tōtara software
Consulting and softwareAuckland and LondonSince 2006

Clear strategy and sound architecture for data and AI.

We set data, analytics and AI strategy, and design the business and solution architecture that delivers it.

We also build the AI Control Plane, so the AI you adopt stays visible, governed and evidenced.

01What we do

Four disciplines, one line of sight from strategy to delivery

Strategy that can't be delivered is a wish list, and architecture without a strategy is guesswork. We do both, so every design decision traces back to what the organisation is trying to achieve.

DisciplineWhat it gives you
Data and analytics strategyPlatforms, governance, operating model Where your data needs to go, what has to be true to get there, and the order to do it in: the data platform, ownership and quality, and the reporting and analytics people actually use.
AI strategyUse cases, risk appetite, governance Which uses of AI are worth it, which aren't, and the guardrails to adopt them safely: an inventory of the AI you already run, a risk appetite the board can sign, and the controls that follow from it.
Business architectureCapabilities, value streams, change A shared picture of what the business does and how it changes: capability maps, current-state views and roadmaps that tie investment to outcomes.
Solution architectureAssessments, designs, decisions Designs and product choices that fit the estate you have: impact assessments, solution assessments and decision records that make the trade-offs explicit and the risks owned.
02Industries

Deep in two industries where data and regulation decide everything

Water

Water, wastewater and stormwater

Water organisations run long-lived assets, big capital programmes and models that inform decisions worth billions. We help them get the data foundations and architecture right.

  • Data and analytics strategy for asset-intensive organisations
  • Capital delivery: systems of record and asset handover
  • Water-quality, modelling and decision-support platforms
  • Architecture principles, assessments and roadmaps
Banking

Banks, credit unions and financial services

Twenty-five years of building and architecting financial-services platforms, from investment-banking risk and reporting to core banking, payments and financial crime.

  • Platform selection, target architecture and migration
  • Regulatory, prudential and risk data
  • AI adoption under prudential and conduct regulation
  • Our own banking platform, at totara.global
03Our story

Twenty years of hard data problems, from trading floors to water networks

Tōtara Software started in 2006. Its roots go back to 1999 at Peace Software, tuning what became the highest-volume utility system in the world.

For most of the next decade we worked on the trading floors of the big investment banks: Deutsche Bank, Barclays Capital and Nomura. We built risk and P&L systems, real-time analytics and the calculation grids beneath them, running billions of calculations a day across London, New York, Tokyo and Singapore. For Deutsche Bank we built a monitoring system that went live on four of the world's exchanges. For Barclays Capital we built a derivative analytics platform that replaced a dozen legacy systems.

At BP we moved market and credit risk to the cloud and put machine learning to work on risk scenarios. We worked on graph databases and Spark in Sweden with 51 Zero, and built a technology team from nothing as contract CTO for Harmoney. At Barclays Group we led architecture and development for official risk and P&L, with a hundred people across six countries.

Back in New Zealand we set the data, analytics and insights strategy for Watercare, with an IoT strategy and a digital twin of its network. Its drought modelling informed the Mayor and central government. We architected the launch of a new Australian investment bank for Jarden. At Westpac New Zealand we led the data strategy that brought in Snowflake. We were also the architects for separating Westpac NZ's treasury and financial markets from Australia under the Reserve Bank's BS11, and then for the Deposit Takers Act.

In short

Fifteen countries, two industries we know from the inside, and one lesson: the hard part is rarely the technology. It's knowing what to do, and then getting it done.

Why we built a banking platform

Inside the big banks we saw the same thing again and again: good people held back by inertia and legacy systems. Meanwhile smaller banks and credit unions, whose customers like them, were being squeezed out by technology budgets and regulation they couldn't match. So we joined five other financial-services professionals to build Totara, a modern banking platform with governance and regulation built in. Building it taught us how to turn every regulation into a control that runs on live data. That is the engine of the AI Control Plane.

Why AI makes this a dangerous moment

AI is moving faster than any technology we've worked with in 25 years, and the picture changes month to month. New models arrive every few weeks, and tools behave differently after an update. Staff use AI their employer never approved. The rules differ by country and keep shifting: Europe has pushed back its high-risk AI rules, the United States has no federal AI law, and New Zealand and Australia apply existing privacy, conduct and operational-risk law to AI rather than writing new law.

