MashnLearn started doing automated content work using IBM Watson in 2018. At the time, Watson was one of the few enterprise-grade options for natural language generation at scale. It worked, within limits, and the limits were significant.
The core limitation was that Watson's NLG capabilities were template-driven. You defined the output structure, and the model filled it in. For structured data like weather reports or financial summaries, this was adequate. For product descriptions — which require genuine variation in phrasing, the ability to foreground different product attributes depending on category context, and output that does not read like it was written by a machine — the results required heavy post-editing.
The second limitation was cost. Watson's pricing model was per-API call, with a structure that made high-volume content generation expensive enough to limit who could justify it. For most e-commerce businesses outside the enterprise tier, it was not accessible.
What changed between 2018 and now is significant in two directions. First, model capability improved by a large margin. The large language models available today — Anthropic Claude, xAI Grok, and the open-source alternatives — produce product descriptions that require far less post-editing than Watson-era output. The generation quality for unstructured creative and commercial text is categorically different.
Second, AI content generation cost has fallen dramatically. The per-token pricing on current commercial models is a fraction of what Watson-era API access cost for comparable output volume. Open-source models run on commodity hardware and have zero per-token cost once the infrastructure is in place. A content pipeline that was economically viable only for large retailers in 2018 is now accessible to mid-market e-commerce businesses.
The IBM Watson alternative question comes up less often now than it did three or four years ago, because most businesses that were on Watson have already moved. But the pattern we see in clients who delayed the migration is consistent: they stayed on Watson or similar legacy NLG systems because the migration felt complex, and eventually reached a point where the economics forced the move.
Migration from a legacy NLG system to a current LLM-based pipeline is not as complex as it appears from the inside. The data structures are different, but the integration pattern is the same. What changes is the output quality and the cost per piece of content. We completed our own migration from Watson in 2018, which is why we have been building on current-generation models for longer than most integrators in this market.