Why vendor AI looks unavoidable
Every major ERP and CRM vendor now bundles an AI layer. SAP has Joule. Salesforce has Einstein and Agentforce. Oracle has the Fusion AI Apps. Microsoft has Copilot inside Dynamics. The pitch is the same in every account review: the AI is here, it understands your data, switch it on for an additional licence fee.
The arithmetic looks straightforward. You already have the platform. The AI is incremental. The vendor handles updates, model swaps, prompt safety, evaluation. You pay per seat and ship faster.
The arithmetic also hides what the architecture is actually doing. The agent runs against the schema the vendor controls, with prompts the vendor maintains, against models the vendor selects, with retention policies the vendor sets. Every constraint your security team writes is a constraint the vendor must agree to enforce. You buy reliability and lose every degree of freedom that matters.
What "build it yourself" actually means in 2026
The build-vs-buy conversation changed in 2025. Three things shifted:
- Foundation-model quality compounded faster than vendor AI features. The frontier models a procurement team can access today are stronger than the models embedded in vendor AI products. The gap widens at every release.
- Tool-use and structured outputs became table stakes. Agentic workflows that required custom scaffolding in 2023 are now first-class in the model APIs. Less plumbing, fewer brittle integrations.
- Open evaluation harnesses matured. You can measure quality on your own data, on your own benchmarks, with reproducible results. The vendor evaluation black box stopped being the only option.
Together, those shifts mean a competent platform team can stand up an AI agent against the same ERP data in roughly the timeframe the vendor takes to provision their bundled solution — but with full ownership of the model selection, prompt strategy, tool definitions, and the eval suite that proves it works.
What it costs to do well
The honest number for a single high-value agent against an enterprise ERP — say, an order-to-cash exception triager handling 10,000+ items per week — is a typical mid-market Middle East enterprise project, fully owned, transferred to your team, with an eval suite, observability, and twelve months of operational handover support. Compared against the per-seat fees a vendor agent at the same load level extracts over a 3-year horizon, the build-and-own number is materially lower. The model is available on request.
The other line of the comparison is the line nobody on the vendor side will draw for you: the carrying cost of a second AI vendor relationship. Every AI feature you switch on creates a renewal cycle you cannot influence, a roadmap conversation you cannot lead, and a model swap policy that does not consult you.
The saasinator perspective
The vendor AI bundle is rational for low-frequency, low-stakes assistance use cases. The bundle stops being rational the moment the workflow involves your business logic, your customer-facing decisions, or your competitive differentiation. The line between "okay to rent" and "must own" is sharper than most procurement teams treat it.
What we look for in scope
The agents we build for enterprise CTOs share three traits:
- High decision frequency. The economics flip when an agent handles thousands of low-cost decisions per week, not when it handles a few high-stakes ones.
- Auditable inputs. The agent's source data must be data your team already owns and can explain. We do not build against opaque vendor APIs that change shape every quarter.
- A clean handover surface. Every agent ships with the eval suite, prompt versioning, and observability hooks your platform engineers will operate. We are not the enterprise. Your team is.
If you are reviewing a vendor AI quote inside an existing ERP contract, the question worth asking is not whether the vendor AI is good — it usually is. The question is whether the workflow it covers is one where ownership pays back faster than a per-seat subscription. For most high-frequency workflows, it does.
What to bring to the diagnostic
Bring the vendor AI quote, the workflow it covers, the decision volume that workflow carries in a normal week, and the source systems the agent would read. The diagnostic is ten working days. The output is the workflow recommendation, the build-versus-rent line drawn on your own numbers, and the first-quarter scope. Start an ignite build at /diagnostic.