Our approach
An AI strategy that starts with your problems, not the technology
Every client walks in with their own mess, but the mess abstracts to a consistent set of problems. We name them first, who bleeds from each and where each comes from, before any technology appears on a slide.
Origins
Where each problem comes from
Every problem carries an origin. Some were there before AI. More than half are amplified or outright introduced by adopting it. We say so up front, because we govern the problems we bring.
The context problems
Context
what the organisation knows and how it works
The bottleneck people
Something only works because of one specific person, with no shared standard behind it
Everything routes through a few people
Critical knowledge is oversubscribed: a few people hold it, and every decision routes through them.
Technique in one head
Each engineer's technique lives in their head. It isn't shared, isn't repeatable, and leaves with them.
A different path every time
Similar work follows a different path every time: different gates, different artefacts, different quality.
Locked away
It exists, but the wrong people can't reach it
Logic locked in code
The business logic only exists inside the code, unreadable to anyone who doesn't program.
No shared place to ask
There is no one place where engineering, product and leadership read and ask. Context moves through meetings and chat archaeology.
Nobody can tell what's true
You found an answer, but can't verify it
Documents that drift
Documents drift from reality and compete. Nobody can say which version is the truth.
Confident wrong answers
The people who most need answers are the least able to verify them, and a fluent AI makes confident wrongness worse than silence.
What gets lost in the handoff
It existed once (an event, a task) and evaporated before it could move to the next person
Experience never becomes knowledge
The organisation pays for its experience twice: once to have it, once to lose it. Lessons from projects and incidents pile up unread, or get summarised into reports that change nothing, so the same mistake gets solved again next quarter.
Work loses value at every handover
Work loses value every time it changes hands: between colleagues, between shifts, and now between AI sessions and agents, at machine speed. The next person, session or agent re-derives what the last one already knew.
The governance problems
Governance
how safely, at what cost, and which rules actually hold
Rules that don't actually rule
A policy is words on a page. If nothing technical enforces it, and the shortcut is faster than the approved path, the words lose.
The shadow shortcut
If the approved path is slower than the shadow one, people take the shadow one. No policy fixes that. The approved path has to win the race.
Instructions are not controls
Most of what gets called “AI policy” is words the AI may or may not follow, and when it doesn't, nothing tells anyone. A document, an email or a ticket it reads can just as easily give it instructions.
No one to point to
When something goes wrong, there's no name attached to who did it, what they touched, or why.
Who accessed what?
Nobody can answer “who accessed what, when” or “why did it answer that”, a standing compliance gap that AI calls make acute, and that turns every audit and incident into archaeology.
Agents act without an identity
An AI assistant wired into company systems can retrieve what the person asking it could never open themselves, and when something goes wrong, the logs point at no one.
The foundation moves under you
What the system is built on shifts and leaks: the vendor underneath it, and the data crossing its boundary.
Built on shifting vendors
The whole capability rides a fast-moving vendor landscape. Models deprecate, prices shift, and lock-in threatens an investment meant to outlive any single tool, while a prompt or context edit ships to production with no version and no review.
Data out, poison in
Data crosses the boundary in both directions: customers' personal details reaching models on the way out, poisoned inputs on the way in, and centralised knowledge becomes one queryable door.
Nobody's counting the cost
Nobody set up a way to measure this at all, so there's no baseline to say whether AI is helping or hurting.
Invisible cost, unproven value
Cost is invisible and value is unmeasured. There is no baseline to prove whether AI helped at all.
Too fast to catch
The problem is already happening; it just doesn't surface until it's too late to catch cleanly.
Faster than review
AI produces work faster than humans can review it. The bottleneck moves to verification, and ungated AI action on a revenue path is the nightmare scenario.
Reliability decays along the chain
Ten steps at 95% reliability each succeed about six times in ten, and each step hands the next something plausible, so a failure surfaces at the end, far from its cause, or in front of a customer.
The adoption problems
Adoption
whether any of it sticks
Loses to the deadline
If it's not clearly faster right now, people go back to whatever already works.
Tried once, forgotten
Tools land, habits don't change. If the first experience isn't visibly better than working unassisted, people rationally go back.
Never urgent enough
Knowledge work always loses to delivery pressure. It's the first thing to slip when a deadline looms, and the system is worthless until someone puts what they know into it.
Bad incentives, top and bottom
At the bottom, the experts fear becoming replaceable. At the top, the people funding and governing it don't understand what they're deciding on. Different people, same failure: the ones who matter most have bad reasons to get it wrong.
Experts holding back
The experts whose knowledge the system needs most have the strongest rational reason to withhold it: capture reads as “making me replaceable.” The system has to make them more valuable, not interchangeable.
Decisions ride the vendor narrative
The people deciding what to fund and govern often can't describe the technology they're deciding about, so the strategy defaults to whoever tells the best story: a vendor deck, an analyst quadrant, a headline.
Depends on the wrong things
The system leans on something it shouldn't have to: a technical skill most people don't have, or one person or vendor who might not stick around.
Too technical a door
The way in is too technical for most of the organisation. The capability can be perfect. The interface alone caps adoption. For most people, the right door is chat.
Leaves when they leave
The capability stays with the vendor, or with one enthusiast, and dies when they leave. It must outlive any one person or vendor, including us.
Where AI projects fail
Most AI projects fail in the same two places.
Either the AI never learned how the company actually works, so its answers are generic and your team ends up fixing them by hand. Or nobody ended up using it. We fix those two first, because everything else depends on them.
The ladder
Start using AI in steps, with results from the first one.
No big launch at the end of a long project. Each step is useful on its own from the day it goes in, and we only start the next one when you actually need it.
Build your playbook
“Everyone keeps asking the same person.” “Training new staff takes months.” “When someone leaves, what they knew leaves with them.”
We capture how your best people do the job, right where the work happens, and measure where you stand today. The job gets done the same way whoever does it, new staff get up to speed faster, and later you can prove what improved. Everything after this builds on it.
Give your team one source of truth
“We have documents everywhere and nobody can find anything.” “It's quicker to just ask someone.” “The AI sounded sure, and it was wrong.”
Your playbook becomes a knowledge base anyone on the team can ask, with answers from your business, not from the internet. At this step it can answer and suggest, but it can't change anything on its own.
Keep your business data safe
“Is ChatGPT safe for our business data?” “Is anyone pasting customer details into it?” “What are we actually spending on AI?”
It can be, if it's set up properly. Your passwords stay out of the AI's reach, people only see what they're allowed to, private details are hidden, and every action is recorded. You also see what it costs, team by team. Nothing grows across the business until this is in place.
Scale it across the business
“We set it up and it never got any better.” “Keeping it up to date is a job in itself.” “We don't trust it enough to let it do more.”
It keeps itself up to date as your team works, and reaches more of the business as it proves itself. It's only trusted with more once its record shows it gets things right, and you can see that record.
Beyond the software
You can't install a habit.
The software goes in quickly. Getting people to use it is the part most AI projects skip, and it's why most of them quietly stop being used. So it runs alongside every step, from day one: we work with your team on real jobs, show them how to add to the playbook themselves, and talk openly with your most experienced people about what worries them.
Start here
Which of these sounds like your business?
Book a 30-minute call. We'll tell you honestly which problems you have and where to start.