Introduction
Konnichiwa! Welcome to the AI Automation Dojo. Today we’re telling the story of the most underestimated department in corporate history – the one everyone dumped their most boring work on, assuming it would stay boring forever. It didn’t. Over thirty years your Shared Services Center quietly climbed from “process the invoice” to “tell me what the invoice means.” And AI is about to make that exact same climb, in roughly the time it takes you to onboard a contractor.
I’m your host, Andrzej Kinastowski, one of the founders of Office Samurai – where we’re fairly sure the org chart’s least fashionable box is the one hiding all the money.
So whether you’re an operations leader with a budget, a mandate, and absolutely no idea where to start, or a CFO who wants a real number before anyone says the word “pilot” again, you’re in the right place.
Now grab your favorite katana (or the org chart nobody’s dared update since the last reorg), and let’s get to it!
Why Shared Services exist, and how they got promoted
Somebody in your company already figured out what you should automate with AI.
They did it years ago. They didn’t call it AI – they called it “the Shared Services migration” and they were very stressed about it. There were spreadsheets. There was a consultant with a badge.
And the document they produced – the one that lists exactly which processes are boring, repetitive, and safe to ship off to a cheaper location – is sitting in a SharePoint folder nobody has opened since the second Obama administration.
That document is your AI roadmap. You already paid for it.
SSC, GBS, BPO – take your pick. They’re all a lot of letters for “the place your company sends its least glamorous work to be able to focus on core processes”.
Quick refresher, for anyone who’s never had the pleasure.
A Shared Services Center is what happens when a company looks at forty different teams all doing accounts payable slightly differently, and someone finally asks: “why are we paying for this forty times?”
So you rip the work out of every business unit, you standardize it, and you run it once, from one place. Usually a cheaper place. Kraków, Manila, Bengaluru. Hello from one of them.
Here’s the part that matters today. When these centers were born, they only took a very specific kind of work.
High volume. Repetitive. Rule-based. Standardized. And – this is the important one – work that doesn’t need you to be in the room. Nobody has to physically walk to the warehouse to process an invoice.
Accounts payable. Payroll. The IT helpdesk. The boring trenches.
If that list sounds familiar, it should. High volume, repetitive, rule-based is the exact same list of things AI is good at right now. Same filter. Nobody planned it that way. It just worked out that the stuff you sent to Kraków is the stuff you can hand to a model.
But here’s what actually happened over thirty years, and this is the whole episode.
The center did the boring stuff well. Cheap, reliable, nobody got fired. So the business did what businesses always do when something works – they gave it more.
And not just more. More complicated.
First it was invoice processing. Then the whole finance close. Then somebody said, “well, they’re good with data, let’s give them the reporting.” Then the analytics. Then, God help us, the actual analysis.
Single-function became multi-function. Multi-function became “Global Business Services”, which is the same thing with a bigger org chart and a director who says “value creation” a lot.
The center climbed. It started as a foot soldier doing grunt work, and over three decades it quietly earned rank. It went from “process the invoice” to “tell me what the invoice data means.”
And AI is making that same climb. Just faster.
It ate the transactional trenches first – invoice status, ticket routing, copy-paste. And right now, this year, it’s climbing into the judgment work. We’ve got a case study where an AI sorts legal regulations more accurately than the experienced human who used to do it. We’ll get there.
So the thirty-year journey your center took up the value chain is a time-lapse of where your automation is going. What moved first, you automate first.
Two quick housekeeping notes before we walk the floor.
One: we’re not doing the “how ready is your organization” ladder today – the foundations, the governance, the crawl before you run. We made a whole episode on that. Go listen to it. Today’s about where to point the blade, not how to hold it.
Two: assume a human still approves every important click in everything I’m about to describe. I’ve talked your ear off about human-in-the-loop for months now. You get it.
Let’s go tower by tower.
The Map: Finance & Accounting
Green light means draw the blade now. Yellow means it works, but keep a human close. And I’ll flag the one tower that fights back.
Finance and Accounting – the anchor tenant
This is the big one. It’s the first thing every center takes, and it’s the most outsourced function on the planet. So let’s break it into its actual pieces instead of treating “finance” like one grey blob.
Procure-to-Pay. Everything from “we bought a thing” to “we paid for the thing”. Green light, all day.
Three examples of how to use AI here.
- First, an invoice-status agent that answers “hey, did my invoice get paid?” so a human accountant doesn’t have to. Think about what that question costs you today. Someone stops what they’re doing, logs into SAP, digs through two or three screens, cross-checks a report, and comes back with an answer twenty minutes later. The agent does that lookup itself and answers in plain language, right inside Teams. We built one for a client. And the important bit: it doesn’t guess. It runs a read-only query and reads back the real number, so it can’t cheerfully invent a payment that never happened.
