What's new

Changes that affect how we use the Workbench and what its figures say, newest first. Developers will find the technical detail in the repository's changelog; these notes cover what users see.

6 October 2026

Where the savings land

• Savings by group. A new staff-list tab shows people, cost, saving, saving as a share of cost and FTE saved by business, division, department, role, location, tenant or any extra grouping, with a second grouping nested inside the first, a heat map, any scenario and a CSV download. Every grouping adds up to the report's figure, and the tab says so. See Savings by group.
• Each person's saving multiplier. The Explorer shows a Multiplier column (the moderate saving as a share of the person's cost) and a Department filter, and its CSV adds sub-division, department, extra groupings, Savings as a cost multiplier (moderate), FTE and FTE saved (moderate). See Explorer.
• FTE and extra groupings on upload. Map an FTE column (between 0 and 1; it counts FTE saved and never changes cost) and tick up to five Extra groupings, such as an IFA firm. A column that looks like names, identifiers, people (such as managers or advisers), emails or contact details, or that has a different value for almost everyone, is refused, in the browser and again on the server. See Uploading a staff list.
• The finance workbook adds the multiplier, FTE and FTE saved to Raw Data, FTE saved to every group sheet, and By Division, By Department and By (grouping) sheets, each adding up to the same total.
• Report wording. The engagement lead can add a paragraph to the end of any report section, replace the executive summary's value-concentration point and swap words (such as Workforce cost to Cost to serve) on the engagement's Settings tab. It appears in the HTML report, its Markdown and the deck, and never changes a figure: swaps match whole words whatever their capitals, can't contain digits or currency, and never touch the methodology and sources. The wording can be changed only while the engagement is active, and purging the engagement clears it. See Your own wording.
• Clearer standard wording for everyone: scenarios describe productivity improvements and capacity creation rather than guaranteed cash savings; the value-chain line gives capacity creation, productivity improvement and the total impact as a share of the lifecycle workforce; the adoption curve, offshore estimate and role explorer each explain themselves.
• New methodology entries: Saving multiplier and FTE saved.

IDs are never emails or names

• A column of email addresses, or one that is mostly people's names, can no longer be mapped as Employee ID or Manager's employee ID. The mapping says why and which field the column belongs in, and the upload waits until it's changed. The Workbench also refuses an email as an ID on its own side. Staff lists already uploaded are unchanged.

Spans and layers report

• A full Spans and layers report on Exports: summary, data quality, spans, layers, by business, division and department, the cost of managers and every manager, as HTML with a presentation mode and Markdown, a PDF through print, and a CSV of managers. IDs and job titles only, never names. See Spans and layers report.
• Disconnected managers are counted. The report lists managers who are named but missing from the staff list, and shows the figures as reported and with each missing manager added back (adjusted layers are a minimum). A manager reference that looks like an email or a name is never printed. A staff list with no reporting lines says so rather than offering the report. New methodology entries: Disconnected managers, Cost of managers, and Narrow and wide span thresholds (an assumption).
• The Organisation tab's narrow-spans table says how many it isn't showing (it lists the 15 most expensive) and links to the report. See Organisation.
• Managers named by email who aren't in the list are kept as opaque numbered stand-ins (never the email) on upload, so they can be counted. Staff lists uploaded earlier show those people as having no manager.

3 October 2026

Client-branded deliverables

A client can now have a brand: its logo, colours and font, set on any of its engagements' Settings tab by an administrator, the client's creator or one of its engagement leads. When we export, the new Brand card offers Davies, Co-branded with the client, or the client's own brand with a Prepared by Davies credit. It applies to every HTML report, PowerPoint deck and Excel workbook. See Exports and Client brand.

Also new

• Add your name in the account menu: single sign-on gives the Workbench our email address but not our name.
• Rename a staff list from its header.
• Open drop-down lists use the Davies colours in Edge and Chrome.
• When a new version is deployed, an open Workbench offers Refresh now. See When a new version is released.

