AI sales assistants: what they do, and what they don't

An AI sales assistant is software that takes over the administrative work around selling — finding phone numbers, logging calls into the CRM, drafting follow-ups, booking meetings — so that a rep spends more of the day in conversations and less of it in data entry. It does not sell. It clears the path to selling.

That definition is uncontroversial and almost useless for buying, because at least four genuinely different products are sold under it. They bill on different meters, they fail in different ways, and a team that buys “an AI sales assistant” without separating them usually ends up paying twice for one of the four and not at all for the one that was actually the bottleneck.

The four products sold as one category

1. Research and enrichment

The assistant builds and cleans the list: finds direct dials and emails, checks them against multiple providers, removes the records that have gone stale, and flags accounts against buying signals like hiring, funding or a job change at a former customer.

This is the category with the most honest value and the least honest pricing. It runs on a data waterfall — the tool queries provider A, then provider B if A came back empty, and so on — and it bills an enrichment credit for each attempt, including the attempts that found nothing. We worked that arithmetic through on a 5,000-record run in our Clay alternatives comparison; the short version is that the credit meter, not the seat price, is what determines the bill.

2. Writing

Drafting emails, LinkedIn messages and follow-ups from whatever the tool knows about the prospect. Usually priced per seat, usually the easiest of the four to evaluate, and the one where the ceiling is lowest — a first draft is a genuine saving, and it is a saving of minutes, not hours.

3. Execution

Actually placing the calls and sending the sequences. This is where a parallel dialer sits: dialling several lines at once, dropping the numbers that ring out before a rep ever hears them, and connecting the moment a human answers.

Execution is the only one of the four that changes the number of conversations rather than the time around them, which is worth holding onto — it is the reason the arithmetic below breaks the way it does.

4. Analysis

Call recording, transcription, scoring and forecasting. Recording and transcription are close to solved. Scoring and forecasting are not: both are learned from a CRM’s history, and a CRM’s history is a record of what a team used to believe about its pipeline, faithfully reproduced.

Most vendors sell one or two of these well and the rest as checkboxes. The useful first question in a demo is not “what can it do” — it is which of these four is this, really.

What an assistant is actually worth, worked through

Take a rep on a forty-hour week who loses six of those hours to CRM logging, list hygiene and scheduling back-and-forth. That is a plausible number and it is an input, not a finding — substitute your own.

An assistant that removes three-quarters of it returns four and a half hours, or about eleven per cent of the week. Now put those hours back on the phone. At twenty-five manual dials an hour, four and a half hours is roughly 110 extra dials. At a five per cent connect rate, that is five or six extra conversations a week.

Five or six. That is the honest size of an admin-automation win, and it is worth having.

Now change one variable. Give the same four and a half hours to a dialer running four lines instead of one, and the attempt count per slot is four times higher — that is arithmetic on the line count, not a performance claim. The conversations follow the attempts.

The point is not that one number is bigger. It is that the two wins are not the same size, and teams routinely buy the smaller one first because it is the one the category name describes. Admin automation gives a rep back their Tuesday afternoon. Execution changes how many people they speak to on it.

Where AI sales assistants reliably fail

Three failures come up often enough to plan around, and none of them appear in the category’s marketing.

Confident summaries of bad audio. Call summarisation degrades quietly. On a clean recording it is excellent; on a call with crosstalk, a speakerphone or a poor mobile connection it produces a fluent, well-structured summary of things that were not said. Nothing in the output signals which kind of call it was, and a manager reading fifty summaries has no way to tell.

Lead scoring that launders old assumptions. A model trained on closed-won history learns the shape of the deals your team already knew how to find. If your best segment was mid-market manufacturing because that is where your first two reps had contacts, the model will find you more mid-market manufacturing and present it as a discovery.

Personalisation the prospect can see through. An opener assembled from a funding announcement and a job title reads as scraped, and a prospect who has received four of them that week recognises the format before the second sentence. Our outbound call scripts go into why a specific-but-obviously-automated reference performs worse than an honestly generic one: the tell is not the wrongness, it is the machine-assembled particularity.

There is a fourth failure that is not the assistant’s fault and matters more than all three. If the number is showing as Spam Likely on the handset, none of the preparation reaches anyone — the call is judged before the rep speaks. Our guide to avoiding spam-likely labels covers what actually causes it, and it is mostly not what teams assume.

What to check before you buy one

Four questions, in the order they cost you money.

1. Which of the four is this? If the answer is “all of them”, ask which one the company was founded to do. That is the one that works.
2. What is the meter? Per seat, per record, per credit, per minute. A tool billing on two meters at once can quote you a low number on either.
3. Does a failed lookup bill? For anything doing enrichment, this single question predicts the invoice better than the headline price. Published rates for the dialer category are collected in our dialer pricing index.
4. Where does the output land? An insight that lives in a second dashboard is an insight a rep will not see. This is the most common reason a tool with good output produces no change in results.

Where Pharo fits

Narrowly, and it is worth being exact about it.

Pharo is category three. It is a parallel dialer: it dials two, four or eight lines at once, drops the dead numbers before they reach a rep, bridges the audio in under 200ms when a human answers, and writes the outcome back to the CRM. It enriches records through a multi-provider waterfall and passes that data through at cost — the providers are graded against each other rather than resold at a markup, and every record shows what it cost. Automatic spam rotation is on the Pro plan and up, not on Starter; the pricing page has the full split.

It does not do lead scoring, and it does not forecast. If those are the gap, the tool you want is in category four and it is not this one.

That is a smaller claim than the category usually makes, and it is the one we can stand behind. The rest of it — which of the four you actually need — is a question about where your reps’ hours are going, and you can answer it this week with a spreadsheet and no software at all.