- Second, matching invoices to purchase orders. An invoice arrives, and somebody has to check it against what you ordered and what actually showed up – the classic three-way match. AI reads the invoice, even an ugly one, lines it up against the PO, and only ever bothers a human with the ones that don’t agree. The clean ninety percent flow straight through while your people handle the arguments.
- Third, the vendor mailbox. Every finance team has one: a shared inbox where suppliers pile in asking about payment status, disputed amounts, missing remittances. AI reads each message, works out what they actually want, and drafts the reply for a human to send. You go from writing two hundred near-identical emails a week to approving them.
Order-to-Cash. The other direction: getting money in the door. Also green.
You can auto-match incoming payments to open invoices. A customer pays one lump sum against six invoices, with a payment reference that’s technically fiction, and untangling that is a job so tedious it’s basically a hazing ritual. AI is very good at exactly that kind of puzzle.
You can have it rank which overdue customers to chase first – weighing how much they owe against how likely they actually are to pay and draft the polite-but-faintly-threatening reminder emails. And you can have it pull a credit-risk summary on a customer together from a dozen scattered sources – filings, news, payment history – instead of an analyst tabbing between fifteen browser windows to build the same picture by hand.
Record-to-Report. The month-end close. Reconciliations, journal entries, reporting. Green, with a yellow stripe.
AI does the reconciliation grunt work – match thousands of ledger lines against the bank statement, and surface only the handful that don’t tie out. Instead of a person eyeballing a spreadsheet until their vision goes, they review ten flagged exceptions and go home.
It also drafts the management commentary – the “here’s why travel spend jumped twelve percent this quarter” paragraph – for a human to correct and sign. The number stays the human’s. The typing does not.
And then there’s FP&A – forecasting, planning. That’s the frontier. AI can gather the numbers, spot the variances actually worth explaining, and write the first draft of the board deck. What it can’t do yet is own the assumption underneath the forecast – whether sales will really grow eight percent next year. That’s a judgment call with a person’s name on it. That’s the top of the finance mountain, and we’re all still climbing it.
Human Resources – hire to retire
Two flavors, both green.
The helpdesk. Your HR team answers the same forty questions every week – how much leave have I got, what’s the lunch expense limit, when does open enrollment close. Point an assistant at your own policy documents and let it answer, grounded strictly in those documents, with the source attached so people can double-check.
That grounding is the whole trick. It answers from your actual handbook, not from whatever it happened to read on the internet in 2023. Regular listeners know about the parrot. I’m not relitigating the parrot. The parrot has been through enough.
That same assistant fills in forms – it collects the details in a normal back-and-forth, generates the completed PDF, and hands it back. And it can run onboarding, spinning up the new-hire checklist and firing off setup requests to IT and facilities before the person’s even shown up on day one.
Recruitment. This one’s fun. We built an agent that screens CVs. It pulls the structure out of hundreds of them, checks each against the job spec, and flags the mismatches – even when a candidate’s CV and their application form quietly disagree with each other. On one project it took screening from two weeks down to forty-eight hours. It writes tailored interview questions aimed at the gaps in someone’s experience, and it answers applicants’ emails so nobody gets left on read.
To be clear: the AI reads and flags. A human still decides who to hire. We’re sorting resumes, not letting a language model pick your next colleague.
Payroll gets a yellow. It’s great for catching anomalies – “this bonus is ten times the usual amount, are we sure about that?” – before they land in someone’s bank account. It can flag the odd ones, draft the correction, and queue it up. Just don’t let it press the button unsupervised. It’s people’s rent. Nothing torpedoes trust in an automation program faster than the month it quietly underpays half the company.
IT Service Desk – the cleanest win in the building
If you take one thing from this episode, start here.
A master agent reads every incoming ticket and sorts it: is this a “how do I do X” question, or a “please do X for me” request? That one split decides everything downstream.
The questions get answered straight from your own documentation. The requests – password resets, VPN problems, system access – get handled by an agent that reads the details out of the ticket, works out exactly what’s being asked, and runs the flow: it logs in, makes the change, closes the ticket, tells the person it’s done.
And here’s the number that gets a CFO to sit up: at some clients, access requests and revokes are up to forty percent of all tickets. Forty percent. That’s not really a helpdesk. That’s a very expensive game of “have you tried turning it off and on again”.
Procurement – Source-to-Pay
Green, and gloriously unglamorous.
