Services
We automate four kinds of work.
Each engagement starts the same way: we sit with the people doing the work, count what they
actually do all day, and pick the highest-volume action with the smallest blast radius. That is
automation number one. Everything else follows it.
PRACTICE 01
Support automation with an audit ledger
Your support team spends most of its day on four or five
questions and a short list of account changes. We build agents that handle both halves — the
answer and the action — and we put a review gate in front of the action until it has
earned its way out.
What gets built
- Retrieval over your help centre, past resolved tickets, policy docs and product data,
with citations on every claim
- An action layer wired to the systems that matter: your helpdesk, billing, ERP, shipping,
subscription management
- The approval queue itself — proposed action, diff, reasoning, one-click approve or reject
- An append-only ledger: who or what did it, why, on whose approval, and how to reverse it
- Per-action-type accuracy dashboards driving the trust ladder
- Abstention: on low confidence, contract risk, or an angry customer, it stops and routes to a person
What it is not
- Not a chat widget bolted to your marketing site
- Not a deflection metric that just makes tickets harder to file
- Not a rip-and-replace of your helpdesk — it works inside Zendesk, Intercom,
Freshdesk, HubSpot or a homegrown queue
- Not a model with standing credentials to your production database
See the reference build →
PRACTICE 02
AI call answering, intake and scheduling
For any business where a missed call is a lost job:
trades, clinics, property management, professional services, logistics. The agent answers on the
first ring, every hour of every day, and its job is to book the appointment — not to read a
script and take a message.
What gets built
- A voice agent with sub-second interruption handling that does not sound like an IVR
- Real qualification: service type, urgency, address, access notes, insurance or account status
- Live writes into your scheduling system — ServiceTitan, Jobber, Housecall Pro,
Google/Microsoft calendars, or your own
- Overflow, after-hours or full-time answering — your choice, changeable per hour of the week
- Warm transfer to a human the moment the caller asks for one or the agent is unsure
- Every call transcribed, summarised, tagged and pushed to your CRM within seconds of hang-up
Where the money is
Most owner-operated service businesses miss between a quarter and a
half of inbound calls at peak, and the majority of those callers dial the next result instead of
leaving a message. Recovering even a third of them usually pays for the whole engagement inside
a quarter.
Run your own numbers →
See the reference build →
PRACTICE 03
Business process automation — n8n, Temporal, or plain code
The unglamorous middle of the company: a person reading a
PDF and typing its contents into a system. A spreadsheet that three departments email around. An
onboarding checklist that lives in someone's head. This is the cheapest, safest automation you will
ever buy, and it is where we usually start.
Document intake
Invoices, POs, certificates, claims, applications, statements.
Extraction with per-field confidence; anything under the bar goes to a human with the source
document highlighted, not to a guess.
Workflow orchestration
Idempotent, replayable, observable. Built in n8n when your team will
maintain it, in Temporal or code when the reliability bar is higher than a visual builder can hold.
System integration
The connector nobody sells: your 2011 accounting package, the vendor
portal with no API, the AS/400 report that arrives as a fixed-width text file at 4am.
See the reference build →
PRACTICE 04
Transforming the development process itself
Buying every engineer a code-completion licence is not an
AI strategy; it is a line item. The gains are in the loop around the code — specification,
review, testing, and understanding the system you already have.
- Spec-to-PR — an agent that turns a well-formed ticket into a reviewed draft
pull request with tests, running in CI on a branch, never merging itself
- Review agents tuned on your defects — we mine your last two years of incidents
and postmortems and teach the reviewer to look for those, not for style
- Test backfill — characterisation tests generated against legacy behaviour so
the scary module becomes changeable
- Codebase comprehension — grounded question-answering over the repo, the wiki,
the ADRs and the git history, so a new hire's first week is not archaeology
- Guardrails — agent identity, scoped credentials, mandatory human review,
full provenance on generated code, and a policy your CTO can show a customer
- Measurement — cycle time, review latency, escaped defect rate, before and after.
If it does not move, we say so
See the reference build →