We become your IT department and own the day-to-day.
›Managed IT Your whole environment, owned ›Managed Devices Endpoints, patched & tracked ›Microsoft 365 The workplace, managed ›Cloud & Azure Run properly, costs watched ›Connectivity Internet & networks that work ›Business VoIP Phones on any device ›Mobile Plans Business mobile, one billSecurity operations you can verify, and show.
›24-Hour CyberSOC Security ops that don't clock off ›Identity & Access MFA enforced, takeover stopped ›Email Security Phishing & fraud, stopped ›Endpoint Protection EDR on every device ›Backup & Recovery Provably restorable ›Security Bundles Per-user, no hidden feesHonest strategy, budgets you can plan around.
›IT Advisory & vCIO Honest, predictable plans ›Compliance & Governance Audit-ready evidence ›Projects & Consulting Build what you can't buy ›Procurement Right kit, fair price, visible ›Licensing Right-sized, never wasted ›IT Asset Finance Spread the cost, monthly ›Data Centres & Hosting Sovereign, servers in minutesAI, automation and software, where they genuinely help.
›AI at Belton All four phases, and the prices ›AI Readiness What is safe to turn on, from $1,499 ›AI Training Copilot, Claude & ChatGPT, from $299 ›Monthly AI Plans Managed & improving, from $1,495/mo ›AI Projects Bootcamp & custom builds, from $15k ›Build with AI We train your team to build it ›Business Automation Hand work to the platform ›Software Development Custom builds, done properly ›Client Portal AI-first insights & budgets ›White-Label / Partners Our bench, your brandOrganisations that run on trust, time and confidential data.
›Accounting Ledgers & client data ›Legal Privilege & matter files ›Government & Councils OIA, LGOIMA & records ›Business Advisory Plans & reporting ›Mortgage & Finance Loan files & AML/CFT ›Insurance Claims & client records ›Insurance Brokers Cover & premiumsSite, studio and floor, where uptime is the product.
›Construction Sites & project files ›Architecture Drawings & large models ›Quantity Surveyors Cost data & big files ›Manufacturing Floor & ERP uptime ›Healthcare Patient data & privacyFast-moving operations that can't carry downtime.
›Logistics Fleet, orders & tracking ›Retail & Wholesale POS & storefronts ›Hospitality POS, EFTPOS & guest WiFi ›Technology Cloud-native teams ›Distribution Inventory & systems ›Franchise Every location, one platform ›Global Operations Offices across time zones ›General SMB Don't fit a box? Start herePlain-English playbooks for the controls that matter.
›Essential Eight ASD's eight, explained ›DIY Cybersecurity Lock down the basics ›Cyber Standards Guide ISO, SMB1001 & more ›M365 Selection Guide Pick the right licencesSelf-serve checks that map where you stand.
›Security Assessment Score your posture ›Network Assessment Find the weak spots ›M365 Security Checklist Find the gaps ›Cyber Insurance Readiness Be claim-readyGet more from the tools you already pay for.
›SME Tech Outlook 2026 Where NZ tech heads next ›Copilot Learning Hands-on with Copilot ›FAQ Straight answers ›CEO's Blog Plain-English viewsA senior team with the skills and scale for any project.
›About Belton Who we are, plainly ›How We Work Our delivery process ›Who We Work With The fit we look for ›Leadership Jason & Amy Agnew + teamThe evidence behind the claims, and ways to build with us.
›Case Studies Real outcomes, named ›Accreditations The proof behind the practice ›Industries Sector-specialised ›White-label / Partners Our bench, your brandReal people, fast, no ticket-queue runaround.
›Contact Us Talk to a human, today ›Book a Session A 90-minute discovery ›Find Us Newmarket, Auckland ›Remote Support Get help nowYour own team building and using AI in your business, on a framework that keeps you out of trouble, with the licensing right, the running cost understood, and somebody alongside them while they learn. We are not trying to become the department you cannot fire. The point of this programme is that in a year your people are building things we never touched, and the only reason we are still in the room is that you find the catch-ups useful.
