
Beyond AI pilots: what it actually takes to build an AI-centric operation
Over the past two years, most mid-market and enterprise organisations have run the same experiment: buy seat licences for a general-purpose AI assistant, fund two or three departmental proofs of concept, and wait for productivity gains to appear in the management accounts. Almost everywhere, the result has been the same. Individual staff draft emails or code snippets faster, while end-to-end cycle times, unit costs and EBITDA margins barely move.
Recent research from McKinsey's Technology, Media and Telecommunications practice on The AI-centric imperative highlights why. In its survey of software and technology executives, organisations that rebuilt their products and core operations around AI reported 20 to 40 per cent reductions in operating costs and 12 to 14 percentage-point improvements in EBITDA margins. Meanwhile, teams that reported 30 to 50 per cent productivity gains in isolated developer or departmental pilots rarely saw those gains reach the bottom line.
This perspective is for CEOs, CTOs, COOs and private-equity operating partners. It sets out why bolting AI onto existing workflows stalls, the three architectural patterns replacing conventional software, and the five engineering foundations required to run AI reliably in production.
Why individual productivity does not automatically become operational impact
A workflow moves at the speed of its slowest handover, its dirtiest data feed and its most bottlenecked approval queue. If an analyst uses a language model to summarise a case in two minutes instead of twenty, but that case still waits four days in a shared inbox for compliance sign-off and has to be re-keyed into two legacy systems by hand, the operation has not become faster.
Converting individual speed into organisational capacity requires redesigning the workflow end to end: how intake is structured, where evidence is pulled from, which steps an agent can prepare, where a human reviewer must sign off and how the decision is written back to the system of record. That is an engineering and operating-model change, not a software licence roll-out.
Three archetypes for AI-enabled software and workflows
Across both software vendors and internal enterprise platforms, McKinsey identifies three distinct patterns for how agents are being built into systems. When we design AI and Decision Systems and Governed Workflow Platforms with clients, choosing the right pattern is one of the first architectural decisions we make.
| Archetype | How it works | Where value sits | Best suited for |
| --- | --- | --- | --- |
| 1. Agents as augmentation | Agents act inside an existing system of record, automating repetitive steps such as document extraction, form preparation or case routing within the current process | Access to underlying system data and reduction of manual handling time | High-volume intake, triage and document-heavy operations built around established core systems |
| 2. Agent-centric architecture | Operators work through a unified workflow or agent interface that orchestrates multiple back-end services, APIs and data stores directly | Workflow experience, tool orchestration and speed across previously fragmented systems | Cross-system operational processes where staff currently copy and paste across four or five tools |
| 3. Domain-expert systems | Systems shaped around deep vertical rules, proprietary historical data and specialist review standards (for example, legal matter analysis, lending underwriting or property compliance) | Proprietary domain data, calibrated evaluation benchmarks and encoded expert judgement | Regulated or specialist services where generic models fail on edge cases and auditability is mandatory |
In practice, the most durable moat is no longer a list of user-interface features. Features are now cheap to replicate. Defensibility comes from proprietary operational data, the domain rules encoded into the workflow, and the feedback loop captured every time a qualified reviewer accepts, edits or overrides an output.
From buying seats to buying governed outcomes
As software shifts from passive screens to systems that complete work, the economics of buying and building software are changing in two ways:
- Unit economics move from seats to completed tasks. When an automated workflow handles routine triage or document preparation, fewer people need to log into a screen for hours each day, while compute and inference costs scale with volume. Whether you are building an internal platform or selling a digital product, the metric that matters is the total cost per completed, verified case.
- Service-as-software replaces generic tooling. One-size-fits-all horizontal tools struggle when a process depends on sector-specific regulation and local data quirks. Organisations are moving toward tailored platforms combined with forward-deployed engineers who wire the data, configure the guardrails and stay accountable for the operational metric.
Redesigning delivery: smaller, senior teams across the full lifecycle
Adding coding assistants to an unchanged, handoff-heavy development process only moves the bottleneck from writing code to reviewing pull requests, testing integrations and fixing broken deployments.
An AI-centric engineering lifecycle changes the process itself:
- Discovery and scoping on real data: using structured diagnostics and rapid vertical slices to test feasibility against live records in week one rather than after months of specification writing.
- Specification-driven build and automated evaluation: pairing agentic development with strict type boundaries, deterministic tests and a domain golden dataset so every code or prompt change is scored before release.
- Compact, multi-disciplinary teams: replacing large, layered delivery pyramids with small teams of senior engineers who cover data, software, AI and cloud architecture together and work directly with domain operators.
The five-part infrastructure foundation for production AI
McKinsey estimates that for most established organisations, roughly 30 per cent of shared platform capabilities need meaningful renovation to support agentic systems at scale, while 80 per cent of executives expect cloud and inference spend to rise by at least 10 per cent (and over a third expect increases above 20 per cent).
Plugging an API key into a prototype takes an afternoon. Running agentic workflows in production requires five infrastructure layers working together:
1. A governed data and context layer
Models cannot reason reliably over contradictory spreadsheets or unindexed document folders. Production systems need maintained pipelines, resolved entity identifiers and retrieval that respects source document versions and user permissions at query time.
2. Permissions and governance in code
Governance cannot live in a PDF policy beside the system. Role-based access control, human review thresholds, source citations and an immutable audit trail of who approved each consequential decision must be enforced by the application platform itself.
3. Security for untrusted inputs and tool calls
When a system reads external documents or emails and has permission to call internal tools, prompt injection and over-privileged service accounts become operational risks. Every tool call needs least-privilege scoping, schema validation and separation between untrusted text and system instructions.
4. Evaluation harnesses and agent operations
Before any workflow goes live, it needs an automated evaluation harness built from real historical cases, including messy scans, missing fields and cases the system must refuse. In production, orchestration must track step-level failures, retries and reviewer override rates so drift is caught before it harms decisions.
5. FinOps and model portability
Token and infrastructure consumption must be visible per workflow, per team and per completed case. Because the cost of reaching a fixed capability level is falling rapidly while reasoning models can consume more tokens per run, putting models behind a clean internal interface lets you route simple classification steps to smaller models and switch providers with an evaluation run rather than a rewrite. Our perspective on why falling AI costs make model portability a priority covers this discipline in detail.
Where to start without boiling the ocean
Becoming AI-centric does not mean launching a three-year infrastructure programme before delivering any value. The organisations making durable progress move in a disciplined sequence:
- Pick one consequential, end-to-end operational route where delays, manual document handling or inconsistent decisions carry a clear commercial cost.
- Embed senior forward-deployed engineers alongside the operators to build the first governed workflow and the five platform foundations required to carry that route safely.
- Measure impact in operational terms (cycle time, throughput, error rate, cost per case and reviewer overrides), then reuse the same data, evaluation and permission foundations for the next workflow.
If you are ready to move past isolated AI pilots and build production capability your team can own, explore our work in Applied AI Engineering and Governed Workflow Platforms, or talk to us.
Sources
- McKinsey & Company, The AI-centric imperative: Navigating the next software frontier, Jeremy Schneider, Joshan Abraham, Matt Linderman, Naveen Sastry, Mónica Perestrelo and Shayan Salam, October 2025.
- McKinsey & Company, What is a forward-deployed engineer?, Holger Hürtgen, Martin Harrysson and Nick Robson, October 2026.
Survey and benchmark figures are quoted from McKinsey & Company as published.




