Enterprise AI and agentic ERP: a practical guide

In short: enterprise AI pays off when it works inside your processes: it reads real data such as orders, stock and tickets, acts within set permissions and leaves key decisions to a person. A powerful model cut off from your ERP, CRM and online store stays a generic assistant. Agentic ERP applies the same logic to the core system.

Key takeaways

The value of AI in a company depends less on the model you pick and more on three conditions: reliable data, a well-described process and clear rules on who approves what.
Agentic AI is AI that decides its own steps and uses tools such as the ERP, email and CRM to complete a task, within set permissions.
The first use cases that pay back are repetitive and measurable: orders received by email, order-status and returns support, supplier follow-ups, reading production data.
For recruitment and employee evaluation, the EU AI Act applies its high-risk rules, so human oversight has to be designed in from day one.
Before building an agent, try a simple automated workflow: it costs less, behaves more predictably and is often enough.
The main risk with an agent is that it can do more than it needs to: keep tools and permissions to a minimum, require human approval for important actions and filter incoming messages.
Since 2 February 2025 the EU AI Act has required organisations using AI systems to build their staff's AI literacy: it is a legal duty, and also what keeps the project working.

Why operational context matters now

According to Eurostat, 19.95% of EU enterprises with 10 or more employees used AI in 2025, rising to 55.03% among large firms. Italy is catching up fast: Istat reports that AI use among companies with at least 10 employees doubled from 8.2% in 2024 to 16.4% in 2025, and from 32.5% to 53.1% among large companies.

Adoption is moving fast, but licences alone deliver little. In many companies AI arrived as a browser chat window, cut off from the ERP, the catalogue and the customer history: individual help that nobody measures and that does not change how long an order takes to fulfil.

For an online retailer or an established B2B company, the useful question is not which model to use, since a new one ships every few months. It is where AI fits into the workflow, which data it reads, what it is allowed to change and who checks it. That is the operational context, and it is the thread running through the articles in the AI for business section.

What is enterprise AI, and what does operational context mean?

Enterprise AI is the use of artificial intelligence inside a company's systems and processes, with controlled access to its data and measurable goals. What separates it from someone using a chatbot on their own is the connection to the software that already runs the business: ERP, CRM, e-commerce platform, help desk.

Operational context: the processes, data, roles and rules a company works with every day. It covers who receives an order, where it is recorded, which checks it goes through and who signs it off.
ERP (enterprise resource planning): the system that records orders, inventory, purchasing, production and accounting. For many B2B companies it is the source of truth for customers, items and prices.
Agentic AI: a system in which a language model decides which steps to take and uses external tools (reading an email, looking up an item, creating a document) to reach an outcome. Anthropic distinguishes it from workflows, where the model and tools follow a path already written in code.
Human in the loop: the rule that a person reviews and approves any step that affects customers, suppliers or employees.

What is the difference between generative AI and agentic AI?

Generative AI produces content when someone asks for it: a text, a summary, a draft reply. What happens next is up to the person. Agentic AI is given a goal and chooses its own steps to reach it, using external tools: it looks up data in the ERP, opens a document, prepares an action.

Anthropic describes the basic building block of agentic systems as a language model "augmented" with three capabilities: retrieval, tools and memory. The model writes its own search queries, picks the right tool and decides what information to keep. That is what turns a text generator into something that carries out tasks.

Take an online store. Generative AI writes a product description from the spec sheet, and an editor reviews it. An agent, on the other hand, reads an email from a customer reporting a damaged parcel, checks the order and tracking, looks up the returns policy and drafts both the return authorisation and the reply. The two work together: the agent uses the model's generative ability as one of its tools. For the generative side of AI in business, see the guide to generative AI in business strategy.

Why does AI only create value when it is connected to processes and data?

Because a model can reason about text, but it does not know your stock levels, price lists, payment terms or shipment status. Without that data it can only give generic answers, and without a defined process nobody knows what to do with its output.

