10 Business Processes Companies Should Automate With AI
- Why being selective matters more than moving fast
- The five-point fit test
- The ten
- 1. Support ticket triage and routing
- 2. Invoice and expense processing
- 3. First-pass contract review
- 4. Recruitment screening and interview scheduling
- 5. Sales research and CRM hygiene
- 6. Customer onboarding and KYC document checks
- 7. Internal knowledge retrieval
- 8. Financial close reconciliation and variance detection
- 9. Content and localisation production
- 10. Data cleanup and entity resolution
- What a well-built automation actually looks like
- Five processes to leave alone
- How to actually start
- Frequently asked questions
- The short version
- Sources
Most articles with this title are a list of departments. Customer service. HR. Finance. Marketing. You finish reading and you know no more about what to do on Monday than you did before.
The problem was never identifying which department is annoying. Every company knows where the pain is. The problem is that painful and automatable are different properties, and most failed AI projects come from confusing the two.
So this list is built differently. Each process comes with what to automate inside it, what to keep human, the number that tells you whether it worked, and the way it usually fails. Before the list there's a test you can apply to any process in your own company, including the nine hundred I haven't listed.
Why being selective matters more than moving fast
McKinsey's August 2026 survey of 1,719 organisations across 97 countries found something worth sitting with. Eight in ten respondents report individual productivity gains from AI. Only 37% say their organisation sees enterprise-level EBIT impact, and that share is flat year on year. It's the same shape of gap I found in the product management data when I wrote how AI is changing product management in 2026: near-universal individual adoption, and a strategy gap sitting right underneath it at the organisational level.
Why being selective matters
McKinsey, The State of AI in 2026, surveying 1,719 organisations across 97 countries.
The gap is not a technology problem. It is a selection problem: most of that effort went into processes that were never good candidates.
Read those two numbers together. People genuinely are getting faster. The money is not showing up on the finance line. That is not a model capability problem, and at this point it isn't really a tooling problem either. It's a selection problem, compounded by a habit McKinsey names directly: companies layer AI on top of legacy workflows instead of redesigning the workflow around what AI can actually do.
The companies getting returns are not automating the most processes. They picked fewer, picked better, and rebuilt the process rather than bolting a model onto it.
The five-point fit test
Before any project, score the process out of five. One point each:
The five-point fit test
Score any process out of five before you write a line of code.
Volume
It happens hundreds of times a month, not five. Automation economics are volume economics.
Structure
Inputs consistent enough that "good" is describable. If you can't describe good in five bullets, the model can't produce it reliably.
Ground truth
Someone can look at an output and say whether it's right. Without this you cannot measure, and without measurement you're guessing.
Cheap failure
A wrong answer costs minutes, not a customer or a lawsuit. This is the criterion that gets skipped most often, and hurts most.
An existing review step
A human already checks this work today, which means oversight isn't a new cost you're introducing. It's a cost you're reducing.
The failed pilots almost always score two and get funded anyway, because the process was painful rather than suitable.
Score 4 or 5 and automate it. Score 3 and build an assistant rather than an automation, meaning the AI drafts and a human always ships. Score 0 to 2 and leave it alone, however painful the process is.
The failed pilots I've watched almost all scored two and got funded anyway, because somebody senior found the process irritating. Irritation is not a business case.
The ten
The ten, mapped
Business impact against implementation difficulty. Start bottom-left, plan for top-right.
Build these first.
Real value, but budget properly.
Build these first.
Real value, but budget properly.
Navy: build these first. Coral: real value, but budget properly.
1. Support ticket triage and routing
Not resolution. Triage. Classify the incoming ticket, detect urgency and sentiment, attach the relevant account context, and route it to the right queue with a suggested first response drafted for the agent.
This is the highest-scoring process on the test in most companies: enormous volume, structured input, obvious ground truth, and a failure that costs one misrouted ticket. It's also where the temptation to overreach is strongest. Full autonomous resolution is a much harder problem with a much worse failure mode, and the triage layer captures most of the value at a fraction of the risk.
Keep human: anything the classifier flags as low confidence, plus all escalations, cancellations and complaints.
Measure: routing accuracy, first-response time, and the percentage of drafted replies agents send without editing.
Fails when: you skip the confidence threshold and quietly route 8% of tickets into the wrong queue for a month before anyone notices.
2. Invoice and expense processing
Extract line items from an invoice, match to the purchase order, code to the right account, flag exceptions. Accounts payable is one of the few genuinely solved AI automation problems, and companies still process invoices by hand at scale.
