Thought leadershipQuarterly briefingQ3 2026

Adoption is universal. Impact is not.

Almost everyone is using AI and a minority can find it in the P&L. The separation this quarter comes from four things: redesigning the operating model rather than the tooling, running agents that complete work under real controls, owning the intelligence layer, and behaving like a company that builds software rather than one that rents it.

12 min read

The quarter in three figures

  1. 75%

    of AI's economic gains accrue to the top 20% of companies.

    PwC · 2026 AI Performance Study

  2. 2.6×

    more likely that AI leaders redesign their business model — and two to three times more likely to aim AI at new revenue rather than cost alone.

    PwC · 2026 AI Performance Study

  3. <12%

    of companies have AI-ready data. For most organizations the binding constraint is not the model — it is the data foundation.

    Precisely & Drexel LeBow · Data Integrity Research

The architecture

Three tiers. Only one of them is yours.

The model tier is becoming interchangeable and the systems of record were here long before any of this. The layer between them — where work is actually executed — is the only place a decision compounds into advantage. Everything below is an argument about that middle tier.

  1. AI models

    Increasingly interchangeable

    • LLMs
    • Reasoning
    • Coding
    • Voice and chat
    • Vision
    • Images and media
    • Multimodal

  2. The operating layer

    The system of action — where advantage compounds

    • Owned intelligence
    • Workflows
    • Agents
    • Permissions
    • Evaluations
    • Approval rules
    • Bespoke applications
    • Workflow state
    • Audit trails

  3. Systems of record

    Persist beneath

    • ERP
    • CRM
    • HCM
    • CIS
    • VMS
    • Transactional networks
    • Deep vertical platforms
AI models
Provide the intelligence. Substitutable, and getting more so.
Systems of record
Hold the truth. They were here before AI and they will outlast this cycle.
The operating layer
Turns both into execution — and it is the only tier where a decision compounds into advantage.
Where trends 02, 03 and 04 land in the architecture. The middle tier is the emphasis because it is the one you build; the tiers above and below it are largely bought.

AI strategy is now operating-model redesign.

Workflow redesign — not tool adoption — is the factor most consistently associated with measurable returns from AI. Adding an assistant to a process that was designed around its constraints leaves the constraints in place.

The gap between how many companies use AI and how many can find it in enterprise earnings is the clearest number in this quarter, and it is not a technology gap. It is the distance between deploying a capability and rebuilding the decisions, approvals, handoffs and execution around it.

Share of surveyed organizations

  1. Use AI regularly

    88%

  2. See enterprise EBIT impact

    39%

The distanceForty-nine points separate having AI from earning anything with it. Nearly all of that distance is operating-model work, and none of it is bought.

Source: McKinsey, State of AI 2025–26, as cited in the Q3 2026 AWALI briefing.

So whatA pilot that succeeds inside an unchanged process has proved the model works and nothing about whether the business does. The redesign is the deliverable.

Agents: AI that completes the work.

AI is no longer confined to generating content and recommendations. Agents now carry defined work through multiple steps — updating records, routing cases, preparing decisions, completing transactions.

The deployments that hold up are the constrained ones. They specify what the agent may reach, what control it has, which actions it may take, and exactly when a human has to approve. That is not a brake on value; it is the reason the thing is allowed to run at all.

So whatAutonomy is not the goal and it is not the metric. A narrow agent that completes its lane every time is worth more than a broad one nobody trusts to run unattended.

Fix the data. Then build owned intelligence.

Models are increasingly substitutable for a large share of tasks. What is not substitutable is the understanding of your customers, your operating rules, your past decisions and the exceptions you have already learned the hard way.

Leaders are converting institutional knowledge into reusable instructions, decision logic and evaluations — the working equivalent of a twenty-year veteran available to everyone at once. Microsoft has named this "owned intelligence", and the name is apt: it is the part of the stack that is yours.

So whatThis is the compounding asset in the whole brief. Every engagement, exception and decision you capture makes the next one better, and none of it transfers to a competitor with a purchase order.

Think like a technology company.

A differentiator is emerging between businesses that build their own value-driven software and businesses that rent generic products. AI tooling has collapsed the cost of building internally — domain experts now prototype in natural language — and shallow vertical SaaS is under real pressure as a result.

Buying generic software is no longer the automatic default. When a tailored solution delivers materially better fit, economics, integration or control, the burden of proof has moved to the renewal.

So whatThis is not an argument for building everything. It is an argument that "buy" stopped being the safe default answer, and that the companies noticing first are quietly rebuilding their cost base.

The conditions it runs in

Four things that decide whether any of it holds.

