Stop the AI Pilot Spiral

Activating AI with Architecture 

By Hugh Evans, Co-founder & CEO

Every organisation we work with is grappling with generative AI. A year ago, the question was whether to experiment; now it's how to stop those experiments from sprawling into a mess of stranded pilots. A recent McKinsey survey finds that 71% of organisations now regularly use generative AI in at least one business function, up from 65% at the start of 2024, and from just one-third around early 2023. This mass adoption is creating a new problem: if everyone uses the same capabilities from the same vendors, there's no competitive advantage. To succeed, organisations must treat AI as a discipline rather than a distraction, and build the muscles to absorb it safely and repeatedly. 

 
Start AI with your people, not pipelines 

AI momentum doesn't come from an idea funnel. It comes from pairing the person who feels the pain with a hands-on builder who can ship a fix fast. Drop them into the live workflow, roll out the thinnest viable solution, and track minutes saved and errors avoided. 

From day one, bake in guardrails, including tested prompts, privacy checks, cost limits, and rollback plans. Each experiment should be quick and designed to prove or disprove assumptions, with a clear call at the end: scale it, tweak it, or stop it. The wins get industrialised with observability, access controls, and run-cost budgets, then turned into reusable assets others can build on. Small, visible victories flip scepticism into a momentum push, turning AI from a pilot graveyard into a compounding engine of outcomes. 

 
Not all AI is created equal 

We think of enterprise AI in three layers, starting from the foundation. At the base is training proprietary models from scratch, which is a capability rarely needed by most companies but fundamental when pursued. Above that sits bespoke, engineered AI, such as customised retrieval-augmented generation (RAG) or predictive models tuned to your data. At the top are vendor-delivered utilities and "agentic frameworks" - embedded assistants in platforms like Salesforce or ServiceNow. Each layer carries different risks, costs, and governance needs. Inference costs can balloon fast if prompts and usage aren't controlled. Pilots accumulate bills without providing any value. These layers should be governed collectively to avoid duplication and to control cost. 

 
Two tracks: disruption and uplift 

In conversations with senior executives, we encourage them to split their AI investments into two tracks. "AI for disruption" sits in an innovation hub led by a business GM. It's there to explore new revenue models and to test business-model breakthroughs. "AI for capability uplift" sits in the operating business and seeks incremental improvement in productivity, customer experience or compliance. The tools may be similar, but the stakes, time horizons and governance are different. When executives conflate these tracks, they end up applying start-up mentalities to critical processes—or, conversely, subjecting innovation efforts to heavy procedural drag. 

Your AI is only as good as its data 

Regardless of the specific AI use cases an organisation chooses to prioritise, trust in the underlying data is non-negotiable. AI systems are inherently dependent on the quality, completeness, and integrity of their inputs. Without trusted data, even the most sophisticated AI models risk producing unreliable, biased, or misleading outcomes. We have seen that a frequent bottleneck in scaling AI adoption is not model capability but data readiness. While it is relatively straightforward to prototype AI solutions using available datasets, moving those prototypes into production requires data that is clean, consistent, and contextually relevant to the business problem. For example, applying AI to finance processes demands an exceptional level of precision; both input and output data must meet accuracy thresholds approaching 100%. Even minor errors are not acceptable in the context of major financial decisions, reporting, and compliance activities. 
 

These challenges emphasise the criticality of robust data governance frameworks, data lineage, and transparency, as well as automated data quality monitoring, as part of any truly production-grade AI system. 

 
The role of the forward-deployed architect 

In this model, architecture is no longer a bureaucratic bottleneck; instead, it's a partner embedded in the business. We advocate a two-layer structure. A portfolio-level leader with a deep data/AI background looks across all solution design and execution to ensure coherence. Underneath, a forward-deployed architect collaborates with business teams to develop prototypes, assemble pipelines, and iterate quickly. 

This role is about more than technical delivery. Forward-deployed architects are context engineers: they curate the right knowledge, chunk it effectively, and embed guardrails so AI solutions are both usable and trustworthy. Their job is to create ‘bankable’ prototypes with the correct information, in the right shape, then hand them to delivery teams. Architects, with their enterprise mindset, are uniquely placed to do this if they develop hands-on prototyping skills, strong stakeholder engagement, and a deep grasp of the business context. 

