Most software sold as “AI sales coaching” does not coach during a sales call. It records the call, transcribes it, scores it against a rubric, and surfaces the result somewhere between an hour and a week later.
That is a genuinely useful category. It is also a different product from one that helps a rep while the customer is still on the line, and the two get shelved together in every buying process we have seen. This guide is for separating them, and for evaluating the second kind properly.
It contains no vendor rankings. Feature sets in this category change monthly, and a guide that asserts what a named competitor cannot do is out of date the week after it publishes. What does not change is the set of questions that expose whether a tool will survive contact with a live call.
The three categories, and why the distinction matters
Almost everything in this market falls into one of three groups.
Conversation intelligence / revenue intelligence. Records calls, transcribes them, analyses them in aggregate, and reports patterns to managers and RevOps. The buyer is usually a sales leader, the unit of value is the pipeline, and the output is a dashboard. Strong for forecasting, coaching programmes and deal inspection.
Meeting notetakers. Join the call as a participant, produce a transcript and a summary, push notes into the CRM. The buyer is often the rep. The unit of value is time saved on admin.
Real-time assistants. Run during the call and put something on the rep’s screen while the conversation is happening. The buyer is the rep or a front-line manager. The unit of value is the outcome of the call itself.
The three overlap on the surface — all of them touch call audio, all of them produce a transcript — and they are architecturally quite different. A tool built to analyse a finished recording is optimised for accuracy over latency, because nothing depends on speed. A tool built to help mid-call has to be useful within the few seconds a rep has before they must say something, which constrains everything about how it is built.
Neither is better. But if the problem you are solving is “my reps freeze on the objection,” a dashboard delivered on Friday does not solve it, and if the problem is “I cannot forecast,” a mid-call prompt does not solve that either. Most disappointing purchases in this category are a category mismatch, not a quality problem.
Start by writing down which problem you have
Before any demo, write one sentence describing the outcome you want to change. Then check which category can plausibly change it.
- “Our forecast is wrong” → conversation intelligence.
- “Reps spend their evenings on CRM admin” → notetaker, or the post-call half of a real-time tool.
- “New reps take two quarters to ramp” → could be either; the deciding question is whether you want to change what happens in the calls or how you review them afterwards.
- “We lose deals on the same three objections” → real-time, if you want the save to happen on the call. Conversation intelligence, if you want to build a training programme.
- “Discovery is shallow and we find out too late” → real-time.
If you cannot write the sentence, you are not ready to demo anything, and every vendor will happily fill the gap with their own framing.
The seven criteria that actually separate real-time tools
Once you are evaluating tools in the real-time category specifically, these are the axes where they genuinely differ. Feature checklists mostly do not capture them.
1. Latency, measured on a real call
The entire category rests on this. Guidance that lands after the rep has already answered is not guidance, it is a transcript with extra steps. Ask for the median and the p95 time from a customer finishing a sentence to something useful appearing on screen, and then measure it yourself during a trial rather than accepting the number.
2. Whether anything joins the meeting
Some tools attend the call as a visible participant. Others run locally and are invisible to everyone but the rep. This is not a cosmetic difference. A bot in the participant list changes the conversation — prospects behave differently when they can see they are being recorded by a named tool, and some enterprise buyers will refuse it outright. Ask specifically: does the customer see anything, in the participant list or on screen share?
3. What the rep is actually shown
There are two philosophies here and they produce very different behaviour on the floor. Some tools show the rep a score — a number, a health rating, a talk-ratio gauge. Others show guidance — the question to ask, the answer to the objection, the thread that has gone unclosed.
A score mid-call is at best neutral and at worst actively harmful: it gives a rep something to feel judged about at the exact moment they need to be present with a person. Guidance is actionable in the ten seconds it appears. Ask to see a real mid-call screen, not a marketing render, and ask what a rep is expected to do with what is on it.
