AI automation, wired into real systems

Work that a person currently does by hand, handed to a system that does it continuously and correctly. In practice that splits three ways. There is enrichment, where a model generates or cleans content at a scale nobody would attempt manually, such as product descriptions and category structure across thousands of items. There is triage, where incoming work gets read, classified, and routed to the right place automatically. And there is plain workflow automation, where systems talk to each other without a human copying data between them. Most engagements are a mix, and the least glamorous parts usually produce the most value.

The same fixed-price model as any other project, quoted after a discovery call. Smaller automations that connect existing systems and add a model where it earns its place start around $10K. Custom applications with agents, retrieval, and their own admin interface usually land between $25K and $50K. Platform work spanning several systems runs $50K to $100K and up. Ongoing model usage costs are separate and are billed to the client's own provider accounts, so the running cost stays visible rather than buried inside a retainer.

Where the work is high volume, rule-shaped, and currently done by a person who would rather be doing something else. Two live examples. Salem Wine & Liquor, a family-owned shop trading since 1966, has thousands of products enriched by an AI pipeline that writes accurate descriptions, sources bottle imagery, and sorts everything into a coherent category tree, work that nobody was ever going to finish by hand. BridgeLift Capital, a US lender, uses custom AI agents that read incoming deals, classify and route them, and feed the pipeline, replacing manual triage and data entry.

Claude and OpenAI models for generation, classification, and extraction, with retrieval pipelines where the model needs grounding in a client's own data. Orchestration runs either in application code, usually Laravel with queues so the work is retryable and observable, or in workflow tools like n8n and Make.com when the job is genuinely a connector problem rather than a software problem. Cloudflare Workers handle edge tasks where they fit. The choice is deliberately unromantic: whichever combination is cheapest to run and easiest for someone else to maintain later.

By treating a model as an unreliable component and designing around it rather than hoping. That means grounding output in the client's own data instead of the model's memory, validating results against known-good sources before anything is written to a live system, keeping a human approval gate on anything expensive or irreversible, and logging every automated decision so it can be audited afterward. Confidence thresholds decide what applies automatically and what waits for review. When a model cannot answer reliably, the correct behavior is to fail loudly rather than produce a confident guess.

Most land in the same 4 to 12 week window as other builds. A focused automation connecting existing systems can ship in 3 to 4 weeks. An enrichment pipeline running over a large catalog usually takes 6 to 10 weeks, because the real work is validation and edge cases rather than prompting. Anything touching several systems at once can run a quarter or more. Eyal Gantz has been shipping LLM features in production since 2021, and the automations he has built now remove more than 1,000 hours of manual work a month across his clients.