What Can AI Automate in My Business: 6 High-Impact Areas

Most businesses automate the wrong things. They pick trivial tasks or force automation where human judgment matters, then wonder why the investment doesn't pay off. The question isn't whether AI can automate work - it's which processes deliver measurable ROI fast enough to justify the effort and cost.

This article identifies specific workflows that are proven automations across sales, operations, customer service, and finance: lead qualification that cuts sales cycles, invoice processing that eliminates data entry errors, customer inquiry routing that improves response times, inventory forecasting that reduces waste, contract review that speeds legal workflows, expense reconciliation that closes books faster, scheduling that eliminates coordination overhead, and compliance monitoring that catches issues before they become violations. These processes can be implemented with current technology and can pay for themselves within months.

Automating customer support triage

AI triage systems read every customer ticket the moment it arrives, extract intent and entities like product names or error codes, search the knowledge base for relevant solutions, and either automatically send perfect-match answers or route complex issues to the right department with full context already pulled, eliminating manual queue sorting and reducing first-response times from hours to seconds.

Many support teams today devote substantial time to ticket triage, i.e. determining where each ticket should be assigned and waiting for the assigned staff member to respond to the ticket. This wasted time can be avoided.

AI triage is NOT about getting humans out of the loop entirely. It is about the AI reading every ticket the moment it arrives and then either immediately solving the problem or routing it to the right human being with all context already pulled for them. Modern LLMs excel at deciphering messy customer-generated content to work out intent and then matching that intent to a human-created and structured knowledge graph (i.e. database of known fixes for known problems). It does not have to be perfect - it just has to be faster and more consistent than a human manually clicking through a support queue at 9am on a Monday.

A good way to process AI-powered customer service tickets is to set up a system that processes all of the tickets as they are submitted through APIs (such as email, chat, forms, etc). Within the body of each ticket, the system processes the intent and entities (such as product names, account IDs, error codes, etc) to classify the tickets by level of urgency (based on language patterns it has been trained to spot).

Once the intent of a customer's question or problem has been established by the AI, the system can proceed to search the organization's knowledge base for documents, past solutions to similar problems, and product guides relevant to the customer's query. The knowledge base can be structured in many different ways, but for searching by intent it is ideal if the knowledge base is stored as a database of vectors which can be searched using vector search methods. The AI reads through all of the potential solutions found by searching the organization's knowledge base and establishes a potential solution to the customer's problem. The solution can then be automatically sent to the customer if it is a perfect match for their question or problem. For example, if the customer has asked how to reset a password and the solution found by searching the knowledge base is a step by step guide to resetting a password, then that guide can be automatically sent to the customer. A solution that is not a perfect match can still be automatically sent to the customer after review by a human for verification before it is sent to the customer. Alternatively, the AI could have established a summary of the customer's question and then passed it along to the correct department (i.e. sales, technical support, billing, etc.) for human assistance in providing a solution to the customer's problem.

The ticketing system will be an API in for the LLM (email, chat, forms, etc.). And the biggest component will be the knowledge base, which must be accurate enough to trust. The AI part is a very small part of this whole system.

Automating data entry and CRM enrichment

AI monitors email inboxes, Slack channels, or uploaded files for prospect information, uses OCR or parsing to read text from any format, extracts entities like company names and deal size with LLMs, validates against existing CRM records to prevent duplicates, and writes clean data to the CRM via API in seconds, reducing per-record processing costs and eliminating typos and missing fields.

Reading in messy text and then writing out clean CRM records is one example of work automation in businesses today. A sales person's email inbox could be automatically read for new prospects.

Reading in all kinds of messy formats and then automating the writing out of clean CRM data is one of the clearest answers to the "what can AI automate in my business" question.

So, let's get into the implementation details. First, a monitored source of data is needed. This could be an email inbox, a Slack channel, or a folder full of uploaded files (for example, scanned contracts). As new items are added to the monitored source, the AI reads the text within them. This could involve the use of Optical Character Recognition (OCR) to read images of paper, or simply parsing digitally created documents. The text is then passed to the LLM to determine the entities of interest (e.g. company name and contact details of person who signed the contract, size of deal, priority, contract terms, etc.). The extracted data is then validated for things like duplicate entry or conflict with an existing CRM record(s). Finally, the AI writes the clean, validated data to the CRM system via API.

