AI
AI that pays for itself — found, built, and proven
We find where AI genuinely pays in your business, rank it with costs and payback, and build only the two or three you choose.
- Home
- AI
AI
Most businesses are being sold AI. Very few are being shown the arithmetic.
We spend two to three weeks finding where AI genuinely pays in your business, hand you a ranked list with costs and payback, and then build only the two or three you choose. Anything without payback inside twelve months gets parked — including things we would enjoy building.
Problems we solve
What usually brings
people to us.
01
Everyone is proposing AI, nobody is costing it
Three suppliers, three visions, no numbers anyone can check.
02
A pilot ran and quietly died
Usually because there was no evaluation set, so nobody could tell whether it was working.
03
The data is not ready and nobody said so
This is where most AI projects fail, and it is knowable in week one.
04
Compliance stopped it
Because nobody could write down which model saw what, and where it ran.
05
It works but nobody owns it
The prompts live in a supplier's account and the person who built it has moved on.
06
The team is worried
Fair enough. Agents that draft for a human land very differently from agents that act alone.
What we do
The work,
in plain terms.
| Service | What it is | Typically |
|---|---|---|
| AI Opportunity Audit | Two to three weeks. Shadow the work, score every recurring task, audit the data, deliver 5–10 ranked use cases with costs and payback. | Fixed price |
| Agent Pilot | One agent, four to six weeks, measured against one metric agreed before we start. | Fixed price per agent |
| Agent Operations | Running, monitoring and improving live agents: evaluation runs, model changes, cost per task, review queue. | Monthly, tiered |
| Fractional Head of AI | One or two days a week inside your team, for businesses building the capability in-house. | Monthly |
| Handover and training | We train your team to maintain the simpler agents themselves. This is a service, not a courtesy. | Included or scoped |
The method
Find first.
Build second.
Two to three weeks, fixed price, ending in a document you could hand to a different supplier. That is deliberate.
-
Shadow the work
For a week we watch the business as it actually runs. We export tickets, emails, CRM notes and the spreadsheets nobody admits to. Remote, and quiet.
-
Score every recurring task
Volume, minutes per instance, how rule-based it is, and what a mistake costs. Four numbers, applied consistently, so the ranking is not a matter of taste.
-
Audit the data reality
What exists, what is inconsistent, what has to be cleaned before anything can be built on it. This is where most AI projects quietly die.
-
Rank and cost the use cases
Five to ten candidates, each with estimated hours saved or revenue gained, a build cost, and a risk rating. In writing, with the workings shown.
-
You pick two or three pilots
Anything without payback inside twelve months gets parked, including things we would enjoy building. Pilots run four to six weeks at a fixed price against an agreed metric.
Sample output — an audit runs these numbers on your business
Rank
Candidate use case
What it moves
Est. year 1
Confidence
01
Inbound lead qualifierReads, enriches, scores, drafts a reply and books the call
Response time, 6h → 8min
412 hrs saved
02
Invoice and PO processingExtracts, validates, posts to the accounting system
Finance admin hours
310 hrs saved
03
First-line support assistantGrounded in your docs and order data, escalates with a summary
Tickets deflected
46% of volume
04
Proposal drafterBuilds a first draft from the rate card and past proposals
Time to proposal
9 hrs / week
05
Monthly reporting agentReads analytics and writes the narrative for a human to edit
Reporting effort
22 hrs / month
06
Content localisation EN→ARAdapts and stages copy for approval, never publishes alone
Translation spend
£1.4k / month
Inbound lead qualifier · reviewed by a human before send
Inbound — 21:14 GST
“Hi, we are a 40-person facilities firm in Dubai. Our site is old and we get maybe 6 enquiries a month. Do you do this kind of work?”
Agent — 21:14 GST · 38 seconds
Enriched the company from its domain, matched it to the professional-services segment, scored it 82/100 against the fit rules, drafted a reply that answers the question and offers two call slots, and put it in the human review queue.
Account lead — 07:02 BST
Approved with one edit at 07:02, ten hours after the enquiry arrived — with the draft already written and waiting.