It's a dangerous time to use AI. It can be confidently wrong, drift without anyone noticing, tell you what you want to hear, and leak data through tools no one approved.

It's a more dangerous time not to. Competitors who use AI well are already faster, cheaper and better informed, and the gap compounds. Standing still is a decision too, and usually the most expensive one.

So we start from the business outcome, not the technology. We use AI where it adds value: try it, measure it, then scale it or stop. People stay accountable for every decision AI informs, and AI is governed with the same discipline as any other risk. That is why we built the AI Control Plane, below, and why every conversation with us starts from the five types and six terms that follow it.

04Product

The AI Control Plane

We've built a product to solve the AI challenges facing every industry. The AI Control Plane is one place where a company sees, governs and evidences all its AI. It runs as a Snowflake Native App inside your own Snowflake account.

The answer

Every model, vendor, agent and shadow tool on one register, a risk appetite that switches the right controls on, and evidence a board, auditor or insurer will accept.

1 · Coverage

See all of it

Every AI system, model, vendor, agent and shadow tool, registered and traced to the laws and standards that apply, down to the clause. New uses come through an intake gate: request, classify, approve, then deploy.

2 · Appetite

Set the risk you'll accept

Set a risk appetite for each area and the matching controls switch on, including the good practice nobody mandates yet. Change the appetite and you see exactly what changes.

3 · Evidence

Prove it, honestly

Every control shows where its evidence comes from and how fresh it is. Stale evidence never shows as green, so the board sees the real position, not a hopeful one.

Your AI
Models and agents on Snowflake
AI vendors and SaaS tools
Shadow AI and new requests
What you get
Board packs
Exceptions and deadlines
Evidence auditors accept
The AI Control Plane runs inside your own Snowflake account. Where your AI runs on Snowflake, evidence is collected automatically from the telemetry Snowflake already records.
Where the evidence comes fromWhat it means
MeasuredCollected automatically from inside your Snowflake account, every time.
IntegrationFed from the systems your AI runs in, through a connector.
AttestedSigned off by the owner, once or on a schedule, with maker-checker.
StaleOverdue evidence is shown as overdue, never as green.

Built for regulated work, ready for any industry

The engine underneath is the one we built for banks, where every control has to stand up to a regulator. Controls map to the frameworks your auditors use, including ISO/IEC 42001, the NIST AI Risk Management Framework and the EU AI Act, and to the privacy, operational-risk and conduct rules regulators apply to AI in New Zealand, Australia and beyond. A cross-industry core comes with sector packs, so a bank and a water utility start from the controls that matter to them.

Also included: impact assessments, incident clocks, AI literacy records, third-party AI, and a dated regulatory horizon with a "what if this law commences" view.

05Using AI

Five ways to use AI, each with its own controls

"Using AI" covers very different things. We sort every use into one of five types, from the most controlled to the most autonomous, because each needs different systems and different controls.

TypeWhat it isWhat runs in productionControlsRisk
AI that buildsA forecasting model, a data pipeline, a rule set, an automated workflow AI helps build code, workflows, rules or models. People test and release what it builds. Deterministic. The code or model runs, not the AI: versioned, tested, validated. Code review, testing before release, version and lineage recorded. The tool, not the AI, is the evidence. Low to medium
AI agents for data insightExtracting obligations from contracts, classifying requests, flagging anomalies in readings or transactions Prepared agents that find, prepare and interpret data, structured or not, inside a designed process. Not deterministic, but outputs are structured (fields, scores, classes), so routines and people can check them. Guardrails on data and actions, output checks, accuracy measured against samples, human review by risk tier. Medium to high
AI you ask"Which assets failed most often after last week's storm?" A person asks in plain language, and the AI answers from the data and functions it's given. Staff, or the public. Not deterministic: an answer in words, with its sources. Answers cite their data and aren't used alone as evidence. For the public: published data only, topic limits, hand-over to a person. Medium; high for the public
Workplace efficiencyAssistants for documents, meetings and email AI helps people draft, summarise, search and organise their own work. A draft a person owns and edits. Acceptable-use rules, approved tools only, no sensitive data in unapproved tools. The person owns the output. Low
AI that actsRaising work orders, sending messages, changing settings, approving transactions AI changes systems, sends messages or makes decisions on its own, without a person approving each step. Not deterministic, and it acts before anyone checks. Limits on every action, approvals, the ability to undo, full audit, and a maturity most organisations don't have yet. High
Which type?
  1. Will the result be used as evidence, or run again and again in a regulated process? Use AI to build a tool, test it, and run the tool: AI that builds.
  2. Does a system or routine need to use what the AI produces? Give the AI a fixed output format, check its outputs, and keep a person in the loop by risk: AI agents for data insight.
  3. Is a person asking questions to understand something? AI you ask. If the public is asking, add the public guardrails.
  4. Is it a person's own document, email or meeting? Workplace efficiency, with approved tools.
  5. Would the AI act on its own? AI that acts. For most organisations today, redesign it so a person approves each action.

