One AI tool or several? Why multi-tool setups usually fail

The short answer: one standard tool plus one deliberately permitted second tool. Anything beyond that costs more in coordination, training and data protection effort than it returns in quality. Not because variety is bad, but because the effort per additional tool does not grow linearly — it grows with the number of combinations in which people have to explain why they did what with which.

Last checked: September 2026.

Disclosure: I earn my living advising on and training people in Claude. An article arguing for monoculture would be in my interest — which is why I explicitly do not argue for it, and why the counter-position gets a full hearing in this text.

The short version

• One standard tool for breadth — the one that gets trained, documented and reviewed.
• One second tool, deliberately permitted for a named purpose, with a named group.
• From the third tool onwards costs rise faster than value — and not in licence fees.
• The power-user counter-argument is valid, but applies to individuals, not organisations.
• The worst state is not variety but unregulated variety.

Why this becomes a question at all

The question rarely arises at the start. It surfaces after about a year, when the situation in many businesses looks like this: management introduced a tool eight months ago. Marketing has been using a second one for four months because it produces better images. Development uses a third because it codes better. And part of the workforce still uses personal accounts, because they knew those before anything was introduced.

Nobody decided this. It happened. And the question is no longer “which tool” but “how many of them can we afford”.

What an additional tool actually costs

The licence is the smallest item — for ten people we are talking a few thousand euros a year. Four other things are expensive.

Training. Every tool needs its own two to three guided hours per person. Two tools do not mean two training sessions — they mean two sessions plus the explanation of when to use which. That third part is precisely the one missing in practice, and it accounts for most of the friction.

Data protection. Every tool needs its own review, its own data processing agreement, its own line in your processing register, and its own answer to who removes which account when an employee leaves. That effort arises once per vendor, regardless of whether three or thirty people use the tool.

Knowledge that does not travel. The underrated item. What your marketing department builds in good templates and approaches does not transfer one-to-one into the specialist department’s tool. You pay for the same learning twice — and, more seriously, the two groups can no longer help each other.

Decision load in daily work. The item that never appears in a table. If someone has to pause before every task to consider which tool is right now, that costs little time — but it costs something on every single task. Across a year and a whole workforce it adds up to more than the second licence ever cost.

The counter-position — and what is right about it

The most common argument against my recommendation runs: “Experienced users run two or three tools in parallel anyway and switch by task. That is the most productive way to work.”

That is true — for individuals. Anyone working with these tools daily develops a feel for which suits what, and genuinely gains from switching. I work that way myself.

The error lies in transferring it to an organisation. What works for one person, because they bear the cost of switching and collect the benefit themselves, works differently in a business: there the organisation bears the cost — training, coordination, data protection — while the individual collects the benefit. That is not a moral problem but a question of incentives. By the same logic, businesses do not run seven word processors, even though individual people are perfectly capable of using several.

So: a second tool, yes — but for a named group with a named purpose, not as a general permission.

What a sensible second tool looks like

Three conditions, all of which in my view must be met:

1. A named purpose. Not “for everything the other one does less well”, but “for visual material in marketing” or “for development”. If you cannot state the purpose in one sentence, you do not have one.
2. A named group. Four people in marketing, not “whoever wants it”. Otherwise you do not have a second tool; you have two standard tools.
3. A written demarcation. Three lines in the policy suffice: which tool is standard, which is the second tool, for what exactly. Without that sentence you get the friction described above.

Combinations that have worked in projects: Copilot or Gemini for breadth plus Claude for the department doing demanding text work. Or Claude as standard plus ChatGPT for marketing because of image generation. Both are clear divisions along the work, not along preferences.

The worst state

Variety is not the problem; unregulated variety is. The state in which three tools are in use, none of them officially, nobody is responsible for any of them, and part of it runs through personal accounts.

It looks harmless because it costs nothing. It merely costs elsewhere: you cannot say which company data was processed where. You cannot give auditors a defensible answer. And you cannot secure the knowledge built up, because it sits in personal accounts that disappear with the next departure.

If you are reading this and recognising the situation: the order is stocktake, then decision, then business accounts, then policy. In that order — a policy issued before the business accounts exist is a request, not a rule. More in the GDPR comparison.

What to do concretely

• Count how many AI tools are genuinely in use — including personal accounts. Ask without threatening sanctions.
• Name a standard tool. The one training is built around and the policy points to.
• Permit at most one second tool, with a purpose and a group.
• End the rest — with a transition period and an explanation of why.
• Secure the knowledge from the departing tools into your own store before you switch them off.

Frequently asked questions

Is this not simply convenience for the IT department?

Partly yes, and that is a legitimate objection. The difference lies in whether the limit is justified. “We permit only one because otherwise we cannot manage the data protection review” is a justification. “We permit only one” with nothing added is convenience — and will be understood as such.

What if a department rejects the standard tool?

Then listen before you decide. Often there is a genuine professional reason behind it — and that is exactly the case the second tool exists for. Sometimes, though, it is only habit, and then training helps more than a second licence.

There are three of us. Does this apply to us?

In a weaker form. With three people the coordination effort is small and you can afford more variety. What remains is the data protection part: even with three people you need a contract per vendor and an answer to what happens to the data.

How often should we revisit the decision?

Once a year is enough. More frequently and the workforce stops taking the decision seriously — and a tool change always costs more than the improvement initially promises.

What about specialist tools that are not assistants at all?

A transcription service or a translation tool is not a second assistant and does not fall under this rule. The data protection part applies unchanged, though: separate vendor, separate review, separate line in the register.