Most automation lists are written backwards. They start with what AI can do and work toward a business that might want it. The result is a catalogue of impressive demonstrations that nobody implements, because the tasks chosen are the ones easiest to show rather than the ones most expensive to keep doing manually.
The better starting point is a question about where money leaks. In almost every business, there is a category of work that is simultaneously high-frequency, low-judgement, and performed by someone whose time is worth considerably more than the task. That work is usually invisible on any org chart. It is the twenty minutes a salesperson spends assembling context before a call, repeated four times a day. It is the accounts assistant retyping numbers from a PDF into a ledger. It is the three separate people who each read the same support email before anyone answers it.
None of that is glamorous. All of it is expensive, and most of it is automatable now with tools that are widely available.
What follows is twenty-five specific automations, grouped by function. For each one, you get what the manual process actually costs, how the automated workflow runs, where a human should stay in the loop, and what tends to break. Some of these are near-certain wins. Several come with real caveats, and those are stated rather than skipped. At the end there is a short section on the automations that are widely promoted and generally not worth building.
A note on how to read the list. The value of an automation is roughly frequency multiplied by the cost of the time it replaces, minus the cost of checking its output. That last term is the one people forget. An automation that saves ten minutes but requires eight minutes of verification has saved you two minutes, not ten. Throughout this list, the automations that survive that arithmetic are the ones where either the output is easy to check quickly, or a downstream system catches errors automatically.
Getting and Qualifying Demand
1. Inbound lead capture and enrichment
Doing it manually today: A form submission arrives as an email. Someone reads it, copies details into the CRM, searches the company on LinkedIn to work out whether it is worth pursuing, and adds notes. Ten to fifteen minutes per lead, done inconsistently, often skipped when busy.
The automated flow
- Form submitted
- AI extracts name, company, role, and stated need from free-text fields
- enriches with public company data (size, sector, location)
- deduplicates against existing CRM records
- creates or updates the record
- writes a two-line summary of what the prospect appears to want
- assigns to the correct owner by territory or product
Human stays in: Nothing here needs approval. This is a read-and-write task against your own CRM with no customer-facing output.
What goes wrong: Duplicate creation when the matching logic is loose. Insist on matching by email domain plus company name, and route ambiguous cases to a review queue rather than guessing.
Best for: Any business with more than about thirty inbound leads a month.
2. Lead qualification and prioritisation
Doing it manually today: Everything is treated equally, or prioritised by whoever shouts loudest. Sales time gets spent on prospects who were never going to buy.
The automated flow
- The system scores each lead against criteria drawn from your actual closed-won history rather than from a generic template,
- then explains the score in one sentence so the salesperson can disagree with it.
Human stays in: Always. Scores should influence ordering, never suppress a lead entirely. A system that silently hides leads will eventually hide a good one and nobody will ever know.
What goes wrong: Scoring models trained on thin data reproduce your existing biases about who a good customer looks like. If you have fewer than a hundred closed deals, use simple explicit rules instead of anything learned.
Best for: Teams where lead volume exceeds capacity to call everyone within a day.
3. Lead follow-up sequences
Doing it manually today: The first follow-up happens. The fourth rarely does. Most deals are lost to silence rather than to a competitor.
The automated flow
- Drafted follow-ups that reference the specific conversation and the prospect’s stated problem, scheduled at sensible intervals,
- paused automatically when the prospect replies.
Human stays in: Approve the first message to each new prospect. After a relationship exists, later touches in a sequence can go automatically if they are genuinely useful rather than filler.
What goes wrong: The failure mode is volume without relevance. If your automated follow-up would embarrass you to receive, it is damaging the brand faster than it is generating pipeline.
Best for: Businesses with sales cycles longer than two weeks.
4. Customer reactivation
Doing it manually today: Dormant customers sit in the database untouched because nobody has time to work out who is worth contacting or what to say.
The automated flow
- Identify accounts with no purchase in a defined window
- retrieve their purchase history and past interactions
- segment by likely reason for lapse
- draft a message specific to what they previously bought
- route to the account owner for approval
- send and log
Human stays in: Approval before sending, particularly for higher-value accounts.