Autonomy is earned
Draft first. Act later. Always logged.
Every agent ships with an evaluation set of 50–200 real examples drawn from your own records, prohibited actions, spend limits, escalation triggers, full logging and a human-review queue. Autonomy expands as the evaluation scores earn it — and it can be pulled back the same way.
Browse the agent libraryHow agents are built and run
The engineering,
stated plainly.
| Decision | Our position |
|---|---|
| Model-agnostic | Claude, OpenAI and Gemini via API, plus open-weight models on Azure UAE or AWS where data has to stay in region. |
| Orchestration that fits | n8n or Make for lightweight flows. Code-based frameworks such as LangGraph or a vendor agent SDK when the agent is genuinely complex. |
| Integrations | HubSpot, Salesforce, Shopify, Zendesk, Xero, Microsoft 365 and Google Workspace. |
| Retrieval by default | Agents read your content at run time. We fine-tune only when the volume and consistency clearly justify it. |
| Evaluation before autonomy | Every agent ships with 50–200 real examples as an evaluation set. Autonomy expands only as the scores earn it. |
| Guardrails and logging | Prohibited actions, spend limits, escalation triggers, full logging and a human-review queue — from the first day it runs. |
How it is delivered
Owned by you,
from the first week.
- You own the prompts, the code, the evaluation sets and the integrations
- The contract names the models, the regions and the human reviewer
- In-region hosting available on Azure UAE and AWS UAE/Bahrain
- Pilots are fixed price against one agreed metric — not time and materials
- Every output is logged and traceable to its inputs
- A handover-and-training track for teams who want to maintain agents themselves
- If the numbers do not work, we say so and stop
Packages
Where most engagements
start.
| Package | Model | Indicative price |
|---|---|---|
| Launch Site + AssistantA marketing site plus one grounded site assistant, live in weeks | Fixed scope | from £4,500 |
| AI Opportunity AuditTwo to three weeks, ending in 5–10 ranked and costed use cases | Fixed price | £6,500 |
| Agent PilotOne agent, four to six weeks, against one agreed metric | Fixed price per agent | from £9,000 |
| Agent OperationsRunning, monitoring and improving live agents | Monthly, tiered | from £1,500 / mo |
| Fractional CMO / Head of AIOne or two days a week, inside your team | Monthly | from £4,000 / mo |
These are starting points. Every fixed-scope quote states exactly what is in and what is out before you sign, and the number in the quote is the number on the invoice. See all pricing.
In practice
How this runs
in a real engagement.
Professional services
Intake that stops consuming partner time, and expertise that is findable.
In most advisory and legal firms the bottleneck is not demand. It is that every new matter waits on a senior person before anyone else can touch it.
What it is judged on
Time from enquiry to matter summary
Median, measured in working hours.
Summary acceptance rate
What proportion is accepted with light edits, tracked against the evaluation set.
Senior hours returned
The number the engagement is actually judged on.
Questions
Asked before the first call.
Do we have to start with the audit?
If you already know which process you want to automate and can produce a year of examples for it, you can go straight to a pilot. Most businesses cannot, which is what the audit is for.
Which models do you use?
Whichever fits the task, the budget and the data-residency requirement — Claude, OpenAI and Gemini via API, or open-weight models on Azure UAE and AWS where data must stay in region. The contract names them.
What if the agent gets something wrong?
Every agent ships with prohibited actions, spend limits, escalation triggers, full logging and a human-review queue. Agents start by drafting for a person and earn autonomy by clearing an agreed score on the evaluation set.
Can our team maintain the agents afterwards?
For the simpler ones, yes — there is a handover-and-training track for exactly that. Complex multi-step agents usually stay on Agent Operations, and we will tell you which is which.
What does a pilot cost and how long does it take?
From £9,000 per agent, four to six weeks, fixed price, measured against one metric agreed before we start.
Next step
Tell us what you are trying to move.
Thirty minutes, no deck. We will tell you what we would do first, what it costs, and whether we are the right people for it.