The whole loop runs in seconds.

LLMs can achieve high accuracy for reading very clean text in a very well defined format and then inserting all of the entities in the correct fields in a CRM for a single deal. However, for LLMs reading handwritten notes that are often very vague and also are outside of the LLM's training data, accuracy falls rapidly for all LLMs on the market today. To combat this, it is recommended that organizations set aside time each month for a human to spot-check all of the information inserted by an LLM for accuracy. This human will also be responsible for making any necessary changes to ensure that everything is correct and up to date. In addition to spot-checking work for accuracy, the human should also set up the rules for the LLM for extraction work in order to better read and extract text. This work will require a period of training and refinement per LLM and complex schema.

Employees typically spend time adding information to CRM records, including fixing typos, missing fields and poor formatting. In contrast, AI processing can run at a lower per-record cost depending on complexity and API usage. Thus for organizations processing large volumes of records per month, there can be sufficient savings to offset the cost of setup within a few months. If processing very few records then the cost of setup for automating such a small amount of data entry work typically will not be worth it unless there is a very large and growing data entry workload.

There are so many businesses that deal with vast amounts of information gathered from sales calls, contracts processed by legal or procurement, support tickets, and so much more. The fact that there is so much manual work going into entering information for a CRM and that people are complaining about the current process is enough indication that this particular problem can be solved via automation.

Automating document processing

AI document extraction uses OCR to convert scanned invoices into text, applies LLMs to identify vendor names, totals, dates, and line items, maps extracted data to the correct ERP fields via schema, and routes high-value or low-confidence items to human approval queues, eliminating manual data entry for high-volume accounts payable workflows at a per-page API cost.

Many businesses in accounts payable open up dozens of PDFs every month, read the line items for each, type in the numbers in their accounting system, find the total and mark up any mis-numbered total lines. This process can be automated by having AI document extraction read in the invoices and pull out the key information such as the vendor name, the total dollar amounts for each invoice, the date the invoice was created and all of the line items for each invoice.

There are several technologies that are used within AI Document Extraction in order to process documents. First, the scanned documents are processed using Optical Character Recognition (OCR) that converts the scanned document into text. Next, an LLM (Large Language Model) is used to read the text in the document in order to determine which parts of the document should be extracted, in what order the extracted data should be presented, and how the extracted data should be mapped to the correct fields in the user's ERP (Enterprise Resource Planning) system. The extracted data is then put through a schema that maps the extracted data to the correct fields in the user's ERP system. The typical cost for using AI Document Extraction is charged by the page for API calls, as well as integration costs to integrate the document storage with the user's accounting platform.

For very high volume of similar structured documents (e.g. dozens of invoices, hundreds of receipts, etc.) the math only works if processing them manually is too expensive.

Simple invoices and receipts: For simple invoices and receipts, such as purchase orders, W-9s and standard forms of other documents like NDAs, simple document extraction can achieve high accuracy provided the documents are good quality clean scans. Even with slight changes in layout for very similar documents the model is able to recognize strings of text, for example "Invoice Number:" followed by numbers. Simple document extraction is reasonably robust for extraction of information from similar documents.

Difficult to extract: Very redacted documents. Poor quality scanned images (faded thermal paper, crumpled paper, etc.). Non-standard documents (e.g. embedded key data within large body of text). Complex legal documents (interpretation required). A 3 page services agreement with interrelated clauses and conditions would likely require human review whereas a simple vendor invoice with a subtotal, tax, and total would be automatically extracted and approved.

Accounting AI document extraction also has errors which follow a predictable mode of failure. A threshold can be set on these errors and have them automatically route the document to the approval queue for example if the total amount on an invoice is above a confidence threshold. These thresholds will typically get tweaked every few months as the system learns and starts to process more clean data on a monthly basis. This typically starts out conservative and gets looser and looser over time.

As a first pass, the model would be tweaked based on how much risk an organization is comfortable with. Usually over the course of a couple of months, until the accuracy reached a stable plateau. After that, the amounts in the individual fields would simply be automatically approved.

