This is the segment of the panel for which the shift away from "chatbot" matters most: a conversational exchange is, at most, how a person hands one of these platforms an instruction, not a description of what happens afterwards, which may involve multiple tool calls, intermediate planning steps, and autonomous decisions a user never directly sees. Typical use cases are workflow-shaped rather than answer-shaped: processing an application end to end across several internal systems, running a multi-step research task and assembling the resulting report unattended, or handling a customer or citizen case from first contact through to resolution without a human touching every step. Because these systems act rather than merely answer, they represent the newest and least mature segment of this year's panel, the one most likely to show significant movement, in either direction, from one monthly reading to the next, and, in both government and industry, the one that most urgently needs clear oversight, audit, and rollback mechanisms wrapped around it before it is trusted with a consequential task.

​

Multi-Variant Platforms

​

Twelve of the 33 platforms, spanning Sub-Panels 2A, 2B, and 2C, in fact offer two or more model variants rather than one: typically a speed-optimised tier for quick, low-latency answers and a depth-optimised, or extended-reasoning, tier for harder problems. Gemini is the only platform with three variants rather than two, reflecting both a speed-versus-depth split and a version upgrade, 3.1 Pro to 3.6, running at the same time. Each variant is tested and scored separately, exactly as if it were its own platform, before being averaged into a single platform-level score; this keeps the speed-versus-depth gap visible as a finding in its own right, while still allowing fair, like-for-like comparison against single-variant platforms elsewhere in the panel.

​

A Note on Terminology

​

"Chatbot" accurately described the field this study first measured in 2023, text-in, text-out systems whose only function was to converse, and it still accurately describes Panel 1. It stopped being accurate once Panel 2 platforms began retrieving live information, calling external tools, and reasoning through multi-step problems, and it becomes actively misleading in Panel 3, where a platform's real work happens after the conversational exchange, not during it. This report therefore uses "AI platform" or "AI system" as the general term throughout, and reserves "chatbot" specifically for the low-agency, conversational-only systems in Panel 1, where the word still means what it says.

​

Value of a Consistent Design

​

The chief merit of this design is discipline: holding the instrument fixed, across all three panels and regardless of a platform's agency level, means every monthly reading is directly comparable to the one before it, to the same month a year earlier, and, since 2023, to a platform's own starting point four years ago. That consistency underpins meaningful comparisons both across very different platforms at a single point in time, a low-agency companion sitting alongside a high-agency autonomous agent in the same table, and across the evolving versions of a single platform over time. Because the four competency clusters are weighted equally, no single domain can quietly dominate the overall picture, and the resulting composite score reflects a genuinely rounded assessment of each platform's underlying cognitive abilities, independent of how much autonomy it is designed to exercise. Spanning four full years, the study has moved past isolated jumps in capability to reveal sustained patterns of development, or the lack of them, across successive model generations and across the low-, medium-, and high-agency tiers alike, giving a clearer read on the actual pace of technological change than any single year could offer on its own. The fixed ten-item allocation per cluster also keeps the exercise practical to repeat every month across an expanding platform list without diluting its analytical value, and it prevents any one competency area from crowding out the others in the final score.

​

Acknowledged Constraints

​

No design of this kind is without trade-offs. Treating four competency clusters as a stand-in for real-world usefulness inevitably simplifies a much richer picture of what intelligence, human or artificial, actually involves, and this premise remains open to legitimate challenge. Running an unchanged item set on a monthly cycle for four years also raises the risk that certain platforms come to perform well on these particular items specifically, rather than on the broader skills the items are meant to represent, a form of overfitting to the instrument rather than genuine growth in underlying ability. Language and reasoning are also inherently sensitive to context in ways that a standardised battery cannot fully capture, and the even ten-item split per cluster, while convenient and easy to administer, may understate how these cognitive skills interact with and depend on one another rather than operating as neatly separable domains. Extending the panel to Agentic AI introduces a further limitation that should be stated plainly: this battery measures underlying cognitive skill through a fixed set of questions and answers, not the planning, tool use, or autonomous multi-step execution that actually define a high-agency platform's day-to-day usefulness, so a strong composite score should never be read as a guarantee of reliable autonomous execution, and a weaker one should not be read as proof that an agent cannot get real work done in practice. Finally, the current instrument says little about the ethical dimensions of AI behaviour, including bias, safety, and fairness, which matter as much to real-world deployment and societal impact as raw cognitive performance, and remain outside the direct scope of this methodology.

