MMOOZZGG.com

Why AI does not work

What to expect from AI implementation: realistic and unrealistic goals

5 min read2 sources

What to expect from AI implementation becomes clear from tasks, not from presentations: AI takes routine work, and people keep the hard cases. According to a 2025 MIT NANDA report, 95% of organizations saw no measurable return from generative AI, so set the goal before the start and in numbers. Below: what AI does well, what stays with people and how to set a goal you can check.

On this page · 4 sections

What AI does well

AI does well with work that repeats hundreds of times a day and follows rules: taking a request and entering it into the CRM, booking a customer, answering a standard question, building a report from several systems.

It answers around the clock, with no days off and at the same speed when volume grows. A request at nine in the evening gets an answer at once, not in the morning.

It does not get tired or forget the rules by the end of the day. The answer to the hundredth request is the same as to the first.

It prepares drafts: an email, a commercial proposal, a check of a contract against company details, meeting minutes. A person checks and sends them.

It finds answers in large volumes of text. Contracts, instructions, a year of correspondence: it finds the right clause in seconds if it has access to these documents.

The sign of a suitable task: you can describe it with rules, and you can check the result. If a strong employee can explain to a newcomer how to do it, AI can usually do it too.

What stays with people

Complex negotiations, a dispute, a complaint, a large customer with unusual terms. AI passes such cases to a person together with the dialog history, so the manager does not start from zero.

Decisions about people and money: hiring, evaluation, dismissal, payments. An agent can prepare and calculate, but a person must send the money and make the decision.

Work without rules: a new product, a one-off task, a process that lives in one person’s head. First it gets described, then it goes to AI.

Strategic decisions also stay with people: which product to launch, whom to work with, where to invest. AI can collect information and options, but the manager makes the choice.

Responsibility also stays with people. The agent works on behalf of the company, and the company answers for its reply to a customer, so a person approves the agent’s rules.

So in practice the question “should we replace a manager” sounds different: how many managers you need if an agent takes the flow of standard requests. A comparison by money and speed is in our analysis “AI agent or manager”.

Where inflated expectations come from

Demos show a lucky case: a clean question, a ready answer, no exceptions. In a real process, many requests do not look like the example.

AI sellers promise to “replace a department” and do not say how much work the rules, system connections and error review will take in the first weeks.

News stories write about models that pass exams and write code. A company applies this to its own processes and expects the same from the bot on its website.

Another source: the expectation that AI will understand the process by itself. In fact, an agent needs your rules, prices, terms and sample conversations; otherwise it answers correctly and uselessly.

The result shows in the numbers: according to the MIT NANDA report, 5% of custom AI tools reach working use. The rest never get there, and the company never learns what AI can do for its business.

Inflated expectations cost a lot: the project is closed after the first errors, although it could have paid off on routine work.

How to set a realistic goal

1. Choose one process with routine work and a large flow: requests, bookings, support questions.

2. Measure it before the start: how many requests a day, how much time an answer takes, how many people are busy with it.

3. Write the goal as the share of work AI takes. For example, for an AI support agent the pilot goal can be up to 80% of standard questions without people. This is a goal you check by measuring, not a promised result.

4. Decide in advance what to do after the results: keep, expand or stop. How to choose a metric and write it into the contract is covered on the page about the metric before the start.

5. Appoint a person who checks AI answers every day in the first weeks and fixes the rules. Without this, the goal stays on paper.

A realistic goal sounds duller than promises: not “replace a department” but “take part of the standard questions off the operators and answer requests at night”. But you can check it and count it in money.

If you have several processes and do not know where to start, the AI solutions map sets the order: where AI brings money, with which solution and what it costs.

Sources and fact-checking

Checked by the editors on 13 October 2026.

  1. MIT NANDA: The GenAI Divide, State of AI in Business 2025 (PDF) (mlq.ai)
  2. Virtualization Review: MIT report finds most AI business investments fail, reveals GenAI Divide (virtualizationreview.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us