Strange thing for an AI company to say, we know. But we've now watched enough businesses buy this stuff before anyone worked out what it was actually for that we'd rather get it out of the way early. AI works. We use it every day, it's changed how we run, and that's a big part of why the way it's currently being sold gets on our nerves.
Nearly all of it runs on fear. Move now or fall behind, your competitors are already ahead of you, the window's closing. It sells beautifully. What it mostly produces is a lot of expensive software wandering around a business looking for a job to do.
Some businesses should just wait, and if yours is one of them we'd honestly rather say so than take the money. There are four reasonably reliable signs, which we'll get to. If none of them land, the second half of this is what's been working for us, and it's a good deal less complicated than you'd expect.
Everyone has started. Almost nobody has finished.
McKinsey put this to 1,993 people across 105 countries last November, and the drop-off between starting and getting anywhere is the whole story.
Source: McKinsey & Company, The state of AI in 2025, November 2025. The 7% comes from their follow-up analysis in December.
Bain arrives in the same place from another direction, with fewer than a fifth of organisations having scaled generative AI in any way they'd call meaningful. None of which says the technology is a dud. It says most attempts die somewhere between the demo that impressed everyone in the room and anything that changes how the work actually gets done.
Hardly any of this is a technology failure. It is tools bought before anyone named the job.
Which is worth sitting with for a second, because it changes what you think you're buying. AI isn't an answer, it's a tool, and a good one. It makes a team sharper at problems they already understand and quicker at work they already do. It won't hand you a strategy you haven't got, unbreak a process that's broken, or tell you which parts of the business are worth fixing in the first place. Aim it at something well understood and it's genuinely brilliant. Aim it at "we should be doing AI" and you'll end up in the 61% with nothing to show at enterprise level.
Four signs it isn't your turn yet
These are the ones where we'd tell you to go and spend the budget elsewhere. Not for ever. Just not this quarter.
Two or more of those and we'd point you at whatever's genuinely holding the business back instead. That isn't us being coy about wanting the work. A decent slice of what we get asked to build shouldn't be built, and working that out in a first conversation is a lot cheaper for everybody than working it out in month four.
What it looks like when it does stick
We're an AI-native company so we're an easy case, and you should discount our numbers accordingly. The bit we didn't see coming was where the time came back.
That caveat stays attached, because anyone quoting you a tidy productivity percentage for something this new is telling you a story rather than a finding. What we're confident about is the shape of it. The gains landed hardest on the people who spend their days turning context into documents, which is most of a business and almost none of its engineering.
So how do you get there
We gave a version of this at the Civitas AI Conference recently, to a room full of local businesses who wanted to adopt AI and mostly weren't sure where to begin. Five things came out of it.
1. The part everyone skips is context
Your competitor can license the same model this afternoon, for the same money, and have it running by dinner. What they can't get hold of is your clients, your pricing, the reasoning behind decisions you made three years ago, or the way you actually win work, which isn't written down anywhere and certainly isn't what the website says. Take that away and you get generic output, which is exactly what people mean when they say the first month was underwhelming.
The model is rented. Your context is owned.
2. Your knowledge comes in two kinds
Treating both halves the same way is the thing that quietly kills these projects about six months in. They want opposite handling.
3. Two loops, and you can start the first one this week
You don't need a finished knowledge base, you need a first one and some way for it to grow. That's loop one, and it's an afternoon's work. Loop two is the one that stops the whole thing rotting once real work starts moving through it.
4. Build skills on skills
Teach it your house style once and never explain it again. Everything built on top of that inherits it for free, so each workflow skill only has to do one actual job.
5. Why our team adopted it so fast
This is the one that surprised the room, because it wasn't training and it wasn't enthusiasm. Nobody on our team has unlimited usage, and that turned out to be the whole trick.
We didn't design it that way on purpose, which is a bit embarrassing given it turned out to be the cheapest training programme we've ever run. The other half was letting people mess about first, share whatever worked, and only then hardening the good ones into something everybody gets.
References
- McKinsey & Company. The state of AI in 2025: Agents, innovation, and transformation. November 2025. mckinsey.com
- McKinsey & Company. AI at work but not at scale. Week in Charts, December 2025. Source of the 7% figure. mckinsey.com
- Bain & Company. How Do Companies Create Value with AI? June 2026. bain.com
- The 61% is our own arithmetic, being the share that did not report enterprise-level EBIT impact in the McKinsey survey.
- Per-role time savings and licence mix: riivo internal, self-reported and illustrative. Seat prices are Anthropic's published monthly rates; annual billing is lower.