​

Contending with AI Drift

 

Assessing platform cognition over a period as long as four years brings a further complication: AI drift. This term captures the gradual, often unannounced, shift in a system's outputs and behaviour that can occur even when its name and version number stay the same, driven by changes to training data, fine-tuning routines, or the infrastructure behind deployment. It is a phenomenon our methodology must contend with at every reading rather than treat as a one-off nuisance, and one that plays out differently depending on a platform's position on the agency spectrum: a low-agency companion may drift mainly through changes to its underlying language model, while a high-agency agent can also drift through updates to the tools, orchestration logic, or retrieval pipelines it calls upon, quite apart from any change to the model powering it.

​

Drift makes it genuinely difficult to draw firm conclusions from a long observation window. A platform that scores well on language and logic tasks in one quarter may look meaningfully different two or three years later, without ever having been formally "re-released." That inconsistency complicates the search for stable benchmarks and undermines confidence in any single trend line, since a dip or a jump in the data could reflect a genuine change in capability, an unannounced adjustment behind the scenes, or simply noise.

​

Drift is not always visible in the headline numbers. A platform can hold a steady composite score while quietly picking up new biases or reasoning quirks that only surface under closer, more qualitative scrutiny, or, for an agentic system, while a change to the tools it calls on shifts how reliably it completes a task even though its underlying reasoning score has not moved at all.

​

Its causes are numerous and often intertwined: shifts in the mix of text used for training, tweaks to model architecture or optimisation routines, changes to the tool integrations and orchestration layers that sit around an agent, and even the feedback loops created by how users interact with a system day to day, which can subtly reshape its future responses. Because large language models behave in a highly non-linear way, small internal changes can occasionally produce outsized shifts in what the model actually outputs, making the relationship between a change upstream and its downstream effect on performance hard to predict.

​

Together, these dynamics make multi-year cognitive assessment a genuinely difficult analytical problem. It calls for evaluation tools that are periodically re-checked, ongoing vigilance for drift between readings, and methods capable of isolating a specific cognitive skill from the noise that drift introduces, so that decision-makers are not left mistaking incidental fluctuation for a genuine trend, or the reverse.

​

Instruments for Data Collection

 

The five instruments in each of the four clusters below are administered identically to every platform in the study, from the low-agency conversational systems in Panel 1 to the high-agency autonomous agents in Panel 3. That uniformity isolates the underlying cognitive skills common to all of them, however much autonomy they are built to exercise, and keeps scores comparable across the full 33-platform panel. Because this report is read as much for its operational implications as for its academic ones, each instrument below is described first in terms of what it measures, and then in terms of where that skill actually shows up in day-to-day work, in government and in industry, and what is at stake, in speed, in volume of work done, in the time spent on human oversight and correction, and in the risk of error, when a platform performs well or badly on it.

​

Linguistic Proficiency Tests

​

When we set out to probe Linguistic Proficiency, we asked a cluster of practical questions. Can a platform pick up on subtle shades of meaning, tell apart concepts that are easily confused, and interpret language with precision? Is it able to pull out the essential meaning of a passage, spot the main idea buried in supporting detail, and use that understanding to solve problems or reach a decision? Does it recognise words that mean the same thing, and separate those that sound alike but mean something different, all while expressing itself accurately and without wasted words? Can it build sentences that hang together, communicate clearly, tell a story well, argue persuasively, and carry complicated information without losing the reader? And finally, does its vocabulary grow and vary rather than repeat itself, in a way that keeps its writing sharp? These questions guided the design of our Linguistic Proficiency instruments.

