AI that changes how the work gets done

We help lean, data-heavy organisations get the output of a much larger team without building an IT department. Training that sticks, automations that hold up, and system builds you own at the end.

Book an introductory call See how we work

Senior-led
Fixed-fee
Client-owned at handover
Security-led data handling

Positioning is not the point

We don’t sell tools. We change how the work gets done.

Most AI projects fail in a predictable way. Someone buys licences, runs an all-hands demo, and six months later the tools are unused and the same three people are still doing the same manual work at month-end.

The problem is rarely the model. It’s that nobody mapped where the time actually leaks, nobody decided what the AI is allowed to do without a human, and nobody built the habit before building the system.

Hunter McKenzie works the other way round. We start by finding what eats your week and what your data will actually support. We train your people on their own real work, so capability and discovery happen at the same time. We prove value on one narrow thing before anyone commits to a programme. Then we build module by module, each one earning the next.

Every engagement is senior-led. You work directly with the people doing it — two decades in second-line risk and information security at global banks, a regulated trading firm and critical-infrastructure operators, now building the systems rather than just the policies.

AI prepares. A human approves. Correctness lives in code, judgement lives in the model, sign-off lives with a person.

The method

How an engagement runs

The same six-step shape, scaled to the organisation. A solo founder might run it in three weeks. A trading firm runs it over four quarters.

1
Enablement
Function-specific training on the team’s own live work. Builds capability, secures the sponsor, and surfaces what’s actually worth building.
2
Discovery
Where the data really lives, what your licences permit, who approves what. Produces firm, fixed prices per module. We don’t quote a build blind.
3
Proof of value
One narrow, real win on real data. It de-risks the decision to go further, and it’s yours whether or not you do.
4
Phased build
Module by module, each proving itself before the next is funded. Payment tied to gates, not to the calendar.
5
Handover
Runbooks, training, and a clean transfer into accounts you own. Handover is a deliverable, not the last week of the project.
6
Stewardship
An optional retainer covering the evaluation and regression discipline that keeps an AI system working as your data and your business move.

Enablement leads, deliberately. Training people on their real work is the fastest honest discovery there is. You find out in an afternoon what a two-week requirements exercise would have guessed at — including which processes are genuinely broken and which people will actually adopt something.

The proof gates the money. No module is priced or started until the one before it is accepted. You commit to what you’ve already seen work.

What we do

Four service lines

Commission any one on its own, or run them in sequence as a programme.

AI Enablement & Capability Building

Your people are already using AI. Most of them are typing five words and hoping. That gap is free capacity sitting on the table.

Beyond Search: Using AI as a Practical Work Partner — an 80-minute live session on how to brief AI like a capable assistant, control the output, set up reusable context, pressure-test a decision, and apply plain safety rules to real work. No code.
Function-specific programmes — one-to-one sessions where each participant builds a working tool on their own real files. Delivered for crude marketing, accounts and administration, structured finance, and executive teams.
Executive and board briefings — what AI does, what it doesn’t, and what leaders are now accountable for.
Follow-up clinics two to three weeks after the build sessions, because adoption is the risk, not the technology.

What participants leave with

A six-part prompting brief · an output-control checklist · their own AI work profile · a continuity file for carrying context between sessions · one reusable work recipe · a verification checklist · and one specific thing to try on Monday.

AI Automation for Small & Mid-Sized Businesses

Small businesses don’t lose time to the hard parts. They lose it to the same five repeating jobs every week — the follow-up nobody sent, the invoice raised by hand in Word, the customer detail that lives in one person’s phone, the status update that takes four calls to assemble. Those jobs are automatable now. What they need is someone who’ll be honest about which ones are worth automating and which ones aren’t.

Discovery and pain audit — a structured 60 to 90 minute conversation. You get a written summary back confirming what we heard. That document becomes the brief, and it’s the first thing we deliver.
Data classification before design. Green, amber, red. Red-tier data sets the architecture before anything gets quoted, not after something gets built.
Automation build — client onboarding, invoicing with an approval step, lightweight CRM, morning briefings, quotes and follow-ups, document handling, reporting. On a deliberately boring stack.
Concierge pilots — where the value is genuinely unproven, we run the workflow by hand first and prove the output is worth having before building the pipeline. Sometimes that’s the whole engagement.
Handover and review — a walkthrough, the controls, an obvious off switch, and a scheduled checkpoint against the numbers we set at the start.

