Strategy

Planning Agents Need Assumptions You Can Challenge

A planning agent recommends hiring four people next quarter. The forecast looks sensible. Revenue rises, delivery capacity catches up, and the margin stays above target. Then someone asks whether the forecast assumes every open sales opportunity closes on time. Nobody can tell from the recommendation. That is a planning system with a missing interface.

Agents can help assemble a plan, investigate constraints, and recalculate alternatives when conditions change. But a recommendation is only useful if the people responsible for the business can challenge the assumptions underneath it. Those assumptions need to be part of the output, not something a reviewer has to extract through follow-up prompts.

Start with a decision someone can make

Consider a hypothetical services company deciding how to cover next quarter’s workload. It has 12 delivery employees, each with 30 billable hours available per week, and a 13-week planning period. That gives it 4,680 hours of nominal capacity. Signed work requires 4,200 hours. Another 1,200 hours sit in the sales pipeline.

At a 50 percent conversion assumption, expected demand is 4,800 hours: a 120-hour gap. At 80 percent, demand reaches 5,160 hours and the gap becomes 480 hours. The agent should show both calculations. Neither conversion rate is a fact about what will happen.

Even that comparison leaves questions open. Do the 30 hours already account for leave? Can all 12 employees perform the required work? Does the demand land evenly across the quarter? A quarterly total can hide a severe capacity shortage in week two. Hiring may not solve it if recruiting and onboarding take eight weeks.

Define the decision before asking for scenarios: should the company authorize hiring, reserve contractor capacity, or negotiate later delivery dates? Name the person who can approve the choice and the date by which a decision is needed.

Give each assumption a source and an owner

Keep an assumption register alongside the forecast. For each material input, record its value, unit, source, effective date, and accountable owner. Label whether it is an observed fact, a forecast, or a management choice. Include a plausible range where one is available.

  • Signed demand: 4,200 hours, tied to a dated export of approved project schedules.
  • Pipeline conversion: 50 percent in the base case, owned by the sales lead and supported by the relevant historical cohort.
  • Available capacity: 30 hours per employee per week, confirmed by the delivery manager against leave and existing commitments.
  • Hiring lead time: eight weeks for this example, an estimate that needs validation before the plan relies on it.

The numbers above are illustrative, not a client result or an industry benchmark. In a deployed workflow, the agent should link each sourced input to the underlying record and flag missing evidence. It should not quietly fill a missing hiring lead time with a plausible number and present the result as approved capacity.

A reviewer should be able to change one assumption and see which recommendation changes with it.

Compare scenarios on the same basis

A hiring scenario with optimistic sales and a contractor scenario with conservative sales do not tell you which staffing option works better. They change two things at once. Evaluate each option against the same demand cases, planning horizon, labor rates, and service commitments. Show the cost of doing nothing as well.

Treat budget ceilings and required qualifications as explicit constraints. If an option exceeds the cash available or needs skills the team lacks, mark it infeasible. Do not leave it in a ranked list with an attractive score and a caveat buried below.

Separate the language model from the calculation engine. Let the agent collect records, explain differences, and propose questions. Use versioned formulas or application code for capacity arithmetic and financial totals. Store the input snapshot with the result so a reviewer can reproduce the comparison later.

Find the assumption that changes the decision

In the simplified example, the team has 480 hours of capacity left after signed work. Dividing that by the 1,200 pipeline hours puts the capacity threshold at 40 percent conversion. Above that level, expected demand exceeds nominal capacity. That threshold is more useful than a confident sentence saying the company should hire.

It still does not establish that hiring is the right response. Break demand down by week and skill, account for onboarding, and compare the incremental costs of the available options. Run combined downside cases too: a delayed hire and earlier customer start dates can occur together. Varying one input at a time will miss that interaction.

Report which assumptions drive the recommendation and where the evidence is weak. Avoid invented precision. A model’s confidence in its own prose is not a probability that a customer will sign a contract.

Keep approval separate from execution

The review packet should identify the selected scenario, rejected alternatives, unresolved questions, and named approver. Record the exact plan version approved. If someone changes a material input afterward, show the difference and return the affected decision for review.

Approval to explore a hiring scenario is not permission to create job requisitions or commit contractor spend. Give those actions their own authorization boundaries. The agent can prepare the work, but the execution workflow must check what was approved, by whom, and within which spending limits.

After the decision, compare actual demand and capacity with the approved assumptions. Set review triggers with the decision owner, such as a major contract signing or a hiring delay. Preserve the original forecast instead of overwriting it with the latest numbers. Otherwise, the team loses the evidence it needs to improve its next planning cycle.

Start with one recurring planning decision

Pick a decision the team already makes, such as monthly staffing capacity. Connect the smallest useful set of source records. Build the assumption register and a reproducible scenario comparison before adding execution permissions. Measure how long review takes, how often reviewers find unsupported inputs, and whether forecast errors become easier to explain.

Foundation builds agent workflows around the systems and decisions a business already owns. Talk with us about your planning workflow. Bring one recent plan, the spreadsheet behind it, and the person responsible for approving it.