Most AI engagements sell you an outcome and keep the knowledge. That works right up until the model changes, the person who commissioned it leaves, or somebody asks a question nobody internally can answer. This one is built the other way round. The deliverable is your people, and the things they build are the evidence.
These are not modules in a course. They are the six things that decide whether AI in a business ends up as a capability or as an unexplained line on the card statement. Most clients are already part-way through two or three of them without having named them, which is usually where the trouble is.
Rules for AI that a person can actually follow and an auditor can actually read.
Read this one ›Stop paying for seats nobody opens, and stop good people working on personal accounts.
Read this one ›The bill follows the work, and your team can see it move before the invoice arrives.
Read this one ›An evidence-based answer to whether any of this should run on your own hardware.
Read this one ›Your team can stand one up, and knows the times not to.
Read this one ›A live backlog, a cheap way to try things, and permission to kill the ones that do not work.
Read this one ›A framework is not a document nobody reads. It is the small number of decisions that stop the predictable accidents: confidential material pasted into a personal account, an agent given the permissions of whoever happened to build it, a tool in daily use that nobody ever approved. Written down once, these take an afternoon. Discovered after the fact, they take a lawyer.
The most expensive AI licence is the one bought for everybody because deciding was harder. The second most expensive is the one your best people avoid, so they quietly keep using a personal account with none of the protections you are paying for. Getting this right is mostly a matter of being specific about who does what.
Here is the field as it stands. Vendor pricing moves without notice, and several of these are billed in US dollars, so we link to each vendor's own pricing page rather than repeat a figure here that would drift. We go through the current numbers with you when we look at who actually needs which seat.
Agents that live where your people already are, inside Teams and Microsoft 365, working on your SharePoint and Graph data with the identity and permissions you already run.
You are billed per credit, so a popular agent gets expensive quietly rather than loudly. What it can reach is bounded by what the connectors expose, and that boundary is where most projects meet their surprise.
Helping individual people inside the apps they already have open. It is the per-seat assistant layer across Word, Excel, Outlook and Teams.
It is an assistant, not an agent platform. It waits to be asked. If you want something that runs on its own and hands work back finished, this is the wrong line item.
Work that is thinking rather than formatting: long documents read properly, analysis that holds up, and processes built once and reused. It is not tied to one software suite, which is why it handles the jobs that do not fit inside an app.
It is not natively wired into your tenancy the way a Microsoft product is, so connecting it to your systems is real integration work rather than a checkbox. Billed in USD, so your actual cost moves with the exchange rate.
An agent you own outright, that runs on your schedule against your systems, with no seat count and no dependence on what a product happens to expose.
Usage priced, so the bill follows the work rather than the headcount. That is honest but it needs a budget and a ceiling set deliberately at build time.
Broad general capability with the widest staff familiarity, which lowers the training cost of getting a team started.
Enterprise is quote only and carries a large seat minimum, so the entry point is a procurement exercise rather than a subscription.
The plumbing. Moving work between systems on a trigger, with the steps visible on a canvas that a non-developer can actually follow and audit.
It orchestrates, it does not think. It is the frame you hang a model on, and the cloud tiers are priced by execution volume, which climbs faster than people expect.
The genuine constraint cases: data that is not permitted to leave your control, a regulator who wants a straight answer about residency, or volume large enough that per-token pricing stops making sense.
The hardware is the small number. Published analyses put the maintenance at 10 to 20 engineering hours a month, and below roughly 2 million tokens a day an API is usually cheaper than the box, before anyone is paid to look after it.
We are a Microsoft Solutions Partner and a Claude partner, and we will still tell you when the honest answer is fewer seats. Check the vendor's own page before you commit to a number, and see licensing for how we handle the rest of your subscriptions.
