We are exploring how AI-assisted forecasting can contribute to legal advice and shared organisational knowledge. The question is practical: can it help an organisation recognise a consequential development early enough to respond? The examples below are fictional and describe a proposed approach. They are not client results or evidence of forecasting performance. Sources reviewed 17 September 2026.
Begin with a decision.
Tim Harford argues that forecasting has value both as a test of judgment and as a shared process of thinking through uncertain futures. Financial Times: AI can forecast the future. Should we let it?
A supplier renewal, a product launch or a change in distribution creates a decision with a limited time to act. Start by identifying that decision, the person responsible and the developments that could affect it. This makes it possible to decide which information is worth gathering and what an assessment needs to establish.
Keep three judgments clear: what might happen, what it would mean legally and commercially, and which response is worthwhile. A forecast of likely enforcement does not establish an organisation’s legal obligations. A forecast of disruption does not establish that a particular intervention will prevent it.
People with operational knowledge should help frame the questions and challenge the assumptions. Record their initial assessment before showing an AI forecast where an independent comparison would be useful. The discussion should leave them better able to recognise and respond to change.
A supplier decision in the next three months.
Consider a fictional supplier whose recent deliveries have been late. The customer has eight weeks before a renewal decision, while qualifying an alternative could take twelve. Procurement has delivery records, Finance has cost information, and counsel holds the contract and amendments. Public information may help explain conditions across the supplier’s industry.
Define an observable question: will the supplier miss the next agreed delivery date? Examine the historical record, current capacity information and alternative explanations for the delays. A late delivery, a contractual breach and supplier insolvency are different propositions. Each needs its own evidence and assessment.
The legal work examines the operative terms, notice requirements, possible remedies and transition constraints. Commercial options might include a shorter renewal, additional assurances or qualifying a second source. Negotiating better continuity terms may be worthwhile even while the forecast remains uncertain.
The brief should identify what would change the recommendation and when another check is useful. Once the delivery date passes, record the outcome against the original question. Record any intervention separately, because it may have changed what happened.
A market opening over the next year.
Imagine a proposed regulatory change that could make a new service commercially attractive. The organisation also sees public announcements suggesting that a competitor is preparing to enter the market. These observations raise separate questions: whether the proposal will reach a defined legislative milestone, whether the competitor will launch, and whether the organisation can be ready in time.
Track the official legislative process and distinguish proposals, adopted rules and effective dates. Treat competitor statements, job advertisements and product notices as evidence of what they disclose; a claim about strategy remains an inference. Use information obtained and shared lawfully, within the agreed purpose.
The resulting legal work could concern a licence, product terms, a partnership or participation in a consultation. The opportunity is to prepare useful options while uncertainty remains. Set review points around observable developments, and state which assumptions would need to hold before committing further resources.
Prepare for several longer-term futures.
The OECD’s foresight toolkit uses scenarios to challenge assumptions and test strategies against possible disruptions. Its application to the commercial examples here is our proposal. OECD: Strategic Foresight Toolkit
Over two to five years, a business model may face different combinations of technological change, regulation and industry consolidation. Several plausible scenarios can help reveal where today’s commitments create exposure and where preparation would retain useful choices.
Consider a company deciding whether to sign a five-year distribution agreement. Examine what happens if platform access becomes more expensive, interoperability improves or a new regulatory requirement changes the arrangement. The scenarios inform whether to accept exclusivity, negotiate a break right or preserve the data access needed to use another channel.
Identify actions that remain useful across several scenarios and indicators that would justify a different response. Scenarios need coherent assumptions and periodic review. A precise probability should be used only where its basis can be explained and assessed.
Run a premortem before committing.
Mitchell, Russo and Pennington studied explaining a future event as though it had occurred. Their findings concern how people generate explanations; they do not establish improved forecasting accuracy for this proposed exercise. Back to the future: Temporal perspective in the explanation of events
A premortem asks participants to imagine that a proposed decision has failed and work backwards to plausible causes. For the supplier renewal, assume that six months later continuity has suffered despite an acceptable price. People from Procurement, Operations, Finance and Legal first write their explanations independently. They might identify an untested recovery plan, insufficient acceptance capacity or a notice date with no clear owner.
