Most AI projects don’t fail at the model level. They fail one step earlier, when a team discovers it has never written down what good work looks like. That makes AI readiness a clarity problem, and small teams are better placed to solve it than they think.
Until three years I had never written a technical scope of work, an SOW. If you haven’t either don’t panic. This is something that engineers lived by and most probably still do today. It is in their wheelhouse. As a communications consultant I focused on strategy documents, tenders, proposals and decks. The terminology is different but the objectives are a lot more simliar than you may think. With AI, you now have a scope of work partner.
However, how can you assess what work needs to be completed if you are not sure about the roadmap and the story that you are trying to tell. So there you go, back to the drawing board and starting with the narrative will help you to set the tone and pave the path towards success. I have seen too many projects build without knowing where they are going. Take the time to think about your story.
It is a remarkably common trap. We spend so much time debating which large language model to deploy, or how to secure our data pipelines, that we completely overlook the human foundation of the work itself.
When organizations try to automate a process, they usually start by mapping out the steps. But as we explored recently in our analysis of Stanford’s New TaskNet, a job is poorly described as a mere list of tasks. A task list tells you what is happening. It tells you nothing about how well it needs to happen.
The system, not the person
In my experience working with clients, most small teams have never actually written down what good work looks like from a users viewpoint.
For years, quality control has been managed through tribal knowledge. A junior employee drafts a report, a senior manager reviews it, applies their invisible, internalized rubric of good, and sends it back with edits. The system works because the human brain is highly adaptive. But AI is not an employee you can simply mentor through osmosis.
“You cannot automate a standard that only exists in someone’s head.”
When a team tries to plug an AI agent into a workflow without a documented standard of excellence, the project stalls. There is a skill to be harnessed here and it has become known as context engineering.
Context engineering is now an emerging discipline that has become critical for internal workflows. While prompt engineering teaches you how to issue commands to a model, context engineering is the rigorous work of structuring the environment around the prompt.
An AI cannot infer your company’s tone of voice, its historical preferences, or its strategic constraints out of thin air. By engineering the context, curating the exact guidelines, past examples, and boundaries the model needs to reference, you are effectively translating your team’s tribal knowledge into a machine-readable standard of excellence.
The evaluation bottleneck
This isn’t just an anecdotal observation. Research is consistently pointing to evaluation as the critical bottleneck in AI adoption.
Take the MIT NANDA research (recently highlighted by Virtualization Review), which underscores the complexities of measuring knowledge work. Or look at the AI evaluation frameworks from METR. In their updated February 2026 study, they highlight that the ability to autonomously evaluate the success of a task is one of the most significant hurdles in AI development.
Even at the enterprise level, the companies seeing the most dramatic shifts, such as Klarna, which reported massive AI-driven shifts in customer service and marketing workflows to Fortune last year, are succeeding because their outputs are highly measurable. They know what a successfully resolved customer ticket looks like.
If you don’t have that clarity, you don’t have an AI strategy. Use AI today as a writing test. How well do you know the beginning, middle and end of your story?
What it means for builders
This is where the narrative flips. We often assume that AI adoption favors massive enterprises with bottomless R&D budgets. But when it comes to the fundamental requirement of AI readiness, clarity, agility wins. If a five person team can write down what good looks like in an afternoon, a 50,000 person company needs a full programme.
For smaller teams and independent builders, defining excellence doesn’t require a six-month consulting engagement. It just requires sitting down and translating your intuition into a rubric.
Let me give some tangible examples or what you can look for. For a startup communications team, good means a press release with a succinct boilerplate, references from the founding team, containing zero passive voice, and always tied to the companies key features as outlined in the product roadmap.
Here is another one. For your paid media strategy good means creating X number of campaigns, referencing key product attributes, using the documented brand voice and leading to X amount of traffic or leads. Both of these examples replace subjective instruction and lead with concrete guidance and rules for the AI to follow.
Once you have that definition, everything changes. You can write better prompts. You can evaluate model outputs objectively. You can finally stop blaming the technology for failing to read your mind.
The future of work belongs to the teams with the sharpest clarity and understand the value of communication.