​

Antonym Recognition Test. Here, platforms are given a word and must supply its opposite, a task that puts their grasp of word meaning and semantic contrast on display. It is a direct probe of vocabulary depth and the capacity to spot contrasts in meaning.

​

This skill sits behind more official work than it might appear to. Government departments and regulated industries alike depend on getting words like "shall" versus "may", or "mandatory" versus "recommended", exactly right in legislation, permits, contracts, and technical standards; a platform that blurs these distinctions when drafting or reviewing such text can quietly change what a document actually commits an organisation to. Strong performance supports faster first-draft turnaround on policy notes, contracts, and technical documentation, cutting the volume of manual redlining a legal or compliance team has to do. Weak performance has the opposite effect: it pushes more of the drafting and review burden back onto specialist staff, slows the document cycle, and, in the worst case, lets an imprecise word choice through into a binding text, with the legal or financial exposure that follows.

​

Language Comprehension Test. This instrument presents a passage of text followed by comprehension questions, a format familiar from language-proficiency exams and professional certifications alike. It tests the ability to extract meaning from written input under realistic conditions.

​

Reading and correctly digesting long documents, incident reports, tender submissions, audit findings, is a daily task across both government and industry, and one that consumes a disproportionate share of skilled staff time. A platform that comprehends well can triage inboxes, summarise reports for a briefing pack, or flag the three paragraphs in a hundred-page submission that actually matter, multiplying the volume of material a small team can get through in a day. A platform that comprehends poorly does the reverse: it produces summaries that miss the point, forcing staff to read the source material anyway, so all the usual oversight time is spent with none of the expected time saved, and, if a missed clause slips through unnoticed, the risk shifts from lost time to a genuine misjudgement.

​

Word Substitution Assessment. Platforms are asked to condense a phrase or a longer description into a single word that preserves its meaning. The task probes vocabulary range, precision of expression, and the ability to compress complex ideas without losing their substance.

​

This is the skill behind good headlines, case titles, and metadata tags, the small labels that determine whether a document, a support ticket, or a case file is easy or impossible to find again later. In large public administrations and corporations alike, where searchable records are now as much a compliance requirement as a convenience, a platform that reliably compresses meaning into precise labels speeds up classification of high-volume inflows, correspondence, service tickets, procurement files, without a person tagging every item by hand. Poor performance produces vague or misleading tags, which does not show up as an error on the day, but resurfaces months later as time lost searching for a document that technically exists but cannot be found, or as duplicated work because nobody realised a relevant file was already on record.

​

Syntactic Organisation Assessment. This test asks platforms to reorder a jumble of words into a sentence that is both grammatical and meaningful. It is a useful lens on how well a system has internalised the syntax rules of a language.

​

Public-facing communications, official correspondence, customer notices, technical manuals, need to read as though a competent professional wrote them, first time, with minimal editing. A platform that constructs sound sentences reliably can draft or clean up this kind of text at volume, cutting the copy-editing workload that otherwise falls to communications teams and freeing staff for substance over phrasing. A platform that struggles here produces text that reads as awkward or unclear, which might seem a minor issue until it appears in a public notice, a safety instruction, or a shareholder communication, where unclear phrasing is not just an aesthetic problem but a source of genuine confusion, complaints, or, in regulated sectors, a compliance finding.

​

Synonym Recognition Test. Platforms are given a word and asked to supply or select a near-equivalent, revealing the breadth of their vocabulary and their sensitivity to fine shades of meaning. This is a well-established way to gauge vocabulary range and the recognition of semantic overlap.