Three conditions make the difference:

Reliable, reachable data. Clean customer records, consistent item codes, complete product descriptions. An agent searching a catalogue with thin descriptions gets it wrong more often; Microsoft says so explicitly in the documentation for its order-handling agent, noting that the quality of product information affects how well it works.
A process written down step by step. Who receives the request, which checks apply, which exceptions exist: a customer on credit hold, a discontinued item, a special price. If the process only lives in the head of the person who has done it for years, write it down before you automate it.
Permissions and approval points. AI should only read and write where it needs to, and every outward-facing action goes through review until results are stable.

With those three in place, AI becomes a measurable part of the process: fulfilment time, error rate, tickets solved on first contact. To pick the right indicators, see the guide to KPIs for measuring the ROI of AI.

Agentic AI in ERP: what does it do, and where does it stop?

ERP vendors are already building agents into their products. One well-documented example is the Sales Order Agent in Microsoft Dynamics 365 Business Central. It watches a dedicated mailbox, identifies the customer from the sender's address, looks up the requested items in the catalogue, checks availability, prepares a quote as a PDF and handles the email back-and-forth to fill in missing details.

Three design choices behind that agent apply to anyone bringing agentic AI into their own ERP, whatever the software:

It works with the permissions of a role assigned by the administrator, like any other user: it only sees the tables and actions it needs.
It always creates a quote first, even when the customer asks for an order, because a quote has no effect on planning, availability or cash-flow forecasts. It is a safe intermediate document.
Every outgoing message is reviewed and approved by a designated user before it reaches the customer.

The agent does not replace the ERP; it uses it. The ERP remains the source of truth for customers, prices and stock, while the agent removes the manual work between the customer's email and the recorded document. You can build the same pattern on ERPs without native agents by connecting the ERP, the mailbox and the online store through an orchestrator such as n8n or Make, with a language model called via API. For more on how agents relate to traditional software, read how AI agents are changing the SaaS model.

Which ERP systems already have built-in AI agents?

Microsoft is not alone: other ERP vendors document agents that are already available. Two verifiable examples, at very different price points:

Oracle Fusion Cloud Applications. Oracle presents its Fusion Agentic Applications as teams of agents that plan and execute processes across finance, HR, supply chain and customer experience. Published examples include a collections workspace with contextual risk analysis and a batch manufacturing workspace that detects recipe issues and batch deviations and guides the resolution.
Odoo 19. Its documentation describes configurable agents with a system prompt, topics that define their tasks, tools assigned to each topic and sources they draw information from. Creating and customising agents requires the AI app. The standard "Ask AI" agent, by contrast, cannot change the database: it can open views and display reports, but it cannot create leads or alter data.

The lesson is the same as with Microsoft: even with native agents, permissions, assigned tools and data quality decide the outcome. Availability, versions and pricing of AI features change often and depend on your contract, so check them with the vendor or your ERP partner before planning the project. If your ERP has no native agents, AI connects from the outside.

How do you connect AI to your ERP and other business systems?

There are four routes, often combined in the same project. The choice depends on what your ERP offers and on what the AI needs to do: only read, or write as well.

Method	How it works	When it makes sense	Watch out for
ERP APIs through an orchestrator	n8n or Make read and write data through the APIs and call the model only where it is needed	Cloud ERPs, Shopify, WooCommerce, help desks with documented APIs	Credentials with minimum permissions, API rate limits
Exports to a supporting database	Data is copied at regular intervals into a separate database that the AI reads from	ERPs without APIs, analysis and reporting	Data is not real-time; writing back to the ERP stays a separate step
Retrieval from documents	Before answering, the AI searches company documents for the relevant passages	Policies, manuals, spec sheets, terms of sale	Documents kept up to date, each with an owner
Model Context Protocol (MCP)	An open-source standard for connecting AI applications to data sources, tools and workflows	Several systems to expose to the same assistant or agent	Each connector adds capabilities and permissions that need limiting

The official documentation compares MCP to a USB-C port for AI applications: a standard way to plug them into external systems. Anthropic mentions it as one way to give a model tools and data. For a B2B company the most common route is still an orchestrator on top of the APIs, because every run stays in the orchestrator's execution history and you decide precisely what the AI can write, and where.

