ERP or AI copilot? How to decide what to build first
One gives you a reliable record of the business. The other helps people think and act faster. Building them in the wrong order is one of the most expensive mistakes we see.
Two different jobs
The question usually arrives as “should we buy an ERP, or build something with AI?”. It sounds like a choice between old and new technology. It is really a choice between two different jobs.
A system of record, such as an ERP, an operations platform, or an inventory or trading system, answers the question “what is true?”. What did we buy, what do we hold, who owes us what, and who approved it.
An intelligence layer, such as a copilot, an assistant or an AI agent, answers “what does it mean, and what should we do?”. It reads, summarises, flags, drafts and recommends.
The second job depends on the first. A copilot asked “which suppliers are drifting on price?” is only useful if purchase prices are recorded somewhere, consistently.
Signs you need the operational core first
- The same number has different values depending on who you ask.
- Key processes run on spreadsheets that only one person understands.
- Approvals happen in WhatsApp or email, and cannot be traced later.
- Month-end takes days because data has to be gathered and reconciled by hand.
- Simple questions, such as current stock or margin by product, need a project to answer.
In this situation, a copilot will mostly produce confident answers from unreliable data. The better investment is a focused operational platform that captures the work as it happens, designed around how your teams actually operate rather than a generic ERP template.
Signs a copilot can come first
- Your core records already live in an ERP, accounting or CRM system you trust.
- The bottleneck is reading: contracts, reports, research, tenders or regulations.
- Experienced people spend hours searching for answers that exist somewhere in your documents.
- Decisions are slow because analysis takes time, not because the data is missing.
Here, a copilot that searches your documents, cites its sources and drafts first versions can save hours per person every week, without touching the systems you already run.
Side by side
| System of record | Intelligence layer | |
|---|---|---|
| What it fixes | Fragmented data, re-keying, untraceable approvals | Slow reading, searching, analysis and drafting |
| What it needs | Clear processes and agreed data definitions | Trustworthy data or documents to work from |
| Time to first value | Typically 6 to 16 weeks for a focused core | Often 3 to 8 weeks for a well-scoped use case |
| Main risk | Building too much before anyone uses it | Answers that sound right but rest on bad data |
| Who uses it | Everyone who records or approves work | Analysts, managers and specialists |
The hybrid path most companies take
In practice, the best answer is rarely pure. The pattern we build most often is a lean operational core with AI at the edges, where it removes the manual work of getting information into the record.
In a cafe inventory and accounting system, AI reads supplier invoices to reduce data entry, while accountants approve entries before they reach the books. In a multi-store restaurant platform, handwritten sales bills are converted into editable drafts that staff verify. In both, the AI does not replace the system of record. It feeds it faster, with fewer errors.
Once the core is reliable, the intelligence layer has something solid to stand on: explaining trends, flagging unusual transactions, forecasting demand.
A five-question test
Answer these honestly for the problem you want to solve:
- Could you produce the data the AI would need, today, in under an hour?
- Would two people in your company give the same answer to “what is our current position?”
- Is the slow part of the work recording information, or interpreting it?
- Are approvals and changes traceable after the fact?
- If an AI gave a wrong answer, would anyone notice?
Mostly “no” to questions 1, 2 and 4, with recording as the slow part: build the core first. Mostly “yes”, with interpreting as the slow part: a copilot can come first. Mixed answers usually point to the hybrid path.
What to avoid
- A copilot on top of chaos. It will answer confidently, and some answers will be wrong in ways that are hard to spot.
- The two-year ERP. Big implementations that try to model everything before anyone uses them. Start with the processes that hurt most, and grow.
- AI as a checkbox. If you cannot name the hours saved or the decision improved, it is not ready to build.
Where to start
Pick the one process that causes the most friction, and run the five-question test on it. The answer tells you whether your first build is a record, an intelligence layer, or both. If you would like help reading the answer, we are happy to look at it with you.