What an AI-native operating model is, and what it asks of the enterprise

When tools stop being the change

The first wave of enterprise AI often looked and felt like upgraded productivity software: faster search, better drafting, more automation around documents. What is emerging now is closer to a new operating logic for knowledge work, and that has implications for cost, control and organisational design.

In many functions the limiting factor is no longer access to information; it is the time and effort required to turn information into action without losing quality or accountability.

Take a high-volume employment dispute practice. A new matter arrives with a claimant statement, an employment contract, workplace policies, email and chat records, and a timeline that may already be contested. In an AI-native set-up, the first pass is not simply “someone reads it and writes a note”. The workflow routes the intake through a structured triage: the system extracts key dates and obligations, drafts a chronology and issues list, highlights potential procedural risks, and produces an initial case summary linked back to the source documents it relied on. It can also draft a first response outline and a practical plan for what information will be needed next. If this compresses the first meaningful case view from hours to under an hour, the immediate question is not which model is best. It is: who is accountable for the triage outcome, what confidence thresholds trigger senior review, how sensitive data is handled at ingestion, and how the reasoning record is kept so a supervisor can stand behind the approach if the matter escalates.

Or take disclosure in a commercial dispute. Instead of a one-off document review sprint, an AI-native approach treats disclosure as a routed pipeline. New material is continuously ingested; the system deduplicates, clusters documents by issue, flags likely privileged communications and confidential third-party information, and proposes redactions with reasons. Low-risk items flow through light-touch checks; high-risk items, privilege, regulatory sensitivity, anything novel or ambiguous, are automatically escalated to a named reviewer. The hard problem isn’t the model’s speed. It is designing the exception path and the audit trail so that the supervising lawyer can credibly explain, “this is how we protected privilege, reduced error, and maintained proportionality”, and can evidence it quickly if challenged.

This is the same pattern again: the technology speeds up the first pass, but the operating model decides whether the organisation can trust the output.

Procurement shows a similar pattern. If sourcing teams can produce draft supplier comparisons and risk summaries in hours rather than days, leaders still have to decide what becomes standardised, what remains judgement-based, and how auditability is preserved when the work is machine-assisted.

This is the heart of “AI-native”: not new tools, but new assumptions about how work is composed, where decisions sit, and what we choose to reward. The opportunity is thoughtful redesign, change that is intentional and humane, rather than disruption for its own sake. These questions of AI strategy, implementation and organisational change are also examined in LSI’s  online MSc AI for business transformation.

A practical definition of AI-native

“AI-native” can become a branding term, so it is worth defining it in a way that supports real operating model decisions and sensible investment governance.

A practical definition is this: an AI-native operating model is an organisational design in which AI is treated as a persistent capability embedded into workflows, governance and economics, with explicit accountability for outcomes, risk and learning.

The word “embedded” matters. If AI lives in a lab, it tends to remain a pilot. If it sits in the flow of work, it changes unit economics, and it also changes what we mean by a decision, a record, and a rationale. That question, what something is, not just what it does, sits close to my own interest in philosophical ontology, and it is surprisingly relevant here.

In practice, AI-native redesign often shows up as a shift in what becomes standard. Customer service can move from static knowledge articles to policy-grounded response generation with confidence measures and clear triggers for human takeover. Finance can move from manual reconciliations to automated anomaly detection with narrative explanations, provided the organisation agrees what counts as a “material” exception.

And we start to see different “assets” becoming valuable: prompts and dashboards help, but the enduring assets are decision logs, evaluation harnesses, controlled access to data, and policies that translate risk appetite into executable rules.

In my own work building AI-supported learning experiences, the lesson has been consistent: impact comes less from the model and more from the surrounding system that measures understanding, captures feedback and improves continuously. Enterprises tend to need the same discipline.

AI in Business: Strategies and Implementation

Artificial intelligence is transforming how modern businesses operate, compete, and grow. But its power lies not in blind adoption, but in strategic discernment. This module moves beyond the hype to explore how leaders can make informed decisions...

Learn more

Enterprise value moves to workflow economics

The business case is often presented as “productivity”, but I have found a more reliable lens is workflow economics: cost per case, time to resolution, error rates, and the cost of control.

Productivity claims become credible when we translate them into unit economics. A practical approach is to choose a small number of workflows that are high volume, high labour cost, or high risk, and then model a realistic range of impact.

In a regulated contact centre handling 2 million interactions a year, a 15–25% reduction in average handling time could represent meaningful capacity release. But value only appears if staffing models, scheduling and quality assurance change in response; otherwise we simply get faster work, not cheaper or better work.

In B2B sales, the story may be less about minutes saved and more about win rates and reputational risk. Shaving a day off proposal drafting can be marginal; reducing compliance errors that previously caused bid disqualification can be decisive.

The trade-off is uncomfortable but important: more automation can bring throughput and consistency, and it can also increase the speed of failure when governance is weak. That is why AI-native models treat the “cost of control” as a first-class metric, not an overhead to squeeze.

A helpful test is: if AI increases output, what must change in staffing, controls and incentives so that the benefit shows up in the P&L rather than as a busier organisation?