Which processes should a B2B company or online store hand to AI first?

Start with processes that are repetitive, run at steady volumes, already have their data in your systems and produce a measurable result. Here are the most common cases and who keeps the final say.

Area	Data needed	What AI does	Who decides	What to measure
B2B orders by email	Customer records, catalogue, price lists, stock	Reads the request, matches items, drafts the quote	The sales rep approves sending	Time from request to quote, correct lines
E-commerce customer service	Orders, tracking, returns policy, ticket history	Answers order-status and returns questions, drafts replies for complex cases	A support rep handles refunds and exceptions	First response time, tickets solved without escalation
Manufacturing	Machine data, production plans, failure history	Flags anomalies and delays, summarises job status	The production manager	Unplanned downtime, late deliveries
Logistics and deliveries	Orders to ship, available vehicles, delivery windows, tracking	Proposes the delivery sequence, flags shipments at risk of delay, drafts customer notices	The logistics manager	On-time deliveries, customer chasers
Forecasting and reordering	Sales history, stock, promotions, supplier lead times	Estimates demand per item, proposes reorders, highlights deviations	The buyer or planning manager	Stock-outs, average inventory
Purchasing and suppliers	Purchase orders, confirmed delivery dates	Drafts follow-ups for late deliveries, proposes new dates	The buyer	Delays caught early
HR	Internal policies, handbooks, calendars	Answers recurring employee questions, prepares documents	HR, always for hiring and evaluations	Requests handled without intervention, response time

Orders and quotes

Take a scenario built for illustration: a spare-parts distributor receives dozens of email requests a day, often with mistyped codes or vague descriptions. An AI-assisted flow reads the email, finds the customer in the ERP, matches the items and drafts the quote; the sales rep checks it and sends it. The gain is hours of manual data entry removed and fewer coding errors.

Customer service

Customer service is one of the areas with the strongest evidence so far. A study published by the National Bureau of Economic Research, covering 5,179 customer support agents, found that a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and less skilled staff. For an online store, the first step is connecting the assistant to orders and tracking: an answer to "where is my parcel?" is only worth something if it reads the real data.

Manufacturing

In manufacturing, AI works well as the first reader of the data: it summarises job status, flags gaps between plan and actuals, and highlights machines with recurring anomalies. The decision to stop a line or reschedule production stays with the manager.

One step further is predictive maintenance: if machines have sensors recording parameters such as vibration, temperature or power draw, a model can recognise the patterns that come before a failure and raise an alert, so maintenance is planned rather than endured. The prerequisite is a long enough, well-kept history of sensor data and failures; without it, you start from anomaly alerts. More in the article on AI in industrial machinery.

Logistics

In logistics, AI helps plan deliveries: by combining orders to ship, available vehicles, time windows and, where connected, traffic and weather data, it proposes a delivery sequence and recalculates it when something changes. Within easier reach for many companies is exception handling: spotting stuck shipments, warning the customer before they get in touch, drafting the claim to the carrier. For the data and tools to start with, see the article on digitalising logistics for SMEs.

Forecasting and decision support

An ERP holds years of history on sales, purchasing and lead times. AI can read it to estimate demand per item, suggest reorders and prepare a weekly summary for management with the deviations worth a look. The value lies in the time saved searching and cross-checking data; decisions on budgets, prices and purchasing stay with the people accountable for them. For the estimates to be useful, the history has to be consistent: promotions, stock-outs and item code changes need flagging, or the model will mistake them for real demand.

HR

In HR, AI is useful for recurring employee requests (leave, expenses, policies) and for preparing documents. Hiring and evaluation are a different matter: the EU AI Act classes systems used for recruitment and worker management, such as CV-sorting software, as high risk. According to the European Commission, the rules for these areas apply from 2 December 2027. An HR flow designed today needs human oversight and traceability built in; the details are in the guide to AI in HR processes and compliance.

Automated workflow or AI agent: which should you choose?

