AI assistants are now capable enough to change how a seller spends the week. They read call transcripts faster than any person, draft a spreadsheet in seconds and will play a sceptical finance director at eleven at night. Used well, they take much of the grind out of building an evidence-based case. Used carelessly, they produce confident numbers nobody has checked, which is the opposite of what the case needs.
Our rule is short: agents gather, draft and monitor; reps verify; judges still decide.
Where an assistant helps
Evidence of the problem. An assistant can search call transcripts, emails and support tickets for the same issue raised by different people, and group the pains that share a root cause. It can turn a pain register into a cost model, labelling every input as customer-supplied, benchmark or the seller's own estimate, and flag the figures the total is most sensitive to.
Pressure-testing. Asked to act as a sceptical finance director, an assistant will find the figure most likely to be attacked before the real meeting. In our worked example it challenged a 20% discount-leakage assumption; the seller raised the point first, and the customer's analyst settled on 15%.
The judges. From an organisation chart and call notes, an assistant can propose who the three judges are, how confident it is, and the question that would confirm each one. It can prepare interview questions and draft the written play-back of each judge's criteria.
Evidence of the solution. It can design the smallest experiment that would produce a judge's evidence, track results against the agreed measure, and tell the seller what to raise with the judge now rather than at the end.
Review. Given the evidence pack and the judge's criteria, it can deliver a strict mock verdict. In the worked example that review, run the day before the finance meeting, found internal time shown as “0.5 FTE” when Finance's template asked for named roles and day rates. It was fixed that evening.
Where it does not
An assistant does not know whether a figure is true. It does not hold the relationship with the customer, and it cannot give a verdict on anyone's behalf. Every number it produces has to be checked by the seller before it reaches a customer, because the value of the whole approach rests on the customer being able to trace each figure to a source.
There is a subtler risk as well. The more polished a draft looks, the less likely anyone is to question it. A cost model the customer has corrected is worth more than one an assistant has made immaculate, so we use assistants to prepare for the challenge, never to avoid it.
Setting it up
We suggest one workspace per opportunity, holding the workbook, call notes and cost model, connected to the CRM, call recorder and shared drive where available. Placeholders are replaced with real names and figures, and every output is reviewed before it is used.
Done this way, the assistant makes the seller quicker at investigating and more consistent at recording what they find. The judgement stays where it belongs.