How to talk to your boss about AI: making the case for a tool, budget, or pilot

The short version

Your boss does not need to be convinced that AI matters. They need to be convinced that your specific request is safe, cheap, and reversible. So do not pitch "AI." Pitch one task, one tool, one plan tier, a small group, a date the pilot ends, and what you will report when it does.

This guide gives you the framing, answers to the four objections you will hear, a one-page proposal you can copy, a script for the conversation, and what to do when the answer is no.

Before you ask: build your own evidence

Walk in with two weeks of your own experience, not an article you read. Pick one recurring task that eats your time and involves no confidential data: drafting internal updates, summarizing public documents, reformatting notes, writing first drafts of job posts or process docs. Do it with a free or personal AI plan on green data only (the red/yellow/green test is in our privacy guide), and keep a simple log.

The log is three columns: the task, how long it used to take, how long it takes now with review. After ten entries you have something no slide deck can match: "I did this fifteen times, here is what it saved, here is what it got wrong." The mistakes matter as much as the savings, because they prove you are reviewing the output rather than trusting it.

Keep the log honest. If the tool saved twenty minutes on a draft but you spent ten minutes fixing it, record ten minutes saved. Your boss will trust a modest, careful number over a large, rounded one.

Frame the ask around a task, not a technology

Managers say no to vague requests because vague requests are impossible to approve. "Can we get AI tools?" has no budget line, no owner, and no end date. "Can three of us pilot a team plan for four weeks to draft ticket replies, with me reporting back on the 15th?" has all three.

The one-sentence ask

Fill in this sentence before you do anything else:

"I want [NUMBER] people to use [TOOL AND PLAN] for [TASK] for [DURATION], and I will report [WHAT YOU WILL MEASURE] on [DATE]."

If you cannot fill in every blank, you are not ready to ask. The common gaps are the plan tier (you want a business or team plan, not personal accounts, because it excludes your data from training and gives an administrator controls) and the measure (minutes per task, errors caught in review, turnaround time).

Choose the task with the best ratio

The best pilot task is frequent, text-heavy, low-risk, and easy to measure. Drafting replies, summarizing long threads, converting notes into structured documents, first-pass research, and rewriting for tone all qualify. Anything involving customer personal data, financial results, legal matters, or HR complaints is a bad first pilot even if the savings would be larger, because the security conversation swallows everything else.

The four objections and how to answer them

Your boss will raise some version of these four. Answer them before they come up and you look prepared. Wait for them and you look like you had not thought about it.

Security: "Where does our data go?"

This is the objection that kills most requests, and the answer is the plan tier. Consumer accounts may use conversations to improve the vendor's models unless the user opts out. Business, team, and enterprise plans generally exclude your data from training by contract, add single sign-on and admin controls, and let an administrator set retention. Say that in plain words, then add the rule that makes IT relax: "During the pilot we use it only on non-sensitive data, and nobody pastes customer records, credentials, or anything under NDA."

If your company already has a policy, quote it. If it already has an approved tool you have not been using, ask for that one. A request to expand an existing approval is far easier than a request for something new.

Cost: "We don't have budget"

Team plans are priced per seat per month, in the same general range as other per-seat business software. Compare the seat price to the loaded hourly cost of the people using it. If a seat costs less than one hour of an employee's time each month, the tool only needs to save one hour a month to break even. Your two-week log should show far more than that.

Then shrink the ask. Three seats for four weeks is a rounding error in any department budget, and you can offer to cancel on day 29 if the numbers disappoint. A pilot you can cancel is a decision your boss can make alone. An annual contract for the whole team is a decision they have to escalate.

Quality: "It makes things up"

Agree, because it does. Then explain how your workflow handles it: the assistant drafts, a person reviews every output before it goes anywhere, and anything with a number, a name, a date, or a claim gets checked against the source. Bring an example from your log where the tool was wrong and you caught it. Nothing reassures a skeptical manager like proof that you are skeptical too.

Point out that the same review exists today for a new hire's first drafts. AI does not remove review. It moves the review earlier and makes the first draft cheaper.

Jobs: "Are you asking to replace people?"

Sometimes this is asked out loud. More often it sits under the conversation. Answer it directly: you are proposing to reassign hours, not people. Name what the freed-up time goes to, in your team's own terms: faster response times, the backlog nobody gets to, the improvement project that keeps slipping. If your company has a stated position on AI and headcount, use its language. If it does not, your pilot report is a chance to help set one.

Worked example: a support lead pitches a pilot

Say Priya runs a six-person support team at a regional software company. Each rep answers a few dozen emails a day, and the slowest part is the first draft of any reply that needs explanation rather than a canned answer.

She spends two weeks drafting her own replies with a personal assistant account, using only tickets she has anonymized first. Her log shows a typical explanatory reply dropping from roughly twelve minutes to roughly five including review, and two cases where the assistant invented a product feature that did not exist, both caught in review.