In most cases, start with a workflow and move to an agent only when the process has too many variations to capture in rules. Anthropic's practical guide to agents recommends finding the simplest solution possible and adding complexity only when needed, because agentic systems trade latency and cost for better performance on harder tasks.

Criterion	Automated workflow	AI agent
Path	Fixed, written in advance	Chosen by the model case by case
Predictability	High	Medium, needs monitoring
Cost per run	Low	Higher, due to repeated model calls
Best for	Stable processes with few exceptions	Free-text requests with many variations
Example	Shopify order posted to the ERP with a confirmation email	A customer email listing twelve items in free text, turned into a quote

The best answer is often a hybrid: a workflow in n8n or Make handles the predictable steps (receive the order, record it, notify) and calls the model only where text needs interpreting. See also the guide to business process automation.

What are the risks of agentic AI, and how do you reduce them?

The typical risk is not AI "going rogue" but AI being able to do more than it needs to. The OWASP Top 10 for LLM applications, published by the OWASP Gen AI Security Project, calls this excessive agency and traces it to three root causes: excessive functionality, excessive permissions and excessive autonomy.

A second risk applies precisely to agents that read emails and attachments: indirect prompt injection. According to OWASP, it happens when the model takes in content from external sources, such as websites or files, containing instructions that alter its behaviour. In practice, an email from a fake customer could carry text designed to extract a confidential price list or change an order.

The controls you need, in order of priority:

Only the tools it needs. If the agent has to read orders, it does not get a tool that can also delete them.
Minimum permissions and a dedicated identity. The agent logs in as its own user with a limited role, never with an administrator's credentials.
Human approval for high-impact actions: messages to customers, changes to prices and terms, payments, employee data. OWASP explicitly recommends human-in-the-loop control for these cases.
Filters on incoming messages. Microsoft's order agent checks every email and attachment against a relevance model that filters out messages outside its scope, such as offers to sell, discount requests or attempts to access internal data.
Testing in a separate environment, and an action log. Anthropic notes that agents' autonomy means higher costs and the potential for compounding errors, and recommends extensive testing in sandboxed environments with appropriate guardrails.

For security more broadly, see the article on AI and cybersecurity.

How to apply it in your business: first steps

Pick one process and measure it for four weeks. Volumes, average times, errors. Without a baseline you cannot tell whether AI helped.
Map the data and systems involved. Where customers, items, orders and tickets live; which systems have APIs (many cloud ERPs, Shopify, WooCommerce and the main help desks do); which data needs cleaning first.
Write down rules and exceptions. What AI may do on its own, what it must only draft, and what always goes to a person.
Run a 30 to 60-day pilot with human approval on every outward action, in a narrow scope: one channel, one customer group, one product family.
Compare results with the baseline and decide. If the numbers hold, cut approvals on simple cases and expand; if they do not, fix the data and rules before switching tools.
Keep the system up to date. Catalogues, price lists, policies and teams change: review rules, instructions and permissions every time they do, and recheck results at regular intervals. An AI system is run as a living process, not installed once and forgotten.

In projects like these, the order of the steps matters more than the tools. I use n8n or Make for orchestration, OpenAI models to read and write text, Supabase when a supporting database is needed, and the Shopify, WordPress or ERP APIs to read and record data. The e-commerce and B2B solutions I design always start from this map of processes and data.

How much does it cost to bring AI into business processes?

There is no figure that fits everyone, and a sensible rule is to be wary of anyone who quotes one before seeing your processes and data. What you can know in advance is what the cost is made of:

Licences and subscriptions: the ERP's AI features, where they are paid extras, and the orchestration tools.
Model usage: models called via API are billed on usage, by the tokens (chunks of text) processed as input and output. An agent makes more calls than a workflow for the same task, so each run costs more.
Integrations: often the most variable item, because it depends on the ERP. With documented APIs the work is straightforward; closed systems need exports or custom development.
Data clean-up: customer records, item codes and product descriptions to fix before you start.
Training for the people who will use and oversee the system.
Maintenance and monitoring: rules, instructions and connections to update when price lists, catalogues and policies change.