Automating meeting notes and summaries

Real-time speech-to-text APIs like Deepgram and AssemblyAI transcribe Zoom meetings as they happen, LLMs such as GPT-4 and Claude summarize conversations into bullet points with action items and decisions, and the formatted output is automatically sent to Salesforce, HubSpot, Asana, or Slack, saving sales reps manual note-taking time and improving CRM accuracy and pipeline forecasts.

The APIs process real time transcription from Zoom meetings into a summary of the meeting, the action items from the meeting, the decisions made in the meeting, etc. This can be output in simple bullet points or as a number of CRM fields to be automatically logged into a CRM.

The output can then be sent to any area of a business where it is needed. Salesforce, a project management tool or even a Slack channel for example. All of this can be done in the background whilst on the call.

First, real-time speech-to-text APIs from Deepgram and AssemblyAI produce live transcription of speech in real-time. Whisper (speech-to-text model released by OpenAI) is extremely fast and accurate on already recorded audio but not suitable for live streams.

The transcription is then passed to the LLM in the API, which can even be set to create a summary of the conversation in bullet points. APIs such as GPT-4 and Claude are particularly well suited to these types of processing tasks, because they are able to keep a very strong memory of very long conversations and summarize them effectively. The processed output can then be requested to be formatted in a wide variety of ways and sent to APIs such as Salesforce, HubSpot, Asana, and Slack to add to projects and records etc.

The processing cost for real-time transcription will vary by provider and usage volume, but for most teams the value of automatically documenting each meeting far exceeds the cost of the API calls.

Sales teams get the biggest ROI from this type of processing because every call is logged in their CRM automatically. The actions and next steps for deals are also logged by the API automatically. This can save sales reps time after every meeting that would otherwise be spent writing notes to a CRM manually.

However, there are even greater benefits of automatically transcribing calls. For one, there is immediate transcription and therefore greater accuracy because it has been written whilst the call is still fresh, rather than several days later for example when many sales people record notes for previous calls. Additionally, by accurately recording calls in a sales CRM all the time, sales forecasts and pipeline will also improve allowing sales people to be more effective with their time.

Automating email sorting and response drafting

AI connects to email APIs like Gmail or Outlook, uses LLMs to classify incoming messages by intent and priority, applies workflow rules to auto-reply to repetitive questions like password resets or invoice requests, and routes complex emails to the appropriate person with CRM context, saving hours daily for companies receiving high volumes of sales, support, and info emails.

The automation of high volume email is very time-consuming for companies, mostly for emails sent to sales@, support@, info@ and many more email addresses. After a while, these emails are mostly repetitive and can be answered automatically for a substantial portion of the emails that a company receives on a daily basis.

The setup of an email automation like this is generally pretty straightforward. First, connect the email API (e.g. Gmail, Outlook). Then, use the LLM to classify each email as it comes in. Then, set up the workflow with rules based on the email's category and other factors. For example, automatically send a reply or flag for human review with a label. The cost here is typically based on the number of email API calls per email, which for most companies is modest until reaching tens of thousands of emails per day.

However, in order to determine whether or not AI should be used to automate email in a company, current email needs to be examined for patterns. Are most of the emails in the inbox for things like 'reset my password' or 'where is my invoice'? If so, then using AI to automate the majority of repetitive emails will save hours every day. On the other hand, if the emails in the inbox consist of things like 'negotiate this deal', 'handle that customer', or 'deal with internal politics', then teams will just end up spending as much time reviewing the drafts that the AI generated as they would have spent just reading and answering the emails themselves.

An LLM is able to read subject lines and email bodies to determine intent, categorize emails, and score priority for each message. It can then forward the email to the appropriate person based on the message's labels, forwarding rules, or even CRM assignment.

The best email processing systems are also integrated with a CRM (customer relationship management) system, so the email can be processed within the context of the customer. Does this email belong to an existing customer? Is this a sales lead from last week's demo? These are the kinds of things that would be considered when handling a support question. A support question from a large customer would be handled differently from a support question from a trial user that signed up yesterday.

Teams commonly establish rules for routing messages to different categories, such as: new lead, billing questions, technical support issues, spam, internal emails (e.g. to and from other employees of a company). A useful rule of thumb is to start off with a manageable number of different categories for which there are good training examples. Over time, edge cases that are outside of the existing set of rules will emerge for which more can be done. But this is not a "set it and forget it" type of system. It will continue to need to be fine-tuned in order to have greater and greater accuracy for routing emails.

