Operational Lessons from Applying LLMs in Mid-Market Companies

Lessons from optimizing operations across several mid-market companies ($40M – $300M) with LLM help at DevriX:

1. Operational efficiency IS possible when workflows are proven to work

2. Revenue generation is only possible when data and brand MOATs already work and pipelines are self-managed by now (not for new GTM initiatives that require a lot of human touch)

3. New agentic frameworks, MCPs, A2As, Clawdbots are mostly too early for any established business

4. The most impactful results are seen in complex workflows with multiple people and 10+ steps, where an existing automation (deterministic) is tied to one or two non-deterministic steps via language models

5. Provisioning out-of-the-box agents and chatbots may lead to severe security breaches and significant token use (even accidental runs were $1,000 – $2,500 per single looped run across several different tests)

6. Vibe use is definitely helpful (v0, Lovable, Replit for demos and prototypes). Not for public-facing systems exposed to the web + sensitive data (these are built in-house after prototyping)

7. Image generation should be treated carefully. Some media is obviously coming from LLMs, and this may lead to both brand and contractual violations

Lastly, our teams can’t operate effectively without Cursor, a handful of LLM APIs, support in Google Drive or BigQuery from Gemini, and MCPs to HubSpot. This transition is taking place for good.

But maximizing value creation through EBITDA optimization and governance is a delicate act that delivers the strongest output in late 2025 and 2026.