Can AI help me forecast whether my team has capacity to take on more work?

13 October 2026

What should a capacity forecast actually tell you?

Yes. AI can help you forecast whether your team has room for more work, but the useful output is a range and a decision, not a confident-looking percentage. Start with the work already committed, the hours or days each job will genuinely need, the people and skills available, and the date by which each commitment must be delivered. The forecast should show where the bottleneck is, what assumptions it used, and what changes if a new order arrives. A six-person service team might have spare hours overall while the only qualified reviewer is booked solid. An average utilisation figure hides that problem.

Treat capacity as three separate questions: can we accept the work, can we deliver it by the requested date, and can we do it without damaging existing commitments? A dashboard that says “72% utilised” answers none of those on its own. Ask for a weekly view by role and by individual, then show committed hours, likely interruptions and a realistic buffer. For a small business, a useful first version may be a spreadsheet refreshed each Friday. You do not need a custom model before you understand the maths.

Imagine a job worth £4,000 that requires 40 delivery hours and 8 review hours. If delivery staff have 50 hours free but the reviewer has only 3, the honest answer is “not by Friday”. The forecast becomes valuable when it points to that constraint and lets you test moving a deadline, finding cover or declining the job. It is a planning aid, not a substitute for talking to the team.

Which data do you need before you involve AI?

You need a small, trustworthy operating dataset. Gather open jobs, agreed deadlines, estimated effort, the person or role required, work already booked, leave, known training time and recurring internal duties. Record what actually happened after each job closes. If the estimate says 10 hours and the job takes 18, that difference matters more than a sophisticated forecasting algorithm. Keep the data at the level needed for planning. A job ID, stage, role, due date and effort estimate may be enough; you usually do not need to send client documents or staff health information to an AI service.

UK employers also need to account for ordinary working limits. GOV.UK says almost all workers are legally entitled to 5.6 weeks of paid holiday each year. A forecast that assumes 52 fully available weeks is wrong before it starts. GOV.UK explains the usual 48-hour average weekly working limit, subject to exceptions and opt-outs. Do not treat overtime as your default capacity plan. Leave, sickness, training, management work and customer interruptions need explicit allowances.

If you connect calendars, a CRM or project software, check the permissions first. Give the forecasting tool read-only access where possible, limit fields and users, and decide how long the data is retained. The ICO AI and data protection guidance is the relevant UK starting point if personal data is involved. Staff should know what is measured and how it will be used. A workload forecast should support staffing decisions, not become a hidden scorecard of individual performance.

How do you build a first forecast without buying a new platform?

Start with four weeks of history and the next six to eight weeks of commitments. For each type of job, compare estimated effort with actual effort. Use a median or sensible range rather than the best case. If five similar jobs took 9, 11, 12, 14 and 22 hours, planning at 12 hours with a clear buffer is more defensible than assuming 9. Mark the 22-hour outlier for review: was it a one-off problem or an overlooked step that can happen again? Then calculate available hours per role after leave and recurring work, and subtract committed work week by week.

A simple spreadsheet can do the arithmetic. AI is useful for cleaning inconsistent job descriptions, suggesting which historical jobs are comparable, spotting a rising backlog and turning the numbers into a plain-English exception report. It can also run scenarios: “What if the client brings the deadline forward by one week?” or “What if the reviewer is away for three days?” Keep the actual calculation visible. If an AI assistant says the team can accept another project, you should be able to trace the answer to rows and assumptions.

For a low-cost starting point, use the reporting already available in your project tool, or a spreadsheet in Microsoft 365 or Google Workspace. Microsoft Planner, Asana and other project platforms can be useful if your work is already recorded there; none will rescue incomplete timesheets or vague job stages. The purchase decision should follow a measured pilot, not a demo. A two-week pilot can compare forecast delivery dates against what the team actually completes, while a manager records why estimates missed. That is a more useful test than asking whether the chart looks polished.

What numbers should you show the owner each week?

Use a short weekly view: committed hours by role, available hours after leave and routine duties, work at risk of missing its date, unallocated new requests and the range of hours likely to be needed. Add forecast accuracy. If you predicted 60 hours of work and it took 75, the error is 15 hours or 25% of the original estimate. Track whether that error is concentrated in one job type. A forecast that is often wrong in a known direction should be corrected before it informs a hiring or sales decision.

