How to Plan Team Capacity Using AI (Without New Software)

What Capacity Planning With AI Means

AI capacity planning is the practice of using an AI assistant to compare your team’s genuinely available hours against the hours already committed, so you can see over-allocation and staffing gaps before they turn into missed deadlines. The AI does not predict anything mystical. It does arithmetic, applies rules you set, and formats the result — fast, and without you rebuilding a spreadsheet every Monday.

Three words get used interchangeably in this space and they mean different things. Getting them straight is half the battle in a leadership conversation.

A capacity plan answers the first question. A workload map answers the second. You need both, because a team can sit at healthy average capacity while one person is drowning — averages hide the exact problem you are trying to find.

Most guides on how to plan team capacity using AI stop at the concept and then recommend a platform. That is the wrong shape of answer for a team lead with eight people and no budget line. What follows is the workflow instead.

How to Plan Team Capacity Using AI

The Real-Hours Ledger is a four-column weekly record: each person’s ceiling (hours genuinely available after leave and standing commitments), their load (everything actually assigned, including work nobody counted), the gap between the two, and a test of whether incoming work fits that gap.

Why Four Stages, Not One

You could ask an assistant to “sort out my team’s workload” in a single prompt. It will produce something. It will also silently invent the missing pieces, because you gave it no ceiling to work against and no threshold to flag against.

Splitting the job into four passes gives you a checkpoint after each one. If the ceiling looks wrong, you catch it before it poisons every downstream number — and each stage is short enough to eyeball in seconds.

The order matters more than it looks. Most managers start at stage two, tallying assigned work against a 40-hour week, and end up with a plan that has already collapsed by Wednesday. Ceiling first, always.

Set the Ceiling Before Anything Else

Your team’s ceiling is not 40 hours per person. It is what remains after leave, standing meetings, and the ordinary friction of a working week — inbox, approvals, interruptions, the ten-minute questions that arrive between calendar blocks.

Pick a planning buffer and apply it consistently. A common convention is to plan project work at 75 to 80 percent of nominal hours, but the number matters far less than using the same one every week so you can compare across weeks and tune it against what actually happened.

Here is the calculation in order, which is also exactly what you will hand the AI:

1. Start with each person’s standard weekly hours
2. Subtract booked leave, holidays, and training
3. Subtract recurring meetings that happen every week regardless
4. Multiply what is left by your buffer percentage
5. The result is that person’s ceiling for project work

You are helping me calculate realistic weekly capacity for my team.

For each person below, calculate their available project hours using this method, in this order:
1. Start from their standard weekly hours.
2. Subtract booked leave and holidays.
3. Subtract their recurring weekly meeting hours.
4. Multiply the remainder by 0.8 to leave a buffer for admin work, email and interruptions.

Team:
- Person A: 40 standard hrs, 0 leave, 5 hrs recurring meetings
- Person B: 40 standard hrs, 8 hrs leave, 3 hrs recurring meetings
- Person C: 20 standard hrs (part time), 0 leave, 2 hrs recurring meetings
[add the rest of your team in the same format]

Return a table with columns: Person, Standard Hours, Leave, Recurring Meetings, Ceiling.
Show the arithmetic for one person underneath so I can check your maths.
Do not round anything up.

That last instruction is not decoration. Asking for the working on one row is the cheapest verification step you will ever run, and it catches the specific failure mode covered later in this guide.

Build the Weekly Ledger Step by Step

With the ceiling set, the rest is a repeatable four-step pass. It is worth saving the whole sequence somewhere you can reach on a Monday morning — a note, a saved project, or a custom assistant.

You need three things: who is off, what is assigned with rough hours, and anything that arrived by request rather than through the task board. That third bucket is the one that sinks plans. Copy them into a plain list or export a CSV from wherever your tasks live.

If your export comes out messy — merged headers, inconsistent date formats, half-empty columns — clean it before it goes anywhere near a prompt. Our guide to cleaning messy data with AI covers that pass, and it takes a couple of minutes.

Structure beats volume here. A tidy table of 30 rows produces better arithmetic than a rambling paste of 300 lines. If your source is a spreadsheet, our walkthrough on analysing a spreadsheet in ChatGPT covers the upload route.

This is the core of the system. It takes the ceiling from stage one, adds up the load, calculates the gap, and flags anyone outside the thresholds you set.

You are building a weekly team capacity ledger for me.

Below are two datasets: each person's ceiling (available project hours), and every task currently assigned with an estimated hour cost.

CEILINGS:
[paste the table from the previous prompt]

ASSIGNED TASKS:
[paste your task list: Person | Task | Estimated hours]

Do the following:
1. Total the assigned hours per person.
2. Calculate utilisation as assigned hours divided by ceiling, shown as a percentage.
3. Calculate the gap in hours (ceiling minus assigned).
4. Flag each person: OVER above 100 percent, TIGHT between 85 and 100 percent, OK between 60 and 85 percent, LIGHT below 60 percent.

Return one table: Person, Ceiling, Assigned, Utilisation, Gap, Flag.
Sort by utilisation, highest first.
Use only the numbers I gave you. If a task has no hour estimate, list it separately as "unestimated" rather than guessing a value.