4. Whether it knows anything about your business
Generic objection handling is available free from any language model. The difference in this category comes from whether the tool knows your product, your ICP, your competitors, your pricing structure and the objections your reps actually hear. Ask how that knowledge gets in, how long it takes to configure, who maintains it, and what the tool does on day one before any of it exists.
5. Call-app coverage and how it is achieved
Zoom, Google Meet and Teams are table stakes; the interesting question is how. An integration built on a per-platform API breaks differently — and gets deprecated differently — than one that works at the OS audio layer. If your team also sells on phone calls or in person, ask what happens there.
6. The privacy and consent story, in writing
Live call audio is the most sensitive data your sales org handles. Establish, in writing: what is processed in memory versus stored, how long anything is retained, whether your calls train models used for other customers, who inside the vendor can access a recording, and what a deletion actually deletes. Then check the answers against the vendor’s published privacy policy and DPA rather than the sales call.
You are also responsible for consent law in every jurisdiction your reps dial into, and that obligation does not transfer to the vendor. Two-party consent states and the EU have materially different requirements.
7. What survives the call
A real-time tool that leaves nothing behind creates a second admin problem. Ask what lands in the CRM without the rep typing it, what the manager can see afterwards, and whether the post-call artefact is a transcript or something a human would actually act on.
Questions worth asking in the demo
Vendors are ready for feature questions. These tend to be more revealing.
- Show me a call where the tool was wrong. What did the rep see, and what happened?
- What does the rep see in the first thirty seconds of their first-ever call with this installed?
- Which of these does the product do today, not on the roadmap?
- How many of your customers have a rep using this on every call, versus having bought seats?
- What is your median guidance latency, and how is it measured?
- Can we run a trial without connecting our CRM?
- What happens if your service is down mid-call?
That last one matters more than it sounds. A tool that fails loudly during a customer conversation is worse than no tool.
How to run a trial that tells you something
Most evaluations fail because they are run by the wrong people on the wrong calls.
Pick reps who are not your best. Your top performer already knows the answer to the objection. The value of real-time coaching concentrates in the middle of the distribution and among new hires, so a trial staffed with your two strongest AEs will under-measure the product.
Use real calls, not internal role-play. The behaviour that matters — a rep glancing at guidance while maintaining a conversation — does not reproduce in a mock call where nothing is at stake.
Run it for a full cycle of at least twenty calls per rep. The first three calls measure novelty and interface friction, not value.
Decide the success metric before you start and make it something you can observe: objections handled without a follow-up email, discovery questions asked, next steps agreed on the call, ramp time to first closed deal. “Reps liked it” is not a metric, though reps refusing to give it back is a meaningful signal.
Check what happens when they stop. Take it away for a week at the end. What the team complains about missing is the product’s actual value, and it is frequently not the thing the vendor led with.
Pricing models and where the cost hides
Per-seat monthly is the norm. Watch for:
- Minimum seat counts that make a genuine pilot impossible.
- Usage caps on calls, minutes or post-call analyses, especially on lower tiers.
- Annual-only contracts in a category this young. A one-year commitment to a product whose category did not exist three years ago is a real risk.
- Implementation and configuration fees, particularly where the product needs to learn your business to be useful at all.
- The cost of the seats that go unused. The failure mode in this category is not the product being bad, it is reps quietly not using it. Ask how the vendor measures active usage, and insist on seeing it during the trial.
Where salesviking sits
We are in the third category, deliberately and narrowly. salesviking runs as a local macOS overlay during the call, puts guidance rather than a score in front of the rep, and nothing joins the meeting — the customer sees no bot in the participant list and nothing on screen share.
That focus has costs, and you should weigh them. If what you need is org-wide forecasting analytics, a revenue intelligence platform will serve you better than we will. If you need coverage on Windows today, we do not have it.
If you are comparing us against specific tools, those pages are here:
And if you would rather just try it on your next call than read about it, that is the faster answer.