How to Implement Artificial Intelligence Safely

1.10.2024 Reading time: 11 min. Like many others, in recent years we have been closely following AI and introducing AI assistants into our processes.
We covered this, for example, in one of our previous articles→. And here is the paradox we noticed.
Artificial intelligence can be useful in almost any field, from consulting to agriculture.
Up to 97% of executives think ChatGPT could also be useful in their business, but... there are not many real implementations.
In this article, we look at why this happens and how to implement an AI assistant quickly, effectively, and safely.
According to research by Nielsen Norman Group, AI implementation already allows→ productivity gains of 13.8% for support agents, 59% for marketing and HR staff, and 126% for programmers. Here is how McKinsey & Partners estimates→ AI's impact on the global economy in the coming years:

If there are not enough people

Fairly recent labor market figures: 70% of CIS companies are short of staff! In some cases, this means "hands" for the least skilled work.

But more often, companies lack engineers, sales managers, developers, project managers... the people whose work should multiply the company's revenue. Do not count on the best-case scenario.

New employees with the right qualifications are not wheat; they do not grow in the fields in a single season. That means companies have to reshape their talent strategy to recruit and retain people as efficiently as possible: screen new specialists quickly and reliably; train them in the company's business processes quickly; keep the best people, including by offering them more interesting tasks and reducing routine work.

But let's imagine what this looks like in real life. Suppose you have an experienced procurement manager named Rinat.

He knows how to and enjoys closing the toughest orders: he looks for complex product configurations, checks supplier quality, negotiates discounts, and arranges deliveries. Because of him, there were never any production stoppages or unjustified costs. And now you decide that you need not just one Rinat like this, but at least two, preferably three.

So you assign him to train junior managers in all the techniques and methods that Rinat knows perfectly.

Now, instead of solving complex procurement problems, your best employee spends 90% of their time training new procurement staff, listening to their calls with suppliers, and giving improvement recommendations. A month passes, then two, and

Rinat submits his resignation.

When asked, "What happened?" it honestly says it lost interest.

It does not solve complex tasks like "how to get a basket of snowdrops for free in December"

, but trains junior staff, most of whom do not even make it through probation. This case is highly exaggerated.

But in reality, situations like this happen all the time.

Instead of working on interesting, high-margin tasks, the most qualified employees spend more than half their time on routine work, giving feedback, and explaining the company's business processes.

Or let's look at a process everyone knows: calls and online meetings

In practice, they have become standard across most company departments: from sales to finance, from marketing to manufacturing.

Traditionally, complex business processes are built around work calls: employees write follow-ups, schedule the next meetings, search email and review follow-ups from past meetings, analyze the quality of negotiations, and so on. As a result, employees spend about 31 hours a month on meetings alone, according to a study by Atlassian→. That is the average.

In some departments, the figure can reach 80-100 hours per person per month.

When most of the work is calls, handling them takes a great deal of time and effort.

All these routine, time-consuming tasks can be automated by handing them to a virtual assistant, for example an AI that can transcribe meetings and prepare minutes.

The GPT-4 language model, which handles such tasks with ease, was released in spring 2023.

But for some reason, few people still use AI assistants built on it.

The gap between "what should be" and "what actually happened"

To understand this, let's continue with the topic of call transcription.

By the roughest estimate, there are more than a hundred products on the market that can convert speech to text and generate short summaries from that text. TL;DV, Fireflies, Rewatch, Transcriptor - all of these and many other tools easily transcribe calls in major languages, including CIS. And if a sales manager decides to implement any of them, they can easily find a ready-made solution.

At first glance, it seems simple: buy the product, install it, and start using it. But not so fast.

As a rule, such ready-made solutions live separately from the company's infrastructure.

This means that call recordings will have to be uploaded manually to the transcription tool or a special bot will need to be added to every meeting; the recordings will be processed on an external server; all information from the discussion will be stored in that service and, most likely, the language model will use this information to answer other users' requests; the transcript and short protocol will need to be separately retrieved from the transcription tool and saved manually in your own systems.

Will this ultimately improve employee efficiency and eliminate routine work? Possibly, but not necessarily.

Will you be able to protect confidential data about your products, projects, and clients? Unlikely. Or will it not pass?

There are also problems adapting the product to a company's specific needs. For example, the transcription tool may not understand employees' technical jargon very well. A simple "API" can turn into either a "bun" or a "pawn." That means someone will have to reread the transcripts and edit them.

The tool is inconvenient to use.

It is not safe to trust it with confidential information.

It is impossible to use the output

So there is no point spending time and money on implementation, the sales manager concludes after running into all these inconveniences and risks. But is it always like that?

Even a perfect tool still has to be implemented properly

Let's figure out what "proper implementation" means.
Autumn, rain outside, mood is not great. You say, "Alice, turn on the kettle, I'll brew some oolong"
- and you do not need to check that the kettle shuts off at exactly 80 degrees.
You ask it to play music for a sad evening, and it picks the perfect styles and tracks for you.
You ask when the rain will stop, and it does not ask which time zone to use. "Alice" is properly integrated into your life: it knows your preferences and processes. AI should work the same way.
You and your employees should not have to recheck its work every time and repeat a million actions.
Let's take the same call transcription process and break down what proper AI implementation for call handling consists of. Warning:
All screenshots and GIFs used in this article come from real interfaces KT.Team works with.
This is what an implemented AI tool looks like: easy to use and easy to get any data from.

