AI for ERP and CRM: the vendor's built-in features or a layer of your own

Each of the four vendors on this page now sells AI inside its product. Salesforce has Agentforce, Microsoft puts Copilot and prebuilt agents in Dynamics 365, SAP has Joule, and NetSuite has AI features such as Text Enhance and Ask Oracle [1, 5, 8, 10]. The pitch is the same everywhere: the AI already sits next to your data, so you switch it on.

The other route is a layer you own. It is software that reads from and writes to those systems through their published interfaces, calls a model you choose, and keeps its own checks and records. It can be an agent you build, or an assistant such as ChatGPT or Claude connected through the vendor's own connector.

I think the choice usually gets framed as a contest, and it is not one. Built-in AI is good at work that starts and ends inside one system. A layer earns its cost on work that crosses systems, such as a renewal decision that needs the contract from the CRM, the invoices from the ERP and the support history from somewhere else. I would expect a mid-sized business to end up with both, so the useful decision is which work goes where.

The table sets out what each vendor's own pages said on the date below. It is not a ranking. The four products serve different systems, and the relevant one is almost always the one attached to the system you already run. The row that matters most for a layer of your own is the last one.

Two things stand out. The meters differ in kind: Salesforce and SAP count actions, Microsoft counts credits on top of a per-user licence, NetSuite draws down units that come with each user licence, and Salesforce also sells flat per-user access. You cannot compare those without your own volumes, which is why the guide below starts from the work, not the price.

The terms for outside AI now differ more than the built-in features do. NetSuite has opened a connector with published guardrails, Salesforce is building towards any authorised agent reaching its platform, Microsoft charges for that access in some cases, and SAP's policy restricts it unless you go through a pathway SAP endorses. This is not legal advice. What you may do depends on your own agreement with the vendor, which can differ from what the vendor publishes.

Five questions settle most cases. Answer them for one piece of work at a time rather than for a whole system, because a single CRM can hold work that suits each route.

Where the answers split, I would split the work: the vendor's AI for help inside its own screens, and a layer of your own for the few processes that cross systems. Before you design that layer, check the last row of the vendor table for your system, because the access route decides what the layer may do and what it costs.

The vendor route looks cheaper because it arrives on an invoice you already pay. Check what runs the meter. At Salesforce's list prices, a standard Agentforce action of 20 Flex Credits works out at 10 cents [1], so an agent that takes five actions to close a case costs 50 cents a case on that meter alone. SAP prices its agents per action as well [8], and has announced that later in 2026 most generative AI capabilities that now need AI Units move into Base AI at no extra cost [8], so expect that line to change. Microsoft includes Copilot in the Sales Enterprise licence and meters agents in Copilot Credits [5], and NetSuite includes an allowance of AI Units with each user licence and sells more in packs [10].

A layer of your own moves the cost rather than removing it. You pay a model provider per token, you pay for the engineering to build and run the layer, and you may still pay the vendor for access: Microsoft charges Copilot Credits when agents built outside Copilot Studio use its Dataverse MCP tools, unless your licences cover it [7]. Price all three lines before you compare, because the model calls are the only one with a public price list.

The cost nobody prices is agreeing what a correct answer is. In engagement file D-06, connecting an AI workspace to a company's finance data took an afternoon, and the week that followed, agreeing definitions with the finance director, was the actual work. That week is the same whichever route you choose.

Lock-in used to mean the cost of moving your data. With AI it also means where your business logic lives. Instructions, prompts and approval rules written inside Agentforce Builder, Copilot Studio or Joule Studio stay there if you leave. That is a fair trade if you will run the system for a decade, and a poor one if the same process has to span systems from two vendors.

Access is the other half. SAP's API Policy, version 4.2026a, says that outside SAP-endorsed architectures, data services or service-specific pathways, SAP prohibits API use for interaction or integration with AI systems that plan, select or execute sequences of API calls, and for large-scale data extraction [9]. The policy does not list which pathways SAP endorses, so ask SAP in writing which one covers your use before you design around it. The same policy says it does not limit data export the law requires [9].

The other three vendors are opening up, on different terms. NetSuite's connector lets outside assistants act with the user's own role [11, 12]. Salesforce says its Headless 360 MCP Server lets agents in Claude, ChatGPT and other platforms use Salesforce capabilities with the same identity and permissions, though that server is in open beta [4]. Microsoft's Dataverse MCP server works with clients such as Claude and bills some of that use [7]. The Microsoft charge started on 15 December 2025 and SAP's policy is a 2026 version, so I would expect these terms to move again. Keep each vendor's connection in one piece you can replace. This is not legal advice: whether a use is allowed depends on your own agreement, and your legal adviser should read it.

NetSuite's published guidance on outside AI works as a checklist for any of the four systems [12]. It warns that prompt injection and hallucination can lead an agent to make payments or grant approvals the user never intended, to change or delete data, or to disclose sensitive data. It recommends granting access only to users who need it, keeping high-privilege roles away from AI, starting with read-only tools, and asking for confirmation before high-impact actions. Those points, plus three of our own, are what to ask of anyone who builds the layer for you.

Building that layer is what our AI integration and automation service does: AI inside the systems a business already runs, NetSuite among them, with every write checked, traceable and reversible, and never more access than the person it acts for. The closest documented case is the finance connector above, which carries each person's existing permissions so nobody sees more than they already could.

If you have not yet decided which work belongs on which route, our AI Opportunity & Feasibility Sprint scores the candidate processes on value, feasibility and risk over a typical two to four weeks, and gives you the answer in writing. If the scope is already clear, a fixed-scope build can start within a day of sign-off, and a job you have already scoped can be resourced on contract from the associates who work with 1AYM, held to the same standard. US clients can contract through 1AYM's US entity. We have built on the ERP and CRM vendors' own AI tools as well as OpenAI's and Anthropic's, so the recommendation follows your systems rather than the partnership. The next step is a short call, booked from the end of this page.