The right size of AI.
Most failed AI projects die of over-engineering, not under-ambition. A chat window bolted onto everything, an agent where a script would do. The real deliverable is judgment: how much AI does this problem actually need? Sometimes the answer is a full autonomous harness. Often it is twenty lines of Python and one model call.
The wrong default.
Take a nightly digest email. The 2026 reflex is to hand the whole job to an agent: let it query the systems, decide what matters, write the email, send it. Now you have non-deterministic collection, an audit problem, and a bill.
What the problem actually needs: a Python script collects the data (deterministic, testable, free), and one AI call summarises it, the only step where language actually helps. That is the whole system. It runs every night, and nobody has ever had to wonder what it did.
Right-sizing is the discipline. Everything else follows.
The ladder.
The terraces narrow as they climb. That is the point: most problems live on the broad lower slopes, and only a few earn the climb to the top.
-
Plain script
The job is deterministic and the rules are known, however complex the calculation gets. Sophistication is not a reason for AI; some of the most valuable systems here contain none at all.
the composite metric · case study A · the 250-branch reporting engine · case study D
-
One repeatable AI step
Deterministic collection, and a single packaged AI step for the one part that needs language. The nightly digest lives here: a script gathers pipeline history and delivery logs, and the AI only writes the morning email.
-
Multiple skills
Proven procedures packaged as versioned skills with triggers, guardrails, and checklists, then combined. Individual know-how becomes organisational capability.
-
Agentic workflows
Multiple agents dispatched with goals and loops: one decomposes the work, others build, research, and test, and the results converge. Orchestration, not a single prompt.
-
Full agent harness
An agentic operating system: autonomous agents with memory, safe environments to verify their own output, and learning loops that make every run smarter than the last.
AI helped build every system on this hill. The judgment is how much AI lives inside each one.
How an engagement runs.
Where a business sits on that ladder is a decision you earn, not guess. This is the process I bring in, built on operational sense from a decade of customer-facing work before I ever wrote code.
-
Find the leverage
Before any tool: where in this business can AI move a number that matters? Revenue, cost, hours, risk. The impact hypothesis comes first, so success can be measured against it later.
-
Understand the work
Shadow and interview the people who actually do it. Pain points, procedures, workarounds, the things nobody writes down. A decade of hospitality and client-facing operations means I read a workplace quickly.
-
Get the data ready
Gather, analyse, and organise what the business already has, then preprocess it so AI can actually use it. Most AI failure is data failure in disguise.
-
Prototype in a safe environment
Small tools against real workflows, sandboxed. Test the impact before anything touches production.
-
Evaluate with the users, then productionise
Review the results with the people who will live with the tool. Only then build the full version, staged into production.
-
Keep measuring, keep learning
Nothing ships perfect. The implementation is evaluated iteratively, and the AI, the process, and the business all learn from each cycle.
And at every step, the same non-negotiables: evidence before assertions, structured results over free text, repro-gated fixes, fail-closed defaults, staged cutovers. Every one of them is visible in the case studies.