AI agent on top of ERP is not

• Not a chatbot for the back office - Operations teams do not need a chat window. They need the work to flow. Agents act behind the scenes through APIs rather than competing with the ERP UI.
• Not a replacement for the ERP - The system of record stays. Agents add the reasoning layer above. Replacing an ERP is a multi-year project nobody should undertake on the promise of AI.
• Not RPA on top of the ERP - RPA scrapes ERP UIs and breaks on every update. Agents integrate through APIs, reason about goals, and handle exceptions natively. The architectural difference matters most precisely on top of ERP.
• Not magic - Gartner predicts more than 40 percent of agentic AI projects will be cancelled by end of 2027⁵, mostly due to inadequate governance and unrealistic scope. Operations AI needs human-in-the-loop checkpoints, audit logs, and rollback paths.

Six Differences That Matter in Operations

These differences are not philosophical. They show up in how each system behaves when the customer email arrives at 22:00, how procurement evaluates them, and what happens during the auditor visit.

Difference 1: What it is for

The ERP exists to record and orchestrate the business according to defined rules. It is the system of record for what happened. The AI agent exists to reason about decisions across the data the ERP records plus context the ERP cannot see. The ERP owns the truth. The agent owns the reasoning.

Difference 2: How it makes decisions

The ERP decides through configured rules. If the customer is in good standing and the credit limit is X and the discount is on the price list, post the order. The rules are predictable, deterministic, auditable. The AI agent decides through goal-directed reasoning. Given the goal of processing the inbound request correctly and the current state across systems, what is the best next action. Different paradigm, different outcomes on novel situations.

Difference 3: What data it sees

The ERP sees what users post into it and what other ERPs and integrated systems send into it. Structured, transactional, audit-trail-aligned. The agent sees the ERP data plus everything else: email, contracts, portals, spreadsheets, DATEV, the CRM, the shared SharePoint folder. The decision quality depends on the breadth of context, not just on the algorithm.

Difference 4: How it handles exceptions

The ERP routes exceptions to the user or stops the workflow. The exception queue grows in proportion to the variability of the operations environment - and Mittelstand operations are variable by nature. The agent handles exceptions inline: it reasons about what happened, proposes a resolution, and either acts within defined authority or escalates with context. The clerk reviews the agent’s analysis, not the raw exception.

Difference 5: Audit and traceability

The ERP is the audit source of truth. Every financial posting, every inventory movement, every master-data change has a record in the ERP. The agent logs every decision it makes - what it found, what it considered, what it decided, what it escalated - but the business record itself remains in the ERP. The agent creates an additional layer of decision audit, which is often valued during regulatory reviews and EU AI Act compliance.

Difference 6: How fast it deploys

An ERP greenfield deployment takes 9 to 18 months and 250,000 to several million euros in a Mittelstand context. A first focused AI agent on top of an existing ERP takes 8 to 12 weeks and 40,000 to 120,000 euros. The economics flip: ERP is a long-cycle investment in infrastructure; agents are a short-cycle investment in specific outcomes. The pace of improvement is fundamentally different.

Where Each One Wins

The framing “ERP or AI agent” sets up a false binary. The real question is process by process: which one fits which job. The same Mittelstand company will rely on its ERP for some work and lean on an AI agent for other work - and the boundaries run cleanly between them when both are deployed deliberately.

Where the ERP wins clearly

• Financial posting and closing - Booking transactions to the correct accounts, running depreciation, calculating accruals, closing the period. Deterministic, regulated, audit-critical. ERP territory.
• Master data of record - The single canonical version of the customer, vendor, material, employee, chart of accounts. Every other system points to the ERP master.
• Three-way match in procurement - Purchase order, goods receipt, invoice. The ERP enforces the rule deterministically. Built for it.
• Inventory accounting and valuation - Standard cost, moving average, FIFO. The ERP holds the numbers and the rules to calculate them.
• Payroll calculation and posting - Tax tables, social security, payroll runs. Regulated, structured, periodic.
• External regulatory integration - E-Rechnung output, ZM declarations, tax filings, customs filings, DATEV interface. ERPs hold the integrations the regulator accepts.
• Period close and consolidation - Multi-entity, multi-currency closing. Structured, repeatable, audit-aligned.
• Sales order release and credit hold workflow - Rule-based credit check, release strategies, blocking and unblocking by exception code.
• Audit and tax documentation - GoBD requires immutable, audit-ready records of every business event. The ERP is the answer.

