AI Automation
Put AI to work on the jobs nobody wants to do
Not chatbots for the sake of it. I use language models where they genuinely beat a rule: reading messy documents, sorting inbound requests, extracting data, drafting copy and making the small judgement calls that currently sit on somebody’s desk.
Seven years building automation, the last three of them with language models in production. I have also seen where AI is the wrong answer, and I will tell you when a plain rule would be cheaper and more reliable.
The problem
Traditional automation stops where judgement starts
Zapier and n8n are excellent at fixed steps. They fall apart the moment a human has to read something and decide.
What rules handle well
“When a form is submitted, create a CRM record and send an email.” Every input looks the same, every branch is known in advance, and the same action fires every time. This should never involve AI. It is faster, cheaper and more predictable as a plain workflow.
Where rules break
“Read this invoice and pull out the totals.” Every supplier formats differently. “Route this email to the right team.” The meaning matters, not keywords. “Is this lead worth a call?” That is a judgement made from context. Rules here become hundreds of brittle conditions that break on the next edge case.
The honest test: if you can write down every rule on one page and they never change, use a rule. If a person has to look at each item and think before acting, that is where a model earns its cost.
What I automate with AI
Six jobs that used to need a person reading every one
These are the builds I am asked for most often, and the ones where the return is easiest to measure.
Document processing
Invoices, contracts, applications, delivery notes and PDFs read automatically. Fields extracted, validated against your own rules, then written straight into your accounting system or CRM. Typical result: a person who spent two hours a day on data entry spends ten minutes reviewing exceptions.
Inbox triage and routing
Inbound email and form submissions read, categorised by intent, prioritised by urgency and routed to the right person or pipeline with a short summary attached. Sales enquiries stop sitting behind support tickets.
Content generation
Blog posts, product descriptions, proposals and follow-up copy drafted in your voice on a schedule, with a human approval step wherever you want one. This is the same engine behind my blog posting agent.
Data extraction and enrichment
Company and contact records enriched from public sources, deduplicated, normalised and scored. The difference between a CRM you can segment and one nobody trusts.
Decision automation
Lead scoring, eligibility checks, refund approvals and routing decisions driven by a model reading the full context rather than a brittle if-then chain. Every decision logged with its reasoning so you can audit it.
Call and meeting summaries
Calls transcribed, summarised and logged to the CRM with action items assigned to owners. Nothing depends on somebody remembering to write it up afterwards.
Choosing correctly
Rule, model, or both
Most builds I ship are a rule doing the heavy lifting with a model handling only the ambiguous cases. That is usually the cheapest design and the most reliable.
| The task | Best tool | Why |
|---|---|---|
| Form submitted, create CRM record | Rule | Fixed inputs, fixed output. A model adds cost and latency for nothing. |
| Classify inbound email by intent | Model | Meaning matters more than keywords; phrasing varies endlessly. |
| Extract totals from supplier invoices | Model + validation rule | The model reads any layout; a rule checks the maths before it is trusted. |
| Send reminder 3 days before renewal | Rule | Pure scheduling. Nothing to interpret. |
| Score a lead on fit and urgency | Model | Judgement from context that a scoring matrix cannot capture. |
| Move a deal stage when a contract is signed | Rule | A definite event with one correct response. |
How I build it
Built with guardrails, not vibes
An AI automation that is right 80% of the time and silently wrong the rest is worse than no automation. Every build gets the same treatment.
Find the judgement
We walk the process and mark every point a person makes a small repeatable decision. That is where a model pays off, and nowhere else.
Pick the approach
A model, a rule, or both. I will tell you honestly when the cheaper option is good enough, including when the answer is no automation at all.
Add guardrails
Validation on every output, confidence thresholds, and a human review step wherever a wrong answer would be expensive. Low-confidence cases get escalated, not guessed.
Test on your real data
Run against your actual historic records and measure accuracy against known-correct answers before anything touches live systems.
Deploy and monitor
Live with full logging, so you can see what it decided and why. Alerts on failure paths and on accuracy drift.
What it costs to run
The part most people forget to ask about
Build cost
Quoted per project after the discovery call, once the scope is genuinely understood. A single document-processing pipeline is typically a few days of work; a multi-step agent with CRM integration is longer. You get a fixed figure before anything starts.
Running cost
Model usage is billed by volume and is yours directly, not marked up through me. I size this during scoping so you know the monthly figure before you commit, and I design prompts to keep it low.
Ongoing support
30 days included on every build. After that, a monthly retainer covers monitoring, accuracy tuning and changes as your process evolves. Optional, never automatic.
What I build AI automation with
Common questions
AI automation, answered honestly
Related
Often built alongside this
AI agents and chatbots
Conversational agents that qualify leads, book calls and answer support questions on your site, wired into your CRM.
Learn more →CRM automation
Lead routing, follow-up sequences and deal-stage triggers running inside your CRM around the clock.
Learn more →Automation platforms
The n8n, Zapier and Make.com layer that connects your tools and carries the data between them.
Learn more →Ready when you are
Have a job that needs judgement?
Bring me the task that is too messy for normal automation. On a free 30-minute call I will tell you honestly whether AI is the right tool, what accuracy to expect, and what it would cost to run.
Book a free 30-min call →