​

Any organisation producing large volumes of written material, training manuals, marketing copy, multilingual public information, internal knowledge bases, needs that material to read naturally rather than as a mechanically repetitive draft, and needs its search tools to find a document even when a user's search term is not the exact word used in the text. A platform with strong synonym recognition supports both: it varies its own language convincingly, and it can power search and retrieval systems that find the right document however a person happens to phrase a query, reducing both editorial rework and the time staff spend hunting for information that technically exists somewhere in the system. Weak performance shows up as flat, repetitive drafting that needs a human pass to fix, and as search tools that miss relevant results, a quiet but real productivity drag across large organisations.

​

Verbal Reasoning Tests

​

Our Verbal Reasoning instruments grew out of a related set of questions. Can a platform see how different words or objects relate to one another, and sort information into meaningful groups based on shared traits? Is it able to catch a flawed argument, judge whether reasoning holds together, and land on an accurate conclusion? Can it string ideas together in a logical order, communicate them clearly, and hold a reader's or user's attention across a conversation or a piece of writing? Can it put mathematical and statistical thinking to practical use, for instance on questions of cost, output, or efficiency? And can it make sense of spatial descriptions, whether in text or image, including where objects sit relative to one another and to itself? With these questions in mind, we built our Verbal Reasoning tests.

​

Word Categorisation Test. Platforms are given a list of words and asked to sort them into groups by shared characteristics or theme. The task reveals Verbal Reasoning through a system's capacity to spot abstract connections between words, recognise recurring patterns, and work with nuanced meaning, going well beyond simple vocabulary recall into higher-order thinking that underpins both communication and problem-solving.

​

This is essentially the skill of sorting a large, messy inbox of unstructured material, customer feedback, complaints, information requests, procurement submissions, into the right categories without a human doing it item by item. A platform that categorises well can process a high volume of incoming correspondence or case files automatically, routing each one to the right team and giving management an accurate, near-real-time picture of what is actually coming in. A platform that categorises poorly creates a different kind of hidden cost: items land in the wrong queue, complaints get treated as routine enquiries or vice versa, and someone eventually has to notice the misfile and manually re-route it, by which point a service deadline may already have been missed.

​

Deductive Reasoning Test. Platforms are given a set of premises and asked to draw a valid conclusion through logical inference. Doing so well demands close reading and critical interpretation: understanding how pieces of information relate, spotting the underlying pattern, and applying logical rules correctly. Because this leans so heavily on language comprehension and the handling of verbal concepts, it doubles as a strong test of both logical and verbal skill, particularly the capacity to sift relevant detail from noise and to explain one's reasoning clearly.

​

Few of the tests in this battery map more directly onto high-stakes decision work than this one. Determining whether a case meets the criteria for a benefit, a permit, a tax exemption, or a loan approval is, at its core, an exercise in deductive reasoning: given these rules and these facts, what is the correct outcome? A platform that reasons deductively well can pre-screen large caseloads reliably, letting human case officers focus their time on genuinely borderline decisions rather than routine ones, which materially increases the volume of cases a team can clear. A platform that reasons poorly here is dangerous precisely because its errors look confident: a wrong eligibility determination, wrongly granted or wrongly denied, generates appeals and rework, and, in the public sector, reputational and legal exposure, which is why this is exactly the kind of task organisations are right to keep under close human oversight until performance on this test is consistently strong.

​

Linguistic Cohesion Test. This test looks at how well a platform reads the connective tissue of language, transition words, pronoun reference, and the logical links that hold a passage together. It measures a system's ability to follow an argument as it develops, infer relationships that are implied rather than stated, and grasp a text's overall shape, offering a window onto analytical thinking, reading comprehension, and sensitivity to linguistic nuance.

​

This test matters most wherever a platform is asked to produce, or check, a long document rather than a short answer, a policy brief, a due-diligence report, a board paper, an audit report. A platform with strong cohesion can be trusted to draft or summarise a genuinely long document without losing the thread partway through, which is precisely what makes it useful for cutting the drafting time on the kind of substantial reports that otherwise take a skilled analyst days to produce. Weak cohesion shows up as a document that reads sensibly in any given paragraph but contradicts itself, or loses its own argument, over several pages, an error type that is hard to catch on a quick read and can therefore slip past a reviewer, undermining confidence in AI-assisted drafting for exactly the long-form work where the time savings would otherwise be largest.