Scan the totals against your own instinct. If the ledger says your busiest person is at 62 percent, either the estimates are wrong or something is missing from the task list. Both are useful findings. Fix the input and rerun rather than arguing with the output.

Spot Over-Allocation Before Burnout

The people who burn out are rarely the ones on the obviously enormous project. They are the ones absorbing eleven small requests that nobody counted, each of which looked like twenty minutes.

Your ledger catches this only if the unlogged work goes in. The most valuable question you can ask your team each week is not “how are you doing” but this one:

• What did you spend real time on last week that is not on this list?
• Which of your estimates turned out to be badly wrong, in either direction?
• What is arriving next week that nobody has logged yet?

Add those hours and rerun. Once someone is flagged OVER, the fix is a reassignment, not a pep talk. This prompt does the matching:

Here is my team capacity ledger for this week:

[paste the ledger table]

Here are the skills each person can cover:
- Person A: [skills]
- Person B: [skills]
- Person C: [skills]

Suggest a rebalancing plan that brings everyone under 100 percent utilisation.

Rules:
- Only move tasks to someone whose skills cover them.
- Do not move a task if it would push the receiver above 85 percent.
- Prefer moving fewer, larger tasks over many small ones.
- If the numbers cannot balance by moving work alone, say so explicitly and tell me how many hours short we are.

For each suggested move, give me: task, current owner, proposed owner, hours moved, and one sentence on why.

That fourth rule is the one that earns its place. Ask an AI to solve an unsolvable allocation and it will usually produce a confident-looking plan anyway. Telling it that “this does not balance” is an acceptable answer is what turns the output into information rather than reassurance.

Then you still have to have the conversation. Moving work off someone who is drowning often means telling a stakeholder their thing slipped, and our guide on preparing for a difficult conversation with AI is a useful rehearsal step before that meeting. If the reassignment goes the other way and you are handing work to someone new, delegating properly with ChatGPT covers writing the brief.

Pressure-Test New Work Before You Say Yes

The question that used to take two days of spreadsheet work — “can we take Project Alpha next month?” — now takes about five minutes. You have the ceilings. You have the load. A scenario test just adds the proposed work and shows you which week breaks.

Run a capacity scenario for me.

Current committed hours per person for the next 6 weeks:
[paste your baseline, one row per person per week]

Proposed new project:
- Week 2: [X] hrs [skill], [Y] hrs [skill]
- Week 3: [X] hrs [skill]
- Week 4: [X] hrs [skill]

Show me, week by week, what each person's utilisation becomes if we accept this work.

Then answer three questions directly:
1. Which specific week does capacity break first, and for whom?
2. How many hours short are we in that week?
3. What is the smallest change that makes it fit — moving a deadline, cutting scope, or adding hours?

Be blunt. If this does not fit, say it does not fit.

The output of that prompt is the raw material for the answer you give leadership. Turning it into something an executive reads in thirty seconds is a separate job — our guide to writing an executive summary with ChatGPT handles that, and the same numbers slot neatly into a weekly status report built with AI.

Keep Employee Data Out of Trouble

This is where managers hesitate, and they are right to. Capacity data is people data. Leave records, workload patterns, and who is struggling are all things your team would reasonably object to seeing pasted into a chat window.

The safe version of this workflow is also the easier one: strip the identities. The arithmetic works exactly the same on Person A through Person F, and you are the only one who needs the key.

Which tier you are on changes what is possible, not what is wise. Consumer AI accounts may use your conversations to improve models unless you turn that off; business tiers generally do not. According to OpenAI’s ChatGPT Business documentation, business workspaces require a minimum of two seats and carry workspace-level admin controls that consumer plans do not. For the wider question of what belongs in a work chat window at all, our privacy guide for ChatGPT at work is the fuller treatment.

Copilot users sometimes assume a licence means it can simply read the whole team’s calendars. It reads what you can already read. Per Microsoft’s Copilot privacy documentation, it only surfaces organisational data the individual user has at least view permission for — it does not grant new access.

So if you cannot see a colleague’s calendar detail today, neither can Copilot on your behalf. Our breakdown of what Copilot can see at work goes deeper, and Copilot free versus paid covers whether a licence is worth buying. Copilot Chat comes at no extra cost with eligible Microsoft 365 business accounts, which is enough for everything here.

When the Numbers Come Back Wrong

Four failures account for nearly everything that goes sideways here. Each has a fix that lives in the prompt rather than in a longer spreadsheet.

The first row is the common one and worth understanding. Language models produce plausible text, and a plausible-looking sum is not the same as a correct one — the same underlying behaviour behind AI hallucination. Asking for the arithmetic on one row, every time, catches it in seconds.

If your estimates are the problem rather than the maths, get the AI to decompose the work first: hand it a vague deliverable like “update the annual handbook” and ask for realistic sub-tasks with hours before it goes anywhere near the ledger. Structuring these requests well is its own skill, and the Role, Goal, Context, Format system is the shortest route to prompts that behave.

Run the ledger once this week with made-up numbers before you use it for real — twenty minutes with a fake five-person team will teach you more than any amount of reading, and you will find your own thresholds fast. Then run it for real on Monday, share it with your team, and let their corrections make it accurate.