AI at Work

The same tool, the same prompts, an entirely different set of consequences — and the one principle that explains all of them.

• Name what changes when output carries your employer's name
• Explain why accountability does not transfer to the tool
• Identify which of your tasks carry obligations you had not considered

Why unapproved use is so widespread, what it actually costs, and what to do when nobody has told you the rules.

• Explain why unapproved use happens and why hiding it makes it worse
• Find your organisation's position, or establish one where none exists
• Raise the subject constructively rather than confessing

A test that resolves most cases, six scenarios worked through, and wording you can actually use.

• Apply a single test to decide whether disclosure is warranted
• Work through the common workplace scenarios
• Write disclosure wording that is honest without being theatrical

Four questions that sort any proposed use into low, medium or high risk — and tell you which controls it needs.

• Triage a proposed use case with four questions
• Match controls to the risk tier rather than applying one standard
• Recognise the uses that should not proceed at all

What meaningful human review looks like as opposed to a signature, and the seven questions to ask any tool before it touches your information.

• Distinguish meaningful review from rubber-stamping
• Identify decisions that require a human to decide rather than approve
• Assess a vendor with seven questions

The blank page is the expensive part. Give the model your rough thinking and it becomes an editing job, which is far easier to judge.

• Turn rough notes into a usable first draft
• Explain why a blank-page request produces generic writing
• Edit a draft down instead of regenerating it

Getting through the pile without inheriting errors you never read — and the specific requests that make a summary defensible.

• Ask for summaries that expose omissions rather than hiding them
• Use a summary as a reading plan rather than a replacement
• Recognise which documents should not be summarised at all

Structuring a problem, arguing against yourself, and stress-testing a plan — without outsourcing the judgement you are paid for.

• Use it to structure a problem rather than to answer it
• Run a deliberate argument against your own position
• Recognise when you have started outsourcing the judgement

Concrete workflows for the roles that write for a living, with the failure mode each one has to watch.

• Apply specific workflows to marketing, sales and comms tasks
• Avoid the sameness that comes from unedited output
• Recognise the claims that create real exposure

Workflows for the roles where a quiet error is expensive, and the one category where the answer is usually no.

• Apply workflows to operations, finance and HR tasks
• Make arithmetic verifiable rather than trusting it
• Identify the people decisions that must not be automated

Assistants in your documents, mail, spreadsheets and chat — what changes when the tool can already see everything.

• Use built-in assistants for the tasks they are genuinely best at
• Understand what changes when a tool has standing access to your material
• Apply the verification rule to spreadsheet work

The difference between someone who gets good output occasionally and a team that gets it reliably.

• Build a reusable context pack for your organisation
• Turn a one-off prompt into a shared template
• Decide when a custom assistant is worth building

The fingerprints of unedited output, why they matter commercially, and the edit that removes most of them in one pass.

• Recognise the recognisable patterns of unedited output
• Edit efficiently rather than rewriting from scratch
• Explain why this matters beyond taste

Trigger, action, and the judgement step in between — plus the most common mistake, which is using AI where plain automation was better.

• Decide where an AI step belongs in a workflow and where it does not
• Recognise when plain automation is the better answer
• Apply least privilege when connecting tools together

What an agent actually is, where it genuinely saves time today, and the four categories where the answer is no regardless of how good it gets.

• Distinguish an agent from an assistant with extra features
• Design approval gates around irreversible actions
• Apply the approve, gate or refuse decision to a proposal

Why most AI initiatives quietly die, what a pilot worth running looks like, and how to tell whether it actually helped.

• Design a pilot narrow enough to produce a real answer
• Baseline before you change anything
• Measure value in a way that survives a sceptical question

What a usable AI policy contains, how to handle the enthusiast/refuser split, and what actually changes about a role.

• Draft a one-page AI policy people will actually follow
• Handle uneven adoption without mandating enthusiasm
• Think about roles at the level of tasks rather than jobs