Read time: 4 minutes

Leader, welcome back.

Anthropic changed something about Claude this week that reads like housekeeping and is actually a governance question.

Also inside: why "how can we automate this" is usually too early a question, and a research brief you can point at five different problems.

Anthropic updated Claude's memory so that chat and cloud-based Cowork now draw on the same underlying memory. Context you build in one place carries into the other.

Claude also updates memory while you work, rather than only summarising conversations afterwards. Whatever it has kept appears as topics in Settings, where you can read, edit or delete each one.

Two details worth knowing.
On Team and Enterprise plans this is off by default and an owner has to enable it. And Cowork only uses memory when it runs in the cloud, not locally.

For finance teams + wider, this matters more than it first looks.

A useful assistant is not just one that reasons well.
It is one that holds enough context to produce work that actually fits your organisation.

That could mean remembering the line between statutory and management reporting, what counts as an acceptable EBITDA adjustment, which KPIs your board actually looks at, or the threshold below which you stop investigating a variance.

Productivity upside is obvious. The governance question is less obvious.

If a system retains context over time, you need to know what it has stored, whether it is still accurate, and whether it should be there at all.

A bad assumption inside one chat is annoying.
A bad assumption that persists across every future piece of work is something else.

Which is why I think persistent context ends up inside the AI governance conversation rather than next to it. This is closer to standing data or system configuration than it is to a feature. Standing data gets reviewed. So should this.

The wider point: model intelligence is a weakening differentiator. The frontier models will keep leapfrogging each other every few months.

What gets genuinely hard to replace is the layer around them. Your documents, your prior decisions, your terminology, your workflows.

Switching provider today means changing a subscription. Switching in two years may mean abandoning a context layer that spent those two years learning how your organisation works.

Go and look at your memory settings this week. Read what is actually in there.

The most common mistake I see in AI projects is starting with "how can we automate this?"

That question is usually too early.

Processes accumulate unnecessary approvals, duplicate data entry, outdated controls, reconciliation that exists only because something upstream is broken, and reports nobody reads.
Automate before you challenge any of that and you preserve all of it, permanently, at speed.

The order I use is audit, optimise, automate.

Start by watching the process as it actually happens, not as the process document describes it.
Record the handoffs, the spreadsheets, the workarounds, the judgement calls.

Then challenge every step.
Can it be removed?
Can two steps become one?
Can a data quality problem be fixed at source instead of reconciled downstream every month?
Can the report stop being produced?

Only what survives that gets automated.

Monthly variance commentary is the perfect example.
It looks like a writing problem, so the obvious move is to have AI draft the commentary.

Look properly and most of the effort is somewhere else entirely: pulling data out of four reports, chasing immaterial movements, and reconciling definitions that two systems disagree on.

The better answer is to fix the extraction, set a sensible materiality threshold, standardise the definitions, and then point AI at what is left.
Same tool, roughly ten times the return.

Here is the structure I use for research-heavy work.

The example is competitor intelligence, but the shape of the brief is the reusable part.

For finance teams, add one line: identify the likely implications for revenue, margin, headcount, capex, opex or competitive economics.

Now change the role and the subject and the same brief does a different job.

The same shape handles vendor selection, regulatory monitoring and market research. Swap the role, swap the evidence you care about, keep the structure.

The principle underneath it: give the model a role, define the scope, specify what evidence counts, and say how the answer needs to support a decision.
That beats "research this for me" every time.

It is also the point of the RACEF framework. It is not about longer prompts.
It is about giving the model enough structure to understand the job it has been given.