The economics are unusually clean here because the failure is caught downstream. Reconciliation exists. If the AI codes something wrong, the close catches it, which means your cost of error is a correction rather than a loss.
Keep human: anything above a value threshold you set, plus every new vendor's first three invoices.
Measure: straight-through processing rate, cost per invoice, exception rate trending down over time.
Fails when: the pilot runs on clean PDFs from your top ten suppliers and then meets the photographed receipt from a supplier who still faxes.
3. First-pass contract review
Extract the clauses that matter, compare them against your standard playbook, and produce a redline summary of every deviation with a risk rating. Legal reviews the summary rather than the forty pages.
This one is transformative in a specific way: it doesn't make lawyers faster at reviewing, it changes which contracts get reviewed at all. Most companies have a threshold below which nothing gets read properly. AI moves that threshold down to zero.
Keep human: every final judgment call, every negotiation position, and any clause the system rates high risk.
Measure: time to first review, percentage of contracts receiving any review at all, and how often legal disagrees with the risk rating.
Fails when: the system is trained on a generic playbook rather than yours, and produces confident opinions about terms your company has deliberately chosen to accept.
4. Recruitment screening and interview scheduling
The scheduling half is pure logistics and should have been automated a decade ago. The screening half needs care, and I say that as someone building in HR technology.
Automate the coordination completely: availability matching, invites, reschedules, reminders, panel logistics. On screening, automate structured extraction and summarisation, meaning pulling stated qualifications, experience and skills into a consistent format so humans compare like with like. Do not automate ranking, scoring or rejection.
That isn't only an ethical position. Under the EU AI Act, AI systems used in recruitment and worker management sit in Annex III, with obligations applying from 2 December 2027 following the Digital Omnibus deferral. Building a scoring engine now is building a compliance project for 2027.
Keep human: every advance or reject decision, without exception.
Measure: time to schedule, recruiter hours per hire, and consistency of summaries across reviewers.
Fails when: the summariser infers attributes that weren't stated, which is both a quality bug and a discrimination exposure.
5. Sales research and CRM hygiene
Two jobs that look separate and are the same job. Before a call, assemble a brief: company news, recent product usage, open tickets, last three interactions, likely objections. After the call, turn the rep's messy notes into structured CRM fields.
The second half is where the compounding value sits. Every sales organisation has a CRM data quality problem, and it exists because updating records is tedious work that competes with selling. Remove the tedium and your forecasting inputs improve, which improves everything downstream of forecasting.
Keep human: what the rep actually says on the call, and any deal-stage change.
Measure: CRM field completeness, prep time per call, forecast accuracy over two quarters.
Fails when: the brief pulls from stale data and the rep mentions a product the customer churned off six months ago.
6. Customer onboarding and KYC document checks
Verify the document is what it claims to be, extract the fields, cross-check them against the application, flag mismatches. In regulated industries this is high-volume, highly structured, and currently done by people reading PDFs.
Onboarding is also where slowness costs revenue directly. Every day of verification delay is a day of activation lost, and drop-off during onboarding is the most expensive drop-off there is.
Keep human: every rejection, every flagged mismatch, and a random audit sample of approvals.
Measure: time to activation, manual touch rate, and false-accept rate on the audit sample.
Fails when: the audit sample is dropped after the first clean quarter and nobody notices the drift.
7. Internal knowledge retrieval
The "where do I find the travel policy" question. Every company runs an internal helpdesk staffed by IT and HR people answering the same forty questions in perpetuity.
A grounded retrieval assistant over your policy documents, wikis and past tickets removes most of that load. The critical word is grounded: answers cite the source document, and the system says it doesn't know rather than inventing a policy. An HR assistant that hallucinates a parental leave entitlement creates a problem considerably larger than the one it solved.
Keep human: anything involving an individual's specific circumstances rather than general policy.
Measure: deflection rate, citation rate, and how often users click through to the source.
Fails when: the document corpus goes stale, and the assistant confidently quotes a policy that was replaced in March.
8. Financial close reconciliation and variance detection
The highest-value item on this list and the hardest to build, which is why it sits in the top-right of the map.
Match transactions across systems, identify the exceptions, draft the variance explanation, and surface anomalies a rule-based system would miss. Finance teams spend the first week of every month on work that is mostly pattern matching, and the close is a genuine constraint on how fast a company can make decisions.
Keep human: every judgment on materiality, every adjusting entry, and sign-off in full.
Measure: days to close, percentage of variances explained without manual investigation, and restatement rate, which should stay at zero.