Usage billing is honest but it is unfamiliar, and a team that has never been taught to think about it will build something that works beautifully and costs four times what it needed to. Almost all of that is avoidable, and none of it requires switching vendor. This is the part of the programme that pays for itself most visibly, because the before and after are both on an invoice.
| Model | What it is for |
|---|---|
| Claude Opus 5 | The hard reasoning. Worth it when being right matters more than the unit cost. |
| Claude Sonnet 5 | The everyday workhorse. Most production agents should start here. |
| Claude Haiku 4.5 | High volume, low judgement. Classifying, routing, extracting. |
A token is roughly three quarters of a word, counted both going in and coming out. The practical read: an agent that reads a hundred long documents a day costs money you will notice, and one that answers forty short questions a day usually does not. The number that matters is not the rate, it is how much you send it, and that is a design decision made at build time rather than a surprise found at invoicing. Current rates are on each vendor's own pricing page.
This question comes up in every engagement, usually second, and it deserves a real answer rather than a reflex in either direction. Sometimes running a model on your own infrastructure is genuinely right. More often the thing driving the question is a control requirement that a business tier already satisfies, and the honest service is to find that out before anyone buys a GPU. We explore it with your team rather than deciding it for them, because the ones who have seen the working do not have to revisit it every six months.
If the answer does turn out to be your own infrastructure, we run sovereign hosting in New Zealand, so exploring it does not mean handing the question to somebody else.
An agent is not a chatbot with ambition. It is a piece of software that does a job you would otherwise pay a person to do, hands the work back finished, and waits to be told whether it got it right. Teaching a team to build one is mostly teaching them to choose the job well and to leave a person in the loop on purpose, which are judgement skills rather than technical ones.
Waits to be asked, answers, forgets. Copilot in your Outlook is an assistant. It is genuinely useful and it is priced per seat, because the value scales with how many people are sitting there asking.
Fires on a trigger and does exactly the same thing every time. Enormously valuable, far cheaper, and the right answer more often than anyone selling agents will tell you. If the rules never change, you want an automation.
Given an outcome rather than a script, it works out the steps, uses the systems it has been given, and comes back with the work done and its reasoning shown. You want one when the job needs judgement in the middle, and the judgement is the expensive part.
Not the most impressive job, the most repetitive one with a clear right answer. If nobody can describe what a good outcome looks like, an agent cannot either, and that is a finding worth having before you spend anything.
You, in an afternoonAn agent inherits the permissions of whoever runs it, which means an over-shared SharePoint becomes an over-sharing agent on day one. The permissions problems that trip up Copilot trip up everything else pointed at the same files.
Readiness assessmentOne job, one data source, one person using it. The point is to find out where the real difficulty is, which is almost never where it looked from the outside.
Build session or plan monthThe agent proposes, a person approves, and the system records who approved what. This is the difference between an agent you can defend in an audit and one you quietly turn off after an incident.
Designed in, not bolted onSomething built and then orphaned is worse than nothing, because people start relying on it. Name the owner before you build, and be honest if the answer is that nobody has the time.
Your team, or a planThe businesses getting real value are not the ones with the best strategy document. They are the ones running four small experiments a quarter and cheerfully abandoning three of them. That only works if trying something is cheap, if there is a queue of things worth trying, and if stopping is a respectable outcome rather than an admission.
These are not hypotheticals. Each one exists because the manual version of it is already happening somewhere, badly, in a spreadsheet. They are here as a starting queue for your team rather than a menu of things to order from us, and size is relative rather than a quoted duration.
Reads every supplier invoice as it arrives, proposes the account code and cost centre, and puts it in front of a person to approve.
Drafts the follow-up on every overdue invoice, in your tone, ranked by what is worth chasing first rather than by what is oldest.
Takes the transactions nobody can place and searches two years of history for the ones that looked like them, with the reasoning shown.
Compares what each job was quoted at against what it actually cost to deliver, and tells you which kinds of work quietly lose money.
Turns a photographed, scanned or handwritten job sheet into a structured record, and asks a human about anything it could not read.
