Everyone is learning to prompt. Almost no one is redesigning the work. That gap is why most AI pilots impress in a demo and disappear in production.
I lead AI and ERP change at Novartis, and I build AInspire on the side. In both worlds I see the same pattern: teams treat AI as a smarter search box. They open a chat window, craft a clever prompt, get a good answer, and paste it back into a process that hasn't changed at all.
That is not transformation. That is a faster typewriter.
The prompt is the last mile, not the road
Prompt tips are useful. But a great prompt sitting outside your workflow is a great answer nobody asked for at the moment it mattered.
Value doesn't come from the quality of a single response. It comes from where that response lands, who acts on it, and what it triggers next. The prompt is the last mile. The road is the process.
If your AI lives in a separate tab, your people are doing two jobs: their job, and the job of feeding the AI.
Every context switch is a tax. Copy the data out, paste it in, read the answer, decide, paste it back. Multiply that by a team, by a quarter, and the "productivity gain" quietly turns negative.
Design AI into the moment of work
The real question is not "what should I type?" It is "where in this workflow does a decision get made, and can AI be standing right there when it happens?"
Concretely, that means embedding intelligence at the exact step where work already lives:
- Inside the ERP screen, not a separate assistant. When a buyer reviews a purchase requisition, the anomaly check, the vendor history, and the suggested approval are already on the screen.
- On the event, not the request. A ticket arrives, an invoice mismatches, a deviation is logged. The AI reacts to that trigger automatically. No one has to remember to ask it.
- In the flow of a tool people already open — the email, the CRM, the change request form — so adoption costs zero new habits.
When AI shows up at the moment of the decision, adoption stops being a training problem. People don't "use the AI." They just do their work, and the work is now faster.
Three questions before you write a single prompt
Before I let a team touch prompt engineering, I make them answer three things about the process itself.
1. Where is the decision?
Map the workflow and find the exact points where a human judges, approves, routes, or drafts. Those are your insertion points. Everything else is noise.
2. What does "good" trigger next?
An AI output that doesn't move the process forward is decoration. Define the next action: auto-fill the field, route the case, draft the reply for one-click approval. Design the handoff, not just the answer.
3. Who owns the exception?
AI will be wrong. The workflow must have a clear, fast path for a human to catch, correct, and override — and that correction should feed back into the system. If there is no owner for the edge case, you don't have a workflow. You have a liability.
Redesign beats fine-tuning
I have watched teams spend weeks tuning prompts to squeeze out a marginally better answer, when the real leverage was one structural change: moving the AI step upstream, before the data got messy, or downstream, after the human made the call it was best at.
The biggest gains I have shipped came from redrawing the process map, not from a cleverer instruction. Cut the handoff. Kill the copy-paste. Put the model where the friction is.
Prompting optimizes the answer. Workflow design optimizes the outcome. Only one of those shows up on the P&L.
What this means for leaders
If you are running a transformation, stop asking "are our people trained to prompt?" Ask three sharper questions instead:
- In which processes does AI actually touch a live decision — not a sandbox?
- How many context switches did we remove, versus how many we added?
- When AI is wrong, how fast does a human catch it, and does the system learn?
Prompting is a skill. Workflow design is a strategy. The organizations that win the next two years won't be the ones with the best prompts. They will be the ones that rebuilt how work flows, with AI sitting quietly inside every decision that matters.
Don't teach your people to talk to the AI. Design the AI into the room where the work already happens.
Cédric Bignet is an AI & ERP Change Management expert at Novartis and founder of AInspire. He writes about change management, AI adoption and enterprise transformation.