Almost every consultancy now sells AI. The pitches blur together: transformation, copilots, agents, efficiency. Choosing the right partner is less about the deck and more about a handful of questions that reveal whether a firm can actually ship — and operate — something that works.

Start with the outcome, not the technology

The strongest engagements begin with a business outcome, not a technology. Before you evaluate any partner, get specific about what you want to be true in six months: a workflow that runs with less manual effort, a pilot finally in production, a governance regime you can show an auditor. A good consultancy will push you toward that clarity; a weak one will lead with its toolset. If the conversation is all models and no measurable outcome, that is a signal.

Ten questions to ask any AI consultancy

Use these in the first one or two conversations. The quality of the answers matters more than the slides:

  • Who actually does the work? Principals, or juniors after the sale?
  • What is your production track record? Real systems under real load, not demos.
  • How do you build in human oversight? Where does a person approve, and why?
  • What does your audit trail capture? Can you explain any action after the fact?
  • How do you measure ROI? Before the build and after launch.
  • Which models do you use, and why? Look for model-agnostic reasoning, not vendor loyalty.
  • How do you handle our data? Privacy, isolation, and whether providers train on it.
  • What happens when an agent misbehaves? Pause, rollback, incident process.
  • What do we own at the end? Source, runbooks, documentation — and any lock-in.
  • How do we start small? A fixed-scope first step beats a sprawling programme.

Red flags

A few patterns reliably predict disappointment:

  • Full autonomy as a selling point. In production, bounded beats autonomous. Uncontrolled autonomy inside your systems is risk, not value.
  • Guaranteed compliance. No one can guarantee regulatory outcomes; be wary of anyone who claims to.
  • Hours over outcomes. If the model is billable headcount with no outcome attached, incentives are misaligned.
  • No exit. If you cannot run what they build without them, you have bought a dependency.

Generalists, specialists and "AI experts"

"AI experts" is an easy label to claim. The useful distinction is between firms that advise on AI and firms that have built and operated it. For agentic AI consulting in particular, you want a partner whose senior people have shipped governed agents — because the hard problems live in operations, not in the model. A focused specialist with genuine production scars will usually serve you better than a generalist with a broad menu.

How Mach Lilies fits

We are a small, senior, founder-led practice: the principals who scope your work are the ones who build, govern and operate it. We sell governed AI operations, not billable hours; we are model-agnostic; and we hand over clean, documented systems you own. If that is the profile you are looking for, the clearest way to test it is a focused Agentic Operations Sprint — or simply read how we think about AI consulting.