You do not need to replace your ERP or move your data to a new platform to start using AI. Replacement projects are slow, expensive and often disappointing: Gartner predicts that by 2027 more than 70% of recently implemented ERP initiatives will fail to fully meet their original business goals. The safer route is to keep the system that already runs your business and add a connector on top of it, so an AI assistant reads only what each request needs and acts only through actions you have permitted, with a person approving anything sensitive. Below: why replacement projects go wrong, why legacy systems stay, how a connector works and how to start.
Why do so many companies assume AI means a new ERP?
The pitch usually sounds the same: your system is too old for AI, your data has to be cleaned and moved first, and the smart move is a new platform with AI built in. For a company whose ERP has been customised for years, or whose system was built in house, that means a long project before anyone sees a benefit.
Yet when you ask managers what they want from AI, the list is rarely about the ERP itself. They want to ask "which items are below their reorder level?" and get an answer in seconds, a purchase order drafted from that list, or a customer's open balance without opening four screens. None of that requires a new ERP. It requires a safe way for an AI assistant to reach the system you already have.
How often do ERP replacement and migration projects go over budget or fail?
The research is consistent: large system projects are hard to deliver as planned. The figures below are global and many come from large organisations, but smaller companies have less room to absorb a delay, not more.
| Source | What it studied | What it found |
|---|---|---|
| Gartner | Recently implemented ERP initiatives | By 2027, more than 70% will fail to fully meet their original business goals (a prediction) |
| Panorama Consulting Group, 2026 ERP Report | ERP projects at organisations across industries, published March 2026 | More than a quarter of organisations exceeded their project budgets, with additional technology needs the leading cause |
| McKinsey with the University of Oxford, 2012 | More than 5,400 IT projects; "large" means a budget above $15 million | Large projects ran 45% over budget and 7% over time on average, and delivered 56% less value than predicted |
| McKinsey, 2021 | Survey of nearly 450 CIOs and IT decision makers about cloud migration | 38% had seen migrations delayed by more than a quarter, and inefficiencies cost the average company 14% more in migration spend than planned each year |
Two details matter more than the headline numbers. First, the McKinsey and Oxford research found that 17% of IT projects go so badly that they can threaten the existence of the company. Second, the leading cause of budget overruns in Panorama's latest report was not a technical failure but additional technology needs: things nobody had planned for at the start. Replacing a core system touches how every department works, which is why so much turns up only once the project is under way.
What does rip-and-replace really cost beyond the software?
The licence or subscription is the visible part. Most of the cost sits around it.
- Data migration. Years of customers, items, prices and transactions have to be cleaned, mapped to new fields and checked. Every mismatch needs a decision from someone who knows the business.
- Rebuilt customisations. Approval chains, special pricing or local tax handling have to be rebuilt or dropped.
- Reports and integrations. Every report people rely on and every connection to the website, the accountant or the e-invoicing portal must be rebuilt and tested.
- Parallel running. Many companies run both systems side by side for a while to compare results, which doubles the work.
- Retraining and lost speed. Staff who were fast in the old system are slow in the new one for months.
- Frozen decisions. While the project runs, other improvements are put on hold "until the new system is live".
None of this means replacement is always wrong. If your system is unsupported, insecure or blocking growth, replacing it may be the right decision. But replacing a working system only to add AI means taking on all of these costs for a goal that can be reached another way.
Why do companies keep legacy and custom systems?
Because they work. An ERP that has run a business for years holds its real rules: how discounts are approved, how stock moves between branches, which customer gets which payment terms. Staff know where everything is. Custom systems were often built precisely because no ready-made product fitted the way the company works.
Even governments with very large budgets keep old systems running. The US Government Accountability Office reported in July 2025 that the federal government spends more than $100 billion a year on IT, and that agencies typically report spending about 80% of it on operating and maintaining existing systems. The 11 legacy systems GAO judged most in need of modernisation ranged from about 23 to 60 years old, and eight of them use outdated programming languages.
The lesson for a company in Cairo, Riyadh or Dubai is not that old systems last forever. It is that replacement is a large decision with its own timing, and the wish to use AI should not force it. If you are still choosing a system, see our CRM vs ERP guide.
How can AI work with your ERP without replacing it?
