Commercial teams have been shown a lot of AI this year. Some of it is automation with a better deck, which is useful and often overdue. Some of it changes what a twelve-person brand team can cover. A feature list will not tell you which is which.
The test that does
Language models are dependable on work that is high volume, well specified and independently checkable. Examples include reconciling a feed against the prior period, drafting a first-pass territory summary, flagging a variance against a threshold and assembling a briefing pack from governed data. In each case the person receiving it can see within seconds whether it is right, which is why this is where most teams see the first real hours return.
The picture changes where being wrong is expensive and the error is invisible on inspection. Interpreting an ambiguous coefficient, choosing what goes to a customer and judging whether an evidence gap is material enough to delay a submission all fall here. A right answer and a wrong one look the same.
Worth confirming before you sign
| Look for | Be wary of |
|---|---|
| Every output carries the query, source and timestamp behind it | Outputs that cannot be traced to a source |
| A named approver signs anything that leaves the building | An approval step treated as a formality |
| It writes to data, never to the structure of your systems | Recommendations with no stated reason |
| Every run is logged and can be replayed | Anything you could not switch off on a Tuesday |
Deploying it so the team keeps using it
In roughly this order
- Start where the work is routine Reconciliation and drafting. If it returns real hours there, the team will extend it themselves. If it does not, you have learned that at low cost.
- Make the review gate real A named person approves, with the capacity to actually read it. Without that time, the signature is a formality.
- Instrument the rejections Log what got rejected and why. Three months in, the rejection rate is the most informative number you have.
- Expand on the evidence Scope grows when the rejection rate falls, rather than when the quarter turns.
The gain is that the same people spend more of their hours on judgment, which was always the scarce part.
There is a limit. None of this tells you whether the model reasoned well, only whether the output was acceptable. That distinction matters most where you are least able to check it, which is why the boundary between automated and approved is worth drawing generously in favor of the person.
Sources and further reading
- Artificial Intelligence and Machine Learning in Software as a Medical Device, U.S. Food and Drug Administration. Retrieved: to be re-verified.
- News and press releases, European Medicines Agency. Retrieved: to be re-verified.