Vendor master data – cleaning up supplier records, killing duplicates, spotting that “IBM”. “I.B.M.” and “IBM Corp.” are the same company wearing three different hats. AI’s great at it, and it’s the kind of work no human has ever once volunteered for at a team meeting. It pulls key terms and line items out of contracts and POs into structured fields you can actually search. And it classifies your spend against your categories, so sourcing can finally see where the money goes instead of guessing and calling it strategy.
None of this is exciting. All of it saves real money. That’s sort of the whole theme of the show.
Documents and Data – the plumbing under everything
This one’s barely a tower. It’s the pipework running underneath all the others.
Every process I’ve mentioned starts with a document somebody has to read. So this is where the leverage hides.
AI takes a two-hundred-page PDF bundle – invoices, delivery notes, customs forms, all scanned into one file by someone who long ago gave up – and splits it into clean, named, sorted documents, deciding where one ends and the next begins by actually reading them. No training, no thousands of labelled samples. On one project that cut document-prep time by seventy to ninety percent.
It also reads genuinely hostile documents – bad scans, coffee-stained faxes, handwriting, twenty different languages – and pulls the data out anyway. For a logistics client we took a process from a planned thirty percent automation to north of sixty.
One rule we live by: if the data doesn’t go into the system, we don’t extract it. You’d be amazed how much time people burn teaching a robot to read fields nobody uses.
Legal and Compliance – where it gets interesting
This is the tower that, for thirty years, was “too much judgment to centralize”. Lawyers, right? Sacred. And it’s the tower AI is climbing into right now.
We helped build a system for an energy company that watches regulatory websites, reads up to six hundred legal acts a month, and routes each one to the right lawyer based on what it’s genuinely about. It reads the opening of each act, sorts it into one of twenty-five categories, groups them, and emails the relevant person – work that used to swallow a specialist’s entire week. Accuracy came in around ninety-three to ninety-five percent. And in some cases, it beat the experienced human who used to do the job.
Sit with that for a second. A machine, sorting law, better than the specialist.
That’s not the transactional trenches anymore. That’s the climb I told you about, caught on tape.
Green for monitoring, sorting, flagging. Yellow for interpretation – the AI tells you which regulation matters, a human decides what to do about it.
Customer Service – the front office
Now I have to be the guy who tempers the excitement.
Because this is the flashy tower. It’s what every vendor demos on stage – AI answering your customers, so warm, so instant, so on-brand. And the mechanics genuinely do work. Same building blocks as everywhere else on this map.
The AI reads an incoming email and actually understands what the customer’s asking for. Some of those it can just answer – the “what’s your return policy,” “where’s my order” questions – pulled straight from your own documentation. And some it can go further and solve, the exact same way the IT desk does it: it recognizes the request, gathers the details, and runs the action – resends the invoice, updates the address, kicks off the refund.
So far, so good. Here’s the catch.
The variance is brutal. Your IT desk gets password resets and access requests – a fairly short menu of things people ask for. A customer mailbox gets… everything. Every mood, every phrasing, every half-finished sentence typed on a phone on a train. Complaints that are secretly compliments, compliments that are secretly lawsuits. The range of what could possibly land in that inbox is enormous, and huge variance is exactly the thing AI finds hardest.
So by all means, do it – just walk in with your eyes open. This tower is a marathon. A pilot we ran on a shared customer mailbox opened at around sixty-eight percent accuracy, and that’s the whole point: you start rough, you feed it better examples and sharper category descriptions, and you grind it up month over month. The teams that win here are the ones who expected that going in. The ones who assumed it’d be flawless by Friday are the ones filing it under “AI doesn’t work” by the following Tuesday.Close: what moved first, automate first
Close: what moved first, automate first
So there’s the map.
And notice what it’s really telling you. Your shared services center spent thirty years climbing – from processing invoices to interpreting them, from the trenches up to the judgment. It took decades, because every rung needed more people, and people are slow and expensive and occasionally on holiday.
AI is making that same climb in about three years.
What moved to the center first is what you automate first. And the judgment work that stayed behind — the stuff that stayed human because a human simply had to do it – that’s the wave breaking right now. That legal number should tell you it’s already here.
And before anyone emails me a think-piece about the robots coming for our jobs: nobody grew up dreaming of matching remittances to invoices. Let the machine do the meatware. Give your people back the work that actually needs a human.
If you want the deep dive on any of these, we’ve got full episodes on the finance agent, the legal case study, and the recruitment build.
Now go open that dusty Shared Services document. It’s been your automation roadmap this whole time.