Fails when: you automate the reconciliation before your data is clean enough to reconcile, which is most companies.
9. Content and localisation production
Not strategy and not the writing that carries your brand voice. The pipeline around it: adapting one piece into six formats, translating and localising across markets, generating variants for testing, producing alt text and metadata.
This one scores lower on impact than the others and I've placed it accordingly, but it's the easiest to start and it builds organisational confidence. It's a reasonable first project for a team that hasn't shipped an AI workflow before.
Keep human: the original piece, and a native-speaker review of anything customer-facing in a new market.
Measure: production cost per asset, time from source to published variants, and engagement parity between the original and the adaptations.
Fails when: volume becomes the goal and you publish forty mediocre assets where four good ones would have performed better.
10. Data cleanup and entity resolution
Deduplicating records, standardising formats, resolving that "Acme Corp", "ACME Corporation" and "Acme Corp." are one customer, enriching product catalogues with missing attributes.
Unglamorous, and it's the prerequisite for half the other items on this list. If you're planning item 8 and your master data is a mess, this is actually your first project rather than your tenth.
Keep human: merge decisions above a confidence threshold, since an incorrect merge is much harder to unwind than a missed one.
Measure: duplicate rate, field completeness, and the downstream error rate in whatever system consumes the data.
Fails when: you run it once as a cleanup project instead of building it as a continuous process, and the data degrades right back within a year.
What a well-built automation actually looks like
Every item above shares the same shape, and the shape matters more than the model you pick.
What a well-built automated process looks like
The confidence gate is the difference between an automation and an incident.
Corrections humans make become the evaluation set for the next version.
That confidence gate is the difference between an automation and an incident. It's also the piece most often missing from the pilots that fail, because it doesn't demo well. Nobody has ever been impressed by a system that says "I'm not sure, please check this one." It's still the most important component you'll build.
Two other things belong in every implementation. Set a quality bar before you start, and agree that below it nothing ships. And log everything, because when someone asks in eleven months why a specific decision was made, "the model decided" is not an answer that survives an audit.
Five processes to leave alone
Five processes to leave alone in 2026
Not because the technology can't. Because the failure is not recoverable.
Assist the humans doing these. Don't hand over the decision.
The pattern across all five: either there's no ground truth to check an output against, or the failure is one you cannot recover from. Both are disqualifying on their own. Together they're the clearest signal you'll get that a process belongs to a person.
How to actually start
Pick one process from the quick-wins quadrant. One, not three.
Weeks 1 to 2
Score it on the fit test and write down what good looks like in five checkable bullets. Assemble twenty real examples from the last month, including the awkward ones. Have a human do the task on all twenty and keep those answers as your benchmark.
Weeks 3 to 6
Build the smallest version that handles the common case. Run it against your twenty examples. Find the failures, and expect three or four you didn't anticipate. Set the confidence threshold where the failures stop.
Weeks 7 to 10
Run it in shadow mode alongside the humans doing the work today. Compare outputs daily. This is the step everyone skips and it's the one that catches the problems that would otherwise surface in front of a customer.
Weeks 11 to 12
Go live on a slice, maybe 20% of volume, with the human path still fully staffed. Publish the numbers weekly. Expand only when the quality holds for three consecutive weeks.
Twelve weeks, one process, real numbers. That gets you into the 37% rather than the other 63%, and it does more for internal appetite than any amount of vendor enthusiasm. It's roughly the same shape I used building the automation layer inside Humentra.AI.
Frequently asked questions
The short version
The list matters less than the test. Volume, structure, ground truth, cheap failure, an existing review step. Score the process honestly, pick the ones that pass, rebuild the workflow rather than bolting a model onto it, and put a human behind a confidence gate.
Do that with one process and you'll have something to point at. Do it with three and you'll have a capability. Do it with ten at once and you'll have what most companies currently have, which is a lot of activity and a flat EBIT line. If you're the person tasked with making these calls, that judgment is the same core skill I described in AI product manager vs traditional product manager: deciding what's worth building, regardless of what's on your business card.
Sources
- McKinsey, The State of AI in 2026: On the Road to ROI, August 2026
- McKinsey Global Survey on the state of AI, 2025–2026
- MIT research on enterprise AI pilot outcomes, 2025–2026
- EU AI Act, Annex III and the Digital Omnibus implementation timeline
Muhammad Usman Mustafa is an AI Product Manager and Project Manager based in Stockholm, Sweden. He writes about AI product practice at muhammadusmanmustafa.info. Learn more about his AI product management and consulting services.