AI for a defined business task

Start with a task whose result you can assess: drafting a reply, organising material or analysing specified sources. We help prepare instructions, a trial and a way for the team to review the output.

Describe the task for your team

From your material to a draft you can review
Source A

Permitted source

Draft output

Check the facts and their relationship to the source.

Source: A, section 2

A person approves

Uncertainties remain identified.

An example of working with AI: source material, a draft, then review.

Choose a task rather than a promise to automate everything

We establish what a person does today, the information they receive and how they recognise a good result. A repetitive task with defined boundaries and an identified reviewer is a useful starting point. Not every activity needs an AI model; a template or clear rule may be enough.

Drafting, organising knowledge and analysis

We can develop instructions for drafting an offer, organising notes, comparing materials or reviewing editorial work. A knowledge base needs defined sources and update rules. Generated text is material for review, not automatically a verified fact.

Sources, information and responsibility

We select materials permitted for use and agree what must not be passed to the tool. The environment depends on the task and the organisation’s rules. A model’s name or where it runs does not prove confidentiality or accuracy. The responsible person approves data handling arrangements.

A pilot with quality criteria

We prepare typical cases and cases with missing information. Checks cover faithfulness to the source, completeness, uncertainty and usefulness. Where the material does not establish a fact, an appropriate result can be a question or an explicit statement that information is missing.

Demonstration example: a fictional note becomes a list headed “Agreed”, “To check” and “Next actions”. A missing deadline goes under “To check”; the system does not invent a date. The example uses no private documents and does not present a customer result.

People retain the decision

We identify the review point before publication, sending material or another agreed action. AI can prepare a draft but is not given default permission to execute instructions independently. We do not promise self-learning from conversations or unlimited autonomy.

What the team receives

Deliverables include instructions and prompts with quality criteria, a workflow description, review examples and a list of exceptions. Prepare the task, permitted materials, tools in use and a reviewer. Acceptance uses an agreed trial, including incomplete information, with limitations recorded. Further use follows that assessment rather than the mere availability of the tool.

Choosing the tools

Our work can use ChatGPT and Codex from OpenAI, Gemini, Gemma and NotebookLM from Google, and tools such as Notion, Gmail and Google Workspace. The choice depends on the task, available data and operating constraints.

Next step

Describe the task and how your team recognises a correct result. Leave confidential information out of the initial enquiry.

Describe the task for your team

Have a quick question? Get in touch.