What goes wrong: Contacting customers who left unhappy, with a cheerful message that reads as tone-deaf. Exclude anyone with an unresolved complaint before the list is built.
Best for: Businesses with repeat-purchase models and a database older than two years. This is frequently the highest-return automation on this entire list, because the customer acquisition cost was already paid.
5. Pre-call account research
Doing it manually today: A salesperson spends fifteen to twenty-five minutes before each meeting reading the CRM, checking the company website, and searching for recent news. Often it gets skipped, and the call is worse for it.
The automated flow
- A briefing document generated automatically before each scheduled meeting, pulling the account’s history, open opportunities,
- previous objections, recent public news about the company, and the specific questions raised in the last conversation.
Human stays in: The salesperson reads it. That is the whole point.
What goes wrong: Hallucinated details about the company, which is embarrassing on a call. Require that every factual claim in the briefing carries a source link, and treat unsourced statements as unusable.
Best for: Any team doing more than ten external meetings a week. The arithmetic here is unusually favourable, because the input is already stored and the output is checked by a human who knows the account.
6. Proposal and quote drafting
Doing it manually today: A senior person assembles a proposal from previous documents, adjusts scope, calculates pricing, and formats it. Two to four hours for something substantial.
The automated flow
- Requirements captured from the discovery call notes
- matched against previous similar engagements
- relevant case studies retrieved
- scope and pricing drafted from the rate card
- assembled into the standard template
- routed to the owner for review and adjustment
Human stays in: Absolutely, before anything is sent. Pricing and scope commitments are contractual.
What goes wrong: Confident pricing errors. Keep pricing calculation in deterministic code or a spreadsheet formula, and let the AI handle only the narrative sections. This is a good general principle: numbers that must be exact should never be generated by a language model.
Best for: Professional services, agencies, and B2B firms issuing more than a few proposals a month.
Serving Customers
7. Tier-one support deflection
Doing it manually today: A meaningful share of your support volume is the same twenty questions, answered repeatedly by people capable of much harder work.
The automated flow
- A retrieval-based assistant answering from your actual documentation, with an explicit and easy path to a human.
Human stays in: As the escalation path, always visible, never buried. A deflection system that makes reaching a human difficult converts a small problem into a complaint.
What goes wrong: Confident wrong answers on edge cases, and a measurement trap. Never report a single aggregate deflection rate. Segment by question type, because the residual after deflection is systematically enriched with your hardest and highest-stakes conversations.
Best for: Businesses with more than a few hundred support contacts a month and documentation worth retrieving from. If your documentation is poor, fix that first. It is the actual bottleneck and fixing it has value regardless.
8. Ticket triage and routing
Doing it manually today: Someone reads every incoming ticket to decide who handles it. Simple, repetitive, and a real delay in first response time.
The automated flow
- Ticket arrives
- classified by topic, product area, and urgency
- sentiment assessed to flag frustrated customers
- customer tier retrieved from CRM
- routed to the right queue
- priority set
- suggested response attached for the agent
Human stays in: The agent handles the ticket. The automation only decides where it goes and offers a starting draft.
What goes wrong: Mis-routing urgent issues, particularly when customers describe serious problems calmly. Build an explicit override for keywords indicating legal risk, safety, or churn, and let those bypass classification entirely.
Best for: Support teams with more than one specialised queue.
9. Appointment booking and rescheduling
Doing it manually today: Phone tag, or a back-and-forth email thread that takes four messages to settle a time.
The automated flow
- Conversational booking that checks live availability, offers slots, books, confirms, and sends reminders.
Human stays in: Design this so the booking system itself enforces slot uniqueness, rather than trusting the model to reason about availability. Let deterministic software guarantee what it is good at guaranteeing.
What goes wrong: Double bookings, timezone errors, and failure to handle cancellation edge cases. Timezone bugs are more common than any other failure in this category.
Best for: Clinics, salons, trades, consultants, and anyone whose calendar is their inventory.
10. Review requests and response drafting
Doing it manually today: Reviews are requested inconsistently and responded to slowly or not at all, which is visible to every future customer reading them.
The automated flow
- Trigger a review request at the right moment after a successful interaction, monitor incoming reviews across platforms, and
- draft a specific response to each.
