AI for business: where to start without wasting budget

Artificial intelligence for business starts with choosing two or three revenue processes where the pain already has a number, and buying tools comes after that choice. The company that hands out AI licenses to everyone in the first month pays the bill and cannot say what changed, because nobody measured the process before and nobody answers for its use afterwards. The path I follow with executive teams has six steps: choose processes with measured pain, organize the minimum data those processes use, write down the company’s context, define who approves each action, train the team and compare indicators before and after. With these steps in place, a 90-day pilot is enough for the executive team to decide where to expand the investment and where to cut.

At Iasbeck&Co we implement CRM and AI agents in B2B operations and train the teams that work with them. In recent months, the question from executives has changed: they no longer ask whether AI is worth using, and instead ask why the company already spends on AI and the result does not show up in the quarter’s close. I wrote this guide for the executive team, for sales, marketing and post-sale leaders and for HR, and at each step I point to the article in the series that goes deeper.

Why does starting with the tool waste the AI budget?

The scenario I see most often starts with the executive team approving licenses for an AI assistant across the whole company. IT grants access and each person uses it as they see fit: the salesperson asks for emails, marketing asks for posts, finance asks for spreadsheet formulas, and each one writes their own prompt with no access to the systems and no context about the company. Six months later, the bill has grown with the number of users and nobody can point to a process that became faster or cheaper. The individual gain exists, but it spreads across saved minutes that no operational indicator records.

Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. In the same release, the firm recommends pursuing agents only where they deliver clear value or ROI. The three reasons for cancellation map to decisions the executive team can make before buying: how much the company accepts spending per use, which indicator needs to change and who approves what the AI does.

Decision	Starting with the tool	Starting with the process
First step	Buy licenses and grant access	Choose revenue processes with measured pain
Context the AI uses	Whatever each person writes in the prompt	Company context, written and shared
Data	Whatever is at hand	The fields the process uses, reviewed
Who answers for the result	Nobody in particular	One owner per process and one approver per action
Measurement	Satisfaction survey about the tool	Process indicator before and after
Cost	Grows with the number of users	Grows with the uses that showed a return

How to implement AI in the company: six steps before scaling

The six steps fit in a 90-day pilot and do not require switching systems. The order matters: whoever skips choosing the process writes a generic context, and whoever skips the baseline reaches the end of the pilot with nothing to compare against.

Choose revenue processes with measured pain

Start with the processes that bring in or keep money, such as responding to the incoming lead, preparing the sales meeting, following up on the stalled proposal, triaging the ticket and the renewal meeting. For each candidate, ask the process owner for today’s number: hours spent per week, time to first response, share of leads that sales accepts. If the number does not exist, the process can still enter the pilot, and the first task becomes measuring the baseline for two weeks. Prefer high-volume tasks that work on text the operation already produces and where mistakes are cheap to fix.

Organize the minimum data

AI inherits the quality of the data it reads. Instead of opening a project to clean the entire database, which takes months and loses sponsorship along the way, fix the fields the chosen processes use: emails and calls associated with the right deal for meeting prep, category and priority as closed options for triage, an up-to-date job title for nurturing. I detail the effect of a messy database on agents in AI in revenue operations. A quick test for an executive: open the report of open deals and count how many have a close date in the past, because the AI will treat each of them as alive.

Write down the company’s context

The agent only knows what the company sells, to whom, how the team runs a sale and what tone messages should have if someone writes it down. In the workshop I ran in October, I called this work phase zero of the implementation. The material usually exists in pieces, such as the sales playbook in a shared drive and the ICP document marketing put together two years ago. Leadership’s job is to review this material, resolve the contradictions between areas and name who updates each part when the process changes.

Define who approves what

Before turning on any AI that writes to customers or changes a record, each area leader builds three lists: what the AI does on its own, what it does and waits for approval, and what it does not do. Summarizing a call and classifying a ticket usually go on the first list, messaging a customer in negotiation goes on the second, and granting a discount or changing a close date go on the third. These lists form the basis of the AI usage policy, which also defines which tools the company authorizes and which data can go into them. For AI agents, which execute tasks instead of suggesting text, the list needs to name the approver and the spending cap for each agent.