Five questions, in order: the first "yes" gives the type. A use case can mix types, and each part takes its own type's controls.

06Six terms

Six terms everyone using AI should know

In plain words

Deterministic

Same input, same answer, every time. A calculator is deterministic: 2 + 2 is always 4.

Most AI is not. Ask it the same question twice and you can get two different answers, both reasonable. That's fine for a draft, but regulated work needs answers you can repeat and explain. So for regulated work, use AI to build a tool, test it, and run the tool, not the AI.

Watch out for

Hallucination

AI can state something wrong with full confidence: a number, a reference, a rule that doesn't exist. It isn't lying; it's predicting what sounds right. Check the facts that matter, and ask it to show its sources.

In plain words

Guardrails

The fences around an AI: what data it can see, what it may answer, what it must never do, and when it must hand over to a person. Public-facing AI needs the tightest fences: it sees only published information, never internal or personal data.

Watch out for

Drift

An AI's results slowly change without anyone deciding they should. It forgets what it was told at the start of a long task; the supplier releases a new version; or the world moves on, so what was right last year is less right now.

The cure is to design through specifications: write down what the AI must do, keep it versioned, give it to the AI every time, and test against known answers before and after go-live.

In plain words

Human in the loop

A person checks, approves or decides before an AI's output is used or acted on, more so as the risk rises. That person needs the time, knowledge and authority to say no, not just to click "approve".

Watch out for

Sycophancy

AI leans towards telling you what you want to hear. Models learn partly from people rating the answers they like, so they agree, praise and confirm: they back a weak idea, or drop a right answer when you push back. In 2025 OpenAI withdrew a ChatGPT update because it had become too sycophantic.

So don't lead the question; ask for the weaknesses and the other side; and judge results against specifications and tests, not against how confident the AI sounds.

07How we work

The answer first, then the reasoning

Every piece of work opens with the answer on one page: what we recommend, why, and what it depends on. The detail follows for the people who need it.

We work alongside your architects, data teams and decision-makers, and leave you with documents you can keep using: versioned, traceable to your principles, and written in plain English.

What you getThe question it answers
StrategyWhere are we going, what has to be true to get there, and in what order?
Current-state mapWhat do we have today, and how does it fit together?
Architecture principlesWhat rules do our decisions follow, and why?
Architecture impact assessmentWhat does this initiative change, what does it cost to run, and is it ready to fund?
Solution assessmentDoes this product or design fit our estate, and on what conditions?
Decision recordWhat did we decide, what did we reject, and why?
08People

Directors

Ross Millen

Ross Millen

Director

Ross has spent more than 25 years designing and delivering data and financial-services platforms in 15 countries, from investment-banking risk and reporting systems to enterprise architecture in the water sector. He leads our strategy and architecture work and the AI Control Plane.

LinkedIn
Melanie Gabriel

Melanie Gabriel

Director

Melanie brings corporate and financial accounting experience from some of New Zealand's biggest companies, and has delivered large financial transformation programmes.

LinkedIn
09Contact

Talk to us

Tell us what you're trying to decide or deliver. We'll tell you plainly whether we can help, and how.

Email Ross
Email
ross@totara.software
Offices
Auckland, New Zealand
London, United Kingdom
Banking
totara.global