How I'd approach AI if I were running a company

At EmbedAI, every call, email and Slack message goes into our company AI OS. It knows which client something relates to, which project it’s about and what was agreed last time, and it files it in the right place without anyone having to think about it. My reMarkable notes and the Pocket recordings from in-person meetings go in too. So when I ask it to draft a proposal or chase an invoice, it already has the full picture, and I don’t explain anything.

None of that involved clever prompting. It involved connecting things up.

Goes in

CallsEmailSlackreMarkable notesPocket recordings

→

Company AI OS

Knows which client it's about, which project, and what was agreed last time. Files it in the right place.

→

So when I ask

"Draft the proposal""Chase that invoice"

it already has the full picture.

That’s the part most companies skip, and I think it’s the reason so many of them feel like they’ve “done AI” without anything really changing.

Most companies bought the tool and kept the business

The numbers say the same thing. McKinsey’s latest State of AI survey, published in August, found that nearly nine in ten companies now use AI regularly in at least one part of the business. Only 37% could point to any effect on profit at all, about the same as a year earlier. MIT’s NANDA project went further last year: of the organisations piloting generative AI, 95% were getting zero measurable return.

Most companies use AI. Far fewer see it in their profits.

Use AI regularly in at least one part of the business88%

Say it's scaling across the company44%

See any effect on profit (EBIT) at all37%

The usual pattern is the same everywhere I look. Everyone gets a ChatGPT or Copilot licence, there’s a workshop or two, and maybe a pilot runs somewhere in the business. People use it to tidy up emails and summarise documents. The business itself works exactly the way it did before.

Microsoft 365 Copilot is $30 per person per month on the enterprise plan. Put 200 people on it and that’s $72,000 a year, mostly so that emails come out a bit tidier.

That isn’t nothing. But it’s a personal productivity tool bolted onto a company that hasn’t changed, and it’s priced like a transformation.

The value is in the access

An AI that can’t see your CRM, your inbox, your project files or your finance system can only ever be a fairly clever chatbot. It doesn’t know who your clients are, what you promised them or what’s overdue. Every time someone uses it, they have to paste in the context by hand, and most of the time they don’t bother.

The MIT report is fairly blunt about why. Tools like ChatGPT “primarily enhance individual productivity, not P&L performance”, and the custom projects that should go further stall “due to integration complexity and lack of fit with existing workflows”. That’s not a model problem. It’s a plumbing problem.

The value comes almost entirely from the access. Once the system can see what’s going on, it stops being something you go and ask and starts being something that does the work: filing, chasing, drafting, flagging what’s slipping.

So if I were running a company, the first money would go there. Connect the systems, give the AI somewhere to keep what it learns, and only then worry about what people do with it.

Prompt training is solving the wrong problem

A lot of companies are sending their staff on prompt engineering courses. I think that’s mostly wasted.

Prompting is a workaround for a system that doesn’t know anything. If the AI already has the context and the access, people can just say what they want, the way they’d ask a colleague. Templates and brief-writing are another layer between the person and the work, and that layer doesn’t need to exist.

The training that is worth doing is much simpler: show people what the system can now do for them, and get them in the habit of asking.

Build more than you buy

The cost of building software has collapsed, and most companies haven’t adjusted to it yet.

At FanFinders we’re replacing HubSpot with our own CRM. It’s been built since July, mostly by describing what we need, and it already holds the real data: 75,322 contacts, 10,536 deals and 415,261 emails, synced from HubSpot every night until we switch over. It does what the sales team actually uses, plugs straight into everything else, and when someone wants a change it’s a conversation, not a feature request to a vendor.

I’m not saying rip everything out. There’s plenty you should still buy. But every renewal is now a real decision, and the question is no longer “which tool” but “would we be better off with our own”.

One person, not a department

You don’t need a team of data scientists. For a small business, one person whose whole job is AI is probably the best hire you can make right now: someone curious and practical who sits with each part of the business, works out what’s eating everyone’s time, and builds the systems that take it away.

It’s also worth looking at who you already have. Product owners, designers and people with deep knowledge of the business are in a very good position, because they already know what good looks like and how to explain it. They often get more out of this than the developers do.

Larger companies are better off bringing in a team that’s done it before. That’s what we do at EmbedAI, so take that with the appropriate pinch of salt.

Be sensible about access, not scared of it

People are nervous about giving AI access to everything, and in a company that matters more, not less. But the answer isn’t to lock it out. It’s to set it up the way you’d set up a new member of staff.

At EmbedAI the two founders have full access, and the rest of the team deliberately see less of the founder-level material. The AI works the same way: it sees what the person using it is allowed to see. And anything consequential, like emailing a client or paying an invoice, comes to a person for a yes before it happens.

Lock everything away and you get a fraction of the value. Open everything up and you’ll regret it. The middle is where the work gets done.

The licences aren’t wasted

To be fair to the licences, they do help people. The best evidence I know is a Harvard study of 758 BCG consultants: with GPT-4, they finished 12.2% more tasks, 25.1% faster, and the work was rated higher quality. On a task deliberately picked to be outside what the AI was good at, they were 19% less likely to get it right.

So the licence isn’t wasted. It’s just not the change. It makes individuals a bit quicker at the work they were already doing, inside a business that still runs on the same handoffs, the same chasing and the same copy-pasting between systems.

The jobs change either way

AI is going to change every job in your company. It already has changed most of them, whether or not the people doing them have caught up. I’d be open with the team about that. The way I see it, people and their AI are on the same side, and the most valuable people will be the ones who learn to direct it well. Someone still has to decide what’s worth doing and whether the output is any good.

If I had to boil it down: connect your systems, give the AI the context it needs, and let people just talk to it. Start with the part of the business where everyone’s drowning in admin, get that working properly, and expand from there.

The companies that do this won’t look like they’ve adopted AI. They’ll just look like they run better.

Sources: McKinsey, The state of AI in 2026 (Aug 2026) · MIT NANDA, The GenAI Divide (Jul 2025) · Microsoft 365 Copilot pricing (checked 19 Oct 2026) · Dell'Acqua et al., Navigating the Jagged Technological Frontier (HBS).