Train the team

Platforms have become easy to configure, and configuring them well remains hard. The team needs to know how to ask, with an instruction that includes input, rules and output format, and how to check, recognizing when a summary confidently states something the record does not support. The design of role-based cohorts is in AI training for companies.

Measure before and after

Record each process’s baseline before the pilot and compare, in the same team, the pilot weeks with the same number of previous weeks. Measure behavior and time first, such as meetings with a recorded next step, drafts approved without edits and classifications the analyst corrected, because revenue and retention respond more slowly. Add credit consumption and the hours of configuration and review to the license cost, since they tend to stay out of the calculation. The full calculation is in AI ROI.

Where does AI pay off in each area of the company?

The map summarizes the uses I see working in the areas that ask for AI first. Effort is the data, context and configuration work before the first use; expected return is a relative reading across uses, based on the projects I follow.

Area	Use	Effort	Expected return	Prerequisite
Sales	Meeting prep and call summary	Low	High	Calls and emails associated with the deal
Sales	Follow-up on stalled deals	Low	Medium	Stages with entry criteria
Sales	Revenue forecast	High	High, after months of clean data	Reliable dates, stages and loss reasons
Marketing	Nurturing personalization	Medium	Medium	Up-to-date contact job title and engagement
Marketing	Lead qualification against the ICP	Medium	High	Written ICP and standardized fields
Support and CS	Ticket triage	Low	High	Categories and priorities as closed options
Support and CS	Summary for the renewal meeting	Low	Medium	CS records inside the CRM
Support and CS	Automated reply to customers	High	High	Reviewed knowledge base and handoff-to-human rule
HR	Job description, interview script and onboarding track	Low	Medium	Written competencies per role
Finance	Collections sequence with a message per customer	Medium	Medium	Contracts and due dates in the system
Finance	Expense categorization and reconciliation	Medium	Medium	Chart of accounts and cost centers defined

In sales, the low-effort uses work on what the salesperson already produces, such as the call transcript and the account emails, and the manager starts reviewing what the agent proposes and demanding the records that feed it, the subject of AI for sales managers. In marketing, nurturing personalization pays off when the manager controls the share of sends that go through AI and compares them with a group that receives the standard email; qualification against the ICP returns more and requires an ICP written with criteria that sales recognizes, as I discuss in AI in B2B marketing.

In support, triaging the new ticket is the use with the best ratio of effort to return that I know, as long as categories are closed, and the automated reply to customers goes to the end of the line, because it requires a reviewed knowledge base and a rule for handing off to a person. The path from ticket to renewal is in AI in customer service. HR and finance come after the revenue areas in most plans, because their data tends to live in systems separate from the CRM. In HR, I keep decisions about candidates with a person, for a reason that appears in the risks section.

A 90-day plan for the executive team

I propose this plan for the executive team to move beyond individual use and reach an investment decision with numbers in hand.

Days 1 to 15: choice and baseline. The executive team chooses two or three revenue processes, an owner for each and the indicator that needs to change. The owner measures the process as it works today, and finance surveys how much the company already spends on AI, including the subscriptions areas pay for on their own.

Days 16 to 30: minimum data, context and rules. The CRM owner fixes the fields the processes use. Leaders review the company context, build the approval lists and set the spending cap per use, and legal or the data protection officer reviews which personal data goes into the tool.

Days 31 to 45: training and pilot start. Role-based cohorts practice on the chosen processes, with the company’s data and context, and the pilot starts in a small team with a manager willing to review what the AI produces.

Days 46 to 75: pilot with weekly review. Each process owner reviews the edited drafts, corrected classifications and credit consumption every week, and records who decided each adjustment to the context or rules.

Days 76 to 90: measure and decide. The executive team compares the indicators with the baseline, adds up the real cost and decides, process by process, whether to expand, adjust and repeat, or shut down. The uses that showed a return go into next year’s budget with an owner, a cap and an indicator.

What are the risks of implementing AI in the company?

