How we automated 15 workflows without a data team

The problem with AI adoption isn't access — it's the last metre

Most AI initiatives die before they ship. Not because the technology is bad, or the demos were unconvincing. They die because nobody designed the last metre: the gap between "the model can do this" and "our team actually uses it every day."

We spent six months embedding AI-first workflows across a 100-person professional services organisation. Here's what we learned.

The decision framework

Before writing a single prompt, we ran every candidate process through three questions.

Is there a clear success state? If you can't define what "done correctly" looks like, AI will produce output but you won't know if it's right.

Does it happen often enough to compound? A process that runs once a week has 52 improvement opportunities per year. One that runs hundreds of times daily compounds faster. We prioritised frequency.

Can you verify the output? AI mistakes that get caught immediately are recoverable. Mistakes that propagate quietly through downstream processes are expensive. We only automated processes where output quality was observable.

The filter eliminated roughly half our initial list. The remaining processes were worth building for.

What we actually built

PSR Creation Agent — A ticket creation assistant connected to Jira via MCP. It reads a natural-language request, extracts customer name, product, funding, and timeline, validates all mandatory fields, and creates the ticket. What previously took 10 minutes of manual form-filling takes under 1 minute. The agent refuses to create the ticket if mandatory fields are missing — building quality into the creation step rather than auditing it afterwards.

PSR Routing Automation — A Jira-native automation that re-validates every ticket 5 minutes after creation, checks mandatory fields, auto-assigns to the right SME by product, and returns feedback to the reporter if anything is missing. Freed 13.5 engineering hours per week from manual triage. Nobody noticed the pipeline; they just stopped losing hours to it.

Aurora Readiness Component Analyzer — A dual-AI assessment tool using Claude 3.5 Sonnet and a LibreChat Aurora Agent running in parallel. Identifies every customisation in a customer's environment, maps compatibility, and produces a 20+ column readiness report. The same assessment previously took 1–2 weeks of manual engineering work. It now takes under 30 minutes.

SOW Generator Agent — Connected via MCP to JIRA, Confluence, and four product documentation systems. Pulls the current project request, fetches historical project data, and generates a structured Statement of Work with milestones, deliverables, assumptions, and scheduling. The model doesn't guess — it references actual similar past projects and current product specs.

PS Knowledge Assistant — A knowledge base agent trained on process documentation, pricing rules, and policy. Gives anyone in the organisation instant, source-cited answers to questions that previously required pinging a senior team member. Every answer includes the document name, owner, and last-updated date.

What surprised us

The highest-ROI automations are the invisible ones.

Nobody celebrates the routing pipeline. Nobody sends a message saying "great job, the ticket validation ran correctly again." The system just works, quietly, while the people who used to do it manually spend those hours on work that actually requires judgment.

This is the opposite of how AI projects are typically pitched — where the goal is a visible, impressive demo. The most durable implementations are the ones that disappear into the operating model.

The second surprise was how much the clarity of the process mattered more than the sophistication of the AI. The best results came from processes that were already well-defined but simply tedious. The worst results came from processes where the underlying workflow was unclear — adding AI to a poorly-understood process just automates the confusion.

The executive lesson

AI adoption is an operating model question before it is a technology question.

The bottleneck is rarely the model's capability. It's usually one of three things: the process doesn't have a clear enough definition to automate, the output isn't verified so errors propagate, or the automation is built in isolation and nobody else uses it.

The right starting point isn't "what can we do with AI?" It's "what is consuming our best people's time that follows a pattern?" Find that, define it precisely, build the shortest path to automated output, and make sure someone can verify the result. Then move to the next one.

Fifteen workflows automated, compounding. That's what it looks like in practice.