Where an AI agent wins clearly

• Inbound order processing from email and PDF - The agent reads the email or attached PDF, identifies the customer and ship-to, validates pricing, drafts the sales order for review or posts it directly within authority. Saves hours per day on the order desk.
• Master-data exception handling - A new customer or vendor arrives with conflicting information. The agent searches existing masters for duplicates, reads the source documents, drafts the correct entry, escalates if confidence is low.
• Invoice exception handling and dunning - When invoices fall off the three-way match, the agent investigates: missing receipt, price mismatch, quantity discrepancy. Pulls supplier history, drafts the response, proposes the resolution.
• Credit hold resolution - Order is blocked. Agent reads customer history, payment patterns, outstanding receivables, contract terms, sales rep notes, and drafts a release recommendation for the controller in seconds.
• Customer complaint and inquiry response drafting - Reading the customer email, finding the relevant order, invoice, delivery in the ERP, drafting a fact-based response with the right facts attached.
• Supplier dispute and SCAR letter drafting - Compiling order, delivery, quality, and payment history, generating supplier corrective action requests with attached evidence.
• Cash collection and dunning drafting - Reading the receivables ledger plus customer correspondence history, drafting tailored dunning letters that reflect the actual relationship rather than a template.
• Audit prep across ERP, DMS, and email - Pulling the right records for an upcoming statutory or tax audit. Hours instead of weeks of manual collation.
• Procurement RFQ analysis - Reading multiple supplier quotes in different formats, normalising terms, comparing on a like-for-like basis, drafting the award recommendation.

Where neither is the right answer

• Broken processes - If the underlying operations process is poorly designed, no software fixes it. Map and improve the process first, then deploy the right tool.
• Strategic decisions - M&A, capex commitments, pricing strategy. Human judgement at the executive level.
• Tiny volume, low-value processes - Three transactions a year for a low-margin product. Spreadsheet plus checklist beats automation.
• Politically blocked processes - When the friction is between departments, not in the work. Tooling makes politics worse, not better.

The Cost Comparison Done Honestly

Comparing ERP cost to AI agent cost is the wrong frame. They solve different problems. The honest comparison is the cost of the gap each one leaves and the work each one removes.

What an ERP actually costs in the Mittelstand

• Licence model - Per named or concurrent user, per module. SAP S/4HANA Cloud typically lands between 1,200 and 3,000 euros per user per year for full users; Dynamics 365 Business Central full user around 1,000 to 1,500 euros per user per year. Add modules (advanced manufacturing, advanced finance, supply chain, HR) on top.
• Implementation - The dominant cost line. 250,000 to 1,500,000 euros for a Mittelstand greenfield ERP depending on complexity, number of entities, and modules. Often takes 9 to 18 months. Implementation services routinely cost 1.5 to 3 times the licence fees.
• Annual maintenance - 18 to 22 percent of licence value per year for support and updates, plus internal admin time.
• Integration to peripheral systems - 50,000 to 250,000 euros for DATEV, CRM, e-invoicing, banking integration. Often ongoing as systems update.
• User training and rollout - 50,000 to 200,000 euros depending on user count, modules, and country coverage.
• Periodic upgrade projects - Major version upgrades every 5 to 8 years, or the SAP ECC to S/4HANA migration project, which lands at 30 to 100 percent of original implementation cost.

What an AI agent on top of ERP actually costs

• Platform / agent fee - 2,000 to 8,000 euros per month per active operations use case, depending on complexity, volume, and integration scope.
• LLM inference cost - Cents per task. Even at thousands of tasks per day, total inference cost stays modest - low hundreds to low thousands of euros per month per use case.
• Implementation - 40,000 to 120,000 euros for a focused first deployment. Mostly process mapping, ERP API integration, validation against historical cases.
• Integration maintenance - Lower than RPA or peripheral ERP integration because agents tolerate API drift better. Typically 5,000 to 15,000 euros per year per active use case.
• Monitoring and feedback loop - 0.2 to 0.5 FTE per active agent for review, correction, and continuous improvement.
• No per-user scaling - The number of operations seats does not change the agent cost. Per-process pricing fits per-process value.

Three-year total cost on a representative use case

Take an inbound order processing use case: orders arrive by email and PDF from a long tail of customers. Today, two order-desk clerks spend 40 percent of their time entering these orders into SAP, chasing missing data, and resolving exceptions. The ERP alone cannot help (it can only post orders that arrive in a structured form). The agent reads emails and PDFs, drafts orders, resolves exceptions, and escalates where confidence is low.

Where the cost comparison flips against an agent

• Stable, structured-input operations - When all orders arrive through EDI from a small set of customers and exceptions are rare, the ERP alone handles it cleanly. An agent adds cost without value.
• Very low volume processes - When the back-office work being saved is a few hours per month, agent platform fees do not amortise.
• Bad master data - When the ERP master data is wrong or missing, the agent cannot reason reliably. Fix the data first; deploy the agent on the cleaner subset.
• Pure transactional posting - The work the ERP is built for - posting goods receipts, calculating depreciation, running the close. Agents do not belong in this layer.

