AI is already in use at your company. Nobody chose it

Adoption wasn’t postponed, it was delegated

Waiting makes sense when it means consolidating the data before buying tools: those who bought in a hurry now have disconnected tools and scattered data, and nobody can say whether it paid off. The risk of fragmentation isn’t day-to-day inefficiency: it is that at some point it becomes cheaper to start over.

The problem is another one. In a team that adopts without friction, everyone tries what they find useful. The result is AI used with personal accounts, on company data, with nobody knowing which tools or which data. There is no delay to catch up on: there is a situation already under way and not mapped.

The question to ask before any other

How much spontaneous AI use is there already in the company, with which tools and on which data? It can’t be estimated from outside, but from inside it can be measured quickly. The answer changes the priorities: if use is low, the current pace may be fine. If it is high, one decision becomes urgent and the others scale down.

A few ways to find out without starting a project: ask, explicitly and without judgement; look at expense lines and subscriptions; check which services show up in company sign-ins. The point isn’t to authorise or ban. It is to know.

The minimum that costs almost nothing

• A register of the tools in use. A list kept up to date of what is used and for what. Not to grant permissions, but to have a map.
• A rule on where the output ends up. The work produced goes to a known, shared place, whatever tool made it. Otherwise it stays invisible to the rest of the company.
• A rule on where the data goes. Before adding a tool that produces or consumes company data, the question isn’t what the tool does but where that data ends up.

Manage, don’t ban

A concrete example: someone finds a specialised tool that simplifies their work, but eighty per cent of what it does is already done by the company’s own tool. It is a situation to manage, neither to ban nor to ignore. You need to work out whether the eighty per cent is enough, and whether the remaining twenty is worth a subscription and one more vendor.

Specialised products are often excellent, but they tend to become a feature inside the general-purpose interface people already keep open. Trying them is perfectly fine. Tying a business process to one that can’t later be moved is much less so.

Then choose an interface, with permissions

The next step is to choose a single conversational interface and build a serious net of permissions and access to information underneath it. The risk, if nobody chooses, is that a tool gets chosen and not the permissions: everyone with their own account on company data. It is the worst of the plausible scenarios, and also the most common.

Every month that passes doesn’t just add delay, it adds habits to undo, and undoing a habit costs far more than steering it while it forms. A declaration of interest: we help companies think it through with technical consulting and build agents and controlled access with AI for business, but the map of tools in use can be made on your own, today.