SRG · AI-Native Thinking
AI-native is not a technology concept. It is an organisational rebuild
Eleven reflections from SRG’s building site
What chance does a ten-person company have, head to head, against an institution of 100,000 people that holds decades of data and distribution?
By conventional business wisdom, the answer is: none.
But this time, conventional wisdom may have to be rewritten. Those ten people aren’t necessarily smarter. The race has simply moved to a different track. And their opponent is carrying decades’ worth of brake pads.
The AI-native company’s opportunity lies not in leading on technology but in leading on organisational form. It isn’t that large institutions can’t use AI. It’s that they can’t turn. That lag is the small team’s entire window.
The eleven points below are not predictions. They are notes from the building site. I won’t claim they’re definitively right, but every one of them stands up to scrutiny.
01 | Nobody knows what an AI-native company looks like
Ask what an AI-native company is and you’ll hear three popular answers: a company that uses lots of AI tools, a company with very few people, a company that leads on technology. The three contradict one another, and none has been proven.
The right to define it is still being fought over — consultancies are selling definitions, big tech is selling roadmaps, start-ups are selling case studies. It’s the same position as asking “what does an internet company look like?” in the mid-1990s.
The difference is that AI’s pace of iteration has raised the cost of waiting for an answer: wait for consensus, and by the time it arrives the window has closed.
So for now, practice has to come first: until a standard answer emerges, let your own output be the source of the definition. Definitions aren’t thought up. They’re built.
02 | SaaS is drifting: not an upgrade but a rebuild
Vertical apps, super-apps, headless apps — all three paths have people on them, and it’s too early to say where they’ll converge.
Every few months, AI capability redraws the map of what’s possible, and user habits start over with it. Product form is a function of those two curves. Neither curve has settled, so the function has no solution.
So for now, don’t count on picking the right form. Staying rebuildable is enough: when the form drifts, you can swap out the skeleton cheaply.
Two disciplines
One: don’t compete with the LLM. Training your own general-purpose model to take on the frontier (you lose on both cost and iteration speed), treating the large model as an old engine waiting to be replaced, passing off prompt engineering as a product — all three are dead ends.
The right stance is to ride the LLM: as its capability rises, your product gains value automatically. The test is simple — when the next generation of models ships, are you a beneficiary or a casualty?
Two: a super-app is an orchestration layer. It shouldn’t cram every feature into one app. It should be a layer that connects downwards to specialist capability in vertical domains (one agentic interface per domain), handles general-purpose design in the middle, and delivers upwards in two entirely different forms, depending on who the user is:
For people: the interface is conversation, reports and dashboards, optimised for understanding and decisions;
For agents: no interface, just protocols and structured data, optimised for throughput and fidelity.
These two shouldn’t take the same shape. Mix them and you please neither.
03 | The OPC is wishful thinking: agents can take over tasks, not responsibility
The one-person company (OPC) is an alluring fantasy. Agents can do your tasks for you. They can’t take responsibility, make judgements or build trust for you. Anything that means answering for consequences to the outside world, sustaining long-term relationships or making trade-offs amid ambiguity needs a person.
The model that holds up better is a small team amplified by agents — the same ten people producing the output of several dozen in a traditional organisation (an illustrative order of magnitude, not a measured benchmark). But that depends on designing “work agents can do” and “work a person must own” separately, rather than letting people degrade into agent supervisors.
The boundary is blurring, and that raises a question. Push agents to the limit and a person becomes a kind of “superhuman” — not through exceptional stamina or IQ, but by holding four production lines at once: research, engineering, compliance and delivery.
The trouble is that the traditional “job” was designed around scarce skills. Skills are ceasing to be scarce; what’s scarce is judgement and taste. So the anchor for incentives and promotion has to move: from “what can you do” to “what do you decide, what do you own, what do you teach the organisation”. That’s far harder than choosing tools.
Go one level deeper and an AI-native organisation has to design itself as a system that iterates on itself: every task is a training run, outputs settle into reusable capabilities, and failures settle into rules so they never happen again. That takes four disciplines:
- Results leave a trail and can be replayed and audited;
- Skills have two separate states, “usable” and “trusted”, and are promoted or demoted on evidence from real runs, not on how well they’re written up;
- Nothing goes out without passing a machine quality gate: catch the errors first, then ship;
- Regularly distil scattered process into methodology.
To be clear, this machinery isn’t hard to build. The hard part is sticking with it. The enemy of organisational self-evolution has never been technology. It’s every “let’s make an exception just this once”.
04 | Tool parity is the windfall, but the moat is “data × agents × time”
The biggest structural windfall of this era is tool parity. A ten-person start-up today can use the same class of models, retrieval, compute, data pipelines and automated orchestration as a large institution.
Large institutions have the opposite problem: their history, their organisation and their past success have themselves become drag — processes designed for stability, budgets carved up by department, data locked in silos, incentives tied to the old business. Every one of these is a brake pad on AI transformation.
