Why workshops go wrong
A QRA is only as good as its inputs, and most of its inputs come from workshops. The model cannot compensate for poorly elicited three-point estimates, incomplete risk identification, or social dynamics that produce comfortable consensus instead of honest ranges. The facilitator's job is to create the conditions in which people say what they think, not what is politically convenient.
Four failure modes recur. The dominant voice problem: one or two confident people set the tone and everyone else adjusts. Optimism anchoring: pessimistic values sit too close to the most likely, because moving far enough out feels like admitting the project is failing. Omission: the room captures the risks everyone already knew and misses the less obvious ones. Social pressure: people will not raise risks that sound like criticism of a colleague.
A better spreadsheet or a more sophisticated simulation tool fixes none of these. Workshop design and facilitation do. The numbers in this guide for time and range width are SOMA rules of thumb from our own practice, not published standards, and are labelled as such.
What the three points mean
Before anyone gives a number, fix the convention and write it on the wall. A 90% confidence interval has 5% of the probability in each tail. So the minimum is the P5 (a 5% chance the outcome is lower) and the maximum is the P95 (a 5% chance it is higher). The most likely value is the mode, not the mean, and in a skewed range it sits closer to one end.
Some tools take absolute bounds (the true minimum and maximum of a triangular or PERT distribution). Others take percentile-defined inputs, for example a triangular distribution specified by its P5 and P95. If your tool wants absolute bounds, the P5/P95 you elicit must be converted. Stretching a P5/P95 pair into absolute bounds widens the distribution, so entering P5/P95 values as if they were absolute bounds understates the tails. State which convention the model uses before the session, and use P10/P90 only if the client's guidance says so and everyone is told.
Fix the mitigation basis too. Pre-mitigation impacts assume no further response beyond what is already in the baseline plan. Post-mitigation impacts assume the agreed responses work as planned and are funded. Both are legitimate, but mixing them in one model is not. Record the probability and impact on the basis the client or the model requires, say which it is in the register header, and make sure the cost of a mitigation is not counted twice, once in the baseline and again as a risk.
Pre-workshop preparation
Most of a workshop's quality is decided before anyone enters the room. Three things matter: who attends, what they have read, and an agenda with slack for discussion.
You need people with direct knowledge of the work: the engineers, the package managers, the commercial lead. You also need someone with an institutional view, who has delivered something similar and is not so close to this project that they cannot see it objectively. Do not let the room be dominated by the project leadership team, who have the most to gain from optimistic numbers. Senior leaders should join at the end to hear the outputs, not at the start, where their presence chills junior contributions.
Pre-reading should include the current risk register, the programme summary, the cost plan structure, the assumptions and exclusions list, and lessons-learned outputs from comparable projects. Send it a week ahead and chase two days before. A short covering note explaining what the workshop will ask of people helps them arrive prepared.
As a SOMA rule of thumb, allow at least half a day for risk identification and quantification on a project of about £10m or more, and a full day for a complex portfolio or programme. A two-hour session on a £50m programme will not produce credible inputs. Schedule calibration, the hardest cognitive work, in the morning, and take a proper lunch.
Roles, agenda and remote sessions
Use three roles. The facilitator runs the process and challenges ranges. The scribe enters risks live, in the agreed structure, on a shared screen. The modeller, if separate from the facilitator, listens for correlation, schedule mapping and double counting, and asks the clarifying questions that the model will need. The project manager attends as a contributor, not the chair.
A one-day agenda that works as a template is in the table. For a half day, drop the second identification block and shorten quantification.
Opening with opportunities as well as threats
Many workshops open by asking what could go wrong. That frames the session as a list of bad news and can feel like criticism of the planning. Opening with "what could go better than we have planned?" signals that the exercise is two-sided, so a pessimistic value reads as an acknowledgement of uncertainty, not a prediction of failure. It also surfaces pull-ins, savings and commercial opportunities that belong in the model.
In a cost model an opportunity is a risk with a negative impact. Suppose there is a 20% chance that ground conditions prove better than assumed, with savings of £100k (P5), £200k (most likely) and £350k (P95). Enter it with probability 20% and impacts of -£350k, -£200k and -£100k. Taking a triangular mean of about £217k, its expected value is a £43k reduction in expected cost, and leaving it out biases the output upwards. Clients, including National Highways, have their own rules for how opportunities are treated in a QRA, so confirm that before the session.
Calibration: getting honest ranges
People are poor at estimating uncertainty, and the evidence is consistent. Studies of calibration (Lichtenstein, Fischhoff and Phillips, 1982; Hubbard, How to Measure Anything) find that stated 90% intervals capture the true value well under 90% of the time. The facilitator's role is to slow the estimation down and use techniques that offset the known biases.
The first is the direct challenge, applied to each tail. If an estimator gives a P5 of £400k, a most likely of £500k and a P95 of £600k, ask: "Is there really only a 1-in-20 chance this item costs more than £600k?" Then: "Is there really only a 1-in-20 chance it comes in under £400k?" One in twenty is what the P5/P95 convention means, so say it in those words. Most people revise at least one end outward.
A worked revision. Opening estimate: P5 £400k, most likely £500k, P95 £600k. After the challenge the estimator recalls a similar package that overran by 60% and a favourable supplier quote. They revise to P5 £380k, most likely £500k, P95 £800k. The range between P5 and P95 moves from £200k to £420k, 2.1 times wider, and it is now asymmetric: £120k below the most likely, £300k above. The width change here is one example, not a typical value. The asymmetry is the point.
