Faster, not looser

A draft client report came back from an AI tool with a confident, specific statistic: "According to a 2024 industry study, 67% of mid-market firms…" It sounded authoritative. It read like the kind of detail a senior partner would insert to anchor an argument.

There was no such study. The figure was plausible and entirely invented.

Two minutes of checking kept a fabricated statistic out of a client document. The point is not that AI made up a number. That is known. It is the most discussed failure mode of these tools. The point is what it would have cost if no one looked. A footnote that did not exist. A client decision built on it. A reputation built over years, dented by an autocomplete.

This is the real challenge of adopting AI in a consulting practice:

• AI is fluent, not factual. Output that sounds like an answer is not an answer until it is checked.
• Confidence is uniform. Correct answers and invented ones arrive in the same tone. The reviewer has to know the difference.
• The default failure mode is polished generic. Without specific direction, AI produces competent, forgettable output that drifts off-brand.

We treat AI the way we treat any other powerful tool in the practice: useful in the right hands, with the right guardrails. Seven principles shape how we work with it day to day.

• Lock the standards first. Brand, voice, message, and quality bar are defined before AI touches anything. AI works within constraints. It does not set them.
• Humans hold judgment. AI proposes. We dispose. Every decision that affects a client, a deliverable, or our reputation stays with the person who has to answer for it.
• Verify, do not trust. Every output is a draft. The statistic gets checked. The translation goes to a native speaker. The proposal section gets read by someone who knows the client.
• Transparency with clients. Clients have a right to know when AI was used in work delivered to them, and where. We tell them. Not as a disclaimer buried in fine print, but as part of how we describe our method. A client who understands where the tool helped and where the judgment was ours can trust both.
• Sensitive data stays out of public tools. Client information, contracts, financials, and anything covered by confidentiality does not go into consumer AI applications. We use enterprise tools with appropriate data handling, or we do not use AI for that task. The convenience of a quick prompt is not worth a confidentiality breach.
• Use it where it adds. Skip it where it does not. Drafting, restructuring, ideation, summarising: yes. Strategic positioning, hard judgment calls, the things a client is actually paying us for: no.
• Keep an audit trail. We can explain why every decision was made. If a client asks, we have an answer. If we made a mistake, we can find where.

The discipline pays off in two directions.

• We are faster without being sloppier. AI handles the work it handles well: first drafts, options to react to, reformatting, summarising. We spend our time on the work that is actually ours.
• We bring the method to clients. Every principle above was tested in our own work before we recommended it to anyone. When we advise a client on AI adoption, we are describing an operating model we already run.

AI's value is not autonomy. It is leverage, and leverage only works when something is being levered against. The lever is the operating discipline. Standards locked first, judgment held by humans, every output verified, the use of it disclosed, sensitive data protected, the tool used where it adds and ignored where it does not.

That discipline is what we bring to clients adopting AI. It is also how we work, every day.