Automation of the processing of distinct email types is effective and efficient, but email that could fit into multiple categories (e.g. complaints) and fall within different parameters depending on the view of the reader does not work as well.

Where AI automation fails

AI should not automate decisions that require human judgment on pricing strategy, entering new verticals, B2B sales relationships, legal contract reviews, or employee management, because these tasks carry significant reputational, financial, or legal risk and involve strategic context, stakeholder complexity, and brand voice nuances that AI-generated first drafts fail to capture accurately.

Pricing, new verticals, whether to raise prices, all sales (B2B in particular), all legal reviews, all employee management - all of these need a human to make the decisions. AI can be used to produce a first draft of a pricing proposal or even of notes from a sales call.

Work that carries significant reputational, financial or legal risk will take more time to correct and fine-tune AI generated first drafts than it would have taken to complete the work from scratch in the first place. Thus, while AI can greatly accelerate research intensive work and be used to complete the first draft of such work, it is not appropriate to have AI complete work that will be sent out under one's name. Accountability for the work that is sent out is essential.

Just because a computer can process a lot of information automatically does not mean that the computer has the strategic context to automatically make good decisions with that information. Deciding whether or not to raise prices, or whether to try to sell into a new vertical, are strategic decisions that require human judgment and a deep understanding of the marketplace and customers. They are not things that can be automatically determined by AI.

Even when an AI models out scenarios based on the data provided to it, someone has to decide what tradeoffs to make and who should be making them.

Automating relationship-based sales does not work. Although AI may be able to automatically generate a first draft of a sales pitch (in the form of a deck) and even automatically follow up with customers, the reality is that most sales have many different stakeholders within an organization. As an example, a CFO may be worried about locking into a vendor, an IT department managing vendor relationships and a procurement department owning the process and making all the purchasing decisions. This is a complex process that can't be replicated by AI.

Work that will sound like a brand's voice - creative work, really - is another area where companies may be surprised at just how far off the mark AI first-drafts are. Most companies do a poor job of gauging the importance of their brand's voice until they see the work that has been done for them by AI, side by side with the edited work that has been done to get it to sound more human.

Frequently Asked Questions

What can AI automate in a business right now?

AI can automate customer support by creating a chatbot that can handle tier 1 questions that support teams usually have to answer. AI can automate lead qualification, scoring and preparation for meetings with customers. AI can also support teams with content creation (first drafts of emails, documents, social posts etc.), data entry (e.g. from documents into CRM) or automated invoice processing (e.g. automatically extracting line items and then routing them for approval).

What AI automations should businesses start with if they're new to this?

Two of the most obvious for new AI users are document processing and customer support. Both of these areas have clear success metrics and don't require a lot of custom training.

What are the approximate costs of process automation with AI?

For the Claude API, costs will depend on the features being utilized. Claude API costs will vary per million tokens per month to service conversations per month. This would result in modest monthly API costs.

Custom development, for example building an automation, then connecting it to a CRM, then building a front-end for the automation, then training the model on the data in the automation, can vary widely in cost depending on scope and complexity.

Which business processes should NOT be automated with AI?

Processes that require human judgment on rare or unusual events should not be automated. Processes such as firing employees, reviewing contracts, or assessing financial risk of customers are not yet ready to be automated. AI occasionally "hallucinates" or makes up information that is not in the data that the AI was trained on.

Can AI replace sales teams or customer success teams?

No. Automating the admin work that sales teams currently do on a daily basis will be the best use case for AI. They can then focus on building a relationship with customers and close deals.

How do organizations determine if their business is ready for automation with AI?

Business processes that currently consume substantial hours of a person's time, that have relatively clean data, that can be trained on records (in a CRM, for documents, etc. that are roughly accurate), and for which there is a clear measure of success (e.g. reduction in response time, reduction in errors, processing faster) are good candidates. This only works for very repeatable processes that are stable and for which sufficient and good data is available.

Gable Innovation is a technology consultancy that helps growing businesses assess, select, and implement the right CRM, AI, and automation tools.