The second technique is reference class data, as set out in Flyvbjerg's work on reference class forecasting and in the optimism-bias adjustment in the HM Treasury Green Book (2026 edition). Show the estimator how comparable items performed before they give their range. For illustration only, if ground investigation packages on similar schemes in your own data have ranged from -10% to +150% against estimate, that range, not the estimator's intuition, is the starting point. Replace the illustrative figures with your own data.
The third is to pre-commit before anchoring. Ask each estimator to write down a P5 and P95 on paper before anyone speaks a most likely value. Once a most likely value is said aloud, the other two anchor to it.
Drawing out quiet voices and handling dominant ones
The most senior or most confident people speak first and most often, and their numbers anchor the group. Junior people with operational knowledge often hold the most realistic view and adjust to the room. The facilitator should deliberately invert this.
Use structured individual responses before discussion. Each participant writes a P5, most likely and P95 on a card. Collect them without speaking, then display all values at once. The spread, not the first voice, becomes the starting point.
For a dominant voice, set time limits and invite others by name: "Before we go further, I want to hear from everyone. Name, what is your view on the lower end of this range?" A general invitation will be filled by the same person. This is not silencing expertise. It tests that expertise against what others in the room know.
A pre-mortem is useful for surfacing risks from quieter people (Klein, Performing a Project Premortem, Harvard Business Review, 2007). Ask everyone to write, silently, the reasons the project has failed. Read the responses aloud without attribution and discuss. Writing alone first reduces the pull of the loudest voice, though responses are only anonymous if you read them without names. Risks that appear on several lists but not on the existing register are usually the most important findings.
Risk events, baseline uncertainty and correlation
Keep two things apart. Discrete risk events have a probability of occurring and a three-point impact if they do. Baseline uncertainty is a range on an item that will happen anyway, such as a quantity, a rate or an activity duration, with no probability attached. Eliciting them in one pass produces muddled numbers. Run them as two exercises and label each register line.
For schedule risk, map each risk to the activities or summary tasks it affects, and give a three-point duration impact for those activities. A risk with no mapping cannot move the programme in the model, whatever the register says. Check that mapped activities are on or near the critical paths, and that the programme itself has passed a health check before you rely on it.
Correlation matters because risks that share a driver tend to move together, which widens the output range. Ask in each session which risks share a cause, such as the same supplier, the same ground, the same approval body or the same labour market. Record them in plain words (for example, high, medium or low positive), then agree numerical coefficients with the modeller and test sensitivity afterwards.
Capture assumptions and exclusions in the same session. A QRA result is conditional on them. If the client excludes inflation, a specific third-party risk or a change in scope, say so in the register and in the report.
Wrap-up and turning outputs into register entries
A workshop that produces rich discussion but no clear outputs has wasted most of its investment. The facilitator or scribe should capture entries live in a structured format. Each entry needs a unique ID, a cause-risk-effect statement (use the client's terminology, for example cause, event and effect), a probability, three-point cost and schedule impacts on the stated mitigation basis, mapped activities, an owner, and agreed mitigations.
Entering risks live on a shared screen and confirming them before the session ends is the stronger process. People can correct entries while they are still in the room. Where live entry is not practical, issue the draft the same day and treat it as a draft until confirmed.
Before closing, read back the findings: "The three risks driving the most uncertainty are X, Y and Z. We have least confidence in the estimates for A, B and C. Two risks raised today were not on the register: D and E." This confirms that you have understood the room and tells participants what matters.
Send the draft register, the action list and a short summary within 48 hours. Ask for corrections by a fixed date - two weeks is usually enough - then treat the register as confirmed. Later changes go through a change log, not quiet edits.
Seven checks that your QRA workshop was analysis, not theatre
Use these checks when commissioning a workshop, sitting in one, or inheriting its outputs. Each has a pass and a fail test. As a SOMA rule of thumb, two or more fails mean the model is not giving the assurance it appears to, even if the S-curve looks tidy.
1. The estimates moved. Pass: the final ranges on major packages are visibly wider or more asymmetric than the opening figures, with reasons recorded. Fail: P5 and P95 sit within about 10-15% of the most likely value on a complex package with seven-figure costs or multi-month durations. Narrow ranges look professional and are usually wrong.
2. Someone brought comparable data. Pass: the register or notes show what reference class data each major range was tested against. Fail: ground conditions, commissioning duration or design development were quantified from intuition alone.
3. The room spoke. Pass: most participants contributed to the numbers and the individual cards show real spread. Fail: two or three people account for most of the contributions in a room of ten or twelve.
4. The register grew. Pass: the session added material risks that were not on the opening register. Fail: a half-day session ends with a register that is functionally the one the team brought in.
5. The outputs arrived promptly. Pass: draft register, actions and summary reached participants within 48 hours. Fail: they went into a folder.
6. The basis is stated. Pass: the register header says whether impacts are pre- or post-mitigation and which percentiles min and max represent. Fail: nobody can say.
7. Mapping and correlation exist. Pass: schedule risks are mapped to activities and shared-driver correlations are recorded. Fail: risks are in the register but not connected to the model.
If two or more checks fail, do not re-run the simulation. Re-run the inputs. The model behaves as it should given what it was fed.
Where the workshop sits in assurance
The workshop output feeds the cost and schedule risk allowances in the business case and in the QRA report. On government major projects, NISTA (formerly the IPA) Gateway Reviews are assurance reviews of readiness to proceed, separate from approval of the business case, which follows the HM Treasury Green Book five-case process. On MOD projects, investment approval points are Initial Gate and Main Gate within the CADMID lifecycle. A reviewer at any of these will ask where the ranges came from, so the workshop record is evidence.
How we approach this on a QSRA or QCRA
On a QSRA or QCRA we state the percentile convention and mitigation basis in the first slide, bring reference data we can source, and keep the elicitation cards.