​

Quantitative Analysis Test. This test asks platforms to interpret and reason about numerical information embedded in a problem statement. Though it looks quantitative on the surface, it also draws on Verbal Reasoning, since understanding the written problem, isolating what matters, and building a logical solution path are all verbal skills. It differs from a plain Numerical Ability test by demanding more: real-world application of mathematical ideas, interpretation of data, and problem-solving strategy, rather than speed on basic arithmetic.

​

Budget narratives, business cases, and financial memos routinely mix prose and numbers, a paragraph explaining a decision, with the figures that justify it woven through the text, and getting both right at once is harder than getting either right in isolation. A platform that handles this well can draft or check a budget note, a business case, or a cost-benefit memo where the numbers and the narrative genuinely agree with each other, saving finance and planning staff a meaningful amount of cross-checking time. A platform that handles it poorly produces documents where the prose says one thing and the underlying figures say another, an inconsistency that is easy to miss on a first read and can lead a decision-maker to sign off on a conclusion based on a figure that was never actually correct.

​

Spatial Awareness Test. We wanted to know whether platforms could engage with spatial questions posed in text or image form, while recognising that any such ability would work quite differently from human spatial cognition, which relies on intuitive processing that language models simply do not have. A platform might interpret a written description of a spatial layout, or read an image depicting one, without truly experiencing space the way a person does. Spatial Awareness is usually treated as a separate category from Verbal Reasoning in cognitive testing, since the latter centres on language-based inference and comprehension. Our hybrid test presents spatial problems through written description, engaging both skill sets at once, though this pairing does not make spatial awareness a verbal skill in its own right.

​

Infrastructure, logistics, and construction-adjacent work regularly generates written descriptions of physical layouts, a warehouse floor plan described in a tender document, a site description in a planning application, a network diagram summarised in a report, and someone has to be able to read those descriptions accurately before any decision gets made. A platform that handles this well can support an early read of planning submissions, logistics layouts, or facility descriptions, helping non-specialist staff triage which submissions need a full technical review and which do not, saving specialist engineers and planners from having to look at every single case personally. A platform that handles it poorly risks a plausible-sounding but wrong reading of a spatial layout being taken at face value, a genuine safety and cost risk in any sector where the physical world, not just the paperwork, is what ultimately has to work.

​

Rational Thinking Tests

 

Can platforms sharpen the way we think and help us solve problems more effectively? Are they able to make sense of complicated situations, trace how events connect, spot recurring patterns, and forecast the likely results of an action? And can they think abstractly and creatively in ways that echo human ingenuity, arriving at solutions no one has tried before? These are the questions behind our Rational Thinking tests.

​

Analogies Test. An Analogies Test gauges whether a platform can identify the relationship between a pair of concepts and carry that relationship over to a new pairing. Platforms are shown a related pair, then a third term, and asked to complete the analogy. Because it calls for logical reasoning, pattern recognition, and the ability to abstract a relationship and reapply it elsewhere, strong performance points to flexible, abstract thinking. For an AI platform, doing well here suggests advanced language understanding and reasoning, though it should not be mistaken for human-style general intelligence: it mainly shows the system's capacity to manipulate abstract concepts, which matters for natural-language understanding and generation more broadly.

​

Explaining an unfamiliar regulation, a new technology, or a policy change to a non-specialist audience is very often an exercise in finding the right analogy, the comparison that makes an unfamiliar idea land instantly because it echoes something the audience already understands. A platform that reasons well by analogy is genuinely useful for drafting the plain-language briefings, explainer notes, and public communications that sit between technical experts and decision-makers, work that currently consumes real time from communications and policy staff. A platform that reasons poorly by analogy tends to produce comparisons that are subtly wrong or misleading, which is a quiet risk: a flawed analogy can shape how a decision-maker understands a problem long after the specific words are forgotten, so a bad comparison used in a briefing can do more lasting damage than a bad sentence.

