- Today’s AI excels at four tasks: reading text, writing text, classifying information and extracting data from documents. Most other claims deserve closer scrutiny.
- AI does not replace judgment. It removes repetitive work around decisions that still require judgment.
- The best starting point is not a technology. It is a repetitive task that consumes hours every week.
- Most AI projects fail for ordinary reasons: the wrong use case, disorganized data or a tool that no one wants to use.
- If your board asks the company to “do AI,” answer with a question: which process, exactly?
Why this guide exists
AI is discussed everywhere, but the advice is rarely consistent. A provider promises to cut your costs in half. An article says your profession is disappearing. An employee shows you an impressive demo. Meanwhile, you have to make a sound decision with a real budget.
This guide is for that decision. Its purpose is to help you ask better questions of anyone proposing an AI solution, GXN included.
I have spent more than twenty years implementing technology and a decade working with these systems while pursuing graduate studies in artificial intelligence. This guide draws on real projects, real constraints and real businesses, including projects that did not work.
What AI Really Does
Set the spectacular demos aside. In an ordinary business, modern AI is useful across four main families of tasks. Not fifty. Four.
It reads. Give it a 40-page contract and it can locate the renewal clauses. Give it 200 emails and it can identify complaints. It can work with unstructured, imperfect and multilingual text written by people. That capability changes what businesses can automate.
It writes. Summaries, draft replies, rewrites, translations and meeting notes. The output is useful for preparatory work, but final work still benefits from review. A fast first draft that a person corrects in two minutes often beats writing a polished version from scratch.
It classifies. It can sort incoming requests, categorize documents, route a file to the right person and flag unusual cases. The work is not glamorous, but it can produce meaningful returns.
It extracts. It can pull specific figures and fields from invoices, forms, purchase orders and reports. For many small and midsize businesses, this is among the most valuable use cases, even if it makes for an unremarkable demo.
The common thread is text and documents. If a process contains neither, current generative AI may offer little value. Claims to the contrary should be examined carefully.
Text and documents
| Task family | Appropriate use | Human checkpoint |
|---|---|---|
| Read | Summarize and locate information | Verify the sources |
| Write | Prepare a first draft | Approve the final message |
| Classify | Sort repetitive work at scale | Handle ambiguous cases |
| Extract | Structure fields from a document | Validate critical data |
What has really changed
Businesses have used automation for decades. So what has changed?
Traditional automation required structure. Data had to fit predefined fields in a form, database or fixed format. When information arrived as a free-form email, scanned PDF or voice note, the automation stopped and a person had to take over.
That barrier has weakened. Current models can work with unstructured information, which makes it worth revisiting processes that were considered impractical to automate a decade ago.
The second change receives less attention: cost. Machine processing can now cost only a fraction of a cent per document. Use cases that were uneconomical five years ago may be viable today.
The words you will hear
I'm keeping it short here, a complete guide exists on the subject.
- LLM or large language model. This is the engine that predicts text from text. ChatGPT, Claude and similar products provide interfaces built on top of LLMs.
- Agent. A LLM that is given tools and procedures to accomplish a task in several stages, without a human clicking between each step.
- RAG. A method that lets an AI model answer from your own documents instead of relying only on what it learned during training. It is a core technique behind AI-powered enterprise search.
- Hallucination. When a model presents invented information with confidence. This is not a temporary bug. It is a property of how the system works, so the solution must be designed around it.
Where AI creates measurable value
These are the use cases that repeatedly work well in organizations with roughly 10 to 150 employees.
- Processing incoming documents. provider invoices, purchase orders, client forms and field reports often contain information that someone manually copies into another system. In many organizations, this is the strongest first use case.
- Sorting and routing requests. A shared inbox receives everything, and an experienced employee spends part of each day routing messages to the right person.
- Internal research. Your procedures, contracts, estimates and historical records exist somewhere, but people cannot find them. Twenty minutes spent searching for an existing answer becomes expensive when repeated across a team.
- Repetitive answers. The same questions return with answers rewritten each time. AI can prepare a draft for a person to review before it is sent.
- File summaries. Prepare a decision from 60 pages of documentation.
Where AI is the wrong fit
The poor fits are just as important, but discussed far less often.
- High-stakes decisions without supervision. Credit approvals, diagnoses, hiring decisions and contractual commitments may be technically possible to automate, but doing so without oversight is strategically reckless. Your organization remains accountable when the system is wrong.
- Processes that change all the time. Automating an unstable process is like pouring concrete on sand.
- Small volumes. If a person handles only a dozen documents a month, leave the process alone. Automation may cost more than the problem it solves.
- Pure replacement of personnel. Projects designed primarily to remove people often fail because those same people hold the context the system needs. Strong projects free up their time and redirect it to higher-value work.
