Prompting is a foreign language
4 of 5 sellers weren't using Sales Assist for one simple reason: they didn't know how to create effective, efficient prompts. Prompting is a new way of interacting with AI, and nobody had taught it.
Exploratory research · ServiceNow
Why sellers walked away from their AI assistant.
Sales Assist could answer a seller's question in seconds — yet almost nobody used it. I sat down with sellers one-on-one to find out what was keeping them away, and what would bring them back.
6
1:1 seller interviews
4/5
didn't know how to write effective prompts
5/5
dissatisfied with response accuracy
3/5
found outputs overly verbose
What we asked
Understand the seller before tuning the model.
- How do sellers perceive Sales Assist, and why is adoption low?
- How do sellers write prompts, and how do they judge whether an answer is right?
- Where should the assistant live — and what would make sellers come back?
Mental models
Sellers' conceptualization of, and motivations for, engaging with Sales Assist — what they expected it to be for.
Input methods
How sellers actually enter prompts: most prompted like they were talking to a human, or like they were using search — and didn't know the ideal prompting techniques.
Output feedback loop
How sellers check that an answer is accurate and desirable. Novice sellers often didn't test enough and trusted inaccurate responses.
How we ran it
- Method
- 1:1 interviews over Zoom, walking through Sales Assist live
- When
- Nov 18 – Nov 28, 2023
- Participants
- 6 sellers: Account Executives, Solution Sales Executives, and a Renewals Account Manager
- Geos
- Americas (3) and EMEA (2)
- Focus
- Perception and low adoption, feedback on AI-generated responses, and integration in existing tools
Key findings
Broken answers break trust
All 5 sellers were dissatisfied with the platform's responses, citing accuracy and trustworthiness. That dissatisfaction created a barrier to entry — sellers were discouraged from ever returning.
Meet sellers where they work
3 of 5 sellers recommended integrating Sales Assist directly within Dynamics; others wanted it in the Sales Success Center or Tableau. A standalone platform was an entry point nobody remembered.
Paragraphs don't sell
3 of 5 sellers found the outputs overly verbose — lengthy paragraphs where a tabular layout or bullet points would have been far easier to read and act on.
Novices don't fact-check
Sellers who didn't test the system enough trusted inaccurate responses and shared wrong information with customers — turning an assistant into a liability in front of a buyer.
Accessibility is missing
Key functionality like translation, read-aloud, and spell check was absent, limiting adoption for sellers with different needs and abilities.
What sellers were missing
- Training — no guidance existed to teach prompting techniques, so most sellers never learned the platform.
- Socializing — sellers were unaware of Sales Assist's current capabilities, limitations, and future use cases.
- Fact checking — inaccurate responses forced sellers to other resources and slow manual verification.
- Accessibility — for sellers with disabilities, the missing features alone kept adoption at zero.
In their words
“This is totally inaccurate. I shared this information with the customer and ended up feeling embarrassed, had to clean the mess.”
“This doesn't help me; Enterprise Bing AI gives me a much more detailed response.”
“Would have expected to see some reference links.”
“This is exactly what I would have expected to see.”
What I recommended
Adoption wasn't a feature problem — it was a teaching, trust, and integration problem. Six moves to fix all three.
- 01Run comprehensive training sessions on Sales Assist utilization and effective prompting, tailored to sellers' daily workflows.
- 02Offer output formats sellers prefer — tables and bullet points — to improve readability and usability.
- 03Enhance the tool to retrieve data from Dynamics for more comprehensive, contextually accurate responses.
- 04Let sellers upload transcripts from customer meeting recordings and search for specific information within them.
- 05Integrate a robust fact-checking mechanism to raise the reliability and accuracy of generated responses.
- 06Add accessibility features: translation support, text-to-speech read-aloud, and spell-check, for inclusive usability.
Next steps
Conduct trainings and demos so sellers understand the platform — its capabilities and its limits — and adoption climbs.
Get feedback from more sellers and APJ representatives to map the full range of expectations.
Work with product to prioritize the opportunities sellers themselves suggested.