​

Artificial Language Test. This test presents a platform with an invented language, complete with new words, grammar, and syntax rules that exist nowhere else, and asks it to apply those rules correctly across a range of different situations. Stripping away the familiarity of a real language isolates a system's raw capacity for pattern recognition, rule application, logical reasoning, and cognitive flexibility, since there is no way to fall back on memorised vocabulary or prior exposure. Success suggests fast learning and the ability to generalise rules to unfamiliar contexts, a useful signal of adaptability. That said, performance here is shaped by training data and underlying algorithms as much as by anything resembling genuine language acquisition, and strong results are better read as evidence of sophisticated natural-language processing than of true linguistic creativity, so it is best interpreted alongside other Rational Thinking measures rather than in isolation.

​

Every organisation that adopts an AI platform has its own internal shorthand, coding schemes, form numbers, product names, department acronyms, and regulatory clause references that no general-purpose model has ever seen before deployment. This test is, in effect, a proxy for how quickly a platform can pick up an organisation's own rules and conventions once it is put to work. Strong performance means shorter, cheaper onboarding: less fine-tuning, fewer configuration cycles, and fewer early mistakes while staff are still teaching the system the local vocabulary. Weak performance means the opposite, a longer bedding-in period during which outputs need heavier checking, and a higher ongoing maintenance cost every time an internal process or naming convention changes and the platform has to relearn it.

​

Cause & Effect Test. Here, platforms are shown a scenario and asked to identify the most plausible cause or the most likely consequence of what is described. The test measures the ability to work through a complex situation, separate correlation from causation, and apply logical reasoning to a real-world case, including weighing multiple factors at once, ruling out weak or implausible explanations, and reaching a conclusion that is grounded in the evidence rather than in surface pattern-matching. Strong performance suggests a system can process context, draw sound inferences, and mimic human-style reasoning, though, as with the other Rational Thinking tests, this is best treated as evidence of sophisticated language processing rather than proof of genuine understanding or consciousness. Its practical value lies in what it reveals about a platform's ability to give coherent, contextually sound answers, useful for decision support, troubleshooting, and generating plausible explanations for real phenomena that a user might bring to it.

​

This is close to the core skill required for any diagnostic function, an IT help desk working out why a system went down, an inspector working out why a piece of equipment failed, a fraud team working out why a transaction pattern looks wrong. A platform that reasons well about cause and effect can act as a credible first line of triage, narrowing a wide field of possible explanations down to the two or three worth a specialist's time, which directly increases the volume of incidents a technical or investigative team can get through. A platform that reasons poorly sends investigators down false leads, wasting exactly the specialist hours the tool was meant to save, and in a safety-, fraud-, or security-critical setting, a wrong causal read is not merely inefficient, it can mean the real problem goes unaddressed while attention is spent elsewhere.

​

Logical Problems Test. This test presents puzzles or scenarios that call for systematic analysis, pattern recognition, inference, and sound conclusions drawn from a defined set of given premises. Working through a chain of logical steps exercises deductive and inductive reasoning, the ability to spot logical fallacies, and sound judgement under the constraints of the evidence provided, without room to fill gaps with outside assumptions. A strong result indicates that a system can process information methodically, follow logical rules faithfully, and reach rational conclusions on the basis of what it has been given, though this reflects sophisticated pattern-matching and programming rather than human-equivalent understanding. Its relevance lies in demonstrating a system's readiness for tasks that hinge on careful logical analysis and structured decision-making, which are relevant to a wide range of practical applications.