- Data that doesn't exist. AI cannot recover information that was never collected. It may instead produce a plausible but false answer, which is worse than having no answer.
Three questions every decision-maker should ask
When someone proposes an AI project, ask these three questions. They expose most weak ideas in a single meeting.
First question: what specific task is performed, by whom and how many times each week?
If the answer is vague, the project is still only an aspiration. “Improve customer service” cannot be built. “Prepare replies to 60 weekly order-status requests” can.
Second question: what happens when the system makes a mistake?
Every AI system will sometimes be wrong. What matters is the consequence. An error that takes three minutes to correct may be acceptable. An error sent to a client under your name may not be. This answer determines where human review belongs and shapes the entire architecture.
Third question: who will use this every day, six months from now?
Technically excellent systems can fail within months when they are built without the people expected to use them. Adoption is not a final project phase. It is a design constraint from day one.
The test bench of a use case
Explore a task. This reflection tool does not replace an assessment.
The volume is significant and the error remains correctable. Measure real time before choosing a tool.
How much does it cost, really?
A complete budget has four components, and two are routinely overlooked.
- Implementation. Development and integration. A simple internal assistant connected to your documents costs less. Automation integrated with three systems costs more.
- Use. Every processed document has a usage cost. It may be a fraction of a cent, but at your actual volume over twelve months it becomes a meaningful figure. Ask for that estimate before signing. A provider who cannot produce it does not understand the operating cost of the system.
- Maintenance. Models evolve, your systems change and answer quality must be monitored. Plan for a modest but real recurring maintenance effort.
- Training and adoption. Training and adoption are often overlooked, yet they frequently determine success. Count your team’s time, not only the provider’s hours.
A complete cost guide exists in this section, with ranges by project type.
What I see on the ground
The most common portrait, in 2026, in the Quebec companies that I visit.
The company may have no official AI initiative, yet several employees already use personal AI accounts with business documents without management knowing. They are usually resourceful people trying to work faster, not acting in bad faith.
For many businesses, this is the real starting point: not a strategy, but ungoverned use already underway with privacy risks that no one has assessed.
If this sounds familiar, do not start by buying technology. First understand what is already happening inside the organization. Employees experimenting on their own are revealing the workflow friction worth investigating.
When this advice does not apply
If your business has a more urgent problem, AI can wait.
- If your core data is inaccurate or scattered, fix that first. Connecting an intelligent system to poor data only produces poor answers faster.
- If no process is documented and no one agrees on how the work should be done, start there. A disagreement cannot be automated.
- If your digital platform is unstable or you cannot access your own systems, the foundation is not ready.
- If the volume is low, keep your money.
A sound assessment sometimes concludes “not this year.” In one such case, the client cleaned its database first and the project worked on the first attempt the following year.
A practical starting point
- List the repetitive text- and document-based tasks in your business. Ask the people who perform them, not only their managers.
- For each task, record the weekly volume and approximate time involved.
- Keep the three that consume the most effort.
- For each one, determine what happens when an error occurs.
- Choose the task that combines high volume with low risk. That is your first candidate project.
- Keep the scope small and visible: one process, one team and a result measured within weeks.
Large AI transformation programs that try to change an entire business at once often stall. Small, credible wins create support for the next project.
Frequently asked questions
01Will AI replace my employees?
AI shifts how time is spent; it does not replace judgment. In successful implementations, people spend fewer hours on tedious tasks and more on work the organization already values. Projects designed primarily to eliminate positions fail more often, both technically and organizationally.
02Is ChatGPT enough for a business?
For individual use, often. It is not enough when you need to connect business systems, control where data goes, produce consistent answers or deploy the tool to a team under clear rules.
03Is our business data safe?
Data safety depends on the architecture, not simply the provider’s brand. It is a design decision that belongs at the start of the project. A dedicated guide in this section explains the issue in detail.
04What is a sensible starting budget?
The initial scope should be large enough to test a real use case, measure the result and support adoption. Otherwise, you may fund an experiment with no path forward. An assessment can establish a defensible scope for your situation.
05Should we wait for AI technology to stabilize?
The technology will keep changing, but mature use cases such as document processing and internal knowledge search are already delivering value. A small, carefully governed project is often more useful than waiting for the market to stop moving.
06How long does it take to see a useful result?
A focused first project can produce measurable results within weeks or a few months. Be cautious of anyone promising a complete transformation in two weeks or asking for eighteen months before the first measurable gain.
This guide is the mainstay of the Artificial Intelligence category. Related content is organized between foundations, risks and governance, architecture and models, then investment and implementation.
References to verify for your situation
These references support the external rules and frameworks cited in this guide. They do not replace legal or professional advice tailored to your organization.