Glossary

AEI (Anthropic Economic Index). Anthropic's published data on how Claude is used across tasks and occupations. The Workbench takes the automation share of conversations classified as automation or augmentation (48.6% in the June 2026 release). The research benchmark by occupation group comes from Anthropic's related study, Labor market impacts of AI (2026).

AIOE (AI Occupational Exposure). Felten, Raj & Seamans' measure of how exposed each occupation is to AI; the Workbench uses the Language Modeling index, crosswalked from SOC 2010 to SOC 2018. Each occupation's percentile, which the Workbench works out over the crosswalked occupations, drives its augmentation exposure.

Anchor. The published US occupation a Workbench role takes its exposure from. A role with no anchor uses the rubric, labelled as an assumption.

ASHE. The Office for National Statistics' Annual Survey of Hours and Earnings: median gross annual pay for full-time jobs by UK SOC-2020 occupation, the benchmark salary when the client doesn't supply one.

Assumption. A figure that is a judgement we can change, not a measured value, and labelled as such.

Augmentation. The share of a role's work AI can assist or accelerate: the AIOE percentile mapped onto a 30%–85% band.

Augmentation headroom. Augmentation minus automation: work AI accelerates but doesn't replace. The default value base for the Copilot case.

Automation. The share of a role's work AI could perform outright: augmentation × the AEI automation share.

Baseline region. The region every location factor is relative to (the United Kingdom, factor 1.00, as shipped).

Blind index. A keyed hash of each stored email, so a client's email list can be matched without decrypting anything.

Capacity gain. A scenario's gross annual value of the time AI frees: employer cost × (automation × automation realised + headroom × augmentation productivity). Not a cash saving or a headcount target.

Classification cache. The shared store of AI classifications by normalised title and industry pack, reused across engagements that allow sharing.

Cohort. In a rollout plan, a named rule selecting people to license (by leadership level, reporting line, attribute, fit tier, managers or a named list), with optional exclusions.

Crypto-shredding. Destroying an engagement's encryption key on purge, so stored names and emails can no longer be decrypted from the live database. Backups taken before the purge keep the key and the encrypted data until they age out.

Disconnected manager. A manager named in someone's manager column who isn't in the staff list (they left, were filtered out at upload, or the ID is mistyped). Their reports look like top-level people, which understates layers and managers. The Spans and layers report counts them and shows how far they could move the figures.

Deployed base. In a rollout plan, existing licence holders plus every new licence across the waves.

Employer cost. Salary × the employer-cost multiplier (1.20 by default): an assumption covering employer National Insurance, pension and benefits, above the statutory minimum of about 1.14–1.16 (2026/27).

Engagement. One piece of client work: its staff lists, overrides, plans, team and audit trail.

Executive summary. The board-ready summary on the Overview and in reports: rule-based by default, or AI-written on request, labelled either way.

Extra grouping. One of up to five client columns (an IFA firm or a region, say) chosen at upload so savings can be shown by it. Categories only: a column that looks like names, identifiers, emails or contact details, or that has a different value for almost everyone, is refused.

Fit tier. How central Office applications are to a role: high, medium or low fit, or uncategorised (no benefit modelled). Sets the share of value Copilot captures.

FTE. A person's full-time equivalent from the staff list, between 0 and 1 (1 when not given). It counts FTE saved and never changes the person's cost.

FTE-equivalent. Capacity gain ÷ average employer cost: the capacity expressed as full-time roles. It follows cost, where FTE saved follows headcount.

FTE saved. Each person's saving multiplier × their FTE, added up for a group: capacity counted person by person. Close to, but not the same as, the FTE-equivalent, and not a headcount-reduction target.

GCC. Global capability centre: an offshore hub. The offshore view shows people already offshore and migration candidates.

Industry pack. Industry-specific classification rules, a value chain and Copilot fit lists, chosen per engagement.

Lead, analyst, viewer. The three engagement roles. See Team and roles.

Location factor. A multiplier on benchmark salaries for a region, relative to the baseline.

Migration candidate. A desk-based, location-portable role in a location costing at least 90% of the UK.

Narrow span. A manager with 2 or fewer direct reports (built-in threshold). A wide span is 15 or more.

Override. A consultant's explicit classification of a title for the whole engagement, with a rationale. Wins over every automated source.

PII access. Per-member permission to store, reveal and match names and emails on an engagement. Granted by another lead or an administrator; nobody can grant it to themselves.

Prompt Lab. The Admin screen for editing, testing and publishing the AI prompts.

Provenance. Where a figure comes from: its basis (measured, framework, classified, calibration, derived or assumption) and its sources.

Purge. Deleting an engagement's names, emails, rows and plans (and shredding its key), keeping anonymised summaries.

Realism discount. An optional assumption that scales exposure down for adoption friction without changing the order of roles.

Reference data. The versioned tables and constants the model runs on, published by administrators. Each staff list is pinned to a version.

Report wording. The engagement's own text for its deliverables, set by a lead on the engagement's Settings tab: section paragraphs, a value-concentration point and word swaps. It never changes a figure.

Review threshold. The confidence below which an AI or cached classification needs review (60% by default).

Rubric. The Workbench's own exposure estimate for roles with no published anchor; always labelled an assumption.

S-curve. The adoption curve that phases the moderate scenario's capacity over nine half-years.

Saving multiplier. The share of a person's employer cost that a scenario turns into capacity: automation × automation realised + (augmentation − automation) × augmentation productivity. Everyone in a role shares it, and a person's saving is their employer cost × it, so savings by any grouping add up to the scenario total.

Scenario. Conservative, moderate or aggressive realisation rates for automation and augmentation.

SOC. Standard Occupational Classification: US SOC-2018 (the exposure key) and UK SOC-2020 (the salary key).

Spans and layers report. The organisation deliverable on Exports: spans, layers, groupings, the cost of managers, every manager by ID and job title, and the quality of the reporting lines behind them.

Staff list. One uploaded list of a client's employees in an engagement, classified and scored, with its own assumptions and reference data version. Some administrator screens and the developer guide call it a dataset.

Summary only. A staff list whose engagement has been purged: the anonymised summary remains for reports, comparisons and benchmarks, on the Overview and Exports tabs.

Tenant. A Microsoft 365 tenant. Licences are bought per tenant, so rollout plans count them separately.

Value chain. The phases of an industry's core operating model (such as the claims lifecycle) across which each role's time is spread.

Wave. A stage of a Copilot rollout plan, made of cohorts, optionally capped.