The alternative is a connector: a thin layer between an AI assistant and your system that exposes a short list of defined actions, such as "list items below reorder level" or "create a draft purchase order". The assistant does not get the keys to your database. It calls these actions, and each one follows the rules and permissions you set. One open standard for building such connectors is MCP, which we explain in what MCP is and how it connects AI to your business systems. For an ERP, it means:
- Your data stays where it is. The assistant reads only the records a request needs.
- Your rules still apply. If a sales user cannot change prices in the ERP, the assistant cannot do it for them.
- Old and custom systems can often be connected, as long as there is a safe way in, such as an interface the system already offers.
- You can switch it off. Removing the connector leaves the ERP exactly as it was.
| Replace the ERP | Copy data to an AI platform | Connector on top | |
|---|---|---|---|
| Your data | Migrated to a new system | Copied to a second place | Stays in your system |
| Daily work | Every department changes how it works | A new tool plus data syncing | Staff keep their screens and add plain-language requests |
| If it goes wrong | Core processes are affected | Two copies of the data drift apart | Switch the connector off |
| First useful result | After go-live | After the data is copied and cleaned | After the first commands are built and tested |
| Lock-in | A new vendor | A platform holding a copy of your data | Depends on who owns the connector and which AI it uses |
The last row is where to be careful. Ask who owns the connector, whether you can change the AI model behind it and where the logs are kept. Our guide to avoiding AI vendor lock-in lists the questions to ask before you sign.
Which commands should you start with?
Do not try to connect everything at once. Start with three to five commands that are frequent, slow today and low in risk.
- Questions people ask every day: "Which items are below reorder level in the main warehouse?", "What did this customer order last month?", "Which purchase orders are late?"
- Summaries managers want: "Summarise yesterday's sales and returns by branch."
- Drafts a person approves: "Draft a purchase order for the items that are low on stock."
For more ideas, see AI in operations: inventory, orders and demand forecasting and our list of ready-to-use commands by department.
How do you add AI to a legacy system step by step?
- Write down the requests. For two weeks, ask each team to note what they look up or change in the ERP and how long it takes.
- Pick three to five commands with the most repeats and the least risk.
- Map the data and the people. For each command, list the screens or records involved and the roles allowed to use it.
- Start read-only. Let people ask questions and get answers first. Read-only commands cannot damage data, and they show quickly whether the answers match what the ERP shows.
- Add changes with approval. Next, let the assistant prepare changes, such as a draft purchase order or a new delivery date, and have a person approve them. Keep anything involving prices, deletions or messages to customers behind approval for good.
- Log every action. Record who asked, when and what changed. Our guide to letting AI act inside your systems safely covers permissions, approvals and audit logs.
- Measure, then expand. Compare the time per task before and after, then add the next commands. Our guide to measuring AI ROI shows what to track.
How does Amrak add AI to your existing systems?
Amrak is HBS Group's AI assistant that works inside the systems a company already runs, built on MCP. It is designed for exactly this situation: you keep your ERP, website, CRM and dashboards, and your team gives requests in plain Arabic or English. We walk through the full flow in how Amrak works, from one sentence to action.
- A custom connector. HBS builds a connector for each client that covers its own sections and rules, so commands match how your company works.
- No migration and no replacement. Amrak works on top of your current system as it is. Your data stays in your system, and Amrak reads only what a request needs.
- Old or custom systems. In most cases Amrak works with them, and we confirm this for your system in the free discovery session.
- Permissions and approvals. Each person can only reach what their permissions allow, for example editing products but not prices. Messages to clients, deleting anything and changing prices always wait for approval.
- A log of every action: who asked, when and what changed.
- Your choice of AI provider. Your team uses a well-known AI assistant, and we help you choose and set up one that fits your policy. Amrak can be switched off at any time.
For an ERP, a typical first command is "flag low stock and draft a purchase order". Amrak understands the request, checks the person's permissions, prepares the full change, asks for approval if the action is sensitive, then completes the task and logs it. Projects start with a free discovery session and a command map, followed by the build, testing with your team, handover and training.
If you want AI working inside the ERP you already have, without a replacement project or a data migration, book a free discovery session for Amrak and we will map your first commands with your team.