​

Rule-bound administrative decisions, scoring a tender against fixed criteria, checking a form against a compliance checklist, allocating a resource according to a published policy, are everywhere in both government and large organisations, and they are exactly the kind of task where a platform's logical reliability translates directly into staff time saved. Strong performance here supports genuine, safe automation of first-pass decisions at volume, freeing skilled staff for the judgement calls that machines should not be making alone. Weak performance is costly in a specific way: because the errors are rule-application errors rather than obviously wrong answers, they often look plausible enough to pass a quick check, which means every output has to be independently re-verified by a compliance officer or case worker, quietly erasing the time saving the automation was supposed to deliver in the first place.

 

Number Series Test. Platforms are shown a sequence of numbers and asked to identify the pattern governing it and predict the next value. The task exercises pattern recognition, logical reasoning, and mathematical aptitude, requiring a system to work out the numerical relationship at play and apply it systematically. A good result signals strong numerical analysis and pattern-spotting, but this is more likely a reflection of fast processing shaped by training data than of any deeper mathematical understanding, so it works best as one component within a broader battery of analytical tests rather than a standalone measure.

​

Spotting a trend, a demand curve turning upward faster than expected, a budget line quietly drifting off track, a spike in network traffic that does not fit the usual pattern, is a routine part of running a public service, a utility, or a business, and it is often first noticed by a human staring at a spreadsheet or a dashboard. A platform that is good at this kind of pattern extrapolation can flag emerging trends early and continuously, across far more data series than any team could watch manually, which turns monitoring from an occasional manual exercise into an ongoing automated one. A platform that is weak at it produces forecasts that look reasonable but are quietly wrong, a genuine planning risk: budgets, staffing levels, or infrastructure capacity get set against a trend line that never actually happens, and the gap only becomes obvious once it is expensive to fix.

​

Numerical Ability Tests

​

How capable are platforms when faced with multi-step problems that blend arithmetic, spatial reasoning, and data analysis? Do they correctly follow the standard order of operations? Can they simplify and solve equations involving fractions with precision? Are they able to handle engineering or financial calculations? How well do they reason about the position, distance, and movement of objects in two and three dimensions, and can they carry that reasoning into physics, engineering, or navigation problems? Can they read a dataset, spot a trend, and reach a probability-based conclusion, applying statistical thinking to real situations? And can they walk through their working step by step, showing their reasoning as they go? These questions shaped our Numerical Ability tests.

​

Temporal Mathematics Evaluation. This test covers age-related calculations and other forms of time-based arithmetic, from working out how old someone will be at a future date to reasoning through problems that combine several time intervals. It sits within basic arithmetic and logical reasoning, territory that most current AI systems handle comfortably, so results here mainly speak to a platform's ability to manipulate numbers, grasp time-related concepts, and execute elementary mathematical steps correctly, rather than to any particularly advanced numerical skill.

​

A surprising amount of public administration runs on age- and date-based eligibility: pension age, benefit thresholds, licence renewal windows, statutory notice periods. These calculations look trivial individually but are done at enormous volume, and an error rate that would be negligible on a handful of cases becomes a real administrative and financial problem multiplied across a national caseload. A platform that gets this consistently right supports safe, low-oversight automation of a genuinely large volume of routine eligibility work, freeing case officers for exceptions and appeals. A platform that gets it wrong, even occasionally, creates wrongful payments, wrongful denials, and the appeals and audits that follow, precisely the kind of error that is expensive not because any single instance is large, but because it repeats at scale before anyone notices the pattern.

​

Order & Fraction Mathematics Test. This test carries particular weight in evaluating Numerical Ability. It shows whether a platform can apply basic mathematical rules consistently, a foundation for more advanced numerical reasoning, and whether it can work confidently with abstract mathematical notation. Because the problems demand exact answers, the test is a direct check on computational accuracy, and the ability to break a complex expression into steps points to genuine, structured problem-solving rather than guesswork.

