The first process a company should automate with AI is picked through three filters: how much margin it's leaking, how repeatable the decision it makes is, and whether a named person is accountable for its result. Not by how easy it is to connect. What's easy to connect is rarely what's expensive to keep running by hand.
The three filters · where the first one usually hides · the four processes that never go first · why the impact-effort matrix falls short · what to do with the rest of the list once you've chosen.
Seven minutes. At the end, a four-step exercise to run on your own list of processes.
Your first automated process doesn't just decide whether the project works. It decides how everything after it gets measured. If the first one was picked because it was easy, it gets presented with whatever was easy to count: tasks completed, active users, system speed. Every project that follows gets judged by that same yardstick, even if none of those numbers ever reaches the P&L.
Pick the first one by the margin it's losing, and the yardstick changes from day one: money coming back, shorter cycles, customers who stop walking away. That's a conversation a board understands.
Gartner's 2026 Hype Cycle for Agentic AI, drawing on its annual survey of CIOs and technology executives, reports that only 17% of organizations have deployed AI agents so far, while more than 60% expect to within the next two years. Most of what's already running is narrowly scoped. Put differently: over the next two years, most companies will pick their first process. The technology will be available to all of them. What puts some of them ahead of their market is the judgment behind that pick.
In why AI pilots don't reach production, we covered what has to be in place for an agent to make it across. Here we step back to the list of candidates: which one goes first, which ones should never take that spot, and what criteria decide it? Those criteria live in our method; this article brings them down to your list.
The three filters: exposed margin, a repeatable decision and a named owner
A process qualifies to go first when it clears three filters at once: it has exposed margin, it makes a repeatable decision, and it has a named owner accountable for the result. Run them in that order. The first can be checked against a report you already have; the third takes an uncomfortable conversation.
- Exposed margin. How much money walks out the door today because the process is still manual: a collections cycle that keeps stretching, a proposal that goes cold while someone writes it, a mistake you pay for twice. The quick test: if you can put a figure on it using your own accounting data, the margin is exposed. If all you can say is "it eats up a lot of time," not yet.
- A repeatable decision. An AI agent performs where the decision follows rules someone can explain out loud and repeats hundreds of times a month. A decision whose criteria shift case by case, or that rests on one person's gut, isn't first on the list.
- A named owner. Someone who wins or loses when that number moves. If you ask who runs it and three people answer at once, or nobody does, the filter fails.
The filters don't average out. A process with huge margin and no owner doesn't beat one with moderate margin and a director who wants it fixed: the second reaches production, the first reaches a slide deck.
Where the first one usually hides: quoting, collections, reconciliation and replenishment
The first process almost always sits where clear rules, high volume and money on hold overlap: quoting, collections, document reconciliation and replenishment planning. Its cost already shows up in some report leadership reviews.
- Quoting. A proposal that takes three days to go out is competing with one that went out in three hours. The margin leaks through response time.
- Collections. Every day the cycle stretches is working capital sitting still. Who to call, and when, follows rules the team already knows.
- Document reconciliation. Invoices, purchase orders and warehouse receipts that someone matches by hand. Every discrepancy that slips through gets paid for twice: once in the error and again in fixing it.
- Replenishment planning. Buy too much and capital sits on a shelf. Buy too little and sales stall. Both cost money, and both can be measured.
This isn't a list to copy; it's where to look first. The pattern holds in one of our published case studies: at a services company, customer support and proposal writing were overloading the same team, and margin was leaking through response time. That's the quoting case on this list, and it paid back in under 90 days.
The four processes that never go first
Four kinds of process should never be first, however good they look in a demo: the ones with no owner, the ones whose rules change every month, the ones that depend on data the company doesn't capture yet, and the ones that are already broken. Each has a warning sign you can spot in a meeting, and a fix that doesn't require any technology.
- No owner. The sign: you ask who runs it and the answer is the name of a department, or silence. The fix is structural, not AI: assign it before you ask anyone for a quote.
- Rules that change every month. Commission plans being reworked, rotating promotions, credit criteria still under debate. The sign: the current policy lives in a recent email, not in a document. The fix is to wait for a quarter of stable rules; recalibrating an agent every month eats the return.
- Data the company doesn't capture yet. The sign: ask how long it takes and the answer is "depends who you ask." The fix is to start recording that data a few weeks ahead. It's cheap, and it's what later lets you prove the result.
