Every MSP in London has an AI pitch right now. Very few of us are honest about which bits actually earn their keep in a restaurant or hotel back-office, and which bits are still a slide deck wearing a jacket. Here’s a working operator’s view.

I sit in front of hospitality founders, FDs and ops directors most weeks, and the AI conversation has changed shape twice in the last year. Twelve months ago it was “should we even look at this”. Now it’s “we’ve spent eighteen months looking at this and we’re not sure what we got for it”. The honest answer is that AI in the back-office is real, it does deliver ROI in 2026, but the value is unevenly distributed and a lot of the louder promises are still oversold. If you run anything from a five-site bistro group to a forty-room boutique hotel, this is the version I’d want you to have.

The three categories of back-office AI worth talking about

Before we get into what works, it helps to split “AI” into three buckets, because they behave very differently and they cost very differently.

1. Copilot-style productivity AI. This is the Microsoft 365 Copilot, Google Gemini, ChatGPT Enterprise end of the world. It lives inside email, documents, spreadsheets and meetings, and it summarises, drafts and answers questions over your own files. It’s the most visible category, the easiest to license, and the one most often confused for “doing AI”.

2. Process automation. Power Automate, Make, Zapier, n8n and the rest. This is the unglamorous middleware that moves a PDF invoice from a supplier inbox into your accounts package, or fires a Teams alert when a stock variance crosses a threshold. There’s AI in the modern versions - OCR, classification, extraction - but the value is in the plumbing, not the model. In hospitality, this category is quietly the most useful of the three.

3. Specialist AI tools. Demand forecasting, recipe costing, stock reconciliation, menu engineering, dynamic pricing, review sentiment, AI rota builders. Vertical SaaS with a model under the bonnet. Some of it is excellent, some of it is a wrapper around a forecast you could draw on the back of a fag packet.

If you treat all three as the same thing you’ll either overpay or underbuy. They’re not interchangeable.

What each category actually does in hospitality today

Copilot-style productivity AI

Where it works in 2026: meeting summarisation is now genuinely good. If your ops meeting runs forty-five minutes across eight sites, having an accurate transcript with action points in your inbox before you’ve left the room is a real time saver. Drafting emails to suppliers, landlords, environmental health and the bank is faster, and the quality is fine if a human still reads it before send. Document Q&A - “what does our lease say about the extraction unit” - is the sleeper hit. It saves the FD ten minutes here, twenty there, and after a quarter it adds up.

Where it’s still too early: anything that needs to reason over your live operational data. Copilot is only as good as the data it can see, and most hospitality groups have their useful numbers locked in EPOS, stock and rota systems that Copilot doesn’t touch out of the box. If you’ve heard a demo where Copilot “analyses last weekend’s covers”, ask to see it run on the customer’s actual data, not a sample tenant.

Process automation

Where it works: invoice capture and coding into Sage, Xero or NetSuite. Expense receipts off a phone camera into the month-end. Rota draft generation from a demand forecast that already exists in your forecasting tool. New starter onboarding - accounts, mailbox, phone, training assignments - fired from a single HR record. Daily flash reports built from EPOS exports and dropped into the GM’s inbox at 9am. None of these are exciting. All of them save real hours every week, and they’re the work I see paying back fastest.

Where it’s still too early: anything that requires a human judgement call to be encoded as a rule. If your accounts team can’t write down the policy for splitting a mixed-VAT invoice, the automation can’t either, and you’ll spend longer maintaining exceptions than you saved.

Specialist AI tools

Where it works: short-horizon demand forecasting (next 14 days, single site, established trading pattern) is now solid. Recipe costing tools that pull live supplier prices and flag dish margin drift are genuinely useful, especially when input costs are moving. Stock reconciliation tools that match deliveries to invoices to theoretical usage have caught real money in groups I’ve worked with.

Where it’s oversold: menu pricing optimisation, predictive staffing across a whole estate, sentiment analysis from online reviews, and the much-promised “AI concierge” guest chatbot. I’ll come back to these in a minute, because they deserve their own paragraph of scepticism.

