Our full site is on its way
AI consultants who build it, connect it and run it for you.
For established UK businesses: AI does the legwork, and you make the informed decisions.
What yours could look like
Three jobs in three made-up businesses, start to finish. The agents pull data from the systems each business already uses and put it on one page. A person decides, the follow-through happens, and every step is logged. Each loop runs for about half a minute, then repeats. Pause any time, or tap a step to stop on it.
VAT returns, ready before anyone sits down
Quellbrook & Co, an accounting practice of about 20 people
Every client's books pulled in, every return prepared and checked against the partners' own rules, and every fix explained, all before the accountant starts.
ConceptQuellbrook & Co and its figures are made up.
- Mon 07:00. Quarter end. Chloe's agent pulls every client's books, receipts and bank lines, and prepares all 52 of her VAT returns.
- Mon 07:40. By the time Chloe sits down, 41 pass every check, 8 are flagged with the fix drafted, and 3 wait on receipts already asked for.
- Mon 09:15. Chloe opens a flag: VAT claimed on client entertaining, which can't be claimed. The agent shows the boxes it changes, and why.
- Mon 16:20. The client approves on their phone, with the one change explained in plain words.
- Tue 10:05. Helen, the partner, presses File and the return goes to HMRC through Making Tax Digital. The agent can't file on its own.
- Tue 10:06. Every step is logged, with HMRC's receipt. Kestlegate is filed a month before the deadline.
Where the data comes from. Client books from Xero, QuickBooks and Sage; receipts from Dext and client emails; bank feeds; deadlines from the practice system; past returns from HMRC; and the partners' rules from a spreadsheet. All read only: the agent can't change them. It acts in two places: the client portal, once Chloe has checked, and HMRC, only when a partner presses File.
Who decides. Chloe accepts each fix, the client approves on their phone, and only a partner can file with HMRC. The agent can't file on its own.
Which products deserve more ad spend
Dunlark, a coffee-alternative brand (caffeine-free brews) selling on its own store and Amazon
Data from nine systems joined product by product, so the answer is plain: spend more on these, change the ads on those. One decision, and both happen.
ConceptDunlark and its figures are made up.
- Mon 06:30. Maya's agent joins the ad spend from Meta, Google and TikTok with every order from the store, subscriptions and Amazon, and what each product costs, product by product.
- Mon 06:40. Out of 10 products, four bring customers who keep coming back. Three bring one-off buyers who cost more to win than they spend.
- Mon 06:41. The answer forks in two: more spend where it pays back, and a brief for the creative team where it doesn't.
- Mon 08:05. Maya approves both from her phone. The extra spend is capped at what she approved.
- Mon 08:06. Budgets go up within the cap, and the creative team gets a brief that says what's failing and what works on the mesquite brew.
- Mon 08:06. Logged, and checked again in four weeks: did the extra spend bring customers who stay?
Where the data comes from. Ad spend from Meta, Google and TikTok; orders from the Shopify store, the subscriptions app and Amazon; email results from Klaviyo; visits from Google Analytics; and product costs from Xero. All read only. It acts in two places, and only after Maya approves: Meta budgets, never past her cap, and the creative team's Slack channel.
Who decides. Maya, who runs marketing, approves from her phone. The extra spend can't go past what she approved, and the brief reaches the creative team only when she says so.
From 84 CVs to booked interviews
Ferrovale Group, a fire and security group
Everything on one page, so the decision takes minutes. Then one click does the follow-through: the emails, both calendars, the bookings.
ConceptFerrovale Group and its figures are made up.
- Wed 08:00. 84 applications for a fire-alarm service engineer arrive from four job sites, the careers page, the careers inbox and two recruiters.
- Wed 09:00. One page: every CV summarised in a line against Priya's five must-haves, best fit first. She reads; she chooses.
- Wed 09:05. Priya ticks six. One button: tell them, and offer interview times.
- Wed 09:12. The agent finds times that suit Priya and the service manager, offers each person three, and drafts the emails for her to send.
- Wed 15:30. By the afternoon four have picked a time, and the interviews are in both calendars.
- Wed 15:30. The 78 others get a kind ‘not this time’ when Priya says so. Every step is logged.
Where the data comes from. Applications from LinkedIn, Indeed, Totaljobs, CV-Library and the careers page; CVs from the careers inbox and two recruiters; the job spec from SharePoint; and past applicants from the HR system. All read only. It looks up free times in Outlook, but sends emails and books interviews only once Priya says so.
Who decides. Priya chooses the shortlist and approves every email. The agent never turns anyone down on its own.
If you can picture it, we can build it.
And we make sure it's safe and reliable.
These three loops are examples, not a menu. Every build starts with how your business already works, and plugs into the systems you already use.
Safe and reliable, by design
- 1Only the access it needsMost systems are read only, like the “Data from” list in each loop. Anything more is agreed with you first, system by system.
- 2A named person decidesAnything sent outside, any change to spend, every filing. The agent prepares; your person approves.
- 3Your rules, in plain codeLimits like the £3,000 cap or “only a partner files” are enforced by ordinary code, not by the AI, so the AI can't change them.
- 4Logged and watchedEvery step is logged. If a system stops sending data or the figures don't add up, it stops and tells someone rather than carry on.
Tell us how your business runs. We'll tell you where AI fits.
If we find nothing worth doing, we'll say so. We're happy to sign an NDA before you share anything.