So the small team’s opportunity is that it has no brakes on.
Overtaking on the bend is real, but only on one condition: leading on organisational form, not on technology. On the straight — capital, brand, distribution, accumulated compliance — you can’t catch them. On the bend — speed of decision, frequency of experiment, ability to rebuild form — they can’t turn.
This is a time lag, not a moat. They will change, just slowly; and in exponential change, “slow” is the same as standing still. Once they finish rebuilding their organisations, the windfall will narrow. Small teams need to stay clear-eyed: the window won’t be open for ever.
Here’s another one people often get backwards: data is the lifeblood, but “owning data” is losing value. With AI workflows, the cost of collecting, cleaning, structuring and aligning data is collapsing, so simply owning data will soon be a commodity.
What really opens up the gap is step two: data plus agents. The same data compounds only when you pair it with an agent layer that keeps consuming it, generating judgements and turning those judgements into action.
Put bluntly: data without agents is just inventory. The moat isn’t data. It’s “data × agents × time”.
05 | Treating AI-native as a technology is the most expensive misunderstanding
Treat AI as a technology and you fall into three traps: buying tools and calling it transformation, installing systems and calling it transformation, setting up an AI department and calling it transformation. What they share is mistaking the means for the answer.
Treat it as strategy, and it comes down to three questions:
- What judgement is genuinely scarce in our organisation?
- Which steps must a person own?
- What mechanism makes the organisation genuinely a little stronger every week?
The choice of technology is downstream of these three answers.
Be clear about the price, too: you’re choosing a road that has to be upgraded every day. An old organisation gets its stability from “not changing”; an AI-native organisation gets its stability from “the capacity to change”. The hard part isn’t accepting new tools. It’s continually scrapping what you’ve only just built.
Get this far and “should we use AI?” stops being a question — it’s already how you run.
06 | AI-Native Readiness: four questions, four levels
To judge how far a company has come, start by answering four questions (SRG’s diagnostic framework):
- Are you riding the LLM, or competing with it?
- Is your product a collection of features, or a layer of process orchestration?
- Are your people agent supervisors, or the owners of judgement?
- Does your organisation get a little stronger every week, or make an “exception” every time?
- L1 Tooling
- Individuals use AI; the organisation doesn’t change — the windfall belongs to individuals, not the company.
- L2 Process
- AI is embedded in fixed processes — efficiency rises, but form hardens, and model upgrades bring only linear gains.
- L3 Agentic
- Work is re-divided into “can be delegated to agents” and “must be done by people”; output is amplified by small team × agents.
- L4 Self-evolving
- Outputs and failures alike settle into long-term organisational capability; form can be rebuilt, and model upgrades turn into value automatically.
Most companies are stuck at L2 and mistakenly believe they’re at L3. The difference: L2 stuffs AI into existing processes; L3 redraws the processes around AI.
07 | Three things decision-makers can do today
The thing only you can think through, that no one else can do in your place. Tools are downstream of it.
What goes to agents, and what needs a person to own the judgement and the responsibility. Don’t let people degrade into agent supervisors.
Trails, trust levels, a quality gate before anything goes out. Get it running first; refine it later.
08 | Three things not to take on faith
- Don’t fall for “fewer people means AI-native”. Fewer people isn’t the goal; amplified output is. “Fewer people” without agent amplification is just being small.
- Don’t fall for “tools first”. Get the order wrong, and the more smoothly the tools run, the further off course you drift.
- Don’t fall for “data scale”. Data is inventory; add agents and it becomes capability.
Conclusion | Some wait for the answer. Some write it.
Back to where we started: there is no standard answer to this question. The only difference is that some people wait for the answer, and some go and write it.
The last question I’ll leave decision-makers with is more worth answering than “should we adopt AI?”:
If I plug in the model that’s released tomorrow, will my product automatically gain value, or be overturned?
If there’s already one of these four questions you can’t answer — come and run an “AI-Native Readiness diagnostic” with us. Using the same framework, we’ll help you pinpoint where you stand between L1 and L4, and the three things to act on first to reach the next level.
· This is an opinion piece; its frameworks and judgements come from SRG’s own engineering work between August and September 2026 and involve no client data
· “AI-Native Readiness” (four questions, four levels) is a diagnostic framework named by SRG; it does not constitute professional advice or investment advice
· Statements such as “ten people / the output of several dozen” are illustrative orders of magnitude, not measured benchmarks
· The “internet company” passage is a figurative comparison, not a specific historical conclusion
Disclaimer: This article is for information only and does not constitute investment advice or professional consulting advice. The frameworks and judgements in it come from the author’s engineering work at SRG (Standard Risk Global) between August and September 2026 and involve no client data. Views on trends are the author’s own; readers should form their own judgement in light of their industry and their company’s circumstances.