How processes enable AI at scale

The goal of this lesson is to understand why AI needs structured ways of working to scale — and how your process management practice becomes the foundation that makes enterprise AI reliable.

Key learnings

• AI can’t scale on capability alone: Pilots succeed because a small team compensates for unclear processes; at scale, AI exposes every inconsistency, undocumented workaround, and local variation across the organisation.
• Intent engineering is the missing layer: Context tells AI what to know; intent tells it what to want, when to act, and when to stop. Intent lives in decision logic, trade-offs, and escalation rules — not in strategy slides.
• A complete way of working is bigger than a process map: It includes roles and responsibilities, work instructions, case flows, systems and files, templates, and the trusted Q&A that reflects what the organisation actually cares about.
• The winning sequence is Process → Stability → AI: Define outcomes and trade-offs, build the way of working, establish governance, define human checkpoints — then add AI as the accelerator.
• Three technical enablers unlock AI at scale: RAG grounds AI responses in your governed content, MCP lets agents act inside your process graph, and Telemetry keeps intent honest as the business and its processes evolve.

Transcript

Welcome. In this lesson, we’ll look at one of the most important topics in business right now: why AI initiatives deliver impressive pilots but struggle to scale — and what business process management has to do with solving that problem.

By the end of this lesson, you’ll understand why AI needs more than data and context to act reliably, what “intent engineering” means in practice, and how the BPM work you’ve done becomes a strategic AI capability.

Something strange is happening in most organisations right now. Individually, people are adopting AI fast — drafting documents, summarising meetings, cleaning up data. These are real wins. But they’re personal wins. A company can have thousands of people using AI and still be unable to answer: do we actually work differently now? That’s the scalability challenge.

Research from McKinsey and BCG points to the same thing: pilots are easy. Production is hard. Not because the pilot team didn’t try, but because scale is where the organisation shows up — different teams interpreting the same words differently, processes that vary between sites, exceptions that are the rule. At that point, AI becomes a mirror. It reflects the way your company actually runs. And in many organisations, that reflection is blurred and inconsistent.

Here is the core problem. AI cannot scale in your organisation until it understands your ways of working. Not your organisational chart. Not your strategy deck. Not the values poster. I mean: how decisions are made in practice. Which trade-offs are acceptable. What good looks like. What gets escalated. What never gets done, even if it’s technically allowed. How work moves between roles and teams.

In most organisations, this knowledge exists — but it’s not explicit. It’s locked inside experienced employees’ heads, email threads, Teams chats, half-updated SOP folders on SharePoint, and 200-page process PDFs that no one trusts anymore.

When leaders say “we want to deploy AI at scale,” the question becomes: deploy it into what organisational reality, exactly? Into a set of shared, living best practices? Or into a collection of messy local habits, individual assumptions, and undocumented workarounds? The old IT saying holds: if you automate a mess, you get an automated mess. We saw it happen with RPA. AI is no different.

There’s a concept that captures what’s needed here: intent engineering. It was described well by AI practitioner Nate B. Jones in the context of a real case — where the Swedish fintech Klarna replaced nearly a thousand customer support employees with AI agents optimised for resolution time. The numbers looked great. Tickets were resolved faster. Costs dropped. Then customers complained. Turns out, the human tacit knowledge — experience, intuition, knowing when to spend three extra minutes — was missing. The AI had a prompt. It had context. It did not have intent.

Intent engineering is the discipline of making organisational purpose machine-readable and machine-actionable: goals, values, trade-offs, decision boundaries — in a form AI can actually act on. Most enterprise AI projects are solving for context: better data retrieval, more connectors, bigger context windows. All of that matters. But context tells the agent what to know. Intent tells it what to want — and when to stop.

The failure mode isn’t hallucination. It’s optimisation. AI optimised for speed while silently eroding the trust and relationship quality the business depended on. Intent doesn’t live in company objectives or strategy slides. Those were designed for people — they assume human judgment about priorities, trade-offs, and edge cases. Agents don’t fill in gaps. They fill in blanks.

Intent becomes real at the point of work: when a frontline employee decides whether to escalate; when a manager decides whether speed beats quality today; when a team bends a rule to protect a customer relationship; when someone chooses “do it right” over “do it now” — or the opposite. This is why process thinking becomes an AI capability, not a process excellence side project. Business processes are a big part of the answer — they give AI something it can follow: the sequence of activities, handovers, decision points, variants, escalation paths.

If you stop at process diagrams, you still leave a gap. Because in real organisations, people don’t execute diagrams. They execute ways of working. And a way of working is bigger than a process map. A complete way of working answers two questions: what are we trying to achieve here, and how do we actually do it when it’s unclear?

It includes business processes — but also: roles and responsibilities (who is accountable, who can decide, who must approve); job expectations and competencies (what good looks like in this role, and what judgment is expected); work instructions and SOPs (not just that an activity exists, but how it’s executed); case flows (what happens as a case moves through time, not just a static model); systems and files (where the work actually happens — the CRM, ERP, SharePoint, ticketing tool); templates and artifacts (the forms, checklists, contracts, and reports that make the process real); and the questions people keep asking — and the answers they trust. That last layer is often overlooked. It’s where intent shows up in everyday language.

For organisations that want to scale beyond the pilot, three technical capabilities matter.

RAG — making the process queryable. Retrieval-Augmented Generation grounds AI responses in your actual content — your process model, work instructions, approved Q&A, current SOPs — retrieved in real time rather than generated from generic training data. This is what makes AI responses specific to your organisation and explainable. When the source of truth is governed, the AI’s answers stay current as the process evolves.

MCP — letting AI agents act inside your processes. The Model Context Protocol exposes the process graph as a set of callable tools: retrieve a process, get the tasks for a role, create a change request, record a completion, flag an exception. In practice, an AI agent helping resolve an escalation doesn’t just surface information — it retrieves the escalation process, identifies the next step, surfaces the relevant work instruction, creates the approval task, and flags if the case is deviating from the standard path. All within the governed process structure, with a human still in control at every boundary that matters.

Telemetry — keeping intent honest. Intent isn’t static. Best practice evolves. Exceptions reveal gaps. Execution data (step durations, handovers, deviations, completion patterns) is the feedback mechanism that keeps intent current. When execution data shows that a specific approval step is consistently where cases stall, the process owner can see it, investigate it, and adjust both the process and the AI guidance accordingly.

If you’re responsible for process management — or ways of working more broadly — this changes your role in a practical way. You’re not just maintaining documentation. You’re building the foundation that AI will run on. The questions you were already asking — who owns this, how is this step executed, what happens at this decision point, what does a good outcome look like — are exactly the questions that make AI trustworthy. And the governance habits you’ve been building — monthly reviews, feedback loops, version control, living RACIs — are exactly what keeps AI from drifting as the business evolves. Process-first isn’t a constraint on AI adoption. It’s the precondition for AI that actually scales.

Pick one cross-functional process where AI assistance would create visible value. Map the as-is, make the trade-offs explicit at the decision points, assign ownership, write the work instructions at the critical steps, and define where humans must stay in the loop. Then — and only then — bring AI in as the accelerator. That’s intent engineering in practice. And it starts with everything you’ve already learned in this course.