Stop Looking for an AI Project. Pick a Workflow Instead.
The vendors have made up their minds. Oracle announced Fusion Agentic Applications in mid July, agents that run inside the ERP rather than beside it. Gartner now projects that 40 percent of enterprise applications will ship with embedded agents by the end of 2026, up from under 5 percent a year ago. Microsoft quietly made Excel Copilot able to learn a monthly reporting process once and replay it on demand.
Notice what none of these are: standalone AI products. The industry has stopped selling AI as a destination and started shipping it as a layer inside software people already use. That is a useful signal for any small or mid-sized company still asking “what should our AI project be?”
The honest answer is that you probably should not have an AI project at all. You should have a workflow with AI in it.
Why projects fail and workflows stick
We have written before about why AI pilots die, and the short version is that most of them are built next to the business instead of inside it. A pilot with its own tool, its own login, and its own champion has to fight for attention every single day. A workflow does not. If AI sits inside the step where invoices get matched or support tickets get triaged, people use it because that is simply how the work flows now.
The economics back this up. In accounts payable, the combination of AI capture and rule-based approvals takes a typical invoice from roughly 18 dollars and ten days down to about 3 dollars and one day. Yet only around 7 percent of firms use AI here today. The gap between what works and what is adopted is the opportunity, and it favors companies willing to start small.
How to pick the first workflow
Three filters, applied in order.
First, the work must already exist and repeat. Invoice matching, order exception handling, quote follow-ups, support triage. If it happens daily and follows a rough pattern, it qualifies. If you would need to invent a new process to use the AI, it does not.
Second, the failure mode must be cheap. Your first embedded AI will make mistakes. Pick a flow where a wrong suggestion costs a minute of human review, not a shipped order or a mispaid supplier. This is why “AI drafts, human approves” is the right shape for a first deployment, a point we made in our piece on agentic ERP and governance.
Third, the output must be measurable in one number. Hours saved per week, days of processing time, error rate. If you cannot name the number before you start, you will argue about success afterward.
Smaller companies have the edge here
There is a case that SMEs are better positioned than enterprises for this style of adoption. Less legacy integration debt, fewer approval layers, shorter distance between the person who feels the pain and the person who signs off. An enterprise needs a steering committee to change an approval flow. You need a Tuesday afternoon.
We see this pattern in our own ERP work. The most useful thing Etendo’s Copilot does is not conversation, it is sitting inside flows users already run and shaving steps off them.
So skip the AI strategy offsite. Pick one workflow, embed AI into the step that hurts, keep a human on the approval, and measure the number. If you want help choosing which workflow that should be, our assessment is built for exactly that conversation.