  1. AI-native culture is a performance advantage

    Not loyalty to one vendor — a model-agnostic habit of breaking problems down, delegating the right work, challenging outputs, building repeatable workflows and verifying quality. The goal is widespread fluency, plus the discipline to know when not to use it. Microsoft puts organizational factors at more than twice the weight of individual mindset in explaining AI impact.

    ActMeasure managers on workflow reinvention and outcomes — not on license activation or prompt volume.

  2. Governance moves from policy into runtime

    A written policy cannot govern an agent that reads repositories, touches customer data and transacts. The controls have to live in the operating environment itself: agent identity, least privilege, approval thresholds, activity logging, continuous evaluation and immediate revocation. The two most-cited risks are inaccuracy at 74% and cybersecurity at 72%.

    ActGovern AI agents as privileged digital actors, not as ordinary software features.

  3. The AI subsidy era is ending

    As usage scales, cost becomes a material operating concern rather than a line item. AI's share of IT budgets is projected to rise from under 15% in 2025 to nearly 25% by 2027, and enterprise deployments have seen consumption increase by as much as 400%. Real-time cost visibility, workload routing and usage controls stop being optional.

    ActRoute routine workloads to smaller models, limit context and agent loops, and measure cost per outcome rather than tokens or licenses.

  4. No single model is best for every task

    Match the tool to the required outcome rather than issuing one standard and living with it everywhere. Reasoning, synthesis, cited research, multimodal analysis and code generation are different jobs with different leaders, and the leaders change between briefings.

    ActStandardize approved tools by use case, and combine specialists where the workflow requires it.

  • Microsoft, 2026 · share of AI impact from culture, management and talent versus the individual
  • McKinsey, 2026 · most-cited AI risks
  • IBM Institute for Business Value & Oxford Economics, 2026 · WSJ, May 2026

Share of total IT budget

  1. AI's share of IT budgets, 2025

    under 15%

  2. AI's share of IT budgets, 2027 projected

    nearly 25%

The distanceA projected rise of roughly ten points of the IT budget in two years, at the same time as consumption in enterprise deployments has increased by as much as 400%. Cost control becomes an architectural decision, not a procurement one.

Source: IBM Institute for Business Value & Oxford Economics, 2026; consumption figure from WSJ, May 2026. Both as cited in the Q3 2026 AWALI briefing.
Right tool, right application — an illustrative use-case map

Think and create

  • General reasoning assistantsReason through a problem and propose a solution.
  • Long-form synthesisWrite and consolidate across a large body of material.

Research and validate

  • Cited research toolsAnswers that carry their sources, so the claim can be checked.
  • Multimodal searchAnalysis that spans documents, images and structured data.

Build and automate

  • Workflow-embedded copilotsCapability inside the tools people already have open.
  • Agentic coding toolsBuild and automate rather than suggest and wait.

Example workflowResearch with a citing tool → synthesize with a long-form model → build with an agentic coding tool. Three specialists in sequence beat one generalist doing all three adequately.

Illustrative, as of Q3 2026. Named products are examples of a category, not endorsements, partnerships or performance claims — and a list like this dates quickly.

Structural watch-out

The CIO mandate has to expand

AI cannot sit solely inside traditional IT. The CEO sets the mandate, business leaders own the workflow and the P&L outcome, and the CIO co-leads the operating layer — data, integration, governance, security and scale.

70%

of transformation value comes from people, process and operating-model change — not from the technology alone.

BCG · AI Transformation Research

The Q3 mandate

Where to move now.

  1. Redesign, do not retrofit

    Rebuild two economically material workflows around human and AI execution this quarter, incorporating agents or bespoke tooling where they fit.

  2. Stand up the agent operating system

    Bounded agents with identity, permissions, evaluations and accountable owners — before scaling autonomy, not after.

  3. Fix your data

    Assume the data foundation is not AI-ready until proven otherwise, and define measurable readiness improvements to deliver this quarter.

  4. Fund owned intelligence

    Turn institutional knowledge into reusable context, decision logic and evaluations that improve with every engagement.

Sources & method

PwC
2026 AI Performance Study. Economic-gain concentration and business-model redesign.
McKinsey
State of AI 2025–26. Adoption against EBIT impact, and most-cited AI risks.
Precisely & Drexel LeBow
Data Integrity Research. Share of organizations with AI-ready data.
Microsoft
2026 Work Trend Index. Organizational versus individual drivers of AI impact.
BCG
AI Transformation Research. Share of transformation value from non-technology change.
IBM Institute for Business Value & Oxford Economics
2026. Projected share of IT budgets attributable to AI.
The Wall Street Journal
May 2026. Consumption growth in enterprise AI deployments.

Figures are reproduced as stated in the AWALI Q3 2026 briefing and attributed to the studies named above. Where a shorthand would have widened a finding, the narrower original wording is used. AWALI has not independently reproduced these studies.