Human stays in: Approve every response before publication, without exception. Public replies to unhappy customers are reputational documents.
What goes wrong: Generic responses that obviously came from a template, which are worse than no response. If the draft could apply to any review, it is not usable.
Best for: Local businesses, ecommerce, hospitality, and any business where reviews drive discovery.
11. Customer feedback analysis
Doing it manually today: Survey responses, review text, and support transcripts accumulate unread because reading them at scale is nobody’s job.
The automated flow
- Systematic classification of unstructured feedback into themes, tracked over time, with the actual verbatim quotes retained so a
- human can read the underlying examples.
Human stays in: Interpretation and decision-making. The automation counts and clusters; it should not tell you what to do about it.
What goes wrong: Themes that are too broad to act on. Force the output to name the specific product area or process step, not just the sentiment.
Best for: Any business collecting more feedback than one person can read, which is most businesses past a certain size.
Handling Documents and Data
12. Supplier invoice processing
Doing it manually today: Invoices arrive by email in inconsistent formats. Someone opens each one, reads the numbers, and types them into the accounting system.
The automated flow
- Invoice received
- supplier, line items, totals, tax, and PO reference extracted
- matched against the purchase order
- discrepancies flagged
- coded to the ledger account
- posted automatically if matched within tolerance
- routed to a human if not
Human stays in: Reviewing exceptions and approving anything above a value threshold.
What goes wrong: Extraction errors on unusual layouts. The saving grace is that PO matching catches most of them automatically, which is what makes this one of the safest automations available.
Best for: Any business processing more than about a hundred supplier invoices a month.
13. Contract review and obligation extraction
Doing it manually today: Contracts get signed and then filed, and nobody tracks the renewal dates, notice periods, or commitments buried in clause 14.
The automated flow
- Extraction of key terms, dates, obligations, and unusual clauses into a structured register with automatic reminders before each
- notice deadline.
Human stays in: This is assistive only. The output is a first pass that a person reviews. It is not legal advice and should never be presented as such.
What goes wrong: Missed clauses in unusual drafting. Treat the extraction as a checklist that speeds up a human review, not as a replacement for one.
Best for: Businesses with more than fifty active contracts. The highest-value output here is usually the renewal calendar, because missed notice periods cost real money.
14. Document classification and filing
Doing it manually today: Files accumulate in a shared drive with inconsistent names, and finding anything takes minutes that add up to hours.
The automated flow
- Incoming documents classified by type, tagged with the relevant client or project, renamed to a consistent convention, and filed
- in the right location.
Human stays in: Periodic audit of the classification accuracy.
What goes wrong: Confidently misfiling something important, which is worse than not filing it, because now it is lost and nobody knows. Keep an audit log of every move.
Best for: Professional services firms and any business handling client documentation at volume.
15. Structured data entry from unstructured sources
Doing it manually today: Someone reads a PDF, an email, or a scanned form and retypes its contents into a system.
The automated flow
- Direct extraction into the target system’s fields.
Human stays in: Verification against a confidence threshold. Low-confidence extractions go to a human queue.
What goes wrong: Silent errors on fields that nothing downstream validates. Wherever possible, pair this with a check that catches mistakes automatically, such as a total that must reconcile.
Best for: Insurance, logistics, healthcare administration, and finance operations.
16. CRM hygiene
Doing it manually today: Records go stale. Contacts leave companies. Duplicates multiply. Reports get less trustworthy every quarter until someone declares the CRM useless.
The automated flow
- Scheduled identification of duplicates, stale records, missing fields, and contacts whose details no longer validate, with
- suggested merges and updates.
Human stays in: Approving merges. Automatic merging of customer records without review is a reliable way to destroy data.
What goes wrong: Aggressive deduplication merging two genuinely different people with similar names.
Best for: Any business whose CRM is older than three years.
17. Expense processing
Doing it manually today: Employees photograph receipts, someone categorises them, and finance chases the missing ones.
The automated flow
- Receipt image to structured expense line with merchant, amount, date, and category, matched against card transactions and
- flagged for policy exceptions.
Human stays in: Approval of exceptions and anything above threshold.