The four risks that bring down most AI projects come from decisions the company failed to make, and the executive team can address all of them before the pilot.

Bad data

A duplicated company gets two follow-ups from the same agent on the same morning. The report the AI builds on a database with inflated stages arrives with charts and analyst language, and few people in the meeting question the number. The remedy is the minimum data from step two, applied to the pilot processes.

Tools without an owner

A license without an owner turns into a fixed cost nobody reviews, and an agent turned on by someone who left the company keeps consuming credits and writing to customers. Each tool and each agent needs a named owner who answers for spending, configuration and what the AI produces. At the quarterly review, the executive team asks that person what the use delivered.

Use without a policy

When the company authorizes no tool, the team uses personal accounts, and proposals, contracts and customer lists pass through services nobody at the company has evaluated. Banning tends to push that use out of leadership’s reach. The AI usage policy handles the problem with the list of authorized tools, the data that can go into each one and each area’s approval lists.

Data protection law (LGPD)

Brazil’s General Data Protection Law, Law 13,709 of 2018, applies to the processing of personal data that AI performs. The necessity principle, in article 6, limits processing to the minimum necessary for its purpose, which weighs against pasting the entire customer base into a tool to ask for an analysis. Article 20 gives data subjects the right to request review of decisions made solely on the basis of automated processing that affect their interests, including those that define their professional, consumer and credit profile. That is why I keep decisions about candidates, credit and commercial terms with a person and take the rest to legal, as I discuss in AI and LGPD.

AI training: what HR and L&D need to consider

HR tends to recognize the need for training through signals that appear before the formal request: licenses with low use or use concentrated in a few people, a team using personal tools with company data, and leaders who do not know what to approve when an agent asks for authorization. Training that works separates groups by what each one decides. The executive team works on budget, ownership and measurement, and sales, marketing and post-sale leaders design the approval lists and indicators for their areas. Practitioners practice on the chosen processes with the company’s data and context, and the CRM administrator or RevOps team learns to configure agents, context repositories and credit limits. The satisfaction survey measures the class; to measure the operation, track in the following weeks the use of each area’s prompt library, the share of drafts approved without edits and the corrected classifications, against the pilot baseline.

In most companies that come to me, what stalls AI is the lack of people who know what to ask for, what to check and who answers for what they approved, or a CRM that is not yet organized for the agent to work. That is the work of the in-company training on applied AI, built on the company’s processes and data, with cohorts for the executive team, leaders and sales, marketing and post-sale teams. It delivers the prompt library for each area and the agent architecture, with what each agent reads, what it executes on its own and who approves the rest.

Frequently asked questions

Where should a company start with AI?

Start by choosing two or three high-volume revenue processes with a pain that already has a number, such as lead response time or hours of meeting prep. Measure the baseline, fix the fields those processes use, write down the company’s context and define who approves each action. Expanding licenses comes after the pilot.

Is it worth giving AI licenses to every employee?

Broad access without a chosen process produces individual gains that no indicator records and a bill that grows with the number of users. An alternative is to authorize one general-purpose tool, with a clear rule about what data can go into it, and concentrate the larger investment in the pilot processes, where the company can measure the return.

How long does it take for AI to pay off in the company?

Call summaries and ticket triage show behavior change in the first weeks, and 90 days are enough to compare indicators and decide where to expand. Revenue forecasting takes longer, because it depends on months of organized data.

Who should own AI in the company?

Each pilot process needs an owner in the area that runs it, such as the sales manager for meeting prep or the support leader for triage. A sponsor on the executive team approves the budget and demands measurement, and RevOps or the CRM administrator handles configuration.

How does HR know the company needs AI training?

The most common signals are underused licenses, use of personal tools with company data and leaders who do not know what to approve when an agent asks for authorization. When those signals appear, training needs to be role-based, built on the company’s own processes and data.

Do I need an AI usage policy before starting?

An initial version should exist before the pilot, with the authorized tools, the data that can go into each one and each area’s approval lists. It keeps the team from taking contracts and customer data into personal accounts, helps the company comply with the LGPD, and gains detail as the pilot goes on.