The ERP + Agent Architecture That Works

Almost every Mittelstand company that adopts operations AI agents lands in a hybrid: ERP stays as system of record, agents sit on top. The architecture is not a compromise; it is the only configuration that respects GoBD, EU AI Act audit duties, and the operational reality of cross-system Mittelstand workflows.

The four-layer operations stack

• Layer 1: External inputs and channels - Customer emails and portals, supplier emails and portals, EDI, e-invoicing inflow, WhatsApp, fax-replacement scans, shared mailboxes. The unstructured edge of operations.
• Layer 2: AI agents - Read inbound traffic, reason about context, draft drafts and proposals, write back to the ERP through APIs or escalate to humans. The new reasoning layer.
• Layer 3: ERP and core business systems - SAP, Dynamics, abas, proALPHA, plus DMS, CRM, DATEV, banking. The system of record for transactions, master data, and audit.
• Layer 4: External output and obligations - E-Rechnung output, regulatory filings, financial statements, payments, customs declarations. The audited downstream of operations.

The most common hybrid patterns we see in the Mittelstand

• ERP + agent for inbound order processing - Customer emails and PDFs come in. Agent reads, validates against the customer master and price list, drafts the sales order in SAP or Dynamics, escalates on exception. Saves hours per day.
• ERP + agent for accounts payable - Vendor invoices arrive as PDFs and e-invoices. Agent matches against PO and goods receipt, posts the clean cases, drafts the exception responses, proposes the master-data fixes for the rest.
• ERP + agent for credit-hold resolution - ERP routes the hold. Agent reads customer history, ageing, payment patterns, contract terms, recent correspondence, drafts a release recommendation for the controller.
• ERP + agent for customer complaint response - Inbound complaint goes to the service inbox. Agent identifies the order, delivery, invoice, drafts a fact-based response, prepares the credit note if applicable.
• ERP + agent for procurement RFQ analysis - Multiple supplier quotes arrive in different formats. Agent normalises, compares, drafts the award recommendation for the buyer to review.
• ERP + agent for master-data hygiene - Agent reads new customer or vendor requests, searches the existing master for duplicates, validates against external sources (Handelsregister, USt-IDs), drafts the correct entry.
• ERP + agent for audit prep - Statutory audit, tax audit, customer audit. Agent pulls the relevant documents from ERP, DMS, and email, drafts the audit pack.

How the responsibilities split in a hybrid architecture

The Decision Framework

Use this framework job by job to decide whether your existing ERP is enough or whether an AI agent layer adds clear value. The output is a defensible recommendation you can present to operations leadership and IT in the same meeting.

Step 1: Classify the job

• Execution job - Post a transaction, calculate a number, run a workflow, enforce a rule. Structured, repeatable, audit-critical. ERP territory.
• Reasoning job - Decide what to do when reality breaks the plan, draft a document from cross-system data, process an inbound communication, propose a master-data fix. Cross-system, judgement-heavy, exception-prone. Agent territory.

Step 2: Score the data scope

• ERP data only - The decision needs only what the ERP already holds (transactions, master data, configured rules). ERP rules can probably handle it.
• Cross-system context needed - The decision needs ERP plus email, contracts, supplier portals, customer messages, spreadsheets, DATEV. Agent territory.

Step 3: Score the variability

• Low variability - The case looks the same every time. ERP rule logic is sufficient.
• High variability - The cases differ in small but meaningful ways. Rule maintenance becomes a permanent burden. Agent reasoning is more durable.

Step 4: Score the cost of delay

• Low cost of delay - The work can wait for the clerk or controller to handle it tomorrow. No urgency.
• High cost of delay - Customer SLAs, payment deadlines, decision windows that close. Speed matters, and an agent that reasons in seconds beats a queue waiting for human attention.

Step 5: Read the decision matrix

How Superkind Fits

Superkind builds custom AI agents that sit on top of existing Mittelstand ERP, CRM, DMS, and document systems. We do not replace your ERP. We build the agent layer that does what your ERP was never built to do - reason across systems, handle exceptions, draft documents, escalate with context.