What goes wrong: Category misassignment, which creates tax and reporting issues later rather than obvious errors now.
Best for: Businesses with more than about twenty people submitting expenses.
Internal Operations
18. Meeting summaries and action items
Doing it manually today: Someone takes partial notes, or nobody does, and the decisions get relitigated three weeks later.
The automated flow
- Transcription, then a structured summary with decisions, action items, owners, and deadlines, pushed into the task system.
Human stays in: The meeting owner should review before the summary is circulated, particularly for anything involving commitments.
What goes wrong: Summaries that record what was said rather than what was decided. Prompt specifically for decisions and owners, or you get a transcript with extra steps.
Best for: Everyone. This is among the lowest-risk and fastest-payback automations available, because the output is checked by people who were in the room.
19. Internal knowledge search
Doing it manually today: Employees ask colleagues questions that are answered somewhere in a document they cannot find. The cost is two people’s time instead of one, repeatedly.
The automated flow
- A retrieval system over your internal documentation that answers with citations pointing to the source document.
Human stays in: Citations are mandatory. Without them, staff cannot verify and will not trust the system.
What goes wrong: Answering from outdated documents with total confidence. Old policies are the primary hazard. Version control your source material and exclude superseded documents from the index rather than hoping the model notices the date.
Best for: Organisations above roughly thirty people, or any business with substantial procedural documentation.
20. Employee onboarding
Doing it manually today: Onboarding quality depends entirely on how busy the manager is that week.
The automated flow
- Account provisioning triggered, a role-specific reading path assembled, a structured schedule created, and a persistent assistant
- available for the new starter’s routine questions.
Human stays in: The relationship parts. Automate the checklist, never the welcome.
What goes wrong: A polished process that feels impersonal, which produces a worse first week than a chaotic but human one.
Best for: Businesses hiring more than one person a month.
21. Recruitment screening support
Doing it manually today: Hundreds of applications, reviewed inconsistently under time pressure.
The automated flow
- Structured extraction of stated qualifications and experience against the specific requirements of the role, producing a summary
- per candidate.
Human stays in: Every rejection decision. This is the item on this list with the most serious risk attached, and it deserves plainer language than the rest.
What goes wrong: Automated screening can encode and scale discrimination, and doing so exposes you to genuine legal liability in most jurisdictions. Employment law in this area is active and varies significantly by country and, in some places, by city. Use these tools to summarise and organise applications so humans can review more of them properly. Do not use them to reject candidates automatically, and take actual legal advice before deploying anything in this category.
Best for: High-volume recruiting, with the constraints above treated as binding rather than advisory.
22. Reporting and analysis narration
Doing it manually today: An analyst pulls the same numbers every month and writes the same commentary explaining them.
The automated flow
- Scheduled data pull, automated calculation, and a drafted narrative explaining what moved and by how much.
Human stays in: The calculation must be deterministic. Let the AI write the explanation, never the numbers.
What goes wrong: Invented causation. A language model will happily explain why a metric moved without any evidence for the explanation. Constrain the narrative to describing what changed, and let humans supply the why.
Best for: Any business producing recurring management reporting.
Content and Monitoring
23. Content repurposing
Doing it manually today: One good piece of content gets published once and never adapted for other channels, because adaptation is tedious.
The automated flow
- A long-form piece converted into channel-appropriate formats, each drafted natively for its destination rather than truncated.
Human stays in: Editing before publication. Always.
What goes wrong: Mechanical repurposing that produces recognisably thin derivative content, which damages rather than extends the original. The source material has to be genuinely good for this to be worth doing.
Best for: Businesses already producing substantial content that is under-distributed.
24. Social media drafting and monitoring
Doing it manually today: Posting is inconsistent and mentions go unnoticed until something escalates.
The automated flow
- Drafted posts from an approved content calendar, plus monitoring of brand mentions with sentiment flagging and alerts on
- anything negative gaining traction.
Human stays in: Approval before publishing, and all responses to criticism.
What goes wrong: Tone-deaf scheduled posts going out during an unrelated crisis. Build a global pause switch and make sure someone knows where it is.
Best for: Consumer-facing brands. The monitoring half is considerably more valuable than the drafting half for most businesses.
