Why we acquired Rumi

You don’t need a Ferrari to get milk.

By David Brudenell, Executive Chairman, Decidr AI Industries (ASX: DAI)

The word that kept coming up

Yesterday I was at a conference just outside of San Francisco. OpenAI’s CFO was on stage talking about practical adoption, and the gap between what AI makes possible and how companies are actually using it. The word that kept coming up was context. Not models, not compute, not tokens. Context. How do you give the AI enough structured understanding of your business that it can do useful work?

I sat in the audience thinking: that’s our entire thesis. We’ve been building toward this for two years. And Rumi is the piece that makes it real time.

Token maxxing is not the answer

Token maxxing doesn’t work. Throwing more tokens at a problem, running bigger models against unstructured data, hoping intelligence compensates for the absence of understanding. It produces output that looks impressive and falls apart under audit. Enterprises are learning this the expensive way.

Real value inside a business is judgment and organisational memory. A senior underwriter evaluating a complex claim draws on pattern recognition that took fifteen years to develop. A procurement lead knows which suppliers actually deliver on time and which ones pad their quotes because she’s watched them do both. A regional sales director reads a room in Jakarta differently than she reads one in Melbourne, and that difference closes the deal. None of it sits in a database. Almost none of it has been written down. And it’s the thing that actually makes the business work.

How do you capture this tacit knowledge before a foundation model does it for you, on their infrastructure, for their benefit?

The archaeological dig and the living map

Sugarwork was the first step. We acquired it because it performs that archaeological dig, the structured discovery of how work actually gets done inside an organisation. Not how the org chart says it gets done. How it actually gets done.

Sugarwork interviews the people and maps the workflows, then surfaces the tribal knowledge that never made it into a document and puts it into a structured format the machine can reason about. Work in systems, yes, but more importantly, work happening outside systems. Handoffs and workarounds and undocumented exceptions that experienced people carry in their heads.

That workflow map is valuable, independent of anything AI does with it. Most businesses don’t know how they actually work. They know how they think they work, which is a different thing entirely. A Sugarwork deep dive gives a business quantitative visibility into the risk and value of each task within each workflow, and where automation potential actually sits. It’s often the first honest answer an organisation has ever had to the question of what it actually does all day.

Rumi is the second step. It makes knowledge capture continuous.

Sugarwork captures a point-in-time picture. Rumi captures the ongoing signal. Always-on tacit knowledge telemetry from Slack, from meetings, from the places where institutional knowledge is constantly being created and revised as the business grows and markets shift. A business doesn’t stay still. An onboarding process that worked six months ago has already been patched three times by people who never updated the documentation. A workaround someone invented for a broken CRM integration last quarter has quietly become the de facto process. Rumi catches that evolution and feeds it into the workflow structure Sugarwork identified. The archaeological dig becomes a living map.

Tasks, not jobs

AI doesn’t replace jobs. AI replaces tasks within workflows. And when you decompose a workflow into its atomic tasks, something becomes clear quickly: most of them don’t need a frontier model. They need conditional logic or a rules engine or a human with specific judgment. The tasks that genuinely benefit from a large language model are a minority of the total. Everything else is orchestration and governance.

If a frontier model is a Ferrari, you don’t need to drive it to get milk at the corner store. Most tasks inside an organisation are corner store errands. Using a frontier model to send a follow-up email or update a CRM field is burning venture-subsidised compute on work that a sovereign fine tuned model, running inside your own boundary, handles for a fraction of the cost.

Sugarwork discovers the workflows. Rumi captures the ongoing tacit knowledge that flows through them. Together, they build the training corpus for a sovereign model the enterprise owns outright, fine tuned on how that specific business actually operates, running on infrastructure the enterprise controls.

Decidr orchestrates the execution, routing each task to whatever handles it best. Some tasks need model intelligence. Most need rules or human judgment. The routing decision is where the value lives.

Knowledge you can’t get back

There’s a knowledge security dimension to this. Every prompt your employees send to a foundation model reveals how your organisation reasons. Even where enterprise agreements prohibit training on customer data, the patterns of usage expose the institutional logic that took decades to build. What gets asked. How problems get decomposed. Which decisions require human escalation. The foundation model providers are capturing context about how your business works, whether or not they’re technically training on your data.

And as those providers move toward IPO and token subsidies get repriced, the cost of that dependency is going to increase. Anthropic already killed flat-fee enterprise pricing. OpenAI moved Codex to token metering in April. GitHub tightened Copilot limits the same week. Windsurf replaced credits with daily quotas in March. Businesses that wired these models into their operations without first understanding their own workflows face two problems at once: rising costs and knowledge they can’t get back.

The fix is first principles. Understand your business before you automate it. Capture your institutional knowledge inside your own boundary. Fine tune your own model on what you captured. Use frontier models for the tasks that genuinely need them and sovereign models for everything else.

There’s an age-old saying: How do you eat an elephant? One bite at a time.

Enterprises are similarly great beasts. 

So how do you run an enterprise AI transformation? One workflow at a time.