Skip to content
For businesses in Edinburgh

Optimising your business with AI.

We help businesses in Edinburgh and across Scotland put AI to work on the jobs that eat the week: the enquiries, the chasing, the first drafts, the paperwork. Every step is recorded, and a person still makes every decision.

An AI consultancy in Edinburgh, working with businesses across Scotland. Not sure which job yet? A free 30-minute AI health check

Enquiry handling Agent, replayed
  1. Trigger
    Web form, out of hours “Self-employed, buying at about £340k in Leith. Need a mortgage adviser.”
  2. Read
    Qualifying criteria, version 3 Four rules, written by Sarah M. on 2 September. Three met: self-employed, loan size, area. Deposit not stated.
  3. Asked
    One question, in the client’s own channel “Thanks, roughly what deposit do you have? A percentage is fine.”
  4. Trigger
    Reply “About 15%, maybe a little more with help from family. What rate would I get?”
  5. Read
    The criteria, again Four of four met. Qualifies for a first conversation.
  6. Refused
    Not done The rate question. Not permitted to quote a rate or to advise, so it was not answered: it was flagged to Sarah M. with the reply attached.
  7. Read
    Sarah M.’s diary, read-only Tuesday 22nd, 10:00 free. Wednesday 23rd, 14:30 free.
  8. Did
    Two slots offered “Sarah can talk on Tuesday at 10:00 or Wednesday at 14:30, which suits? She will cover the rate question then.”
  9. Trigger
    Reply “Tuesday please.”
  10. Did
    Booked Tuesday 22 September, 10:00. Invite sent. Case 2026-0913 updated with the transcript.
  11. Handed
    Handed to Sarah M., by name The transcript, the four criteria with their answers, and the refused question at the top of the note.
  12. Confirmed
    Sarah M. read the record Moved the call to 10:30 and answered the rate question herself. Nothing else changed.
12 rows in this record The full record
An illustration, replayed from a fixture, not a live system: the names, times and case numbers are invented. The kinds of row are the real ones.

How an agent works

It does the work either side. A person makes the call.

Every agent we build runs the same loop: it picks the job up where it arrives, does the bounded part, writes each step to the record, and hands the decision to a named person.

Where the job arrives

  • Web formAn enquiry at 8pm
  • EmailA reply, with a photo
  • Missed callNobody free to answer
  • Client documentsThe fact-find, the file

The agent

One bounded job, rules written down

The recordSix fields for every step

What it does

  • Answers and qualifiesAgainst your written rules
  • Chases what is missingIn the client’s own channel
  • Books the call backInto the diary you use
  • Drafts the first versionEvery figure traced
A named personMakes the decision

How we build

Scoped narrowly. Recorded completely.

One job at a time, named before we quote, with a record of every step a person can read afterwards.

  • Either side of a decision

    The enquiry, the chase, the first draft, the pack. Never the decision itself: that stays with a named person.

    What we build
  • Every step on the record

    Six fields recorded for every action an agent takes, including what it refused to do. Exportable, and written for a person to read.

    Read a record
  • A flat price

    A fixed fee to build, then a flat monthly to run. Never per task, per minute or per row: a price that climbs with your volume charges you for the thing working.

    How we work
  • Every figure sourced

    Every figure we publish names its source and the day it was read, including the ones we got wrong.

    What we checked

Where agents are

Strong on bounded work. Not yet on the long, open-ended kind.

So we build agents for bounded jobs, and hand everything else to a person. Three figures this rests on, each with its source and the day it was read.

of enterprise apps will feature task-specific AI agents by the end of 2026
40%
Gartner, 26 August 2025
of OSWorld’s 369 bounded desktop tasks completed by the best model, above the 72% human baseline
86%
OSWorld-Verified leaderboard, 86.0%, read 24 September 2026
of long-horizon, real-world computer tasks the best agent completes on its own
31%
OSWorld 2.0 leaderboard, 108 tasks, read 24 September 2026

Every figure on this site is on one register: 485 sources and 31 corrections, the wrong ones included.

How we work

One job, then the next.

Every build runs through the same four phases. The six fields of the record are written into the statement of work before anything is built, so the record is a deliverable, not a feature somebody might add later.

  1. 01 Weeks 1 to 2

    Discover

    Workflows mapped, controls designed, payback ranked. Audit included.

  2. 02 Weeks 3 to 6

    Pilot

    First agent live on real work, with a person in the loop.

  3. 03 Weeks 7 to 12

    Scale

    The rest of your stack ships in sequence; the team is shown how it works.

  4. 04 Ongoing

    Managed

    Monitored, retrained and reported monthly.

Field notes

We checked. Here is what we found.

88 notes on what AI agents can and cannot do for a small firm, what the rules say, and what the figures in the headlines turn out to mean. Every one names its sources, and the day each was read.

All 88 notes

One job, named. Tell us which.

Not “automate the back office”. The specific thing that is eating the hours.

Or write to support@noxia.co.uk

  1. Step 1 You name the job A sentence is plenty: the job you would hand over first.
  2. Step 2 We reply with the questions What you use now, how often the job comes round, and who owns the output.
  3. Step 3 One job, scoped narrowly Named before we quote, with the record written into the statement of work. Or a plain no, if it is not worth building.