Operating model shifts that matter

AI-native redesign affects how teams are organised, how work is routed, and how accountability is assigned. These are governance questions disguised as technology questions, and they are the ones that determine whether AI becomes genuinely helpful or quietly corrosive.

Roles often shift from producing first drafts to supervising, correcting and approving them, with clearer thresholds for escalation. This does not remove expertise; it changes where expertise is applied. A risk specialist, for example, may spend less time on routine review and more time defining policies, shaping exceptions, and monitoring the signals that tell you whether the system is still behaving as intended.

Decision rights also become more explicit, because “who decided?” can no longer be left vague. Who can approve an automated action? Who owns the policy the AI is meant to follow? Who is accountable when an AI-supported decision is challenged by a customer, a regulator, or even a colleague? This is another place where an ontological lens helps: we have to be clear what, exactly, counts as a “decision”, what counts as a “recommendation”, and what counts as “evidence” inside the organisation.

Structurally, many enterprises oscillate between a central AI team and fragmented experimentation. In practice, AI-native models tend to land on a hybrid: central standards for data access, evaluation, privacy and assurance, combined with federated product ownership in the functions that carry the outcome. The trade-off is speed versus consistency, and it needs to be managed, not wished away.

Finally, routing matters. AI introduces a new routing layer based on confidence, risk category, and context. The design of exceptions becomes the design of trust. If 70% of cases are low risk and can be handled with light-touch review, the remaining 30% need a deliberately resourced path, otherwise the bottleneck simply migrates and people lose faith in the system.

The AI-Native Organisation

How might organisations change when intelligence is built into every process and decision? Exploring emerging models of leadership, structure, governance and culture in an AI-augmented enterprise.

Learn more

From pilots to a production portfolio

Industrialising AI is less about scaling one model and more about looking after a portfolio of use cases with different economics, risks, and change requirements.

Pilots often go well because they sit on motivated teams, clean data extracts and informal workarounds. Production removes those cushions, which is why an AI-native operating model needs a steady cadence for moving from experiments to repeatable delivery.

This is where portfolio thinking becomes genuinely useful. Some use cases are assistive: they improve speed and quality while keeping autonomy low. Others are decisive: the AI triggers actions. These categories have different governance and assurance needs, and they tend to deliver value in different ways and on different timelines.

Vendor and platform choices then stop being purely technical decisions and become operating model decisions. A multi-vendor approach can reduce dependency risk, but it can also increase integration work and the ongoing “cost of control”. A more consolidated approach can simplify controls, but increases lock-in risk and may reduce flexibility as regulation evolves. 

Change management is often the hidden constraint. If performance measures continue to reward the volume of manual work, adoption will stall. If managers cannot explain how roles will evolve, informal resistance will accumulate even when the technology performs.

A practical signal that a pilot is becoming a product is when it has an owner for outcomes, a defined evaluation method, a release process, and an explicit rollback plan, so the organisation can learn confidently without gambling with trust.

Measurement, assurance, and earned trust

AI-native organisations treat measurement as a control system, not a reporting exercise. The aim is to learn quickly without gambling with reputation or compliance.

It helps to separate leading indicators from lagging outcomes. Leading indicators might include quality scores against a reference set, human override rates, time to resolution, and the proportion of work routed through an AI-enabled path. Lagging outcomes might include cost per case, complaint rates, regulatory findings, fraud losses, or employee attrition in teams most affected by the change.

Assurance needs to be designed for the operating environment. A marketing copy assistant has a different risk profile from an AI that drafts adverse credit decisions. In higher-risk settings, the non-negotiables often include traceability, data minimisation, access controls and an auditable rationale. In lower-risk settings, the emphasis may be on brand consistency, factual accuracy and intellectual property hygiene.

The mitigations are rarely glamorous but they are often decisive: clear policy constraints, defined review thresholds, red-teaming for misuse, and incident response that treats model failures like other operational incidents.

If there is one point I would underline, it is this: AI-native transformation is a trust project disguised as an efficiency project. Trust is earned when governance keeps pace with capability, and when accountability remains legible even as work becomes partially automated. And, to return quietly to that ontological thread, trust also depends on shared clarity about what our records and rationales are, what we treat as a decision, what we treat as a reason, and what we treat as an acceptable explanation.

Several questions are worth keeping close:

One: If we had to pause it tomorrow, could we?
If the AI-enabled step started producing doubtful outputs, could we switch it off (or fall back safely) without work grinding to a halt, and do we know who has the authority to make that call?

Two: Where, exactly, does a “recommendation” become a “decision”?
At what point does the organisation treat the AI output as something it will act on, and is it clear who owns that handover, and therefore owns the consequences?

Three: What is our exception path, and is it properly resourced?
When confidence is low, risk is high, or the case is novel, where does it go, who reviews it, and have we sized the people, time and expertise needed so exceptions don’t become a hidden bottleneck?

Four: Can we evidence the reasons, not just the result?
If challenged, can we show what inputs were used, what policy constraints applied, what human overrides happened, and why the final outcome was reasonable, in a way that is readable to the people who matter (clients, colleagues, auditors, courts, regulators)?

And perhaps that is the quiet ontological point to close on: if AI is now shaping how work is done, then we must be clear not only about what we do, but about what our decisions, reasons and records are, because trust rests on that shared understanding.