We did not start by adopting AI

We did not start by adopting AI.

That sounds slightly unfashionable, but it is important. The most useful AI systems we have seen did not begin with a company-wide rush to add tools everywhere. They began with a closer look at how work already moved: where context lived, who owned decisions, what people reviewed, which sources mattered, and where handoff broke down.

AI became useful after the work had enough shape around it.

That is the lesson we now bring to professional teams. The question is not "which AI tool should we buy?" The better question is "which part of our work is important enough, repeated enough, and reviewable enough that better preparation would change the quality of the decision?"

The Tool-First Mistake

The common adoption pattern starts with access.

A team gives people a model, connects a document store, runs a few demonstrations, and waits for productivity to appear. Some value usually does appear. People draft faster. They summarise faster. They get help with blank pages and routine analysis.

But the deeper work remains unresolved.

• Which document is authoritative?
• Who owns the next step?
• Which decision needs review before action?
• What is missing from the evidence?
• Which client, matter, vendor, project, or relationship does this belong to?
• What is the system allowed to do with the answer?

If those questions are unclear, AI has to work around the organisation rather than inside it. It may produce fluent text, but it cannot reliably know what the text means in the operating context of the team.

This is why AI transformation is mostly operating work. A capable model is not enough. The surrounding workflow has to become legible.

Start With Work That Already Matters

The best starting point is rarely a grand AI strategy. It is a real workflow that already matters to the team.

For a legal team, that may be the weekly matter review. For finance, it may be renewal review. For a project team, it may be the site decision meeting. For an advisory practice, it may be client follow-up. For a research-heavy team, it may be the first pass through suppliers, precedents, policy, or market references.

These workflows are good starting points because they already have a rhythm. People already gather sources, prepare notes, review options, assign owners, and decide what happens next. AI can help by preparing the work around that rhythm.

The system does not need to replace the professional. It needs to make the professional better prepared.

The First Useful Output Is Often Not An Action

Many teams imagine AI value as action: send the email, approve the renewal, update the record, complete the task.

In professional work, the first useful output is often earlier than that. It is the first draft, the review packet, the source map, the open-question list, the comparison table, or the handoff note.

Those outputs matter because they are reviewable. A person can inspect them, correct them, and decide what should happen next.

For example, an agent preparing a client follow-up review might gather:

• recent notes and commitments;
• open tasks and promised dates;
• relevant documents and prior decisions;
• relationship context that affects tone;
• missing information that should block outreach;
• the owner responsible for the next step.

That is useful even if the agent never sends a client message. The value is not external autonomy. The value is that the team reaches a better-prepared review faster.

This is the practical meaning of the first draft. The draft is not the final outcome. It is the material that helps a responsible person think faster and with better evidence.

Make The Work Legible Before Expanding Autonomy

As AI becomes more capable, the temptation is to grant more autonomy quickly.

We think the better sequence is slower and more durable:

This sequence is not a refusal to automate. It is how automation earns trust.

NIST's AI Risk Management Framework treats AI risk as contextual and socio-technical. The point is not only model performance; it is how the system behaves in a real setting, with real users, real consequences, and real governance. That is exactly why professional AI should begin with bounded workflows, inspectable sources, and clear review paths.

What This Means For Professional Teams

The practical advice is simple: do not ask where AI can be sprinkled. Ask where preparation is painful.

Look for workflows where people repeatedly spend time reconstructing context:

• preparing a weekly matter or project review;
• checking a vendor renewal before spend approval;
• onboarding a new analyst into live client work;
• preparing a proposal from scattered prior material;
• reviewing open commitments across client relationships;
• comparing suppliers, materials, policies, or precedents.

These are good early AI workflows because the value is visible even before automation. The team can compare the prepared packet against the old way of working. It can see what was missing, what was useful, what should be corrected, and what could become repeatable.

That is the adoption pattern we trust: prepare real work, review it with responsible people, improve the system, then expand only where the evidence supports it.

What This Is Not

This is not an argument against AI tools. General AI tools are useful. Chat can help people think, draft, summarise, and explore.

The point is that professional adoption needs an operating layer around the tool. The system needs source access, permissions, owners, review paths, and boundaries. Without that layer, the user has to rebuild the organisation inside every prompt.

It is also not an argument for endless manual review. Some workflows will become stable enough for narrower automation. But they should get there by earning trust through repeated preparation and review, not by jumping directly from demonstration to delegation.