Intent engineering: the missing layer in your AI strategy

An agent meant to run autonomously over days and weeks needs more than a prompt and good context. It needs to know what the organisation actually wants. When it does not, it will optimise for what is measurable at the expense of what is meaningful. That is where intent engineering comes in.

Two disciplines have matured. One is missing.

The last three years of enterprise AI have produced two meaningful disciplines. Prompt engineering came first: individual, synchronous, session-based. You sit in front of a chat window, craft an instruction, iterate the output. It is a personal skill with personal value.

Context engineering followed. Anthropic defined it in September 2025 as the shift from crafting isolated instructions to designing the entire information state an AI system operates within. RAG pipelines, MCP servers, structured knowledge bases. It is where the industry is now, and it is genuinely important work.

But neither discipline solves the underlying problem. Context engineering tells agents what to know. What comes next, intent engineering, tells agents what to want. It is the discipline of making organisational purpose machine-readable and machine-actionable: not as prose in a system prompt, but as structured parameters that shape how agents make decisions when running autonomously over days, weeks, or months.

What intent engineering actually means

OKRs were designed for people. They encode human-readable goals and assume human judgment about trade-offs, priorities, and exceptions. A manager can tell a direct report what matters this quarter and trust that the report will interpret that guidance through months of institutional context, professional norms, and tacit knowledge built over time.

Agents have none of that. An agent does not know your company's OKRs unless you put them in the context window. It does not know which trade-offs your leadership team prefers unless you encode those preferences in a way it can act on. It does not know when to escalate versus when to decide autonomously unless you define that boundary explicitly.

When a human joins a company, alignment happens through a hundred informal mechanisms: the wiki, the onboarding, Slack conversations, watching senior people handle ambiguous situations. None of that works for agents. Agents need explicit alignment, and they need it before they start working, not six months later.

This means organisations need to produce something most have never had to create: machine-readable expressions of organisational intent. Not "increase customer satisfaction", that is a human-readable aspiration. An agent needs a structured definition: which signals indicate customer satisfaction in this specific context, which data sources contain those signals, which actions the agent is authorised to take, which trade-offs it can make, and where the hard limits are.

The three layers no one has built

Intent engineering operates across three infrastructure layers, each at a different altitude.

The first is unified context infrastructure, the systems that make organisational knowledge accessible to agents. Most teams build this ad hoc: one team pipes Slack through a custom RAG pipeline, another exports Google Docs into a vector store, a third built an MCP server connecting to Salesforce but not Jira, and a fourth does not know the other three exist. The Model Context Protocol, which Anthropic introduced in late 2024 and donated to the Linux Foundation in 2025, is the most promising attempt at standardisation, but protocol adoption is not organisational implementation. Which systems become agent-accessible? Who decides what context an agent can see across departments? How do you version knowledge so agents are not operating on stale information? A standard does not answer those questions.

The second layer is workflow alignment: a shared understanding of which processes are agent-ready, which require a human in the loop, and which remain human-only. Right now, every employee rolls their own AI workflow. One uses Claude for research and another tool for drafting. A third built a custom agent chain. A fourth pastes text into a chat window. None of them can describe their workflow in a way that is transferable or improvable. The difference between individual AI use and organisational AI leverage is the difference between one good hire and a system that makes everyone better.

The third layer is intent encoding proper: translating human-readable organisational objectives into agent-actionable parameters. Decision boundaries. Escalation thresholds. Value hierarchies for resolving trade-offs. Feedback loops that measure alignment drift over time. This is what almost no company has built. It is also what determines whether a long-running agent operates in line with what the organisation actually needs, or optimises its way into a public relations problem.

The race is no longer about the model

Deloitte's 2026 state of AI in enterprise survey found that 84% of companies have not redesigned jobs around AI capabilities and only 21% have a mature model for agent governance. These are not technology gaps. They are intent gaps.

Meanwhile, 74% of companies globally report they have yet to see tangible value from AI. McKinsey found 30% of AI pilots failed to reach scaled impact. The investment numbers say something different: 57% of digital transformation budgets flowing into AI automation, averaging $700 million for companies with $13 billion in revenue. Large investment, mixed outcomes. The pattern is not that the models are weak. The models are not the bottleneck. A mediocre model with clear, structured, goal-aligned intent infrastructure will outperform a frontier model in an organisation with fragmented, inaccessible, unaligned knowledge every time.

The AI race has been framed as an intelligence race for three years, who has the best model, the largest context window, the highest benchmark score. That framing made sense when models were the bottleneck. They are not the bottleneck any more. The race is now an intent race.

What this means in practice

The companies that get this right will treat context infrastructure the way they treated data warehouse strategy: as a core strategic investment, not an IT project. They will build a living organisational AI capability map, a shared understanding of which workflows are agent-ready, which are agent-augmented with human oversight, and which remain human-only. And they will develop goal translation infrastructure that converts human-readable organisational objectives into agent-actionable parameters.

For back office operations, the implication is direct. An AI agent running procurement, order management, or customer service without intent infrastructure will inevitably optimise for what is easy to measure and lose what is hard to measure. Resolution time goes down. Relationship cost goes up.

Building that intent layer is the work that determines whether AI automation delivers genuine value, or simply moves cost from one column to another. The models work. The context pipelines are getting better. What remains is the organisational infrastructure that connects AI capability to organisational purpose.

That is not an engineering problem. It is a strategy problem that requires engineering to implement. And it is the gap that separates organisations running AI tools from organisations where AI operates as genuine infrastructure.