The narrow pilot serves this purpose too: it measures the real cost per case handled and sets it against the time saved before you scale up. Public support may be available for digitalisation investments; there is an overview in the article on funding for AI projects in SMEs, written for the Italian market.

Do you need to train staff to use AI at work?

Yes, and not only as a precaution. Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures to build the AI literacy of their staff and of other people operating those systems on their behalf. According to the European Commission's questions and answers, the obligation has applied since 2 February 2025. Following the Digital Omnibus amendments it remains an obligation, but no specific or "sufficient" level is mandated; the supervision and enforcement rules apply from 2 August 2026.

The Commission advises tailoring training to the organisation's role and to the risks of the systems it uses. For a project on orders or customer service, useful training is very practical:

who approves what, and on what criteria;
how to spot a wrong answer or a made-up figure;
which data must never be entered into unapproved tools;
what to do when the AI flags an exception or makes a mistake.

Training is also how you bring AI in without upending the work: the person who keys in orders by hand today becomes the one who checks the drafts and handles the exceptions, and they know the tricky cases better than anyone.

Mistakes to avoid

Starting from the tool instead of the process. A company-wide licence without a defined use case produces scattered experiments and no measurable result.
Connecting AI to messy data. Duplicate records and inconsistent codes turn into faster mistakes, not fewer.
Granting broad write access from day one. Start read-only and in draft mode, then widen access where results justify it.
Skipping governance. Who is accountable if AI gets a price wrong? Logging, review and ownership need deciding before launch; I cover this in the article on AI governance and explainability.
Measuring usage instead of outcomes. Process time, errors and cost matter, not the number of prompts.

Frequently asked questions (FAQ)

What is agentic AI in ERP?

It is AI that works inside the ERP like a user with a specific role: it reads requests, looks up data, creates documents such as quotes or supplier follow-ups and asks a person to confirm before important actions. Unlike classic automation, it decides its own steps based on the case in front of it. It is still limited by the permissions the administrator assigns.

How is enterprise AI different from a chatbot like ChatGPT?

A chatbot used in the browser only works on the text you paste into it and knows nothing about your company's data. Enterprise AI is connected to the ERP, CRM, online store and help desk, runs inside defined processes and has permissions and controls. The underlying model may even be the same; what changes is the connection to data and process.

Do you need a new ERP to use AI in your business?

No. If your ERP offers APIs or regular data exports, AI can connect through an orchestrator such as n8n or Make. Native agents in newer ERPs make the job easier, but the real requirement is clean, accessible data. Replacing your ERP just for AI is rarely the sensible first move.

Which processes should you automate with AI first?

Repetitive ones, with steady volumes, data already in your systems and a result that is easy to measure. In B2B that usually means entering orders and quotes from email and chasing suppliers; in e-commerce, answering questions on order status, shipping and returns. Avoid processes with high-impact decisions on people or money as a first project.

Can AI handle orders and quotes on its own?

It can prepare almost all the work: reading the request, matching items, checking availability and filling in the document. In the early phase it is wise to have a person approve every message to the customer, which is exactly how Microsoft designed the order agent in Dynamics 365 Business Central. Cut approvals on simple cases only once pilot data shows low, stable error rates.

Does the EU AI Act apply to companies using AI in HR?

Yes, if AI is used to screen candidates or to evaluate or manage workers: the AI Act classes these uses as high risk. According to the European Commission, the related rules apply from 2 December 2027 and require, among other things, human oversight and traceability. Lighter uses, such as an assistant answering questions about internal policies, generally do not fall into that category.

How much does it cost to integrate AI with an ERP?

It depends on the process, the volumes and the ERP, which is why a serious figure only comes after an assessment. The cost items are always the same: any licences for AI features, model usage via API, integration work, data clean-up, training and maintenance. A 30 to 60-day pilot on a single process lets you measure the real cost per case before you scale up.

Sources

If you want to find out which of your processes is ready for AI, and with which data, book a free diagnosis.