Her ask: "I want three reps to use a team plan for four weeks to draft explanatory replies, on tickets with personal details removed, and I will report average handling time and error count on October 15." She anticipates the objections in a one-page memo: team plan so inputs are excluded from training; three seats for one month, cancel any time; every reply reviewed by the rep before sending; freed time goes to the knowledge-base rewrite that has been on the roadmap for a year.

Her boss approves it in the meeting, because there is nothing left to decide.

The one-page proposal template

Copy this into a document, keep it to one page, and send it before the meeting.

Title: Pilot proposal: [TOOL AND PLAN] for [TASK]

The problem. [TEAM] spends roughly [HOURS] per week on [TASK]. The slow part is [SPECIFIC STEP].

The proposal. [NUMBER] people use [TOOL AND PLAN] for [DURATION] to [WHAT THE TOOL DOES], with every output reviewed by a person before use.

What it costs. [NUMBER] seats at the current per-seat price for [DURATION], cancellable monthly. Setup time: [HOURS].

What we measure. [METRIC 1, such as minutes per task], [METRIC 2, such as errors caught in review], [METRIC 3, such as turnaround time]. Baseline collected the week before we start.

Data rules. [PLAN TIER] excludes our inputs from model training by contract. During the pilot we use only [DATA CLASS]: no customer personal data, credentials, financials, or NDA material.

Risks and how we handle them. Wrong outputs: human review of every item. Overreliance: the log records what the tool got wrong. Sprawl: only the [NUMBER] named users, only this task.

Decision on [END DATE]. Expand, adjust, or cancel, based on the numbers.

Owner: [YOUR NAME]. Requested from you: approval and [BUDGET AMOUNT OR A PURCHASE CARD FOR THE SEATS].

A prompt that drafts the proposal for you

Paste this into any assistant, fill in the blanks, and edit the result in your own voice.

Act as a pragmatic operations leader who has approved and rejected many software requests from your own team. Help me write a one-page proposal to my manager for a small AI pilot.

Context:
- My role: [YOUR JOB TITLE] at [TYPE AND SIZE OF COMPANY]
- The task I want help with: [DESCRIBE THE TASK AND WHY IT IS SLOW]
- Current time spent: [HOURS PER WEEK, AND BY HOW MANY PEOPLE]
- Tool and plan I want to pilot: [TOOL NAME AND PLAN TIER]
- Pilot size and length: [NUMBER OF PEOPLE] for [NUMBER OF WEEKS]
- Data involved: [DESCRIBE THE DATA CLASS, FOR EXAMPLE INTERNAL DOCUMENTS WITH NO CUSTOMER DETAILS]
- My manager's likely concerns: [SECURITY, COST, QUALITY, OR JOBS, PICK THE ONES THAT APPLY]
- Evidence I already have: [SUMMARIZE YOUR OWN TRIAL, INCLUDING WHAT WENT WRONG]

Write the proposal with these sections: the problem, the proposal, what it costs, what we measure, data rules, risks and how we handle them, decision date, and owner. Keep it under 350 words, plain language, no hype, no claims about industry statistics. Address each of my manager's concerns in one or two sentences inside the relevant section rather than in a separate list.

Before writing, ask me up to five questions if any of the inputs above are too thin to write a credible proposal.

What to say in the meeting

Open with the problem, not the tool: "The slow part of our week is X. I have been testing a way to speed it up and I want to run a small, contained pilot." Then the one-sentence ask. Then stop talking and let them respond.

When an objection comes, agree with the concern before answering it. "Yes, data handling is the thing to get right, which is why I am asking for the team plan and limiting the pilot to non-sensitive material." Managers rarely object to the substance of a well-prepared request. They object to feeling rushed. Offer to send the one-pager to IT or legal yourself so the approval does not become their homework.

Close with the decision date: "Can we agree to look at the numbers together on the 15th and decide then?"

If the answer is no

A no is information. Find out which no it is, because each one has a different next move.

"Not now"

Ask what would need to be true for it to be yes, and when to raise it again. Keep your log running in the meantime on green data, if policy allows, so you return with more evidence rather than the same request.

"Not this tool"

Ask which tools are approved or under review, and rebuild the proposal around one of them. The task and the measurement do not change, only the vendor does. If nothing is approved, ask who owns the decision and offer to write the evaluation criteria for them.

"Not this data"

Narrow the scope. Pick a task with no customer data at all, such as internal documentation or public research, and rerun the ask. A successful pilot on safe data is the fastest route to permission for the data you actually wanted.

"Not ever"

Rare, but it happens, usually in regulated industries or after an incident. Respect it, and do not route around it with a personal account on company work. That is the definition of shadow AI, and it can cost you your job. Instead, ask whether you can join or start the group that owns the company's AI policy, and keep building your skills outside work. Policies change faster than careers.

Next steps

• Plan the pilot itself with How to run an AI pilot, then put numbers on the result with How to measure AI ROI.
• Get the security answer right with AI privacy at work and the AI policy generator.
• Estimate the savings you will present with the time savings calculator.
• Know the terms you are using: shadow AI, human in the loop, and workplace AI policy.
• Compare the business plans in ChatGPT vs Claude vs Gemini vs Copilot.