The honest ROI list for 2026

These are the back-office tasks where I am comfortable telling a hospitality client they will see a return inside two quarters, with the caveat that the underlying systems need to be in reasonable shape first:

  • Meeting summarisation and action capture
  • Drafting routine email correspondence
  • Document Q&A across leases, contracts, H&S manuals and SOPs
  • Rota drafting from an existing demand forecast (not the forecast itself)
  • Invoice OCR, coding and approval workflows
  • Expense management end-to-end
  • New starter and leaver provisioning
  • Daily and weekly flash reports compiled from existing exports

None of that is glamorous. All of it is real. Most of it sits in the productivity and process automation buckets, not the specialist AI bucket.

The honest “still oversold” list

These are the ones I’d ask hard questions about before signing a contract:

  • Menu pricing optimisation. The model can suggest a price. It can’t tell you how your regulars will react, what your competitor down the road did last Tuesday, or whether your GM will actually charge it.
  • Predictive staffing across a whole estate. Single-site, short-horizon, fine. Group-wide, with new openings and seasonal sites in the mix, the error bars get wide enough to drive a delivery van through.
  • Sentiment analysis from online reviews. The score is easy. The “what should we do about it” is the hard part, and that’s still a human job.
  • AI concierge chatbots for guests. A handful of large hotel groups have made this work. For most independent operators it’s a support burden, a brand risk and a distraction from the front desk doing its job well.

None of these are impossible. They’re just not where I’d start, and I’d be wary of any pitch that puts them at the front of a roadmap.

The governance question nobody wants to ask

Where does your data go, who holds it, and what does the contract actually say about training? This matters more in hospitality than people think, because your back-office holds menu IP, supplier pricing, payroll, lease terms and guest data all in the same place. “We use ChatGPT” is not a data protection answer.

The tools we recommend are the ones with clear enterprise terms, UK or EU data residency where it’s offered, no training on customer content by default, and a tenant boundary you can point at on a diagram. Microsoft 365 Copilot in a properly licensed tenant is one of those. A free consumer chatbot with your wage bill pasted into it is not. This is also where your cyber security posture and your AI strategy stop being separate conversations - the moment AI tools start reading your SharePoint, your access controls and your AI controls are the same control.

The “just pilot it” trap

The most common mistake I see is the open-ended pilot. Someone in the leadership team gets enthusiastic, ten Copilot licences appear on the bill, and six months later nobody can say whether it worked. The licences quietly renew. Nobody wins.

A sensible pilot has three things: a written definition of success that a non-technical board member could read, a hard timebox (eight to twelve weeks is plenty for productivity AI), and an honest kill switch that the sponsor has pre-committed to using. If the pilot doesn’t hit the success criteria, you turn it off. Not “extend by a quarter”. Off. The discipline of being willing to kill it is what makes the next pilot credible.

A sensible AI journey for a 10-site restaurant group

If I were sitting opposite the FD of a ten-site group tomorrow and they asked me where to start, this is the shape I’d suggest:

Phase 1 - Productivity, one quarter. Microsoft 365 Copilot for the head office team only. Finance, ops, marketing, HR. Not the GMs yet. Clear success metrics around time saved on email, meetings and document search. Governance baked in from day one - sensitivity labels, access reviews, a written acceptable use policy.

Phase 2 - Process automation, one to two quarters. Pick the two highest-volume back-office processes - almost always invoice capture and new starter onboarding - and automate them properly. This is where the hard ROI tends to land, and where data analytics and reporting start becoming genuinely useful. It’s also where your managed cloud foundations start earning their keep, because the automations need somewhere stable and well-governed to live.

Phase 3 - Specialist tools, selectively. Now, and only now, look at one or two specialist AI tools where you have a specific pain point. Recipe costing if margins are moving. Short-horizon forecasting if your rota costs are out of control. One tool at a time, evaluated on its own merits, with the same pilot discipline as phase 1.

That’s a twelve-to-eighteen-month journey, not a weekend. Anyone telling you it’s faster is selling you the demo, not the deployment.

Where this leaves you

AI is ready to run parts of your back-office in 2026. It is not ready to run all of it, and the parts where it’s ready are not always the parts the marketing focuses on. The operators who’ll do best with it over the next two years are the ones who are quietly methodical: pick the right category for the job, pilot with discipline, kill what doesn’t work, and keep governance in the room from day one.

If you’d like a straight conversation about where AI fits into your stack - and just as importantly, where it doesn’t yet - that’s the kind of work we do every week with hospitality groups across London and the South East. Start with our managed cloud page, or have a look at how we approach hospitality IT support, and drop me a line. I promise not to bring a slide deck.