Which way are you facing?

The question

Ask a professional services firm about AI in 2026 and you will hear a conversation orientated to speed.

Are we moving fast enough? Are competitors ahead? Which tool should we roll out?

The anxiety is real and the questions are legitimate. But they share a hidden assumption: that the direction of travel is already settled, and the only live question is the pace.

A navigator works the other way around.

Before deciding how fast to sail, she decides which way she is facing. A fast boat pointed at the wrong star only gets to nowhere sooner than everyone else.

Not deciding is still deciding

Most organisations have not decided anything about AI. They are still trialling, working with pilots, and setting up steering committees.

The numbers say they are moving anyway.

The Thomson Reuters Institute's survey of professional services (February 2026) found organisation-wide AI use almost doubled in a year, from 22% to 40% — while only 18% of respondents could say their organisation was tracking whether the tools returned anything.

Use is climbing steeply; strategy and measurement are not. That equals acceleration without instruments: no compass, no log, no fix on a star. Just blind hope that speed equals progress.

And the sanctioned numbers are not where most of the use is. Reporting on MIT research in 2025, Fortune recorded workers at more than 90% of companies using personal AI accounts for daily tasks — frequently without approval — while well under half of those companies held an official subscription.

The gap between those figures is where most professional AI use actually happens: in private accounts, on terms nobody senior has read.

So a firm that has taken no position has not abstained. It has already adopted — quietly, ungoverned, and on the worst available terms — by letting its least-considered behaviour become its de facto policy.

The sailing word for that is drift. Drift is a bearing no one would choose if it were put to a vote.

Understanding comes first

You cannot take a sensible position on something you do not understand.

That sounds obvious; it is also, increasingly, the regulators' position.

The Bar Council of England and Wales, updating its AI guidance in November 2025, put it this way: "The best-placed barristers will be those that make the effort to understand these systems and, if appropriate, use them as tools in their practice, while maintaining control and integrity in their use."

The New Zealand Law Society's guidance (February 2026) starts from the same place — a lawyer is not absolved of responsibility for work because AI produced it, which means you must understand the tool well enough to stand behind what it gives you.

Different regulators, same structure: comprehend before you deploy.

That is the premise this site is built on. Not "adopt AI" or "avoid AI" — understand it, well enough that whatever position you take is one you could defend.

Four legitimate bearings

Once you understand the tools, a genuine choice opens up. It is much wider than the adopt-or-fall-behind binary the vendor-driven sales material offers.

There are four defensible bearings:

• Hold — no AI yet, on purpose. The risk is not worth it for your kind of work, or the tools are not yet trustworthy enough for it.
• Restrict — a default no, with narrow, supervised exceptions.
• Adopt with care — a default yes, inside clear limits, with verification built in.
• Lean in — wholehearted, fast, and deliberate.

Each is right for someone.

Principled restraint, in particular, is not a heresy: regulators now expressly sanction graded caution, and for some work effectively require it. That restraint is not free — the efficiency forgone is real — but it is a selected position, not a failure.

The only indefensible bearing is the unchosen one.

A firm that holds back on purpose and a firm that leans in on purpose have both done the work. They have examined the tools and selected a course that works best for them.

The firm that never decided is the one carrying the full downside of adoption — the exposure, the ungoverned use, the unread terms — with none of the upside of having thought about it.

Where to from here

Orientation, in practice, is three steps. This module is built in the same order.

1. Understand how these tools work. What a language model is, why it can be confident and wrong, and what that means for how far to trust it. Start with part one of the series below.
2. See where they can help, and where they can fail. Task by task, with the failure modes stated as plainly as the strengths: where they help.
3. Know what is safe to put in. Tool by tool, plan by plan, kind of information by kind of information. That is our companion module's job: ask the question directly, or read the guidance on confidentiality, privilege, privacy and the professional rules.

Do those three things and you will be better oriented than most, whichever bearing you then choose. The choice stays yours.

The paper behind this page

This essay is the short form of a working paper, Orientation before velocity, which gathers the research — the adoption and ROI evidence, the organisational psychology, the regulatory guidance, and the strategic theory the doctrine rests on — and is candid about where the evidence is thin. If you want the working shown, read the full paper.