Skip to content
AlanVaneAI & Software

Automation · between the tools you already own

The work between your tools shouldn't be someone's job.

Today

  1. an enquiry lands in the inbox
  2. someone opens and reads it
  3. works out who it's for
  4. forwards it
  5. writes the same reply again

Automated

  1. an enquiry lands
  2. sorted and routed
  3. standard reply drafted

Only the unusual ones reach a person — which is the part worth their time anyway.

Most of what gets called “AI automation” doesn't need AI at all — and I'll tell you which parts of yours do. Custom automation, built once, handed to you, left running. I'm Alan: one person, and you talk to the one who writes the code.

What actually gets automated

The pattern is nearly always the same: a person is moving information between systems, or making the same document from the same inputs over and over. Where a product already covers it, buy the product. Where nothing quite fits, custom is usually cheaper to build than people expect.

Before and after

What each one saves in your version is measured in discovery, not guessed at on this page.

Supplier invoices into the ledger

Today

  1. invoice arrives by email
  2. someone opens it
  3. types the lines in by hand
  4. files the PDF
  5. chases the ones that don't add up

Automated

  1. arrives
  2. fields extracted and checked
  3. written to the system
  4. only mismatches reach a person

Routine invoices stop touching a person entirely. What lands in the review queue is exactly what was always worth a human look — the ones that don't reconcile — and that queue is short.

Quotes from enquiries

Today

  1. enquiry arrives
  2. read it, work out what they want
  3. look up the rates
  4. build the quote by hand
  5. send, then remember to follow up

Automated

  1. arrives
  2. requirements extracted
  3. priced against your rate card
  4. draft quote ready for your check
  5. follow-up scheduled

You still approve every quote — that judgement stays with you. What disappears is the assembly work before the decision, which is usually why quotes go out the next day instead of the same hour.

The Monday morning report

Today

  1. export from three systems
  2. paste into a spreadsheet
  3. fix the formatting
  4. write the commentary
  5. email it around

Automated

  1. scheduled run
  2. pulls all three sources
  3. builds the pack
  4. flags what actually moved
  5. in your inbox before you are

The report stops being a job someone does and becomes something that just exists. The useful side effect: it still gets produced when that person is on leave.

Automation vs AI automation

Most agencies blur the two on purpose, because “AI” sells better. It's worth being clear about, because getting it wrong costs money in both directions. Good automation uses AI only where a rule cannot do the job.

Rules-based

Cron jobs, n8n, Zapier, plain code. If the input has a predictable shape and the logic can be written down as rules, this is the right answer: cheaper to build, cheaper to run, and it behaves identically every single time. Where deterministic is possible, deterministic beats clever.

AI-based

AI earns its cost in exactly two situations: the input is unstructured — free text, a PDF laid out differently by every supplier, an email that could be about anything — or the decision is genuinely fuzzy and a person makes the call today. Outside those two, AI is a more expensive, less predictable way to do what a rule would have done.

The mix

Most real projects are a mix: rules do the moving and the scheduling, AI handles the one step in the middle where the input is messy. Being honest about which is which is why my quotes come out smaller than people expect.

Tools, and when each one fits

n8n and similar — connecting systems that already have APIs, on a schedule or a trigger. Fast to build, and your team can see and adjust it later.

Direct code — when the logic is genuinely complicated, has to be fast, or a visual workflow tool would turn into a diagram nobody can read. TypeScript and Go.

Cloud AI APIs — Claude, OpenAI and similar for the steps that need real language understanding. The model is chosen per step; routine extraction doesn't need the biggest one.

Local models — when data can't leave the building, or when volume makes per-call pricing the wrong shape. Runs on hardware you own.

When not to hire me

If an n8n template and an afternoon of your own time solves it, that's what you should do — and I'll say so on the first call rather than quote for it. The same goes for anything a feature you already pay for can handle: a surprising amount of “we need automation” turns out to be a setting nobody switched on. I'd rather lose a small job and be the person you call for the real one.

How pricing works

Not a price list — the shape of one, so you know what you're agreeing to before the call.

Discovery

free

mapping the workflow properly, confirming the systems can actually be connected, and establishing what the current process really costs. Ends with a fixed quote for the build — or an honest recommendation not to.

Build

fixed price

fixed price per workflow, quoted per workflow not per hour, so the number doesn't move while the work happens. Two things drive it: how many systems have to talk to each other, and how messy the input is.

Running costs

usage-based

API usage plus hosting, billed by whoever provides them rather than by me. Estimated for your real volumes during discovery, visible in the logging afterwards — never a surprise.

Support

optional

available if you want it, never a condition of getting the work built. The code is yours and documented — you can also just take it in-house.

Where your data goes

Automation touches the sensitive parts by definition — invoices, customer records, financial data — so where that information travels deserves a straight answer, and “it's secure” isn't one. For most business data, sending it to a provider like Claude or OpenAI under a sensible plan is fine, and it's what I'd normally recommend. For data that genuinely can't leave — regulated, contractual, or just yours — local models keep everything inside your network. That's a real trade-off: local models are cheaper to run and nothing leaves, but they genuinely don't match the frontier ones. Extraction and classification? The gap barely matters. Work needing real judgement? It does. Which side your workflow falls on, I'll tell you in discovery — before anyone buys hardware.

Frequently asked questions

Is this automation or AI automation?

Whichever each step needs. Rules where the logic can be written down; AI for the one step where the input is messy or the call is fuzzy. Most projects are a mix — and saying which is which is why the quote comes out smaller than people expect.

Can't I just build it myself with n8n or Zapier?

If you can, you should. Clean inputs and simple logic are an afternoon's work, and I'll say so on the first call. Custom earns its keep with messy inputs, real logic, or several systems that have to agree.

How much time does it actually save?

I don't guess, and this page has no invented percentages. Discovery maps the process as it runs today and measures what it really costs; you decide from the real numbers.

Do we have to replace our current software?

No. Automation connects the tools you already own — their APIs, databases, file drops. Your systems stay your systems.

How does pricing work?

A fixed price per workflow, not per hour. Two things drive it: how many systems need to talk to each other, and how messy the input is. You know the number before anything starts.

What's eating your week?

Tell me about the process that keeps getting done by hand. I'll tell you whether it's worth automating, roughly what it takes, and whether there's an off-the-shelf tool to try first.

Book a free call →