AI in operations · Updated October 2026
I help owner-managed firms put AI where it moves a figure the owner cares about, with the records, the checks and the rules sorted first. It sits inside my compliance, quality assurance and Fractional COO work, not beside it, because the point of all three is the same: compliance, quality and performance.
If you ask the owner of a thirty-person firm whether they use AI, you tend to get a yes and then a pause. Two or three people pay for a chatbot out of their own pocket. Reception drafts its emails with it. Last spring there was a trial with an invoice tool, and nobody can quite remember how it ended. None of that is wrong. It just isn't connected to anything the board actually watches.
One study worth knowing came out of MIT in 2025. The team behind it, part of a project called NANDA, spoke to people in 52 organisations, surveyed 153 senior leaders and went through more than 300 AI projects that companies had made public. Their headline was blunt: 95% of organisations were getting zero return from their spending on generative AI. What interested me more was why. It wasn't the software. It was that firms dropped the tool on top of the old way of working and hoped. Where money was made, it tended to be in the back office, while around half of the budgets were going to sales and marketing.
I came to AI from running operations, and I still look at it that way rather than as a programmer would. So I begin with the plan. There is a figure in it the owner wants to see go up or down this year, and some job in the building that, done better, would shift it. Once we know which job, I ask if AI can help at all, what records it would have to read, which person signs off on what it produces, and how we would tell after three months whether it paid.
Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Full list of sources below.
Five steps, in this order.
Ten operational areas. For each: what the tool does, the records it reads, the check a person keeps, and the figure we watch.
01
It checks each supplier invoice against the purchase order and the delivery note, and puts the ones that don't agree in front of a person. Nobody gets paid on the tool's say-so.
Data: Accounting system
Control: A person approves every flagged invoice
Measure: Days to pay, duplicate payments caught
02
It drafts the reminder for each overdue account in a tone that fits that customer's history, and tells you who is likely to pay late next month. Legal letters stay with a human.
Data: Sales ledger
Control: Wording agreed once, every letter sent by a person
Measure: Days sales outstanding
03
It reads the supplier contracts, pulls out renewal dates and notice periods, and shows spend by category so you can see where buying happens outside the rules.
Data: Contracts and spend file
Control: The buyer signs off every renewal
Measure: Savings, spend outside agreed suppliers
04
It compares the competence matrix with the training records and drafts onboarding material for new starters. I keep it well away from hiring and firing decisions, which the AI Act treats as high-risk.
Data: HR records
Control: No decision about a person made by AI alone
Measure: Time to competence, staff turnover
05
It watches for regulatory changes, matches them to your obligations register and drafts what needs to change for the owner to approve.
Data: Obligations register
Control: The compliance owner approves each change
Measure: Overdue obligations, audit findings
06
It sorts complaints and nonconformities by cause, so the same fault showing up in three departments stops looking like three different problems.
Data: Complaints log, corrective action records
Control: The quality owner confirms the root cause
Measure: Repeat complaints, rework
07
It answers routine questions from articles you have approved and passes anything else to a person, with the history attached. Customers are told they are talking to AI.
Data: Knowledge base
Control: Approved answers only, clear hand-off
Measure: First-contact resolution, response time
08
It scores enquiries against what has closed before and drafts a first reply for a salesperson to edit and send.
Data: CRM
Control: Consent and opt-outs respected
Measure: Conversion rate, cost per enquiry
09
It forecasts demand from sales history, suggests stock levels and plans delivery runs.
Data: Sales history, stock system
Control: The planner approves every order
Measure: Stock-outs, days of stock held
10
It pulls the figures from each system into one monthly pack and writes the first draft of the commentary. This is the card that joins the other nine together.
Data: One agreed source for each figure
Control: The owner or COO signs the report
Measure: Report on time, decisions taken
Most of what is on those cards happens inside one team, and that is fine as a start. The money I care about more is lost between teams. Credit control usually knows which customers pay late, but in a lot of firms the sales people never hear about it and keep giving those same customers the best terms. Procurement knows a supplier's contract is up in March, and operations find out in March as well, when the deliveries stop. Fixing that is less about software than about letting one team read the other's records, on terms both have agreed and which hold up under GDPR.
When that is sorted, the reading is exactly the kind of work AI is good at. I like to end up with one monthly report that takes its figures from each system and comes with a first draft of the commentary, which the owner then reads, disagrees with in places and signs. To me that report is where strategy and operations actually meet, and it tends to be the page the board reads first.
Three sets of rules shape this work in Malta and the UK. GDPR still governs any personal data the tool touches. The EU AI Act was reshaped this summer by the AI Omnibus, Regulation (EU) 2026/1744, which came into force on 27 July 2026. The duty to tell people they are dealing with AI applies from 2 August 2026, and the rules for high-risk uses, hiring and staff decisions among them, now start on 2 December 2027. Then there is ISO/IEC 42001. It isn't law, but it gives you a recognised way to show you run AI properly. The UK isn't part of the AI Act, though a UK firm whose tools are used in the EU can still be caught by it, so I plan for both.
I don't recommend anything I haven't run myself. My own work runs on an AI set-up I built and keep up: it drafts, researches and checks, under written rules on sourcing, on accuracy and on what must never be mixed. There is more on my projects page. It has made me faster. It has also made me stricter about checking, because a confident wrong answer costs more than a slow right one.
Where the AI work usually sits: part-time operations leadership, one to eight days a month.
Obligations and risk registers to ISO 37301 and ISO 31000, which the AI work plugs into.
ISO 9001 systems, internal audit and the checks that keep AI output honest.
A closer look at one function: where AI pays its way in buying, and where it does not.
Including the AI operating system I run across my own work.
Tell me the figure you want to move, and I will come back on fit and scope.
What all my work is for
Every obligation met, with an owner and the evidence to show it.
Work done the same way every time, checked and improved.
Figures that move: margin, cash, complaints and the time work takes.
Usually not at first. Most firms already pay for tools with AI built in, in their accounting system, their CRM or their office suite. The work is in choosing which job it should do, cleaning up the records it reads and putting a person in charge of what it produces.
In the back office, with a job that is frequent, dull and easy to check, such as matching supplier invoices or chasing overdue accounts. The MIT NANDA study published in 2025 found the clearest returns in back-office work, while around half of the budgets it saw went to sales and marketing.
If you use AI in the EU, parts of it can. Under the AI Omnibus, Regulation (EU) 2026/1744, the duty to tell people they are dealing with AI applies from 2 August 2026, and the rules for high-risk uses such as recruitment and staff decisions apply from 2 December 2027. I am not a lawyer, so where the answer matters we check it with yours.
No. It is part of how I run the Fractional COO, compliance and quality assurance work. If AI will not move one of your figures, I will say so and leave it out.
