Tonkotsu has nineteen sites and is about to open its twentieth, opposite the Roman Baths in Bath. It has no CTO. Its central team is small enough that the finance director also owns the systems. And it still builds its own rotas in Excel.

None of that is an accident, and none of it is a confession. On this week's episode of The Stacked Podcast, Tonkotsu finance director Mike Statham talked us through what actually changed the numbers over the last few years — and the answers are less about buying new technology than about being honest regarding what the technology is for.

Six sites to twenty

Ken Yamada and Emma Reynolds started Tonkotsu in 2012, opening in Soho after a run of ramen pop-ups went better than expected. Mike joined in 2017, when there were six sites. A significant minority investment from YFM in 2019 funded roughly three openings a year until Covid interrupted it. Bristol and Cardiff have worked well, giving the ops team a western cluster to run, and north is on the list.

He describes his early role as being the founders' conscience — "is it okay that I've spent this?" — and, as happens in businesses this size, he picked up the systems along the way because somebody had to.

Start with the restaurant, not the spreadsheet

Asked what advice he'd give a finance director about to change systems, Mike didn't start with requirements documents or ROI models.

"You've got to remember why you're here. We're here because we make great food, great ramen, and we give people a good experience when they're in the restaurant. So you start by getting the tech that helps people do that."

Then the part most people wouldn't say on camera. He's clear that Lightspeed is, in his words, probably not the best enterprise solution — which means his central team carries more work than it otherwise might — but it was much more user friendly in the restaurant, and that was the trade he was willing to make.

That is a genuinely unusual way for a finance function to think, and it's worth sitting with. The cost of a system is not only its licence fee. It's also where you choose to absorb the friction: in the support office, or on the floor at seven on a Friday.

What visibility actually changed

The clearest before-and-after in the whole conversation is not about cost. It's about who can see what, and when.

On Excel, somebody had to update the file. "They're always waiting on someone. There's always a gatekeeper." Now sales and labour go in daily, and area managers and GMs can open an app on their phone and see where their sites are — every half hour, if it's a good week or a new opening. GMs see their own site; ops tend to want the whole estate.

Underneath that sits a weekly cadence: an estimate of where the business will land on EBITDA, produced in-week rather than discovered at month end. It has become accurate enough that Tonkotsu now uses it as a control in the other direction — checking the monthly financials back against the weekly estimate to confirm the accounts are right.

Mike also makes a point that lands harder than it sounds: the shift isn't just from slow reporting to fast reporting. It's from reporting to forecasting. "You're having a conversation when you can still make a change." His ops team is usually the one starting those conversations now, not finance.

Stock: from counting for the sake of it to a managed variance

Tonkotsu used to get surprises at month end. Looking back, Mike is straightforward about why: they were asking teams to manage inventory well without giving them the tools, and possibly without the right recipes either.

The counts happened weekly, but a lot of the time they were counting for the sake of counting. Anyone who has run a patch will recognise the pattern our host described from his own ops days — count Sunday night, still reconciling on Thursday, count again four days later, and a GM who had a bad week with no way to find or fix the cause.

Now the theoretical variance is managed quickly, with an allowance of just under 2.5% in total — production, usage and staff food all included. The first pass at fixing the process took 3–4% out of cost of sales. Wastage has since come down below 0.9%.

The bit worth stealing is what happens when it slips. A recent run of hot weather caused a blip. Because the numbers were visible, the ops team saw it, jumped on it, and came back not just to where they were but slightly better. Visibility on its own does nothing; visibility plus a team with the discipline to act on it is the whole game.

Stock sits on Storekit, chosen partly for how deeply it integrates with the POS.

Build it or buy it

Tonkotsu schedules in Excel. Not because nobody has offered them an alternative, but because they built something after Covid with a cross-functional team, at a point when changing another system wasn't realistic, and it does things the market doesn't.

"You go in, you see workforce management, and you go: do you do this? Do you do this? They go, no. Well, ours does. And I built it."

He's not ideological about it — they'd happily buy one day, and he can see the appeal of updating a forecast and watching the output change rather than exporting, running, re-importing. But the bar is set by something that already works, and his own tool catches bad inputs in ways generic platforms often don't.

Ordering went the other way. They'd built their own, then tested MarketMan and concluded, fairly, that a vendor with fifty developers was going to beat a vendor with one. That's the honest version of build versus buy: not a principle, a case-by-case judgement about where your own knowledge genuinely beats the market and where it plainly doesn't.

With vibe coding now putting working prototypes in the hands of operators, Mike's position is going to get more common, not less. Being able to show a supplier what good looks like — and ask whether they can match it — shifts the power in that conversation.

One site, then four, then all of them

The POS migration is a useful case study in not doing it all at once. One new site went live first. Then four, while eight remained on the old system — which is when the differences became visible in a way a single pilot never shows. Middleware failures were the tell: one stack going down around four times a year against roughly once for the other. Tonkotsu uses Deliverect on the delivery side, and Mike is generous about what that category solved — a rare piece of technology that removed a real problem at exactly the moment the market needed it.

Support mattered too. An introduction to Eposability, who know both the UK market and the platform well, is what built the confidence to go further.

By the time the full rollout came, the team already knew where the problems were, so the effort went into fixing those first. "You change EPOS to solve problems and you just get new ones." Knowing which new ones you're getting is most of the work. Sites went live in the morning and got checked in the afternoon; one had an issue with online orders not being switched on. That was about the worst of it.

"Don't let AI become that accountant"

Mike is measured on AI in a way that's becoming rare. He notes that invoice automation was called invoice automation until quite recently, and everyone would call the same thing AI today. He can see the potential in self-building rotas. He's about to start using the AI functionality in their BI platform.

And then the best line of the episode:

"When I joined, people used to talk about accountants ruining businesses because they were just running it on a spreadsheet. You've got to be careful that AI isn't becoming that accountant."

The risk he's pointing at isn't robots taking over. It's that bad data now travels faster and sounds more convincing than it used to, and that a business can end up making decisions on an answer nobody questioned. The data points that matter most — are the team happy, are the customers happy — are also the hardest ones to automate.

He's similarly level-headed about AI voice in guest comms. If somebody actually phones a restaurant these days, it's rare enough that you probably want to speak to them: a GM can ask the chef something a bot can't.

Three things to take from it

Decide where you're willing to carry the friction. A system that's better in the support office and worse on the floor is a choice, not a default. Make it deliberately.

Get the foundations working before you buy the dream. Can it count? Can it reconcile? Can you record what you need? "You'll get sold a lot of the dream by people, but the foundations are the most important thing."

Pilot properly, then move. One site tells you whether it works. Four sites running alongside the old system tell you whether it's better. And pick your window — nobody should be changing a POS in November.

The full conversation covers a lot more: the £1m electricity quote that arrived out of nowhere, the NI threshold, 20% EBITDA a decade ago versus the fight to stay flat now, and what Covid genuinely fixed. Worth your commute.

Looking at a change like this yourself? Browse the tools operators are actually using, with real usage data and operator reviews, in the Stacked Marketplace.