25. Threshold monitoring and alerting
Doing it manually today: Problems are discovered when someone happens to look at a dashboard, which is usually after they have become expensive.
The automated flow
- Scheduled check across inventory, pipeline, cash position, or service metrics
- compare against thresholds and historical patterns
- identify anomalies
- draft an alert explaining what changed and what is affected
- notify the accountable person with the relevant context attached
Human stays in: Deciding what to do. The automation notices; the human acts.
What goes wrong: Alert fatigue, which is the near-universal failure mode. If people start ignoring the alerts, the system is worse than nothing because it creates false confidence that someone is watching. Tune thresholds deliberately and review the alert volume monthly.
Best for: Businesses holding inventory, managing cash flow tightly, or running services with commitments attached.
What Not to Automate
Some of the most heavily promoted AI automations are, for most businesses, a poor use of money.
Fully autonomous customer-facing responses on sensitive topics.
Billing disputes, cancellations, complaints, and anything involving money moving in the wrong direction. These are precisely the conversations that determine retention, and precisely where confident errors are most expensive.
Bulk AI content production for search visibility.
Publishing large volumes of generated pages is explicitly against Google’s scaled content abuse policy, and Google’s own documentation states plainly that a high quantity of pages does not make a site higher quality or more relevant. This one is not just low-value; it carries downside risk.
Automating a process nobody has measured.
If you cannot state the current cost and error rate, you cannot demonstrate improvement, and you will end up arguing about whether the project worked based on impressions.
Anything where the rules are fully deterministic.
If you can write the rule down completely, write the rule down. A traditional workflow automation costs a fraction as much, runs in an afternoon, and executes correctly every time. Adding a language model to a task that does not require interpretation adds cost, latency, and a new failure mode.
Exception handling, before the main path works.
Exceptions are where the judgement lives and where automation is hardest. Automate the routine ninety percent and route the rest to people. Attempting the reverse is the most common way these projects stall.
Where to Start
If you implement one thing from this list, make it meeting summaries or pre-call research. Both are low-risk, fast to deploy, checked by someone who knows whether the output is right, and they build organisational confidence without exposing customers to anything.
If you implement one thing that materially affects the profit and loss, look at invoice processing or customer reactivation. Invoice processing has a natural error-catching mechanism built into it. Reactivation works on customers you have already paid to acquire.
Then measure. Before you automate anything, write down how long the task currently takes, how often it happens, and how often the human doing it makes a mistake. That last number matters more than people expect, because AI systems are routinely held to a standard the existing manual process does not actually meet. If your team miscodes four percent of invoices today, an automation at two percent is an improvement, not a failure.
The businesses that get real value from this are not the ones that automated the most tasks. They are the ones that picked three or four workflows where the arithmetic was obviously favourable, implemented them carefully with humans in the right places, and left everything else alone.
Frequently Asked Questions
Which of these can I implement without a developer?
Meeting summaries, review monitoring, content repurposing, expense capture, and basic lead enrichment are available as configured products. Anything that writes into your core business systems generally needs technical involvement.
How much do these cost to run?
Highly variable, but the useful comparison is not the software cost. It is the software cost against the loaded hourly cost of the person currently doing the task, multiplied by frequency.
What if my data is messy?
Then start with the automations that do not depend on your data being clean, such as meeting summaries or document extraction. Use the time to fix the data, which has value independent of any AI project.
Should I build or buy?
Buy for anything standard, such as transcription, expense capture, or scheduling. Build only where the workflow is genuinely specific to your business and represents real cost.
Sources and Further Reading
Google Search Central, Optimizing your website for generative AI features on Google Search . Scaled content abuse policy and the position that page quantity does not improve site quality.
Google Search Central, Spam policies for Google web search . Scaled content abuse definition.
McKinsey & Company, The State of AI in 2025 (November 2025). Functional adoption patterns across business areas.
Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025). Cost and governance as the dominant causes of failure.
The tasks worth automating are rarely the impressive ones. They are the invisible, high-frequency, low-judgement jobs being done by people whose time is worth considerably more.
The GosAI
If you would like an honest assessment of which of these twenty-five would actually pay for itself in your business, and which would not, get in touch with The GosAI.
