Ask ten SAP customers what they have actually done with Joule and you will get ten shrugs. Most switched it on inside S/4HANA Cloud or SuccessFactors, asked it two questions, felt mildly impressed, and never went back. The assistant is not the problem. Nobody handed the team a sequence to follow.
So here is one. Not a feature tour, but the order of operations that turns a licence into something a controller or a planner uses on a Tuesday afternoon because it genuinely saves them time.
Pick one workflow before you pick a single prompt
Joule’s demo mode is easy. Ask it something broad, get something broad back, move on. The useful version starts with a task your team already repeats: explaining why a sales order is blocked, chasing purchase requisitions that have sat in approval for five days, or finding out how many maintenance hours were booked against a specific work order last month.
Write down what that task costs today. Who does it, how long it takes, how often it happens. Fifteen minutes a day across eight people is roughly 40 hours a month, which is a real number you can defend later. Leave the sprawling, four-module, judgement-heavy questions for month six.
Three traits make a good first candidate: the data sits in one system, the answer is a sentence or a small table, and someone asks for it at least weekly.
The setup work that happens before your first real question
Confirm what Joule can actually see
Joule does not read your whole landscape by default. It reaches the systems you have connected and the data your role permits. If a user asks about a plant they have no authorisation for, they get a thin answer, not an error message, and that thin answer is worse than a failure because it looks plausible.
Check roles against the questions people will ask
Walk through your chosen workflow as each persona who touches it. A regional sales manager and a central finance analyst asking the identical question about EMEA revenue should get different numbers, and both should be correct. If they don’t, fix authorisations before you demo anything.
Write a two-page glossary
This is the step almost everyone skips, and it is the single biggest cause of disappointing pilots. Your business has its own vocabulary, and Joule answers the question you typed rather than the one in your head. Capture things like:
- What “open order” means here, and whether it includes credit-blocked items
- Whether “late” is measured against the requested delivery date or the confirmed date
- Cost centre naming quirks, merged codes, legacy leftovers
- Customer and product segment codes people type in shorthand
Two pages is enough. Circulate it, then turn it into example prompts.
How to phrase a request so the answer is usable
Weak request: “How are sales doing?” You will get a curve and a polite paragraph that tells you nothing you can act on.
Working request: “Compare net revenue for EMEA in Q3 against the same quarter last year, broken out by product line. Flag any line down more than 10% and show the order count behind each number.”
The second one carries four ingredients: scope (entity and period), a comparison, a threshold, and the shape of the output. That pattern travels across almost every function you can name.
Follow-ups matter more than the opening question. Joule keeps context inside a session, so “break that down by country” or “which three customers drive the variance” costs you four seconds and often finds the actual story. One enormous prompt with nine requirements tends to produce one muddled answer.
Three worked examples you can adapt today
Finance: purchase orders with no goods receipt
Open with: “List all purchase orders above €50,000 created in the last 30 days that have no goods receipt posted, grouped by vendor.” Then follow with: “Which of those are more than 14 days past the expected delivery date?” The second question is the one that goes into a supplier conversation.
Supply chain: components about to run short
“Show components with projected stock cover below 10 days over the next eight weeks, ranked by how many production orders they affect.” A sensible follow-up asks whether an alternate plant carries the same material. This is the kind of request where people start checking whether the answer can be trusted, which is exactly the right instinct. Our guide to getting proof of value from SAP Joule covers how to turn that instinct into a measurement.
HR: voluntary attrition by department
“Show voluntary attrition by department for the last four quarters, excluding interns and fixed-term contracts.” Follow up with exit reasons for the two worst departments. Watch the exclusions clause. Without it, seasonal contracts will make a stable department look like a crisis.
Where Joule stops answering and starts acting
Everything above is assistant behaviour: you ask, it retrieves and summarises. Agent mode is a different animal. A Joule agent can chain steps together, pull a document, check a policy, and prepare an action for approval rather than just describing what you might do.
The practical question is where to draw the line. Let an agent draft a purchase requisition, gather the supporting quotes, and route it. Do not let it release the requisition. Approvals, payments, and anything that touches an employee record should keep a human click at the end for now. Our breakdown of what SAP Joule agents actually do and where they still need you is worth reading before you hand one a live workflow.
Capability here is moving quickly, partly because SAP has been buying rather than only building. The $1.16 billion bet on an 18-month-old German AI lab tells you the roadmap will change under your feet. Build your prompts and glossary so they survive a version bump.
A 30-day scorecard that keeps you honest
Track four things, nothing more:
- Time to answer your chosen standard question, measured before and after
- Share of prompts that produced an answer you could act on without opening the underlying transaction
- How many users came back in week two, which is the only adoption metric that means anything
- Average follow-ups per session. If it sits near one, people are treating Joule like a search box
Also log the questions where it gave you a transaction code instead of an answer. Those point straight at a glossary gap or a data quality problem worth fixing.
When Joule gives you a confident wrong answer
It will happen, usually in the first fortnight, usually in front of someone senior. The cause is almost never mysterious: a missing nuance in the glossary, a role with partial visibility, or a question with an ambiguous time window.
Build two habits. First, ask which objects or tables the answer came from and spot-check the figure in the source transaction. Second, run the same question as two different users and compare. If they diverge, you have an authorisation issue to fix, not an AI problem.
Turn the prompts that worked into a shared library
By day 30 you will have four or five prompts that reliably earn their keep and a dozen that flopped. Keep the winners somewhere people can find them, not in a chat thread that scrolls away. For each one, record the prompt text, the follow-ups that matter, who it is for, and the one caveat to watch.
Then pick the next workflow, hand the library to the team lead who owns it, and start again. Teams that do this three times stop asking whether Joule works and start arguing about which process to give it next.