AI Ambition Is High. The Scope Is Still Too Vague.

Most companies are past the "should we use AI?" conversation. Budgets are approved, board decks mention agents, and someone senior has been told to "do something with AI this quarter." The pressure to move is real.

And yet the projects stall — not because ambition is missing, but because nobody can answer the next question with any precision: what, exactly, are we building first, and how do we know it worked?

The problem is not lack of ambition. The problem is lack of scoped execution. Ambition tells you the destination. Scope tells you the first ten steps, who takes them, and how you'll know you're on the right road. Skip that, and you don't get to AI value faster — you get to an expensive pilot that never leaves the demo.

The ambition is real. The scope is not.

The macro picture backs this up. In McKinsey's State of AI research, adoption is now widespread, but most organizations are still experimenting rather than capturing durable value — and the ones pulling ahead are the ones that redesign workflows around AI, not the ones that bolt a model onto an unchanged process.

BCG makes the same point more bluntly in Are You Generating Value from AI? The Widening Gap: the distance between AI leaders and everyone else is growing, and agentic AI is widening it further. The leaders aren't the ones with the most pilots. They're the ones who scoped a small number of workflows deeply and scaled what worked.

And the caution is warranted. One widely cited 2025 report — MIT's Project NANDA "GenAI Divide" — warned that a large share of enterprise GenAI pilots have yet to translate into sustained business impact. The headline number is easy to sensationalise; the useful lesson is narrower: pilots that aren't scoped against an owner, a metric, and a real workflow tend to die quietly, no matter how good the technology is.

So the gap isn't ambition versus caution. It's vague scope versus scoped execution.

What is a Minimal Viable Agent (MVA)?

The antidote to a vague scope is a small, provable one. That's the job of a Minimal Viable Agent.

A Minimal Viable Agent (MVA) is the smallest useful AI agent that proves one workflow end to end — the data it needs, the human review points, and the business value it creates — before you scale. One job. Done well enough to decide with evidence, not opinion.

An MVA is not a stripped-down version of your eventual platform, and it's not a throwaway demo. It's a deliberately narrow AI prototype designed to retire your biggest risk first. If the assumption that scares you most is "will the model be accurate enough on our messy data?", the MVA proves that. If it's "will anyone actually use this on a Tuesday morning?", the MVA proves that instead.

Agentic AI makes this discipline more important, not less. An agent that takes actions across systems has more ways to be wrong than a chatbot that answers a question — which means the workflow, the data access, and the human-in-the-loop checkpoints need to be scoped before you build, not discovered in production.

The six questions that turn ambition into scope

When ambition is high and scope is vague, teams get stuck on the same six questions. A good scope answers all of them on one page.

1. Which use case is worth building first?

Not the flashiest one — the one where a named owner's metric moves, the data already exists, and the workflow is small enough to prototype in weeks. This is AI use case prioritisation, and it's the single highest-leverage decision you'll make. A structured AI Scoping Workshop exists precisely to force this ranking before money is committed.

2. What should the Minimal Viable Agent actually look like?

Define the one workflow, the inputs, the outputs, and the moment a human stays in the loop. A Prototype Sprint then builds or coordinates that MVA to test the riskiest assumption — instead of over-building a v1 that didn't need to exist.

3. What must be validated — workflow, data, and human review?

Three things break most AI projects: the workflow doesn't match how people actually work, the data isn't as available or clean as assumed, and there's no clear point where a human reviews or overrides the agent. Scope names each one as an explicit test, which usually means some AI workflow redesign before a single model is trained.

4. Which partner type fits the scope?

FDE boutiques, AI agencies, product studios, automation specialists, freelancers, and large consultancies solve different problems. The right AI implementation partner depends on the scope and the binding constraint — speed, compliance, integration, or cost. Get the scope clear first; then independent partner selection is a short, evidence-based exercise instead of a guess. (If you're weighing embedded engineers specifically, our explainer on forward deployed engineering breaks down when that model fits.)

5. What budget is reasonable?

A first step should look like a scoping workshop in the low five figures and a focused prototype in the low-to-mid five figures — not a six-figure platform build before anything is proven. If a proposal skips straight to a large program, an independent AI proposal review will usually find the over-scoping.

6. How do we avoid overbuilding, vendor lock-in, and pilots that never scale?

Stay vendor-neutral until the data tells you otherwise, keep the first build small, and map the wider tooling landscape before committing — our AI Ecosystem map is there so you choose components on merit, not on whoever pitched first. The hidden cost of the wrong AI partner is rarely the invoice; it's the lock-in and the half-built thing nobody can throw away.

Scope now includes governance

In Europe especially, scope is no longer just "what works" — it's "what's defensible." The NIST AI Risk Management Framework and its Generative AI Profile frame AI governance as lifecycle work: evaluation, trustworthiness, and human oversight designed in from the start. The EU AI Act pushes the same direction — AI decisions increasingly need risk-aware scoping, not just experimentation.

You don't need a compliance department to act on this. You need a scope that names where a human reviews the agent, what data it touches, and how you'd explain a decision if asked. That's cheaper to design up front than to retrofit later.

Before you hire the partner, define the scope

This is the line worth keeping: before you hire the partner, define the scope. A vetted partner can build almost anything. What they can't do is tell you — independently — whether the thing is worth building, how small it should start, or whether their own proposal is over-scoped. That's the work that has to happen first, and it has to be independent of whoever will eventually deliver.

Get the scope right and the rest is procurement. Get it wrong and you're buying lottery tickets with a project plan attached.

Key takeaways

• Ambition is no longer the constraint. Vague scope is.
• A Minimal Viable Agent proves one workflow — data, human review, and value — before you scale.
• Answer the six questions (use case, MVA shape, what to validate, partner type, budget, anti-overbuild) on one page.
• Governance is now part of scope, not a later phase.
• Define the scope before you choose the AI implementation partner — independently.