How we price it

Fixed fee, quoted against value rather than hours. Before we propose anything we quantify what the problem currently costs — volume, minutes, loaded hourly cost, error rate, revenue lost to slow follow-up — and show the twelve-month return in the proposal. If we can’t show a return that makes the decision obvious, the scope is wrong and we say so.

What we’ll talk you out of

Automations that create a dependency you can’t escape. Builds that need five hours a week of your time to train when you have two. And any pipeline where the honest answer is that we don’t yet know whether the output is worth having — those get a hand-run pilot instead.

AI Operating-Model Transformation

Some organisations are small in headcount and large in complexity. A twelve-person trading desk can carry the reporting, reconciliation and market-intelligence load of a firm five times its size — in spreadsheets, by hand, at month-end, dependent on whoever knows where the file is. That’s not a tooling problem. It’s an operating-model problem, and AI is the strongest lever available on it.

Paid discovery and design — data inventory, a per-vendor licence entitlement map, an extraction pilot on your real documents, the approval-workflow map, the benefits baseline, and a costed per-module delivery plan. The discovery deliverable is the firm quote.
Platform and access foundations — the unglamorous prerequisite. Connectivity, file layer, access control, endpoint baseline. Most transformation programmes fail here, quietly, and blame the AI.
Module builds, each accepted before the next is priced: finance and operations monitoring · executive intelligence briefing · deal and market intelligence · operations planning and optimisation · document extraction with provenance.
Branded document engines — structured content out of the model, rendered through version-controlled templates. Brand furniture lives in templates, never in prompts, so it can’t drift.
Governance the programme actually runs on — stage gates tied to payment, a RAID and decision log, change control, a named owner and hard date against every client dependency, and a compliance gate that blocks signature rather than sitting in an appendix.

Sector fluency

Our deepest domain depth is physical commodity trading, energy and structured finance — crude grades and differentials, cargo operations and laytime, prepayment and borrowing-base facilities, entitlement and lifting cycles, letter-of-credit document checks, settlement breaks. Generic AI advice is a commodity. Knowing what a demurrage claim looks like when the statement of facts disagrees with the fixture is not.

AI Assurance for What You’ve Already Built

Plenty of organisations already have AI in production — built in-house, bought from a vendor, or assembled by an enthusiastic team on a free tier. The question arriving at the audit committee isn’t whether to adopt AI. It’s whether anybody can evidence what the AI is doing with the data.

Independent review of AI systems already in use — data flows and residency, model and vendor terms, retention and training settings, human-approval boundaries, prompt and template hygiene, evaluation coverage, and the failure modes nobody has tested. Delivered as findings you can act on and evidence you can show.

For full framework work — NIST AI RMF, ISO/IEC 42001, EU AI Act scoping, shadow AI discovery, agentic protocol audits →

How we build

The rules we build to

These aren’t preferences. They’re what twenty years of watching systems fail under pressure leaves you with.

AI prepares, a human approves.

Nothing consequential publishes itself. No reconciliation goes out unreviewed, no order is submitted automatically, no recommendation reaches an executive unread. Arithmetic runs in code where it’s exact and repeatable. Judgement runs in the model. Sign-off stays with a named person.

The data boundary comes before the feature list.

Before asking what a system should do, we ask what data it touches and where that data must never go. Regulated and confidential data doesn’t reach a third-party automation log, a general-purpose database “just for reference”, or a public model tier. That answer sets the architecture, and it changes the price.

Boring, hireable stacks.

We build with components you could hire against or administer yourself, inside tenancies you already pay for. A system nobody at your company can operate has a bus factor of zero, however elegant it is.

You own it at the end.

Built in our environment, transferred to accounts in your name. Keys rotated. Nothing load-bearing depends on a personal account of ours. Runbooks written during the build.

Least privilege, minimum data.

