AI is pulling your organisation apart

Most organisations are adopting AI the same way: unevenly. One team writes requirements with an assistant. Another builds an agent that answers customer questions. A third is still waiting for procurement to approve a tool. Each group is moving faster than it was a year ago, and each one is doing it in its own direction.

The result is something most leadership teams can feel but struggle to name. The organisation is busier and more productive, yet harder to steer. Decisions that seemed settled come back. Teams build confidently on assumptions that nobody else shares. This paper calls that effect drift: it explains why AI makes drift worse, what it costs, and how to fix it without slowing anyone down.

The speed problem nobody planned for

Organisations have always had to keep teams aligned. Planning cycles, sign-off meetings, specification documents and steering groups all exist to make sure that what one team builds fits what another team expects. These mechanisms were designed around a certain pace of work. A requirement might take weeks to write, so a fortnightly review was enough to catch it going wrong.

AI changes the pace without changing the mechanisms. A team using an assistant can draft, revise and act on a requirement in an afternoon. An agent can turn that requirement into code, test cases or customer messaging before the next review meeting is in the diary. The work moves at machine speed, but agreement still moves at meeting speed.

That gap is where alignment breaks. It isn't that people stop talking to each other. It is that the thing they agreed and the thing being built from it separate faster than anyone can notice.

How drift happens

Drift rarely starts with a disagreement. It starts with a copy.

Someone takes the agreed version of a policy, a requirement or a set of rules and pastes it into a prompt, a project document or an agent's instructions. From that moment there are two versions. Over the following weeks each copy is edited for good local reasons: a clarification here, an exception there, a model's suggested rewording that sounds better. Nobody intends to change what was agreed. Each change is small and sensible on its own.

Without a shared baseline

With locked versions

Multiply that by every team and every AI tool and you get the picture on the left. Teams that started from the same kick-off end up in different places. Three things make this worse with AI than without it:

• Every tool holds its own copy. Assistants, agents and retrieval systems each need the knowledge in their own context, so the number of copies grows with every tool adopted.
• Models rewrite fluently. An AI model will happily paraphrase a rule into something that reads better and means slightly less. The change is hard to spot because the new wording sounds authoritative.
• Nobody can see the gap. There is usually no record of which version a team or agent is working from, so there is no way to compare them until something goes wrong.

What drift costs

The cost of drift is rarely a single failure. It is a steady tax paid in rework, risk and lost trust.

Rework. When two teams discover they have been building from different versions, someone has to unpick the work. With AI in the loop there is simply more work to unpick, because more was produced from the wrong starting point before anyone noticed.

Compliance exposure. In regulated sectors the question isn't only "is this correct?" but "can you show who agreed it, and when?" If the version an agent used can't be traced back to an approved source, the organisation cannot demonstrate control, even if the output happens to be right.

Slower decisions. Once people stop trusting that a decision will hold, they hedge. They reopen settled questions, add review steps and wait for confirmation. The time AI saved at the keyboard is spent again in meetings.

Wasted AI spend. Models asked to work from inconsistent or contradictory sources produce inconsistent results. Teams respond by adding more prompts, more checks and more model calls, paying repeatedly to resolve a problem that sits upstream in the knowledge itself.

None of these costs show up as a line in a budget. That is why drift persists: it is everyone's problem and no one's line item.