You don't need an agency to start with AI
A PAYE puzzle across three systems solved in minutes, a travel assistant on WhatsApp, and why most of the business doesn't need an agency to start with AI. The third post in our series, with a data protection warning from our DPO.
Stacey, our CFO, was trying to make sense of some PAYE deductions that weren't clear on the surface. Getting to the bottom of them meant working out the total cost of each employee, pension contributions included, from figures held in our accounting platform, by our pension provider and by HMRC. Every detail had to be fetched from each system and cross-referenced against the other two. It's the sort of job that eats an afternoon.
We handed it to Claude Cowork. Within a few minutes it had collated the lot and cross-referenced the details, which was the part that had been eating her time.
This is the third post in a short series on what AI has changed at TSD. The first two covered our developers and our operations, where there's always an engineer on hand. This one is about the rest of the business, and why most of it doesn't need an agency. Which I get is an odd thing for an agency owner to write.
The jobs that sit between systems
Look at what the PAYE question involved. The answer was already there, split across three systems that were never designed to talk to each other. Somebody had to bring it together and check each figure against the others.

The root cause analysis in my last post had the same shape. Collate from several sources, compare, find the gap. Agents are quick at it, and every business has jobs like this sitting in the spaces between its systems. The supplier statement that won't match the ledger. The stock count that's out against the till. That's where I'd start using AI first.
An assistant that lives in WhatsApp
The other tool I've leant on lately is Instinct, an AI personal assistant you message through iMessage or WhatsApp. You ask it for things the way you'd ask a person, whether that's booking a restaurant or finding travel options.
Being inside iMessage and WhatsApp rather than it's own app is an interesting idea as it feels more like a person in a weird way. It's also really good from what we've found so far helping me with all those "can you just" or "I wonder if" type thoughts on the go.
Team travel is where it's earned its keep for me recently. Getting several people to the same place means comparing trains and hotels, and someone always ends up with a browser full of tabs. Instinct does the legwork and comes back with options, and the team makes the final bookings.

That last part is deliberate. Instinct can apparently book with your card if you let it, and I haven't. Until I've seen it get a few bookings right, the money stays with a person.
Call it the rule from my last post, applied to spending: a person signs off anything that's hard to undo.
Instinct is invite-only for now, and I have invite codes I can share, so get in touch if you'd like to try it.
So do you need an agency?
For this kind of thing, no.
You need someone who has seen it done, and an hour with the person who owns the job. I didn't send Stacey a guide. I sat with her and asked her to describe what she was trying to do in her own words, and what she expected to see at the end. We wrote that down and sent it to Claude, asking it to come back with any clarification questions before it started.
That's most of what an agency does at the start of a project. A decent brief is the problem in the owner's own words and a clear picture of what done looks like. Letting Claude ask its questions first covers the gaps, and none of it needs a developer.

The ONS found that a lack of expertise is one of the most cited barriers to adopting AI, and that businesses hitting it mostly respond by training the staff they already have. That matches what I've seen. One person who's a bit further along, working through a real task with a colleague, gets further than any amount of reading.
Engineering comes back into it once the quick fix turns into something other people rely on. In the first post I described the wall that looks like stone and turns out to be timber. We're now seeing a lot of systems cobbled together from components that look sensible on the surface. Built by people without engineering experience, they often carry enterprise-grade machinery the business would never need, and the people who built them are none the wiser.
A one-off question, like the PAYE deductions, is fair game for anyone. Something that runs every month, touches customers or holds other people's data needs someone who knows what the wall is made of.
The catch: where the data goes
I'm also TSD's data protection officer, so this is the part I'd underline.
Payroll data is personal data. So are the names and contact details you pass to an assistant when it books a trip. We run Claude on a business plan, which is why I was comfortable pointing it at payroll.
Before anyone points an AI tool at either, check whether your plan lets the provider train on what you give it. Check where the outputs end up, too.

Plenty of people won't ask. Over half of employees (55%) say they use AI for work or education, against around a third of businesses reporting that they use it at all. The ONS suggests some of that gap is people using AI tools on their own in their day-to-day work.
Give your team an approved tool and a short note on what not to paste into it, before they go and find their own.
The pattern so far
Across these three posts, the pattern has held everywhere we've used AI. It does the legwork quickly, and a person still decides. Our developers learned to hand over outcomes and review the result. In operations, every root cause stays a hypothesis until someone checks it. Instinct finds the trains, and the team books them.
The benefits at TSD have been real, and most came from somewhere less glamorous than the sales pitch, like a PAYE question answered in minutes but I'm not entirely sold on the hype, yet.

Where I'd start
Pick one job that means gathering information from several places. Sit with the person who does it, get them to describe the job and the result they expect in their own words, and let Claude ask its questions before it starts.
Before anybody pastes payroll into anything, agree what data can go in and use a business account. A person signs off anything that spends money or can't be undone.
If it works, you'll find the next job yourself. And if it starts turning into a system other people depend on, that's the moment to talk to someone who builds them for a living.