We ask for the narrowest access that makes the build work, and if we’re handed more than that we’ll tell you to scope it down. The model gets the fields it needs, not the whole record. Credentials live in a secret store, never in a shared document.

Evaluation, not just deployment.

AI systems drift. A wrong figure in a lender statement or a hallucinated recommendation is a reputational event, not a ticket. Live systems get a small versioned test set and groundedness checks on anything client- or board-facing. That’s what the retainer funds — not maintenance.

What you keep

Reusable assets, not just outcomes

Every engagement leaves your team something they can use again without us.

Discovery questionnaire and pain audit: The same structured conversation, repeatable on the next process.
Data classification framework: Green, amber, red, with the handling rule and paperwork for each tier.
Value quantification model: How to price a process problem before deciding whether to solve it.
Prompt briefs, work profiles and continuity files: The templates from the enablement programme.
Report and brief templates: Brand-compliant, version-controlled, and yours.
Human-review screens and approval workflows: The gate, designed for your actual sign-off chain.
Runbooks and handover pack: How it runs, how to stop it, what to do when it breaks at 3am on a Sunday.
Evaluation harness: The small test set that tells you whether the system still works.

See the full toolkit library →

How to start

Engagement shapes

Live enablement session

An 80-minute practical session for a team already using AI without a method. Fixed fee, one cohort. The lowest-commitment way to see how we work.

Function-specific programme

Individual sessions where each person builds a tool on their own real work, plus an executive session and a follow-up clinic. Fixed fee. Often the front end of everything else.

Paid discovery

A bounded piece of work that ends in a firm price. Data reality, licence position, approval map, value baseline, and a costed plan. The reason we never quote a build blind.

Build engagement

Fixed fee per module, off the back of discovery. Payment tied to gates — spec signed, working demo on your real data, accepted handover. Then an optional monthly retainer.

Everything is fixed-fee. You know the number before work starts, and it doesn’t move because something took longer than we thought. Tool and model licence costs are separate and paid directly by you to the provider, so you keep control of the accounts.

Questions

The ones we actually get asked

Will this replace my staff?

No, and the engagements that try tend to fail. What it does is take the repeating work off people who should be doing something more valuable, and it means a small team can carry a workload that would otherwise need hiring against.

We’ve tried AI tools and nobody used them. Why is this different?

Because adoption is treated as the risk, not an afterthought. We train on your people’s real work rather than a demo, start with one narrow win that earns trust, pace new work to what the team can absorb, and build a check-in into the engagement. Where a client has told us plainly that they download tools and never use them, that goes in the recommendation as the honest risk — with a human follow-up designed to counter it.

Our data is sensitive. Can we do this at all?

Usually yes, and the sensitivity sets the architecture. Regulated or confidential data gets a zero-retention or self-hosted approach, region-locked or entirely within your own infrastructure. We classify before quoting, name every third party your data would pass through, and confirm in writing that the provider won’t train on your inputs. If your own client contracts don’t permit a third-party processor, we’ll find that out before touching anything.

What if the AI gets it wrong?

It will, sometimes. That’s why nothing consequential publishes itself, why arithmetic runs in code rather than in a model, and why live systems carry an evaluation harness. The design assumption is that the model is a capable, fallible colleague — briefed properly, checked before anything leaves the building.

Are we locked in to you?

No. The work transfers to accounts you own, with keys rotated and runbooks written. The retainer is optional and cancellable. Reversibility is a design constraint from the start: if this stopped tomorrow, could the business still run? If the answer is no, we’ve built the wrong thing.

How quickly do we see something?

An enablement session leaves people with a working tool the same day. A small business automation is typically live in days to a few weeks. A multi-module transformation programme runs quarters — but the first proof of value lands early, deliberately, because nobody should be asked to fund a programme on faith.

Start here

Start with the thing that’s costing you the most.

A 30-minute call, no charge. Tell us what ate your week. We’ll tell you plainly whether AI is the right lever on it, what it would take, and roughly what it’s worth — and if the answer is that you shouldn’t build it, you’ll get that too.

Book an introductory call Request an AI opportunity assessment