The five-word brief
“Summarise this for the board.” Five words, handed to someone who has worked in the organisation for a few years, can be a perfectly adequate brief. They know which decision the paper is meant to support, what happened at the last meeting and why the directors will want a proper explanation of that particular number. They know the difference between an issue that needs escalating and one that belongs in the appendix. Most of the instruction never leaves the manager’s mouth, because most of it is already in the other person’s head.
Give those same five words to AI and you can get a beautifully structured summary that misses almost everything that mattered. The facts may be right. The language may be appropriate. It may even sound more like a board paper than the original. But the decision is buried, the sensitive risk gets a passing mention and the thing the board actually asked for has disappeared in the compression. The machine has completed the task as described. The manager is left wondering how something so capable could misunderstand something so obvious.
What went missing between the two
This is the briefing gap: the distance between the instruction we give and the understanding we assume travels with it. People bridge that distance through shared experience. Ask a long-standing colleague to make a presentation more client-friendly and they know to bring the commercial benefit forward, lose the internal jargon and probably cut slide 14. You haven’t issued seven separate instructions. You’ve used a phrase that stands in for years of working together.
AI’s conversational fluency makes it very easy to forget that those years are missing. It can discuss governance, recognise the conventions of a board paper and produce a sensible recommendation, which makes it feel as though it understands the room. Unless that context has been supplied, it is working with general patterns where your colleague was working with particular knowledge. It knows what a board might want. You meant this board, at this meeting, making this decision. The distinction is small in the wording and enormous in the work.
The version of misalignment that gets approved
AI alignment is usually discussed at a much grander scale: whether a powerful system will pursue an objective in ways its designers never intended, and what happens if those intentions and outcomes diverge. Business has its own everyday version. We give the machine an instruction that seems complete to us, it fills the gaps with plausible assumptions, and the result heads somewhere we didn’t mean to go. You don’t need a rogue machine for that to become expensive. An obedient one will do.
The awkward part is how presentable the result can be. An obviously broken answer gets sent back. A fluent, orderly document is more likely to move through the business, particularly when the person reviewing it is busy and the whole point of using AI was to save time. A recommendation can be reasonable in general and wrong for a company that has deliberately chosen a different course. Repeat that across customer correspondence, management reports and planning documents, and the organisation can accumulate a great deal of polished work that quietly pulls against its own strategy.
Brief it like someone who just arrived
Nobody hires a strategist on Monday, says “develop our growth strategy” and reasonably expects them to understand the business by Tuesday. We arrange conversations, explain the history and tell them which apparently attractive options have already been ruled out. With AI, the speed of the answer tempts us to skip the introduction. Then we spend the time we saved correcting assumptions we could have explained at the start.
The practical fix begins with four habits. Explain what the work needs to achieve, including the decision it should help someone make. Supply the context a capable new employee would need. Set the boundaries, especially what must be preserved, questioned or escalated. Then examine the result and ask what assumptions sit underneath it. The model’s account of those assumptions is something to check, but the question itself helps surface gaps that a quick read for spelling and tone will miss.
For the board paper, that might mean specifying that the directors are being asked to approve an investment, that the previous meeting requested a downside scenario, and that any recommendation must sit within the agreed risk appetite. Those details do more useful work than another paragraph telling the model to act as a world-class governance expert. They give it something concrete to work with.
Stop repeating the induction
That still leaves a problem for larger organisations. If every employee has to assemble the same corporate briefing every time they open an AI tool, the business has created a substantial new job for everyone. Some will do it carefully, some will use an old version, and some will decide that getting a passable answer quickly was the point. The quality of the output will depend on who remembered to explain the organisation that morning.
A shared organisational knowledge base changes that arrangement. The board-paper structure, current strategy, approved risk appetite, escalation rules and examples of good work can be maintained centrally and made available to the AI systems that need them. Someone has to own that material, keep it current and make sure access reflects who is entitled to see it. A folder full of documents won’t do this by itself. The useful part is connecting the right knowledge to the work at the point it is needed.
Now “prepare this for the board” can start from an established understanding of how this organisation’s board works. The employee supplies what is specific to today’s paper, and the system brings the standing context. People still have to exercise judgement, but they spend less of their time repeating the company handbook to a machine. That is a capability worth building into the business: an AI system that arrives properly briefed, instead of asking every employee to start its induction again.
