Organised by department, because that is how you will recognise your own. Each one lists what the AI does, where a human stays in the loop, and the metric that proves it worked — because an automation with no metric is a science project.
The list is long so you can find yourself in it, not so you can do all of it. The companies that get returns from AI start with one high-volume, rules-bounded workflow, redesign that workflow before automating it, keep a human on exceptions, and measure a single concrete outcome.
The companies that get nothing start six pilots simultaneously, automate the process exactly as it exists today, and measure enthusiasm. McKinsey's finding is blunt: the roughly 80% who bolt AI onto old processes see no profit impact at all.
Every entry below names where the human stays. That column is not a disclaimer — it is the design. Automations that remove the person entirely from a judgement call are the ones that get switched off three months later after an expensive mistake.
Not an exhaustive list, and not a menu to order from. It is the set of patterns that have paid back most reliably in the businesses we have worked in.
Where the volume is highest and the rules are clearest — usually the first place we look.
Reads every incoming invoice, extracts line items, codes to the right GL account, matches against the PO and receipt, and routes for approval. Exceptions and anything over a threshold go to a person.
Checks every submitted expense against policy, flags out-of-policy items with the rule they breach, and auto-approves the clean ones.
Watches the AR ageing, drafts context-aware chase emails referencing the specific invoice and prior conversation, and escalates by age and value.
Matches transactions across bank feeds and ledgers, groups the unmatched, and drafts the explanation for each variance before anyone opens a spreadsheet.
Drafts the commentary that accompanies management accounts — what moved, by how much, and against which prior period.
Continuously reviews spend for duplicate vendors, overlapping subscriptions, and contracts renewing without review.
Mostly about removing the administrative drag between a lead arriving and a rep talking to it.
Enriches every inbound lead with firmographic data, scores it against your actual closed-won history, and routes it to the right rep with a briefing note attached.
Writes call notes into the CRM from recordings, updates deal stages from what was actually said, and flags deals with no activity against their stage.
Assembles a first-draft proposal from the discovery call, your pricing rules, and the closest prior deals.
Watches support tickets, usage, invoice status and engagement together, and flags accounts drifting toward churn before the renewal conversation.
Produces a one-page brief before every external meeting: who you are meeting, what happened last time, open tickets, invoice status, and recent company news.
Answers the repeatable 70% of an RFP from your prior submissions and approved content library.
The area with the fastest visible payback, and the one where getting the human boundary wrong is most damaging.
Classifies every inbound ticket by topic, urgency and sentiment, attaches the customer's history and account status, and routes to the right queue.
Writes a suggested reply grounded in your knowledge base and the customer's actual history, for the agent to edit and send.
Spots the questions being asked repeatedly with no article behind them, drafts the missing article, and flags existing ones contradicted by recent resolutions.
Reads tone and content across a conversation and flags the ones heading toward a complaint or churn while there is still time to intervene.
Confirms the fix actually held a few days later, and reopens or routes to CS if it did not.
Reads every ticket, review and survey each month and reports what customers are actually complaining about, ranked by volume and revenue at risk.
High-volume, highly repetitive, and the area where bias risk means the human boundary has to be drawn hardest.
Reads every application against the documented requirements and produces a structured summary — evidence for and against each criterion.
Coordinates panels across calendars, sends the invitations, handles reschedules, and chases confirmations.
Triggers accounts, equipment, training assignments and check-ins from the signed offer, and chases whoever has not completed their part.
Answers the recurring questions about leave, expenses and benefits from your actual documented policy, with a citation to the clause.
Assembles the evidence a manager needs — goals, delivered work, peer feedback, prior reviews — into one prep document.
Aggregates what leavers actually say over time and reports the patterns rather than the anecdotes.
The least glamorous category and frequently the most valuable, because this is where undocumented process hides.
First-pass review against your own checklist — flags missing clauses, non-standard terms, and deviations from your template.
Reads every inbound email, form and message, classifies it, enriches it, and routes it with a draft response attached.
Builds daily schedules against skills, location, travel time and priority, and re-optimises when something changes.
Watches stock against demand patterns and lead times, and drafts purchase orders before you run out.
Gathers the artefacts an audit will ask for on an ongoing basis rather than in a panic the month before.
Watches how work actually gets done and drafts the SOP that nobody has had time to write.
Where AI is most oversold. These are the ones that hold up.
Assembles the research, competitive landscape and source material for a piece, and drafts the brief a writer works from.
Pulls every channel into one report, explains what moved and why, and flags what is underperforming against baseline.
Adapts sequence content and timing to what each contact has actually read and done, rather than a fixed drip.
Watches reviews, social and press for your brand, flags anything needing a response, and drafts it.
What teams ask when they start picking from a list like this.
Whichever one your team complains about most and can describe precisely. That combination — high volume, clearly understood, and genuinely irritating — beats any theoretical ROI ranking, because it comes with people motivated to make it work.
If nothing stands out, the audit exists to answer this. We score candidates on volume, rule clarity, error tolerance, and the strength of the metric behind them.
A well-scoped single workflow is typically live in two to four weeks once the process is documented. That last clause carries most of the risk: if nobody can describe how the work is done today, documenting it honestly is the project, and the automation is the easy part afterwards.
It should be caught by the human boundary in the design, which is why every entry above names one. The failure cases that hurt are the ones where a system was given a judgement call it should never have owned.
We also instrument for it: confidence thresholds that route uncertain cases to a person, logging so you can audit what happened, and a defined rollback. If a workflow cannot fail safely, it is not a good candidate.
Honestly, it changes them more often than it removes them, and we would rather say that plainly than pretend otherwise. In practice the work that gets automated is the part of a role people describe as admin — chasing, re-keying, formatting, routing — and the role becomes more exception handling and judgement.
Where headcount reduction is genuinely the goal, say so at the start. It changes what we build and how it is introduced, and pretending otherwise sabotages adoption when your team works out what is happening.
No, and attempting it is a reliable way to get nothing. Most companies we work with have two or three areas where the return is obvious and the rest is noise for now.
Sequencing across departments is a large part of what the Chief AI Officer engagement is for — deciding what happens first, and equally what deliberately waits.
Book a call and we will walk your actual operation, not a list. You will get a ranked shortlist whether or not you engage us.