- Already broken. The sign: people have built their own workaround, a shadow spreadsheet or a group chat where they solve what the system doesn't. The fix is to repair the workflow first, because an agent on a broken process just reproduces the failure faster. That workaround, by the way, tells you for free how the process should actually work.
"Never" here means never first. Once its fix is in place, any of the four can go back on the list.
The impact-effort matrix, and why it isn't enough
The impact-effort matrix sorts candidates well, but it's missing the third axis that decides whether a project reaches production: who signs off on it. Two processes can land in the same high-impact, low-effort corner and behave in opposite ways once the project starts.
If the process lives inside a single area whose head can approve the change in workflow, the effort the matrix showed is real. If it cuts across three areas and none of them has the final word, the effort was underestimated: the hard part isn't building the agent, it's getting three people to agree to change how they work.
That's why, in an Agentic Discovery, every opportunity gets turned into a figure, ranked by impact and effort, and tagged with who signs off on it. If the decision goes up to the board, what to ask before approving picks up where these criteria leave off.
What to do with the list once you've chosen
Once the first one is chosen, the list you set aside doesn't get thrown out: it becomes the order of what comes next. The immediate step is to redesign the chosen process before building the agent, never the other way around.
That redesign looks at the process end to end, even if only one stretch of it gets automated. Not everything gets automated, and it doesn't need to be; what does need to happen is weighing the impact across the whole value chain. Before building, ask two questions: what changes for whoever hands you the work, and what changes for your customer. A quote that goes out in three hours is no help if it forces whoever takes the order to capture twice the data, or if it reaches the customer with a delivery date the warehouse can't meet. An automation that speeds up your stretch and pushes the cost onto the step before or after doesn't improve margin: it just moves the problem somewhere else.
What almost nobody does is keep the rest. Every candidate that didn't make it has its reason written down: no owner, unstable rules, missing data, broken workflow. Each reason is a task someone can take care of while the first agent goes into production. When it's time to pick the second one, you don't start from scratch: you check which candidates have cleared their blocker.
The first result also changes the conversation about the second. A number expressed in money, on a process the committee already knew, is the easiest approval there is: the one for the next phase. In between, what gets underestimated most is preparing the people who will work alongside the agent, and staying with the number until it's met. That's how you scale phase by phase instead of betting everything on one big project.
What to do on Monday
- List the five processes that came up most often in meetings this quarter. Whatever gets discussed every month is almost always what's costing money.
- Next to each one, write down what it costs per month and how long the full cycle takes. The ones you can't quantify move to the bottom of the list; they don't come off it. They're missing data, not importance.
- Cross out the ones nobody claims as their own. Next to each crossed-out process, write why. That note is the task that puts it back on the list.
- From what's left, start with the one with the most exposed margin. Not the easiest to connect, and not the one that gets mentioned most: the one that costs the most.
Before you take your list to committee
Choosing the first process doesn't require hiring anyone yet: it requires criteria and the information you already have. If you want to know where your company stands on AI before you decide, the self-assessment tells you in five minutes, with no commitment to hire.
To go deeper:
- How we work, from assessment to result: our method
- Where your list turns into figures: Agentic Discovery
- Candidate processes by sector: see industries
- What comes after the pick: Why AI pilots don't reach production
Documented cases by industry available under a confidentiality agreement.
Frequently asked questions
Where should a company that has never used AI start?
With a single process that clears all three filters, not with a platform or a company-wide plan. The first case has one job: produce a result in money that can be put in front of the committee. With that result in hand, the second decision gets made on evidence instead of enthusiasm.
Which processes shouldn't be automated?
As a first project, four kinds: those with no owner, those whose rules change every month, those that depend on data the company doesn't capture yet, and those that are already broken. It isn't a permanent veto. Each has a fix that needs no technology, and once it's fixed, the process can go back on the list.
How many processes should you tackle at once?
One, in the first cycle. Several at once split the owner's attention, the team's and the budget, and make it hard to tell what produced which result. Once the first one is in production with its number measured, you scale phase by phase, each phase approved on its own business case.
What if the process isn't documented?
That doesn't stop you from starting. Documenting the process as it actually runs, not as the manual says, is part of the initial work, and it's usually where the workarounds people have already invented come to light. What does stop you is nobody being able to say what it costs the company to keep it as it is.