Reads supplier price lists as they land, flags exactly what changed, and works out what the change does to your margin before you find out at invoicing.
Reads carrier and supplier notifications, and separates the ones a customer genuinely needs to hear about from the noise.
Drafts a written procedure from a recording of someone actually doing the task, then re-checks it each quarter and flags the steps that no longer match reality.
Reads every inbound enquiry, works out who the company is, and drafts both the reply and the suggested next step before anyone has opened it.
Assembles a first draft from your own previous proposals and the notes from the call, so the blank page stops being the reason quotes go out late.
Finds the duplicates, the records owned by someone who left, and the opportunities that have not moved in a year, and proposes the fix for each.
Reads the tender, maps every requirement to the evidence you already hold, and tells you plainly which ones you cannot currently answer.
Turns a start date into the whole checklist, raises the requests across every system, and chases the ones nobody has done by Friday.
Answers staff questions from your actual policies, quotes the clause it used, and says so plainly when the policy does not cover the question.
Flags the timesheets that do not look like the person’s normal pattern, before payroll rather than after it.
Turns everyone’s scattered interview notes into the same scorecard, so candidates are actually compared on the same things.
Takes a security alert, gathers the context from every other system that saw the same device or account, and writes the short summary of what it actually is.
Lists who can currently reach what, and flags the access that does not match the job the person actually does.
Compares what you are billed for against what is actually assigned, and what is assigned against what anyone has used this quarter.
Compares what you are being billed for against what is actually documented, and surfaces the things running in your business that nobody has written down.
Reads the contracts you already signed and pulls the dates, notice periods, liability caps and auto-renewals into one register you can actually search.
Tells you a notice period is closing while you still have time to do something about it, rather than the week it renews.
Works out where one person’s data actually lives across your systems, so a privacy request is a task rather than a fortnight.
Assembles a defensible chronology from logs, tickets and email when something has gone wrong and the question is what happened when.
Nothing in that group yet. Choose another, or tell us the job you actually have.
Training that ends on the last day of training does not survive contact with a busy month. What makes the difference is a standing session in the calendar that somebody has to bring work to. It is the same reason a gym membership is not the same thing as a training partner.
A working session with the people actually building. What is running, what it cost, what broke, what to try next, and what to stop. It is a review of real output, not a status update, and it is where most of the learning lands.
Some decisions do not wait for the next catch-up. A vendor changes its terms, a tender asks where your data is processed, someone wants to point an agent at the finance folder. You get somebody to think it through with, before rather than after.
Licensing against actual usage, spend against what it returned, and whether the framework still matches how people are really working. This is the session where things get switched off, which is a good outcome and a rare one.
Most AI quotes fail because they price the engagement and go quiet about the running cost. There are three separate layers here and they behave differently. We itemise all three, including the ones that are not ours to charge for, and teaching your team to read them is itself part of workstreams two and three.
Paid to Microsoft, Anthropic, OpenAI or whoever else, mostly monthly and mostly per seat. Predictable, and the easiest layer to over-buy. Microsoft bills this in New Zealand dollars; most of the others bill in US dollars, so that part of the bill moves with the exchange rate. We will tell you when you are paying for seats nobody is using.
The one that varies, because it follows how hard the work is. This is the layer people are surprised by, so it gets a ceiling set at build time and a monthly figure you can see rather than infer. Keeping it down is a skill, and it is one your team leaves with.
Ours, and the only layer on this page we set. Fixed in writing before anything begins, with the cadence and the number of people named in it, so there is nothing to discover later.
Priced per business, because it depends on how many people take part, how often you want us there, and how much of the building you want to do yourselves. Ask and we will give you a number, not a range. What we will not do is print a headline figure that has to be padded to be safe, then quietly requote once we understand the job. Tell us the shape of your team and you will get a fixed number in writing.
By submitting this form you agree to our Privacy Policy.