You do not have to break your processes: AI understands and takes them into account

What does your company's call handling process consist of?
An employee schedules a call in the calendar. The meeting takes place in Meet or Telemost.
After the meeting, all participants must receive the minutes.
But employees must receive the full transcript, while external users can be sent only a brief minute. Corporate emails may forbid mentioning specific numbers or client names.
Transcripts must be clearly tied to a specific client or project, and the format of protocol emails must match the company's internal standards.
After the meeting, both the transcript and the minutes should be uploaded to Bitrix in the project charter.
Or into the client's record in amoCRM, depending on the type of call.
Off-the-shelf solutions usually do not account for these specifics.
They are good at only two things: transcribing and writing minutes.
Everything else will still require a person, which means call handling time will not decrease, and neither will the amount of routine work.
If an AI assistant is implemented properly, it handles everything that can be described through prompts and integrations.
It will do more than transcribe the call word for word: it will prepare minutes according to the standards used in your company and send them.
It checks the employee list and determines whether there were external participants in the meeting.
It follows the integration, checks the participant list and subject, and saves the call recording and transcript to the correct client record.
What does an AI assistant need in order to do all this?
Map out the process around the call and define all requirements, constraints, and steps. And, of course, hand all of these details over to the implementation team.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

You are confident that confidential information is not leaking

In spring 2023, Samsung discovered code fragments in ChatGPT's response. It turned out that programmers had been sending the bot their code and asking it to optimize it. The company completely banned employees from using third-party services at work. No one wants to end up like Samsung and find their scripts, NDA-protected project details, or client financial information in public access.

And when using a cloud AI tool, that is hard to avoid. The option of "installing AI on your own servers and isolating it from the outside world"

- workable, but not the only one

For example, you can integrate a data anonymization service between your systems and the transcription tool, or define conditions and rules for encrypting confidential information. During implementation, the integrator will offer solutions that match your security protocols. A packaged product is the same for everyone.

AI speaks your language

We already wrote above about the API bun

Every field has its own terminology and slang. The language model behind the transcription tool is trained on the entire body of information available on the internet.

It does not understand context, so it randomly picks from many similarly sounding words.

What does a properly implemented AI transcription tool look like?

This is a tool that understands your exact terminology and slang

To achieve this, the AI needs to be fine-tuned on your data, using existing texts, transcripts, and protocols.

You need to review the first few dozen transcripts and give the AI feedback:

yes
you were right to choose that word here
and then it got something wrong
in fact, such a term was mentioned

You will need to invest time and effort at the initial stages

For example, when we integrated an AI assistant for call handling at KT.Team, we fed it hundreds of recorded calls and transcripts. As a result, the AI confidently recognizes slang even in unconventional forms.

The tool does not require lengthy employee training

As an integrator, we have seen this situation many times: a company implements a complex, very well-designed tool that can improve performance in certain processes several times over, but in reality employees use at best a tenth of its features.
Everything else requires lengthy onboarding or training, which means no one uses it. AI tools are the same.
If getting a simple transcript requires 10 extra steps when scheduling the meeting and another 10 after it, you can be sure those steps will be ignored or sabotaged. Of course, you could tell employees to use it or leave, but that is not our approach, right? Proper AI implementation means the tool should be intuitive to use.
You already schedule meetings in your calendar, use conferencing software, and record calls, right?
This is enough for the AI to transcribe the call.
Do you already use any online spreadsheets?
This is enough for AI to organize calls and sort them into categories. Do you already have a messenger, personal or corporate?
This is enough to integrate an AI bot for working with call-related information.
And when we talk about a bot, we do not mean a complicated menu with ten levels of nested logic, but the ability to talk to data like you would to a knowledgeable colleague.
Ask it questions in plain language and get clear answers. For example:
What matters is that you should not be limited in choosing the tools behind the scenes.
The language model, the call method, the interface for storing call information, and the chat interface, all of these should be chosen by you.
If an integrator claims that this exact configuration is dictated by the IT tool and there are no alternatives, it may be easier to change integrators.

After each call, you get exactly the information you need

A packaged AI product already comes with a set of prompts built in:

how to transcribe calls
how to write a brief meeting minute
what to consider important
how to write a call subject

The problem is that these are generic prompts.

They are designed for everyone, which means for no one in particular.

But you have your own processes and needs

For example, you want the call subject to include the project name and the main agreement.

So that every item in the meeting protocol starts with a verb.

You need every sales call to be checked for script compliance.

To achieve this, you need to refine the necessary prompts during implementation together with the integrator team.

You define what you want to get, and the team fine-tunes the prompt wording.

You can create a prompt for a short call summary so it always contains only the most important points: decisions, plans, and ideas.

You can also create separate queries to assess what happened during the call, for example how friendly the overall tone of the conversation was.

What comes next?

After implementation, the AI assistant will become a helper for you and your team.
But processes change and are refined.
Does this mean the assistant becomes less useful after every change: it stops giving you answers to the questions you need, writes minutes incorrectly, or cannot exclude the users that need to be excluded from a mailing list? No.
Proper implementation of an AI assistant means that making any changes is easy.
It is not a black-box system that leaves you only two options, to put up with it or tear it out, but a flexible architecture.
During implementation, the integrator must show you where the prompts for all actions built into the AI assistant are stored and how they can be changed; how to add new call categories and scoring rules.

Key takeaways

So a properly implemented AI tool should not dictate changes to your processes, and it should not be rigid.

Discuss with the integrator: which security rules must be followed - the integrator will propose technical implementation solutions; where meeting information should come from and where the transcript should go, and according to which rule; what you want to know about a call besides its content; how you plan to use call information in the future; which corporate tools you already use and what information should appear

which of these tools should receive the data and according to what rules; which interfaces for working with summaries, minutes, and call statistics are most convenient for you.

How to implement artificial intelligence in business effectively and safely

If there are not enough people

Or let's look at a process everyone knows: calls and online meetings