Core capabilities for operations environments

• SAP-native integration - SAP S/4HANA, ECC, Business One. Agents read through BAPI, RFC, OData, IDoc, REST, and SAP CPI. Portable across editions and through the S/4HANA migration window.
• Microsoft Dynamics integration - Dynamics 365 Business Central, Finance and Operations. Agents read through MCP servers, OData, and the Dynamics REST API.
• Mittelstand ERP coverage - abas, proALPHA, Sage, Infor CloudSuite, IFS. Stable interfaces for the long tail of German Mittelstand ERPs.
• DATEV and finance peripheral integration - DATEV interface, banking integrations, e-invoicing portals, customs systems. Agents bring the finance perimeter into one reasoning context.
• Document intelligence - Reads invoices, contracts, customer emails, supplier announcements, RFQ documents. No template configuration required.
• Cross-system orchestration in one workflow - A single agent reads ERP plus email plus DMS plus CRM plus DATEV and produces a single decision or document. No swivel-chair work for the clerk or controller.
• Posting respects audit and release strategy - Agents call the same posting interfaces a human user would. Three-way match, credit hold, release strategies, GoBD logging all stay intact.
• Human-in-the-loop checkpoints - You define which decisions require approval and at what confidence threshold. Agents escalate with context rather than acting silently. Critical for EU AI Act alignment and high-stakes operations decisions.
• Audit trail layered on top of ERP - Every agent decision is logged: what it found, what it considered, what it decided, what it escalated. The agent log complements the ERP audit trail rather than replacing it - aligned with EU AI Act Article 12 logging duties.
• EU deployment and DSGVO alignment - Agents run on EU cloud or your own infrastructure. Data does not leave your defined perimeter.
• 8 to 12 weeks to first production deployment - From process assessment to live operation on a focused first use case. No multi-year transformation.

Superkind vs alternative paths

The 90-Day Plan

This plan covers running the decision framework on a single operations use case, validating ERP data, deploying the agent in limited scope, and reaching first production value. Use it to align your operations leadership, finance, and IT.

Weeks 1 to 3: Use case selection and data audit

• Pick three candidate use cases - Each one with measurable pain (clerk hours, missed payment terms, audit prep weeks, customer SLA breaches). Document current pain in numbers.
• Apply the decision framework - Score each use case on job type, data scope, variability, and cost of delay. Output: one use case clearly fit for an agent.
• Audit ERP master data quality - For the chosen use case, validate that the data the agent will reason on is clean enough. Most agent failures trace back to bad master data.
• Confirm API access - SAP BAPI/OData/RFC, Dynamics OData/MCP, abas or proALPHA APIs for the data the agent needs. DMS, CRM, DATEV integration confirmed.
• Define success metrics - Hours recovered per week, exception resolution time, audit prep duration, error rate. Numbers measurable in 90 days.
• Brief Betriebsrat where applicable - If the agent touches employee data or HR workflow, start consultation early. Most operations agents stay outside this scope.

Weeks 4 to 8: Build and test

• Process mapping detailed - Document inputs, outputs, decision points, system touches, exception types. The work that makes the deployment succeed.
• Agent build against the process map - Prompt and tool design, integration setup, escalation thresholds, human-in-the-loop checkpoints.
• Test against real historical cases - Pull past examples (resolved by the clerk), run the agent against them, compare outputs.
• Validate exception handling specifically - The hardest cases are the agent’s real test. Confirm escalations deliver useful context.
• Confirm GoBD and DSGVO logging - Audit trail captures what the agent did, why, and what data it accessed.
• Train the team - Clerks, controllers, and sales-back-office staff learn how to review agent recommendations, correct errors, and feed back.

Weeks 9 to 12: Production and learning

• Deploy to limited scope - 20 percent of cases, one customer segment, one business unit. Run in parallel with the existing process.
• Review every escalation and correction - Weekly cadence. What did the agent get wrong? Why? What is the correct answer?
• Measure against baseline - Hours recovered, resolution time, audit prep speed, error rate. If numbers are not moving, diagnose before scaling.
• Expand once metrics validate - Two to three weeks of stable operation at limited scope before scaling to full volume.
• Document lessons for the next use case - Where was the framework right, where was it wrong, what would you do differently.

Related Articles

• SAP and AI Agents: The Pragmatic Path Through S/4HANA, ECC, and Business One for the Mittelstand
• MES or AI Agent: Where the Boundary Runs on the Mittelstand Shopfloor in 2026
• Standard Software or an AI Agent: How the Mittelstand Should Choose Where Software Budget Goes in 2026
• Classic Procurement vs. AI Procurement Agent: Where the Boundary Runs in Mittelstand Sourcing in 2026
• RPA vs AI Agents: What German SMEs Get Wrong About Automation
• AI Bookkeeping with DATEV: How German SMEs Cut Close Time and Errors
• E-Rechnung 2025 with AI: How the German Mittelstand Meets the Mandate

Frequently Asked Questions

Henri Jung

Co-founder at Superkind

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.