How Claude and Alteryx Work Together to Automate More Than Just Data
The workflow was never the whole job
Alteryx has been taking the grind out of data work for years. Joins, reconciliations, validations, and reports that used to eat a morning now run in seconds.
But look at what still happens around the workflow.
Someone decides which workflow to run, and when.
Someone deals with the notice that arrives in a layout nobody has seen before.
Someone checks the output, chases the exceptions, and posts the result into another system.
When a workflow breaks, someone digs through it to find out why.
Alteryx is not short of capability here. It parses PDFs, runs OCR on scanned documents, calls APIs, and mines text. What has been missing is the layer in between: the judgement that decides what happens next, and the hours it takes to build, change, and fix the workflows themselves.
Over the past year, our team at Continuum has been pairing Alteryx with Claude, Anthropic's AI model, to close that gap. This post is about what we have built, what we have learned, and why the combination matters for anyone running data-heavy operations.
Alteryx brings rigour. Claude brings judgement.
The two tools are strong in different places, and together they cover far more ground than either does alone.
Alteryx is the engine. Document parsing, OCR, data preparation, governed calculations, and integrations with the systems you already run, all with a clear record of exactly what ran and when. In financial services, that is not optional.
Claude adds judgement and speed. It handles inputs that do not fit a fixed template, works out which workflow a task needs, explains what it found in plain English, and can build or fix the workflows themselves.
Claude does not replace your Alteryx logic, and Alteryx is not asked to make judgement calls. Claude works out what needs to happen and prepares it. Alteryx carries it out, the same way, every time. A person stays in the loop for anything that matters.
Skills: teaching Claude how the work is actually done
Out of the box, AI models know a little about Alteryx. A little is not enough. Ask a general-purpose model to write an Alteryx workflow and you will usually get something that looks plausible and fails the moment you open it in Designer.
So we build skills. A skill is a packaged set of instructions, reference material, and tooling that Claude loads when it needs it. Think of it as the difference between a bright graduate and a colleague who has been on the job for five years. Our skills include:
Workflow builder. Verified configuration patterns for dozens of Alteryx tools, every one taken from real, working workflows rather than guessed. It covers formula functions, engine selection, macros, analytic apps, API calls, and reporting.
Workflow surgeon. For fixing what already exists. It reads the workflow first, changes only what needs to change, and runs a semantic comparison before and after, so a fix in one place does not quietly break something elsewhere.
Alteryx to Python. For clients moving some workloads off Designer, it converts existing workflows into runnable, tested Python.
Process skills. These encode how a specific job is done end to end, such as reading and posting a capital call notice, or turning a set of pasted debits and credits into a journal ready for review.
The skills are where most of the effort goes, and most of the value. They capture what our consultants have learned over many years of Alteryx delivery, in a form an AI can use consistently.
Keeping people in control
The first question every client asks is a fair one: what stops the AI doing something it should not?
The answer is in how it connects. Claude reaches Alteryx through a single, authenticated gateway, using the Model Context Protocol (MCP), an open standard for letting AI models use tools safely. That gateway sets the rules, not the model.
Read-only by default. Database access for analysis is read-only.
Run, then tidy up. Ad hoc workflows run on Alteryx Server and are removed afterwards, so nothing lingers.
An approval gate for anything permanent. If Claude proposes a fix to a production workflow, it is staged and shown to a person. Nothing is published until someone approves it.
Confirm before posting. Before any transaction goes into a client system, Claude shows exactly what it is about to post and waits for a yes.
The result is AI that speeds people up without taking the decisions away from them.
Built with Claude, open to any model
This started as a proof of concept built with Claude and for Claude. Claude was the model we trusted to work carefully inside a regulated process, so it was where we did the hard work of testing, failing, and refining.
What came out of that work is not tied to one model. The gateway uses MCP, an open standard, so any AI model that supports it can connect to Alteryx in the same governed way. The skills are plain instructions, verified reference material, and tooling, so they travel too.
That matters if your organisation has already standardised on an AI platform, or expects to change in future. You keep the Alteryx logic, the controls, and the know-how, whichever model sits on top. Claude remains our first choice, but it is not a lock-in.
In practice: a capital call notice, from inbox to ledger
Capital calls are a good example because they are routine, high value, and unforgiving of mistakes. Notices arrive in every format imaginable, and each one has to be read, checked, and booked correctly.
Here is how the process runs with Claude and Alteryx working together:
The notice arrives. PDF, email, spreadsheet, or a pasted table. Format does not matter.
The notice is read. Alteryx parses the document, running OCR if it is scanned. Claude interprets what comes out, pulls the fund, investor, amounts, purpose split, and due date, and matches them to the right entities and accounts, even when the layout is new.
The figures are checked. Do the lines add up? Is the currency consistent? Is anything missing? Problems are flagged before anything moves.
A person confirms. The reviewer sees exactly what will be posted, in plain terms.
Alteryx posts it. A governed Alteryx app sends the transaction into the fund administration platform.
Alteryx verifies it. A reconciliation workflow checks the posted transaction against the data warehouse.
The audit pack is produced. Source notice, mapping, checks, and results, bundled together for review.
The repetitive parts are automated. The judgement calls stay with the people accountable for them. And there is a full record of what happened at every step.
What we have learned along the way
None of this worked first time. A few lessons stand out.
Verified beats clever. The biggest improvements came from grounding Claude in patterns we had tested in Designer, not from cleverer prompts. When a pattern was wrong, we fixed it once in the skill and it stayed fixed.
Small changes are safer than rewrites. Asking AI to rebuild a workflow from scratch is tempting. Asking it to change the one thing that is broken, and prove nothing else moved, is far more useful in a live environment.
Control is what makes it usable. Clients in regulated industries do not need AI that acts alone. They need AI that does the legwork and hands a person a clear decision. Designing for that from day one has made adoption much easier.
Expertise still matters most. The AI is only as good as the knowledge behind it. Our consultants' years of Alteryx delivery are what make the skills worth using.
Where to start
If your team already uses Alteryx, you have done the hard part. Your logic is governed, documented, and repeatable. Claude, or whichever AI model your organisation uses, can build on that rather than replace it.
The best place to start is usually one process your team finds tedious but cannot afford to get wrong. Notices, reconciliations, journal postings, or the workflow that breaks every month-end.
