An AI agent did half a day of my analyst work in 3 minutes

A few weeks ago I pasted a stakeholder’s question into a powerful internal data analytics agent and watched it hand back a full report in under three minutes. Writing that SQL and Python by hand would have cost me half a day. I felt two things at once: how good this is, and a small cold drop of what does this mean for me?

I’ve done this job for years, and I’ve come to believe it’s changing faster than most of us want to admit. The analysts who learn to work with these tools are about to pull well ahead of the ones who don’t. That belief is why I’m writing this. What follows is the honest version from the inside: three tools, one I built, two that beat it, and what I learned about staying useful while the work gets automated out from under me.

The bottleneck was never SQL

Start with what the job actually was, because it isn’t what people picture. The hard part was never writing SQL. In a warehouse with thousands of tables, the hard part is knowing which one truly answers the question, then trusting it once you’ve found it. Every request runs the same quiet gauntlet: hunt down a candidate table, validate it for nulls and duplicates, sanity-check that the values even make sense (you expect an order count near a thousand, you sum the column, you get ten thousand, and now you have a mystery on your hands), and only then write the query with the right filters and join keys and build a pivot to see whether the numbers follow the business story. A pull I knew well took minutes. One where I didn’t know the table could eat the better part of a day.

Five steps, and AI is coming for four

Every ad-hoc data request runs through the same five steps. Only the first is truly human: turning a vague business ask into the real question, often from a stakeholder who can’t quite tell you which number they need. The other four are pattern work, and pattern work is exactly what these tools have gotten good at. Four of the five steps are already automatable. Step one holds, for now, and only until the context in people’s heads gets written down.

Boko: a bot that knew where things were

I started small. I built a lightweight assistant (I’ll call it Boko) into our internal chat app, on an internal low-code AI-builder (think Coze). It followed a LangChain supervisor-subagent pattern: a cheap router model greeted you and handed off to a per-function sub-agent (Customer Service, Return & Refunds, Seller Operations), and each sub-agent answered only from its own function’s documentation, retrieved from an enterprise vector database. Little code; a lot of documentation.

The answering sub-agents ran on Claude Haiku, which made Boko cheap, light, and fast. It also set the ceiling. Boko was great at “what does user contact actually mean?” or “what’s the calculation logic behind this OKR?” Those are the logic and discovery questions that used to land in my DMs. It could even draft SQL for people to run themselves, though on a small model that came with a real hallucination rate, so I never let it pretend to be authoritative on queries it couldn’t check.

It worked better than I expected. To date Boko has answered 400+ questions from 60+ people, and at its peak it was clearing around a hundred a month. The reach was deliberately narrow (you needed to be at least SQL-curious to get value), but it did something I didn’t plan for: it made the team better at documentation, because now the docs had a job. Other functions started handing me their knowledge bases to onboard too.

Then two better tools beat it, and I was glad

Around April 2026 our data infrastructure team shipped two tools that, frankly, ran circles around Boko. Both used frontier models far better at tool-calling than Haiku, and both did the thing Boko never safely could: run the query.

The first (call it Astro) let you upload the tables you already knew, then wrote and executed SQL on the spot, generated charts, and explained the results in plain language. Watching it turn an upload into a chart in seconds was the first time I really felt the ground move. Its catch was structure: everything was siloed by topic, so a question spanning one person’s tables and another’s meant standing up a whole new topic. Fine for a glimpse of the future; not how analytics actually works day to day.

The second (Clyde) was copilot-style, like Cursor or Claude Code for the warehouse: you could extend it with rules, skills, and MCP connectors. It explored the warehouse, recommended the right table, and wrote and ran the query. It was Boko’s job done properly. I could have felt precious about being out-built. I didn’t. I’d wanted a tool that could execute since the day I started Boko. Its retirement was inevitable, and I was glad to see it come.

What I actually built into Clyde

Here’s the part that matters: Clyde was only as good as what you gave it, and for my users, what it got was mine. I solo-built the skills and connectors that made it an expert on our data.

Four skills, each encoding a piece of that gauntlet from the top. One reads our documented table catalogue before writing any SQL, so Clyde picks from vetted definitions instead of guessing. One is a memory bank that records confirmed resolutions and reuses them, so a question solved once stays solved. One renders our house-styled charts in matplotlib, collapsing what used to be a full day of building a report by hand. And one makes Clyde interrogate the request before answering. That’s the oldest analyst instinct there is: make sure you understand the business question first. Two MCP connectors let it read external context on demand: our shared spreadsheets and internal wiki pages. I stopped writing most of my own SQL.

Notice what carried across all three tools. The models changed completely: a small one, then two frontier ones. What stayed, and what made each one useful, was the documented knowledge and the workflow I’d captured. The moat was never the bot. It was the structure.

The hard part was people

The technology was the easy half. The hard half was getting people to use it, especially the less technical folks and the old hands, who’d still rather DM me than ask a bot. I never fought it. I answer them, and I answer by sending back the Clyde conversation link that already solved it, so they get their number and they see, in one move, that the tool exists and does this well.

So is the data analyst finished? I don’t think so, but the job is not the one I trained for. The pull-the-numbers labour is going to AI, and it should. What survives is the analytical structure a good analyst brings: knowing which table to trust, what to validate, how a metric is really defined. That structure is exactly what these tools need to be any good, and it is portable straight into them. The people who thrive will be the ones who stop guarding that structure and start pouring it into the tools. I spent years being the person who knew where things were. Now I spend it teaching the tools where to look, and honestly, that’s a better job.