Many CRMs and ERPs now come with an AI assistant. A firm that already runs these systems has a choice: switch on the features each vendor offers, or build AI into a system the firm controls.
Both can be right. The choice depends mostly on where the data may go, who has to approve a change, and whether the question needs more than one system. This post explains when a vendor's feature is enough and when AI belongs inside your own system.
What built-in AI features usually do
The features are similar from one product to the next. Vendor documentation for CRMs lists record summaries, recent changes to records, meeting preparation, email drafting and questions in plain language. ERP vendors document assistants that answer questions about company data and suggest values for fields. Some also read incoming supplier invoices, match them against purchase orders and propose bank reconciliation matches for a person to review.
Many vendors now charge for AI by use, in addition to the price per user: by credit, by action or by resolved conversation. Most of these features are designed around the data the product already holds.
When the vendor's feature is enough
A built-in feature is often the right choice. It is usually enough when the following are true:
- the question can be answered from the data of one system;
- you accept that the data is processed in the vendor's cloud;
- the AI drafts or summarises, and a person reads the result before anything is sent or saved;
- the cost per use stays acceptable at the volume you expect.
In these cases, building your own adds cost and gives little in return. The vendor maintains the feature, and your team uses it inside a tool it already knows. The same reasoning applies to software in general: a process that works the same way as in most firms is usually served sooner, and with less risk, by a ready-made product.
When AI belongs inside your own system
Building AI into a system you control makes sense in four cases.
The question needs more than one system
A board asks why invoiced sales fell in a month when orders rose. The answer is in the CRM, the accounting program and the ERP at once. These systems record the same sale at different moments, and gaps open between them: a won deal with no invoice, an order that never reached the CRM. An assistant designed around one product's data usually sees only that product's part.
The data has to stay inside
Some firms need their records to stay on their own servers. Some built-in AI features are offered only in the vendor's cloud. One large vendor's documentation states that its sales assistant is not available for on-premises installations of its CRM. If your systems run on your own servers, check where the AI part runs before you plan around it.
A Turkish guide on AI in ERP sets a rule worth keeping: price lists, contracts, financial data, personal data and trade secrets should not be sent to outside AI services without control.
A named person must approve each change
Some records matter too much for an AI to change them on its own: a payment, a price, an accounting entry, the deal stage a forecast depends on. The same guide says that high-impact processes should keep human approval and business-rule checks, while a report summary can be produced automatically. A revenue operations guide states the rule in two short sentences: "A suggestion is fine. A silent write is not."
Inside your own system, you decide which actions the AI may take alone and which ones wait for a named person.
The process is your own
If the work that sets your firm apart already runs in software built for you, the AI that supports that work belongs in the same system. A packaged AI feature is designed for the process the package assumes.
What to keep buying, and what to build
Keep the systems of record. The CRM, the accounting program and the ERP you rely on stay where they are, and the AI layer reads them through the interfaces their vendors publish. Some vendors license that interface separately, so check the terms before any work is quoted.
Then build the part that has to read more than one system or stay on your servers, and give it clear rules:
- every figure the AI reports can be traced to the record it came from;
- when a system cannot be read, the AI says so and does not guess;
- changes to records wait for a named person's approval;
- the data is cleaned first, because duplicate customers and wrong stock figures lower the quality of anything the AI produces.
Start with one workflow that costs your team time today. A narrow first build shows early whether the approach works on your data.
How we approach it
In the internal tools we build, the rule is that agents draft and humans decide. The agent prepares the work and records what it did. A person reviews the context and authorises the action. Only then does the agent carry it out, and that step is logged too.
Our AI CRM and our ERP follow the same rule for changes: nothing changes in your accounting program or your ERP without a named person's approval. Both include an analysis mode that reads the CRM, the accounting program and the ERP together. Both run on your own servers or on our cloud, and you choose.
If you are deciding which AI features to switch on and which to build, tell us which systems you run and what you want to ask of them.