Three things we see sales AI fail on

Sales AI is landing in portfolio companies at pace, and the technology mostly works. The implementations we are called in to rescue rarely failed because the algorithm was wrong. They failed on conditions that sat outside the vendor’s scope and so went unowned.

Three come up repeatedly, and there are others.

In our experience none of the three tend to be true at the start, and all three tend to fall to you.

A rep-assist tool trained on your CRM history tends to send reps back to the accounts they already serve, with a confidence score attached.

The limitation sits in the input rather than the algorithm. CRM history records what you sold, to whom and when. Potential is an outside-in question, e.g. market size, sector signals, estimated share of wallet, lookalikes against your strongest accounts, and much of that data is usually not in your system yet.

Last month’s German distributor shows the cost. Reps were measured on visits and spread thin across low-value accounts. Train a recommendation engine on that history and you risk industrialising the pattern rather than correcting it.

The work: build the potential view as its own short workstream. It is an AI job too, and often a quick one: enrichment and scoring can give you a defensible first cut in weeks rather than quarters. The sequencing is what matters: it wants to land before the assist tool, because it is what the assist tool steers against.

The binding constraint on a rep’s week is usually decision load rather than information. Most sales AI arrives as visibility, better forecasting, another dashboard, and visibility was rarely what was missing.

The decisions that absorb the week are where to go, when to go, who to contact and what to bring when they get there. That is what a tool can usefully take off them.

The work: map the sales process end to end before anything is scoped, so the bottleneck is identified rather than assumed. In our experience it tends to concentrate in three places: whitespace the account is buying elsewhere, meeting preparation and value-selling material, and the weekly routing decision. Then name the decision the tool will remove. If it cannot be named in a sentence, what is being built is probably a reporting project.

Adoption often turns on the click count, and many solutions lose on it. AI’s promise here is fewer steps, but much of what gets deployed is bolted onto a legacy CRM and inherits its engagement model wholesale, so the AI-enabled path can end up longer than the spreadsheet the rep already uses.

Adoption then tends to fail quietly rather than loudly. Nobody rejects the tool; the rep goes back to the workaround and the licence renews.

The work: treat the click count as a design parameter rather than a vendor feature. Put a rep into the environment during configuration, have them run a real Monday morning, and count against what they do today. Freeze the configuration when it comes out fewer.

None of the three is a technology question, which is why they rarely surface in a technical evaluation and do not arrive with the software. Building the potential view, naming the decision and holding the click count belong in the implementation plan, with owners and budget on your side of the line, ideally before the tool is configured rather than after it has quietly stopped being used.

This piece is the fourth issue of the Hirondl newsletter, one idea a month from inside the businesses we work in. Subscribe here.