GPT-3: first serious tests on client data

We got access to GPT-3 in its first week and ran it on summarizing sales calls. The results were impressive in ways that previous systems had not been. Tone felt natural and appropriate. Format matched what was asked. On clean data it looked like magic and worked smoothly.

The limitation showed up in testing at scale. We tried five hundred real sales calls and found the model invented things. It would hallucinate product names that sounded plausible but did not exist in the catalog. It altered figures. Sometimes it shifted the date of a commitment.

Those errors are not the ones people mention when they talk about GPT-3. They talk about creativity and generalization. But for commercial use, the boring errors are the ones that destroy trust and cause incidents with customers and stakeholders.

We built a verification layer before we showed anything to the client. Every summary got checked against the CRM records. If a product name did not exist or a date did not match the records, the system flagged it and a person reviewed it carefully.

That verification layer was actually the bulk of the engineering work. Building the summary prompt was straightforward and relatively easy. Building genuine confidence that the output was accurate and safe for commercial use was harder and consumed most of our time and effort.

The takeaway: impressive demos and useful production systems are different problems entirely. The gap between them is where you spend money and time if you actually want to deploy this in production reliably.