Service 02  ·  AI automation

AI that earns its place.

AI for the operational jobs where someone is making the same judgment call over and over. Classifying inbound, parsing PDFs, drafting standard replies, routing decisions. Built honestly: where it helps, where it doesn’t, where a human still signs off. No hype, no theatre.

01  —  Who this is for

When the rules run out.
And someone keeps making the call.

Two kinds of teams come here. One: already running automation and hit the wall where rules can’t go further — classifying inbound, parsing PDFs, drafting the same five replies. Two: looking at AI for the first time, or coming back after an earlier attempt that didn’t land. Either way the answer is the same: AI where it earns its place, with humans in the loop, and an honest line about what it can and can’t do.
02  —  Common symptomsIf one or more sound familiar
01A person classifies every inbound enquiry before anything else can happen.
02Invoices, POs, or supplier emails arrive as PDFs and someone re-keys them.
03Support, sales, or ops drafts the same handful of replies every day.
04Routing quality varies depending on who happens to be on shift.
05You’ve watched the AI hype cycle and aren’t sure what’s actually worth doing.
06A previous AI experiment cost money and didn’t ship anything useful.
03  —  What we build

Concrete deliverables.
No retainers in disguise.

  • 01AI classification and routing inside deterministic pipelinesThe judgment node that sits inside the rest of the workflow. The spine it plugs into is the workflow automation we build alongside it.
  • 02Document and email parsingPDFs, scans, attachments. Unstructured input turned into structured records the rest of your stack can read and act on.
  • 03Drafting and triage flows with human-in-the-loopAI proposes, a human approves. Used where the cost of a wrong send is higher than the cost of a moment of review.
  • 04Evals and structured outputsEvery AI node ships with a test suite that scores its outputs against examples your team agrees on. Performance is measurable, not vibes.
  • 05Guardrails and monitoringLogged decisions, sampled review, alerts when behaviour drifts. AI you can debug, not AI you have to trust.
  • 06Documented handoverPrompts, evals, thresholds, monitoring dashboards. Your team owns and tunes the AI without us in the loop.
04  —  How an engagement runs
Map, build, hand over.
Your team owns the prompts from there.
Four phases. Where AI genuinely helps, where it doesn't, and the human checkpoints in between — all written down before anything ships. Evals are part of the handover, not an afterthought.
Phase
00
01
02
03
04
05
06
07
08
09
10
11
12
13+
01
Map
02
Build
03
Handover
04
Filed
◆  Your team runs it. We step out.
After handover →
Retainer
optional · continuous
Run
→
Iterate
→
Design
→
Build
01 · Map
You
Tell us where someone keeps making the same call. One workshop, real examples in the room.
Us
Work out where AI helps, where rules are safer, and what the cost of a wrong call looks like.
Filed
Mapped decisions · AI-fit scoring · phased options
Honest about where AI fits.
02 · Build
You
Provide examples. Approve the eval cases that define good output.
Us
Build the AI nodes, the evals, the fallbacks, and the monitoring. Wire it into your stack.
Filed
AI nodes live · evals passing · monitoring in place
Measurable, not magical.
03 · Handover
You
Name the owner. Walk it with us. Run the evals yourself.
Us
Document the prompts, the evals, the thresholds. Train the owner.
Filed
Prompt library · eval suite · monitoring dashboard
Yours to tune from here.
04 · Filed
◆ MILESTONE
You
Run the evals. Adjust the prompts. Raise the bar when ready.
Us
Grace period covered. Anything after is booked.
Filed
File closed · no ongoing dependency
The point.
↻  Retainer · optional
Iterate on what’s
live, by the month.
One cycle · run, iterate, design, build
01 · Run
Keep it operating.
02 · Iterate
Act on what usage shows.
03 · Design
Scope the next thing worth building.
04 · Build
Ship it. Measure it. Back to run.
Typical follow-on
Usage optimisation, workflow tweaks, feature additions. Shape varies by what was built and what the team needs once it’s in their hands.
Cancel any month · 30 days' notice
Engagement is always the foundation · retainer never required
05  —  Relevant case study
Recruitment agency · File 01

The pipeline AI plugs into.

The recruitment-agency engagement built the deterministic spine — CRM, automation, data hub, custom apps — that AI nodes sit inside. Current AI work specifically is under NDA and unpublished; this is the surrounding workflow it operates within.

7modules
From public surface to infrastructure
Read the case file →
06  —  Smallest useful first move

A one-week AI diagnostic.
Where it helps. Where it doesn’t.

Wherever you are with AI — already shipping, exploring for the first time, or burned by an earlier attempt — the same one-week engagement. We look at three to five real processes in your business and write up where AI genuinely helps, where rules are safer, and what the cost of a wrong call looks like for each. One week, on-site or remote. You get a written plan with phased scope, indicative costs, and an honest assessment per process. If you stop there, you stop there.
07  —  Start a conversation
If this is the kind of work you’re after, here’s how to begin.
Thirty minutes. What’s in the way, what you’ve already tried, whether there’s a useful first move. No deck, no proposal.
Discuss your setupRead a case file