2024 was the year of demos. 2025 was the year of pilots. 2026 is the first year in which somebody asks for the invoice — and that question is healthy. Below are seven ways that invoice comes out negative.
A catalogue, by symptom
None of these traps arrives with a label on it. All of them are recognisable from what you can see in the team after a few weeks:
| Trap | How you recognise it |
|---|---|
| Tool instead of problem | The rollout began with a product name, not with a number you want to improve |
| Verification costlier than the work | The rep reads the generated text longer than it would take to write their own |
| Garbage in | The model gets an incomplete CRM and produces convincing untruths |
| Automating a bad process | You sped up a step that should have been deleted instead |
| Recordings used as control | The team speaks more cautiously in meetings, data quality drops |
| Purchase without an owner | Licences bought, nobody owns adoption, 20% log in after a quarter |
| Building your own agent | Prototype in a week, maintenance forever, sales ends up maintaining software |
The first six can be reversed within a month. The seventh usually costs a year — because nobody wants to be the person who shuts down a project announced at an all-hands.
There is an eighth, unnamed: executive enthusiasm. If the tool was chosen at a conference rather than in a conversation with reps, adoption will be performative — the team will learn to report usage, not to use it.
A test that takes thirty seconds
Before buying any AI tool, there is one question worth asking: can the output be checked in thirty seconds?
If yes, the task suits automation. A call summary, a CRM entry, qualification fields filled in — all of them can be verified at a glance, because each has one correct answer and you know what it is.
If verification takes as long as doing the job by hand, the tool is not removing work. It is moving it — usually onto the rep who was meant to save the time.
The same rule explains why some AI use cases work and others don't. The more a task is about faithfully recording what already happened, the better it turns out. The more it is about taking responsibility — judging a situation, deciding on a discount, answering a customer — the worse. We mapped that separately, along with the list of things that are simply marketing: AI in B2B sales, signal versus noise.
The pilot that never ends
The most common way a rollout dies is not rejection. It is never being decided — and it always looks the same:
| Element | Actual state |
|---|---|
| Pilot start | March, three people, “let's see how it works” |
| Defined success metric | none |
| Baseline measured before start | none |
| Owner on the sales side | “everyone a bit” |
| Decision in June | extend the pilot |
| Status in October | two users left, topic returns at budget time |
None of these lines is about technology. All of them are about a decision not taken at the start. A pilot without a baseline can neither succeed nor fail — it can only drag on, and dragging on looks safer inside an organisation than closing something down.
Three conditions are enough for a pilot to give an answer: a number measured the week before it starts, a decision date in the calendar with the name of the person who decides, and consent to a negative result said out loud at the beginning. Without the third, every pilot ends in success and none changes the number.
What not to measure
Volume of generated summaries, number of model queries, time spent in the tool — those are the vendor's metrics, not yours. All three rise with team engagement and none says anything about revenue.
Four that do say something: admin hours per person, time from meeting to follow-up, qualification field completeness and forecast accuracy. The first is primary, the second reacts fastest, the fourth is decisive — but only after a quarter.
The whole balance, in twenty pages
The above is one chapter. The rest of the playbook is a map of use cases split into “works, limps, harms”, four levels of team maturity, upsides and downsides in one table — including those that apply to tools like ours — and a 90-day rollout plan.
Key takeaway
AI rollouts in sales rarely lose to the technology. They lose to the absence of a number measured before the start, the absence of an owner, and the absence of a date on which somebody has to say “it works” or “it doesn't”.