AI for Disruption

Owner
Innovation hub leader + Forward‑Deployed Architect
Purpose
Prove new revenue models and net‑new product value
Horizon
6–18 months; high uncertainty
Funding
Seeded portfolio bets with capped runway
Governance
Lean gates; bounded data/safety; kill or scale fast; decisions via design authority
Architecture's Job
Reference architectures for agents, security patterns (e.g., MCP auth), compliant sandboxes (e.g., Dev AI)
Metrics
Time‑to‑first prototype; experiments to decisions; pattern maturity; qualified opportunity value
Exit Criteria
Evidence of demand, viable unit model, security/compliance pass
Examples
Agent‑led services; net‑new data products; experimental features built in a controlled environment

AI for Customer, Product and Ops

Owner
Operating business leader + Platform/Architecture
Purpose
Lift productivity, CX, risk and compliance in core workflows
Horizon
This quarter and next; low failure tolerance
Funding
Opex/BAU with explicit run‑cost budgets
Governance
Formal controls; change management; budget discipline; governance as code
Architecture's Job
Standards, templates, prompt/policy packs, cost guardrails; integrated planning cadence
Metrics
Minutes saved; error rate; cycle time; satisfaction; adoption & run‑cost per task
Exit Criteria
Target adoption, quality thresholds, payback within period
Examples
RAG for policy mgmt; Copilot/Cursor for dev; QA test generation

The knowledge engine and corporate memory 

The most undervalued asset in AI is corporate memory: the lineage of decisions, reasons, and tacit knowledge that lives in documents and people's heads. Generative AI systems need this memory to provide answers you can trust. We call the solution a "knowledge engine"—an expert-in-the-loop process that curates and sequences organisational knowledge so AI can answer with accuracy, quality, and creativity. 
 

This is where architects step in. Their role isn't just technical plumbing; it's context engineering - deciding what knowledge matters, how it should be chunked, and how it flows into AI systems with the proper guardrails. By shaping context as well as code, architects turn scattered corporate memory into a reliable knowledge engine that the business can build on. 

Consider Morgan Stanley. Its chatbot for 16,000 financial advisors draws on more than 100,000 internal documents. But the real innovation is governance: it's trained as much on what not to answer as on what it can. Behind that discipline is architectural thinking - designing the boundaries of context, embedding oversight, and ensuring the AI can serve with confidence. 
 

Or take Duolingo Max. Its AI features, 'Explain My Answer' and 'Roleplay', give personalised feedback - but only after curriculum experts refine scenarios and review AI content. Here, context engineering plays out as careful scenario design and human-in-the-loop review, ensuring tone and accuracy stay aligned with learning outcomes. The lesson is that guardrails and architects working on context are what make AI valuable, not raw technology on its own. 
 

Memory, world models and live systems 

These examples underscore another of our themes: AI systems are "live systems." They must be updated, tuned and retrained as the business and its data evolve. That's why we expect consultants and knowledge workers to become "über-premium." Generic models are rarely sufficient; you need domain-specific knowledge and context to maintain an organisational "world model." Without it, generative systems can hallucinate or embed biases. Modern AI knowledge systems must capture your history, decisions and context - and keep them current. 

 
Navigating cost and risk 

All of this doesn't come cheaply. Compute costs for generative AI are substantial, and the hidden costs of ingestion, tagging and governance of unstructured data are often overlooked. Leaders need to consider both the capital costs, such as data-centre expansion, and the operational costs of running models. That's why a disciplined pipeline and portfolio approach is critical. It allows leaders to make conscious trade-offs about where to invest, to monitor spend and value, and to avoid the trap of scattering resources across too many vendor-delivered tools. 

 
The road ahead 

Generative AI is moving fast, but speed isn't the advantage. Discipline is. The winners won't be those chasing every new tool. They'll be the ones building a living system: disciplined pipelines, forward-deployed architects, and domain-specific knowledge engines that stay current. Competitive edge comes from coherence, not chaos. AI won't hand you an advantage. An activated AI architecture will. 


Thinking about how to structure your AI architecture? Let’s talk.

Next
Next

Experience is the strategy