Put two scenarios beside the base case. One can assume jobs take 20% longer than estimated; the other can assume a key person loses two working days. These are planning examples, not universal risk percentages. For a team with 100 available delivery hours and 85 committed hours, a 20% overrun on the committed work creates 102 hours of demand. The apparent 15-hour gap disappears. That is exactly why “we have 15 hours spare” should not automatically become a promise to a customer.

Do not optimise for 100% utilisation. People need time for quality checks, unexpected enquiries, handovers and improvement work. The right buffer varies by business and season; measure your own interruptions before setting it. Review the forecast with the people doing the work. If they say a job stage has been underestimated, record the reason and update the assumption. The best forecast is not the one that impresses the board. It is the one that changes a decision early enough to protect the client and the team.

What can go wrong with an AI capacity forecast?

The main failure is false precision. AI can make incomplete information look authoritative, especially when a manager asks for a single date or a percentage. Other common errors are double-counting people booked across projects, ignoring non-billable duties, treating all staff as interchangeable, and learning from a period that no longer resembles current demand. A new starter may have nominally available hours but need supervision. A senior specialist may be the bottleneck even when the team total looks comfortable.

Another risk is using the forecast as an automated judgement about employees. A person who takes longer on complex cases may appear less productive than someone doing simpler work. If you use personal data, apply the ICO guidance, explain the purpose and give people a way to correct inaccurate records. Keep consequential decisions, such as staffing and performance management, with accountable humans. Do not allow an assistant to send acceptance promises or rearrange deadlines automatically until its recommendations have been tested and there is a clear approval step.

Finally, check for commercial incentives. A sales team may prefer an optimistic capacity number; a delivery manager may add excessive contingency. Agree the assumptions together and record overrides. When a forecast is wrong, ask which input or rule failed rather than blaming the person who raised a concern. This improves the model and encourages honest reporting. A weekly fifteen-minute review often delivers more value than months spent on a more complicated prediction system.

When is this not right for you?

AI forecasting is not the first move if you cannot list your active jobs, owners, deadlines and rough effort. Fix that basic visibility first. If the business has only one or two people and a handful of jobs, a shared calendar and a weekly conversation may be faster and more reliable. If every job is radically different and history is thin, use expert estimates and broad ranges; do not pretend the past provides a precise prediction. The same applies when demand has just changed because of a new contract, product or staffing model.

Do not buy a capacity platform simply because “AI” appears in the pitch. Ask the vendor to demonstrate your real work types, leave rules, permissions and exceptions. Request an export of the underlying assumptions and a way to correct bad inputs. If the system cannot explain why it recommends accepting a job, it is not ready to drive commitments. A consultancy can help when several systems need joining, but an ordinary spreadsheet may be the right answer for a straightforward team.

If your current process is stable, a sensible next step is to compare two weeks of your existing plan with a simple role-based forecast. Note every difference and ask whether either would have changed a real decision. If that exercise surfaces repeated bottlenecks or costly missed dates, then improving the data and automation is worth considering. If it does not, keep the simple process. The purpose is to make better promises to customers and more sustainable plans for staff, not to install AI for its own sake.

Is This Right For You?

This is useful if you manage several concurrent jobs, have repeatable work types and need to decide whether to accept new commitments. Begin with your current tools and a short pilot.

It is not a good first investment if you cannot yet see active work, deadlines or who owns each job. Fix those records first. Keep final customer and staffing decisions with a person.

Frequently Asked Questions

Can AI predict exactly when a project will finish?

No. It can estimate a range using past delivery times and current commitments. Show assumptions and update the range when scope or staffing changes.

Do we need timesheets to forecast capacity?

Not necessarily. Approximate effort by job type and compare it with actual completion is a useful start. More detailed time recording helps only if people can maintain it accurately.

Should we include annual leave and sickness?

Yes. Include booked leave explicitly and model sickness as a sensible historical allowance, not as a judgement about individuals.

Can AI read our CRM and project board automatically?

It can if the tools support secure integration, but start with limited, read-only access and only the fields needed for planning.

What if our team members have different skills?

Forecast by role or named bottleneck, not just total hours. A surplus of general hours does not replace a scarce reviewer or technician.

How much does a first capacity forecasting pilot cost?

Using an existing spreadsheet and staff time can avoid new software fees. Paid tools and consultancy vary; ask for a scoped quote and compare it with the cost of missed dates or unused capacity.

Who should make the final decision to accept new work?

A manager who understands the customer promise, the forecast assumptions and the team situation. AI should advise, not commit on your behalf.