A person in a dark room looking at a large screen with blue lines resembling a neural network

AI in Business: Which Tasks to Start With and How to Measure the Payoff

Requests to "implement AI" usually reach us in one of two forms. The first: "Our competitors are already doing something with AI, so we need to as well." The second: "Three of our people spend all day sorting incoming emails. Can we speed that up?" The second almost always ends with a working tool. The first often ends with an impressive demo nobody uses a month later.

The difference lies in whether the conversation starts with technology or with a process. Below, we look at AI implementation in business from the task side: what to automate first, how to estimate the payoff before development, what to do about data and security, and how to set up a pilot so the budget isn't wasted.

AI implementation starts with a process, not with technology.

Which tasks to start AI implementation with

A good first candidate looks like this: the task repeats hundreds of times a month, it has a clear input and a clear output, and a model error can be spotted and fixed quickly. Five typical areas fit that description.

Routine writing

Product descriptions, templated replies, call summaries, meeting notes, first drafts of proposals. A language model writes the draft and an employee edits it. The gain isn't that you no longer need a person; it's that editing is faster than writing from scratch.

Customer support

The first line: answering common questions from a knowledge base, order status, suggestions for the agent right in the chat window. There are plenty of pitfalls here, from made-up answers to clumsy handoffs to a human.

Request classification

Incoming emails, requests, reviews, tickets: determine the topic, urgency and responsible department, and extract the contract or order number. This is one of the most reliable tasks for AI automation: quality is easy to check against historical data, and an error is usually cheap, since the email just gets moved to a different queue.

Document processing

Extracting details from invoices and statements, checking contracts against a standard template, searching internal policies. The model doesn't replace a lawyer or an accountant, but it takes over chores like "find the penalty clause in forty contracts."

Forecasting

Demand, customer churn, equipment failures. Here, classic machine learning models on tabular data usually do the work, not chat-based AI.

How to estimate the payoff before development

Run the numbers before the first line of code is written. For most tasks, a simple model with five parameters is enough:

Some effects are harder to quantify: faster responses to customers, coverage at night and on weekends, a consistent standard of replies. Capture them as measurable metrics (average first response time, for example), not as promises in a slide deck.

Risks that get less airtime than benefits

None of these risks is a reason to abandon the project. But each one belongs in the plan: who's responsible, how we notice, what we do.

Data: AI for business doesn't work without it

Which data you need depends on the task:

A typical picture at the start: data is scattered across email, Excel files and the CRM, fields are filled in however each person is used to, and change history isn't kept. Sometimes the first useful work on a project is getting data and integrations in order, and only then connecting a model. In essence, it's the same work as end-to-end development of a web application: map the processes, describe where data comes from and who owns it.

It's also worth assembling a test set: 100 to 300 real examples with reference answers. It's used to compare models, prompts and versions of the solution. Without one, the question "did it get better or worse?" is answered by gut feeling.

Data security and model hosting

The key question in any project: where does the data go when an employee or a system sends a request to the model? There are two hosting options, each with its own trade-offs.

Cloud models via API

A fast start, high quality on complex tasks, pay-as-you-go pricing, no hardware of your own. The downsides: data is sent to a third-party provider, and access restrictions or changes to terms are possible. If the provider processes data in another jurisdiction, cross-border transfer becomes an issue too. Major providers offer regional data residency and enterprise terms under which your data isn't used for training.

Your own infrastructure (on-premise)

Open-source models (the Llama, Qwen and Mistral families, among others) are deployed on company servers or in a rented private cloud. Data never leaves your perimeter, but you need GPU servers, expertise and time for setup. On complex tasks, mid-sized open models can fall short of large cloud ones, so check quality on your own test set rather than leaderboards.

Cloud via APIOwn infrastructure
StartFast, no hardware neededRequires GPU servers and setup time
PricingPay as you goOwned or rented servers
DataGoes to the providerStays inside your perimeter
Quality on complex tasksHighMid-sized models may fall short

Personal data and GDPR

If names, phone numbers or addresses of customers or employees end up in requests to the model, data protection law applies. In short, what to keep in mind: you need a lawful basis for processing and a data processing agreement with the provider; transferring personal data to a service outside the EU requires an adequacy decision or safeguards such as Standard Contractual Clauses; and you need appropriate security measures.

Practical techniques that resolve most of the concerns:

A pilot: testing the idea at low cost

A pilot is an experiment with a success criterion known in advance, not a stripped-down version of the product. Here's a setup that works:

  1. One task, one process, one user group. Not "AI for the whole company" but "routing emails from the shared inbox."
  2. A metric and a threshold before you start. For example, classification accuracy or the share of answers that needed no edits. Plus a value agreed in advance at which you scale, and a value at which you stop.
  3. A baseline. A measurement of how the process works today: time, errors, volume.
  4. A prototype on off-the-shelf models. No training your own model. This is enough to see what quality ceiling is realistic.
  5. A human in the loop. The model suggests, an employee confirms or corrects. That way you collect error data without putting customers at risk.
  6. Results review. Testing on the test set and on live traffic, and a decision based on the threshold set before the start.

"It doesn't work, we're stopping" is a perfectly normal pilot outcome too. Finding that out in a pilot is cheaper than after six months of full-scale development.

How not to waste the budget

If a process isn't documented and everyone does it their own way, the model will reproduce the same chaos, only faster.

At Sphere, a digital agency, we build custom AI solutions through the same stages as any product: first an analysis of the process and data, then design, development and testing, then launch and growth. You can see examples of the problems we've worked on in our projects.

FAQ

How much does AI implementation in business cost?

It depends on the task and how the model is hosted. The total is made up of process analysis, development and integration with your systems, model costs (for cloud API requests or your own servers) and support. A pilot on off-the-shelf models costs noticeably less than a full rollout, so it makes sense to start there.

Do you need to train your own AI model?

Usually not. For writing, support and document processing, an off-the-shelf language model, instructions and knowledge base search are enough. Training or fine-tuning your own model is justified when you have a large volume of domain-specific labeled data and off-the-shelf models don't reach the required quality on your test set.

Can you send customer data to ChatGPT?

Be careful with personal data: sending it to an external service means complying with GDPR or your local data protection law, including a lawful basis, a data processing agreement and, for transfers to other jurisdictions, appropriate safeguards. It's safer to pseudonymize data before sending it, use enterprise plans with regional data residency, or deploy an open-source model on your own infrastructure. Agree on the final setup with a lawyer.

Will AI replace employees?

In typical tasks, AI takes over the routine part of the work, while complex and non-standard cases stay with people. In practice, what usually changes is how the workload is distributed: the same employees handle a larger volume or take on tasks they never had time for before.

How much data do you need to get started?

For generating answers and searching documents, you need an up-to-date knowledge base. For classification, historical examples with correct answers, usually at least a few hundred. Either way, it's worth assembling a test set of 100 to 300 real examples so you can measure quality objectively.

Let’s discuss your project
Related reading