For founders
From data to decisions in seconds
A sales call produces a recording, a transcript, a summary and a set of flags. Reps had data, not answers. What I learned designing AI features inside a call-intelligence platform at Ringover: design backwards from the one decision a user has to make.
Prashant Kumar Singh · · 4 min read
In short
- More AI output is not the goal. Users want one decision they can act on in seconds, and the design should start from that decision and work backwards.
- A call produces a recording, transcript, summary and flags; the useful product is what the rep does next, surfaced in one place.
- The questions reps actually ask of a call are simpler than product teams expect: what did they object to, what did I promise, what do I send next.
At Ringover I design AI features for Empower, a call-intelligence platform used by sales teams. A call comes in, the platform records it, transcribes it, summarises it, and flags moments: an objection, a competitor mention, a pricing question. That is a lot of output per call. When I started, the honest assessment was that reps had plenty of data and very few answers.
The work since has been about one shift: from showing what the AI found to showing what the rep should do. This post is what that shift looked like in practice.
The problem with output
A summary is faster to read than a transcript, but it is still something to read. A list of flags is accurate, but it is a list, and after a day of calls a rep has forty lists. Each artefact is correct and none of them answers the question the rep has at 5pm: what do I need to do about today's calls before tomorrow?
The AI was producing more and more of the right information in forms that required the user to do the final step. The final step, turning information into a decision, is the whole job.
Design backwards from the decision
The method that worked was to start from the end. For each call, what is the one decision a rep makes? Usually one of a small set: send a follow-up with a specific attachment, schedule a next step, escalate to a manager, or close the record. Then: what is the minimum the rep needs to see to make that decision with confidence? Usually one line of context (what was promised, what was objected to) and one link to the evidence (the moment in the call).
Everything else, the full summary, the transcript, the flags, stays available one click away. It is not removed; it is demoted. The decision and its evidence come first.
The questions people actually ask
Sitting with reps changed what we built more than any model improvement. The questions they asked of a call were not the ones the product had been designed around. Nobody asked about sentiment over time. They asked: what did they object to, what did I promise, what do I send next. Three questions, all answerable from the transcript, all more useful than a dashboard of flags. Designing for those three made the feature right more often, and being right more often is where trust comes from.
| Before | After |
|---|---|
| Summary, transcript, flags per call | One suggested next action per call, with the evidence one click away |
| Rep reads, then decides | Rep confirms, edits or overrides |
| Insight features designed from the model's output | Features designed from the three questions reps ask |
| Time per call to process: minutes | Seconds |
What this means for your AI feature
Whatever your product's AI produces, ask what the user does with it. If the answer is "reads it and then decides something", the design is one step short. Find the decision, put it first, put the evidence one click behind it, and keep the full output available for the times someone wants it. The model does not need to change. The order on the screen does.
Questions people ask
- How do you design AI features for sales teams?
- Start from the one decision a rep makes after a call (follow up, schedule, escalate, close), show the minimum context needed to make it, and keep the full AI output one click away rather than first.
- Why do users ignore AI summaries?
- Because a summary is still something to read and then act on. Users want the action, with the evidence available if they want to check it. Output that stops one step short of the decision gets skipped.
- What questions do sales reps ask about a call?
- In practice: what did the prospect object to, what did I promise, and what do I send next. Simpler than product teams expect, and all answerable from the transcript.
If you're a founder
If your AI feature ends with the user reading something, send it for the free teardown and I will show where the decision should sit instead.
If you run an AI lab or data team
The same principle applies to grader output: a score with a trace to the failing criterion is a decision a reviewer can act on. A score alone is output.
Sources