​

This is the arithmetic that underpins invoicing, tax computation, procurement cost breakdowns, and engineering specifications, work where there genuinely is only one correct answer, and every subsequent calculation that builds on it inherits any mistake made early on. A platform that is reliably accurate here can be trusted to handle calculation-heavy back-office work, billing runs, tax preparation, cost estimation, at real volume with a lighter audit burden than a human-only process would require. A platform that is not reliably accurate is, in practical terms, worse than no automation at all in this specific area: every output still has to be independently recalculated to be trusted, so the organisation pays for both the automation and the verification, with none of the promised time saving, and a single undetected error can compound silently through every downstream figure that depended on it.

​

Spatial Reasoning Test. This test might, for example, ask a platform to calculate distances and angles between objects, or to solve a physics problem involving motion and force vectors. It bears on Numerical Ability, but the two are not identical: spatial reasoning draws on arithmetic but goes further, requiring a broader grasp of how objects relate to one another in space, rather than just the ability to compute a correct figure. Strong performance shows not just computational skill but the capacity to apply mathematical ideas to real, multidimensional problems, which is a distinct and somewhat more demanding form of numerical competence.

​

Engineering feasibility checks, infrastructure cost estimates, and network or logistics planning all depend on getting distances, angles, loads, and routes right, and an early, rough pass at these numbers is often what determines whether a project proposal even reaches a full technical review. A platform that handles this well can support that early screening stage, filtering out proposals with an obvious technical or physical problem before they reach a scarce specialist engineer, which increases the volume of proposals a technical team can realistically assess in a given period. A platform that handles it poorly is a genuine risk precisely because a wrong technical answer can sound just as confident as a right one, and taking a flawed feasibility read at face value, in construction, telecoms, or transport, is the kind of error that surfaces expensively, later, and on site rather than on paper.

​

Statistical Reasoning Test. This test asks platforms to work with datasets: interpreting statistical measures, drawing inferences, and reaching conclusions grounded in probability rather than certainty. Beyond producing the right number, a platform is expected to explain its reasoning, discuss what the statistical findings actually imply in context, and flag any limitations, gaps, or potential bias in the underlying data, since a correct figure delivered without that surrounding judgement tells us less about genuine statistical reasoning.

​

Public-health reporting, economic forecasting, market research, and quality-assurance sampling all depend on someone reading a dataset correctly and stating honestly what it does and does not support, and this is arguably the highest-stakes test in the whole battery precisely because a wrong statistic, once repeated up a chain of briefings, tends to travel a long way before anyone checks the original source. A platform that reasons well statistically, including being honest about the limits of its own data, earns the kind of trust that lets analysts rely on its first draft of a statistical summary rather than re-deriving every figure independently, a genuine time saving in any data-heavy function. A platform that reasons poorly, especially one that states an unqualified conclusion from a shaky dataset with full confidence, is a serious risk in this context: an overconfident wrong statistic can end up shaping a public statement, an investment decision, or a policy position before anyone traces it back to its source.

​

Time-Related Calculations. This test covers arithmetic involving time units, dates, durations, and time zones, testing a platform's grip on calendar systems and different ways of representing time. It matters for Numerical Ability because it draws on several skills at once: basic arithmetic, plus an understanding of non-decimal units, such as 60 seconds to a minute or 24 hours to a day, among other quirks of how time is measured. Doing this well points to a platform's ability to apply mathematical concepts flexibly, convert between units, and juggle several variables simultaneously, a strong signal for how it will handle the time-sensitive, practical numerical questions people are likely to bring to it in daily life and at work.

​

Any organisation that operates across time zones, and that now includes most public administrations coordinating with international partners and most companies of any size, runs into this arithmetic daily: scheduling a meeting across time zones, calculating a service-level deadline, working out whether a contractual notice period has actually expired. A platform that handles this reliably removes a small but constant source of friction and error from cross-border coordination, scheduling, and compliance tracking, freeing administrative time currently spent double-checking dates by hand. A platform that gets it wrong produces a specific, recurring kind of failure: a missed meeting, a breached service-level agreement, a deadline calculated in the wrong time zone, each individually minor, but collectively a real and entirely avoidable cost once an organisation is coordinating at any scale across borders.