The first thing you automate with AI should be the most boring job in your company, not the most impressive one. Automation is where AI starts saving real money, but automate the wrong thing, or a broken thing, and in our experience it costs more than doing nothing. This post gives you one rule, a four-question gate, the candidates that win fastest, the traps, and a worked example.
The owners who win aren’t the ones who automate the most. They’re the ones who are deliberate about what they automate first, and in what order.
The golden rule: never automate a broken process
If a process is a mess, automating it just makes the mess happen faster and cost more. The same is true of the data underneath it: Gartner names poor data quality as one of the main reasons generative AI projects stall and get abandoned after proof of concept (Gartner, 2024).
“Broken” is easy to spot once you look for it:
- Three people do it three different ways. There is no process, only habits, and an AI trained on habits picks one at random.
- Approvals happen in WhatsApp. The real decision trail lives in chat threads no system can read.
- A spreadsheet column nobody trusts. If your team already re-checks the data by hand, the AI inherits the same doubt, plus speed.
The fix is not a project. Write the process on one page (the trigger, the steps, who decides what, what counts as an exception), get the people who do the work to agree it is true, then automate the page. Amplifying garbage just gives you more garbage, sooner.
Is this a good candidate? Run it through four questions
A good candidate is usually all four of these at once. If a task stalls on the broken-process question, fix that first. If it needs real judgment, keep a person in charge. What survives the gauntlet is what you automate first.
flowchart TD
A(["A task you might automate"]) --> B{"Is the process broken?"}
B -->|Yes| F(["Fix it first, then come back"])
B -->|No| C{"Repetitive and rule-based?"}
C -->|No| G(["Needs judgment, keep a human on it"])
C -->|Yes| D{"Eating real hours, and checkable?"}
D -->|No| H(["Low payoff, leave it for now"])
D -->|Yes| E(["Automate this first"])
So the four things you’re looking for:
- Repetitive. It happens often, the same way each time. Daily or weekly, not “a few times a quarter.”
- Time-consuming. Your people spend real hours on it, and you can name the hours.
- Rule-based or pattern-based. There’s a logic a machine can learn, even if nobody has written it down. “If the invoice matches the PO, post it; if not, flag it” is a rule. “Use your judgment” is not.
- Tolerant of occasional checking. A person can review the output where it matters, and reviewing is faster than doing.
Where owners get fast, visible wins
Four families of work pass the gate almost every time.
- Reading and sorting documents. Supplier invoices, delivery notes, customer purchase orders, expense receipts. The AI pulls out the fields (supplier, amount, date, PO number) and posts them into your ERP or accounting system, whether that is SAP Business One, Odoo, or Zoho. A person only sees the ones that fail a match.
- Answering routine questions. Order status, opening hours, how to return something. For customers on WhatsApp and the website, and for staff asking HR and IT the same few questions. It has to work in Arabic and English, which is harder than vendors admit; we wrote about getting AI customer service right in both languages.
- Moving data between systems. The “swivel chair” work: copying a new customer from the CRM into accounting, keying a signed contract into the ERP, updating a delivery status in three places. Dull, error-prone, and quietly eating hours every week.
- First drafts. Quotes, replies to inquiries, meeting summaries, the weekly report, the first pass at a tender response you found on Etimad. A person edits and signs; the AI removes the blank page.
Where automation goes wrong
The traps are where budgets go to die.
- Anything where a mistake is expensive and you can’t undo it. Sending money, releasing stock, deleting records, filing with a regulator. Automate the preparation, keep a human on the button.
- Work that needs real judgment, empathy, or relationships. A complaint from your biggest client, a salary negotiation, a supplier dispute. AI can draft and summarize here. It should not run the conversation alone.
- Anything touching sensitive personal data without proper care. PDPL took effect on 14 September 2023 and full compliance was required by 14 September 2024, with SDAIA as the regulator (Morgan Lewis, 2024). An automation that reads ID documents, medical records, or salary data is processing personal data, and it needs a lawful basis, a retention rule, and an answer to “where does this data go?” before it goes live.
- Automating before measuring. If nobody knows the task took, say, 160 hours a month, nobody will notice when it takes 30. Baseline first, or you can never prove the win.
A worked example: the boring job beats the impressive one
Let’s put illustrative numbers on it. The numbers are made up to show the shape.
Say you are a distributor and 1,200 customer purchase orders arrive every month as PDFs and WhatsApp photos. Someone retypes each one into the ERP, about 8 minutes each including a check, so roughly 160 hours a month, a full-time person. You add an extraction system: it reads the clean orders, call it 75% of them, and routes the messy 25% to a person, who now spends about 3 minutes on each because the fields are pre-filled. That is about 15 hours a month of exceptions; call it 30 with spot checks. You have taken roughly 130 hours a month off one person’s desk.
Now compare the impressive project you were tempted by: a demand-forecasting model that promises to cut dead stock. It needs two or three years of clean sales history by product, which you don’t have, because the purchase orders were retyped inconsistently. The boring automation is the prerequisite for the impressive one.
The lesson isn’t the exact hours. It’s that a good first automation has a measured baseline, a share of cases it handles alone, a human on the rest, and a saving you can point to inside a quarter.
Sequence it by value and ease
Once you have a shortlist, plot it on two axes: what it is worth per year, and how easy it is to build with the data and systems you already have.
quadrantChart title Which automation goes first x-axis Hard to build --> Easy to build y-axis Low value --> High value quadrant-1 Do first quadrant-2 Do second, plan it properly quadrant-3 Skip for now quadrant-4 Quick fill-in wins "Invoice and PO extraction": [0.82, 0.8] "Arabic and English FAQ assistant": [0.75, 0.68] "CRM to accounting sync": [0.7, 0.55] "Tender first drafts": [0.6, 0.45] "Demand forecasting": [0.25, 0.78] "Meeting summaries": [0.9, 0.28]
Start in the top right, with something that’s both high-value and straightforward. An easy, visible win builds confidence and frees up time and budget for the harder, bigger automations next. Don’t open with your most complex, mission-critical process. Earn your way there.
It’s also the surest way to stay out of Gartner’s numbers: they predict at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs, or unclear business value (Gartner, 2024). All four reasons shrink when the first project is small, measured, and built on a process that already works. If you want the demo to become a system, run the proof of concept so it leads somewhere.
How to start: one process, one month, one number
- Pick one. The top-right candidate from your plot. One process, one team, one owner who wants it to work.
- Measure. Two weeks of honest numbers: hours spent, error rate, backlog. This is your baseline.
- Pilot. Run it on real work for a month, with a person reviewing every output at first and then a sample.
- Count the net. Hours saved minus the software, the build, and the reviewing time. If it pays back inside a year, you have found your model.
- Scale. Widen it to the next team or document type, then pick the next candidate. Do this three times and you have an automation program, not a pilot.
The month-by-month version is in the first 90 days of AI in your company. And if you would rather have an independent view before you commit, that is what our AI strategy and proof-of-concept work is for.
Want to know what to automate first in your business? SDCG helps owners find the automations that pay off fastest, and warns you off the ones that’ll cost more than they save. Independent advice, grounded in what actually works. Book a free 30-minute review.