If you already know the job and just want a straight answer on whether it is worth building, this is the shorter road. Two fields and the idea. We will come back with a yes, a no, or the cheaper thing you should do instead. No obligation and no sequence of follow-up emails.
Priced per business, because it depends on how many people take part, how often you want us there, and how much of the building you want to do yourselves. Ask and we will give you a number, not a range. What you will get back is a fixed number in writing, with the cadence and the number of people named in it, before anything starts.
Sometimes you should just pay us to build it, and we will say so. A single well-defined system that has to work and rarely changes is a scoped build. The programme is the better answer when the opportunity is spread across the business rather than concentrated in one place, because the people who know where the wasted hours are will always spot more of them than we will. It also removes the dependency, which is worth something the first time a vendor changes its terms.
The people who do the repetitive work, not only the technical ones. The best builds we have seen came from a finance manager and an operations coordinator, because they knew which fifty minutes a day were being wasted. You want somebody technical in the room for the plumbing and the permissions, but if the group is entirely IT, the ideas tend to be about IT.
The first small build usually happens in the first few sessions, because we deliberately pick a job that is narrow enough to finish. The framework and the licensing work runs alongside it rather than in front of it, with one exception: if the first idea touches your document estate or Microsoft 365 data, readiness comes first, because an agent inherits whatever over-sharing is already there.
On business and enterprise tiers, generally not by default, and that is one of the real reasons to move a team off personal accounts onto a business plan. The important part is that this is a contractual question with a written answer per vendor and per tier, so it belongs in your framework rather than in a reassurance. We will show you the terms rather than summarise them, and teaching your team to check that for themselves is part of workstream one.
Mostly your team can, once they know what drives it. The biggest savings come from sending less rather than switching vendor: retrieving the pages that matter instead of the whole library, using a cheaper model for the work that does not need judgement, caching the part of a prompt that never changes, and batching anything that is not urgent. We set a ceiling at build time and report cost per run, so an expensive agent is visible rather than absorbed.
Usually not, but it is a fair question and it gets a real answer rather than a reflex. The cases where it genuinely wins are data that is not permitted to leave your control, a regulator or tender wanting a straight answer on residency, or volume high enough that per-token pricing stops making sense. Below roughly 2 million tokens a day an API is normally cheaper than the hardware before anyone is paid to look after it. If the answer does turn out to be yes, we run sovereign hosting in New Zealand.
Models are replaced regularly and the good ones behave slightly differently when they are. An agent built as a single enormous instruction tends to break quietly at that point. Built properly, with the job broken into checkable steps, moving to a newer model is a test run rather than a rebuild. Knowing that difference is one of the things the programme exists to teach, and the catch-ups are where it gets caught.
No, and that is deliberate. General support agreements, whether per user, per device or per site, exclude AI tooling, agents and automations. AI work sits under this programme, under an AI plan, or is billed as time. The reason is honesty about cost: absorbing AI support into a plan that was never priced for it hides a growing bill inside a line item you thought you understood.
The businesses getting value out of AI did not start with a strategy. They started with one annoying, repetitive job, a person who was glad to see the back of it, and somebody in the room who knew what to watch out for.
We do the switch: your current provider, the migration, the handover, all of it. Most teams barely notice the cutover happened.
On the right plan, compliance, reporting and budgets are handled inside the partnership. You run the business; we run the IT underneath it.
Quarterly business reviews are generated automatically from your live environment: spend, posture, recommendations and roadmap, ready for the board, reviewed with your account manager.
The honest bit: the full looked-after experience comes with the right plan. We charge fairly for what we take on, and when costs step up it's because you are taking on more, always moving in the right direction.
Sovereign data centres across New Zealand and Australia, with your data kept onshore wherever it's required. Our team understands New Zealand, and our leaders have built, scaled and secured businesses right across the New Zealand landscape.
Microsoft Solutions Partner in both Modern Work and Security: the two designations covering the platform your business runs on and the security that protects it.