Then consider a successful outcome: service is dependable and the customer has useful alternatives. Work backwards to the conditions that enabled it. This can reveal opportunities in training, transition support, more flexible terms or earlier qualification of another provider. Keep the exercise focused on conditions and decisions; avoid speculative accusations about individuals.
Convert the most consequential explanations into evidence requests, observable questions and possible actions. For each, identify an early indicator, an owner and the last useful time to respond. An imagined cause remains a hypothesis. Its vividness or repetition does not increase its probability; a forecast should change only when there is a defensible reason.
Retain the initial forecast and decision rationale. At the agreed review, compare the scenarios with what happened, examine which signals were available at the time and identify anything important the exercise missed. Distinguish forecast quality, decision quality, execution and luck. Test proposed changes before adopting them for later work.
Connect the organisation’s different levels.
Swiss employee-data and workplace-monitoring rules impose substantive limits. The FDPIC explains purpose, proportionality, information duties and the restrictions on monitoring employee behaviour. FDPIC: Monitoring systems in the workplace
An industry change can affect a company’s product, a team’s workload and the people qualified to perform a required task. Conversely, recurring problems in one process may reveal a wider contractual or organisational dependency. Moving between these levels can make an assessment useful, but an industry trend does not establish what will happen to a named person.
Begin with the work: capacity, authority, qualifications, handovers and knowledge needed to fulfil an obligation. Analysis involving identifiable employees needs a specific lawful purpose, proportionate data use and an assessment under the relevant employment and data-protection rules. Small-team statistics can still identify individuals.
A forecast can also change how a person is treated and help bring about the outcome it predicts. Our proposed starting point is organisational and process-level analysis, involving the people who understand the work. Any individual-level application would require its own justification and assessment. Technical access alone supplies neither.
Understand what the forecasting evidence shows.
ForecastBench adjusts for differences in question difficulty. Its human reference group answered questions in July 2024, while later models answered different question sets. ForecastBench: comparison methodology
Results on public forecasting questions provide a reason to investigate the method. They do not establish its accuracy for a particular company, contract, industry or employee. Performance must be assessed on the questions and information conditions relevant to the intended work.
AI can help retrieve information, structure a question and compare explanations. For repeated numerical outcomes, test simple statistical methods alongside more elaborate systems. A probability generated in conversation needs its own evidence and evaluation.
Preserve the information available when a forecast was made, its date, assumptions, contrary evidence and reasons for revision. Agreement between several systems is not independent corroboration when they rely on the same underlying sources. Where evidence is insufficient, identify what can still be established and whether further information could change the decision.
Test forecasts and decisions separately.
Research on evaluating language-model forecasters identifies problems including information leakage, correlated questions and misleading interpretations of aggregate scores. Paleka and colleagues: Pitfalls in Evaluating Language Model Forecasters
Define the question, resolution date and outcome evidence before forecasting. Record a simple comparison, such as the historical rate or a no-change assumption, alongside the current expert assessment. Use the same questions, information cut-off and scoring rules. Retain initial forecasts and later updates separately.
For questions with a yes-or-no outcome, a Brier score measures the squared difference between the assigned probability and the observed outcome. Across enough comparable outcomes, examine whether events assigned similar probabilities occur at roughly those rates. Include uncertainty in the comparison and avoid treating many related questions as independent evidence.
Assess usefulness as well: did the work arrive before the decision, expose a material omission or help choose a justified response? Count unnecessary alerts, missed developments and the time spent collecting, checking and maintaining information. A short pilot may reveal whether the process helps; it cannot establish general accuracy or validate forecasts years into the future.
If the organisation acts on a forecast, record the action and reconsider the expected outcome. A problem that does not occur after intervention does not by itself prove either that the forecast was wrong or that the intervention worked.
Start with one working area.
A first exploration could focus on supplier continuity or recurring commercial disputes, with one decision owner and an agreed set of company and external sources. Select a small number of questions that can resolve during the trial. Combine the initial assessment with a premortem, an exploration of successful outcomes and a scheduled outcome review. Agree responsibility for legal assessment, source upkeep, review and any action.
The practical deliverable would be a maintained decision brief: what is known, what may happen, the legal implications, possible responses and the last useful time to decide. Supporting sources and earlier assessments should remain accessible to authorised users.
At the review point, examine forecast quality, decision usefulness and total effort. Continue, narrow or stop the work on that evidence. The organisation should retain the questions, sources, decisions and lessons needed to continue with its own people and tools.