2 October 2026

The built-in reference data is up to date with its sources

We checked every built-in table and constant against its primary source. What changed:

What	Before	Now	What it means
AI exposure (AIOE)	Keyed by the older SOC 2010 codes, so 181 SOC 2018 occupations (software developers, data scientists, security analysts, project managers among them) fell back to a default.	Crosswalked to SOC 2018 with the US Bureau of Labor Statistics' official crosswalk, percentiles re-derived: 800 of 867 occupations covered, up from 686.	Roles in those occupations now take their exposure from the published research. The remaining occupations have no published score (military, "all other" catch-alls, data scientists and legislators).
Automation share (Anthropic Economic Index)	0.426	0.486 (June 2026 release)	Automation is augmentation × this share, so average automation rises for every role; augmentation is unchanged.
UK pay (ONS ASHE)	All employee jobs	Full-time jobs, ASHE 2025 provisional	Benchmark salaries now cost full-time equivalents. Unsupported proxies were removed, and Learning & Development and Secretary/PA moved to their SOC 2020 codes.
Copilot licence price	£15 a user a month	£23.10	Microsoft's UK list price for Microsoft 365 Copilot (enterprise), paid yearly, excluding VAT. Still an assumption: use the client's quote where there is one.
Research benchmarks by occupation group	Earlier figures	Re-sourced to Massenkoff & McCrory, Labor market impacts of AI (Anthropic, 2026), both columns employment-weighted	The Against the published research card and the reports' benchmark section quote the new figures; Management's observed coverage is no longer shown as zero.
Employer cost	National Insurance and pension sources for an earlier year	HMRC and The Pensions Regulator, 2026/27	The 1.20 employer-cost multiplier is now labelled an assumption: the statutory minimum alone comes to about 1.14 to 1.16, and 1.20 allows for pension and benefits above it.
Citations		Corrected	Exact titles, wording and links for the AIOE paper, McKinsey, BCG's 10-20-70 and Eloundou et al. Sources are no longer cited for assumptions they don't contain.

What this means for existing staff lists

Nothing moves until we choose. Every staff list keeps the reference data version it was scored with, so figures a client has already seen don't change. On a new environment, the refreshed data is the built-in version 1. On an environment that was already running, administrators publish it as a new reference data version, and each staff list then offers Rescore with version N. Rescoring keeps overrides and AI classifications and calls no AI.

Before rescoring a staff list that a client has already seen, expect higher average automation everywhere (and so, all else equal, higher capacity figures), different benchmark salaries for people without a client salary, and different exposure for roles in the newly covered occupations. Saved rollout plans keep the licence price they were saved with; change it in the plan if the client should see the new list price.

Privacy and security

• Benchmarks are harder to reverse-engineer. A benchmark figure now needs five contributing engagements, counted after leaving out the engagement being compared, and published figures are rounded (pounds to the nearest £500, percentages to whole numbers, spans and layers to the nearest 0.5). Below five, the figure is withheld.
• Nobody can grant themselves access to names and emails, administrators included. The person who creates an engagement is still its first lead with access.
• Turning AI off at upload is remembered. If we untick Classify job titles with AI, choosing Continue on an interrupted upload never sends that staff list's titles to the AI.

More accurate figures

• Changing an engagement's industry pack rescores its staff lists, so the new pack's rules apply at once.
• The realism discount now applies to roles scored by the Workbench's own rubric too, not only to roles with a published anchor.
• Rollout plans on a staff list without reporting lines no longer license everyone by accident: the starter plan uses high-fit roles instead, and a notice explains why leadership and reporting-line selections find nobody.
• The finance workbook groups exactly as the app does, so Sales and SALES are no longer double-counted in Excel; its locations sheet quotes the offshore threshold and the factor each person was priced at.
• Copilot payback beyond ten years reads Over 10 years instead of being capped at 120 months.
• The value-chain map says which roles the AI placed.
• Overrides are never overwritten by a classification finishing at the same time, and refresh every view that depends on them.
• Salaries such as 45.5k are read correctly, and implausible ones (below £1,000 or above £10m a year) are left out with a warning, so the benchmark salary applies instead.
• Compare warns when two staff lists use different reference data versions or assumptions: These staff lists aren't scored the same way.

Clearer, more consistent screens and reports

• "Dataset" is now staff list throughout: the engagement's first tab is Staff lists.
• Closed engagements are read-only everywhere and say so, with one notice naming the real reason (closed, purged or viewer); inputs that can't be saved are disabled.
• The row-level tabs of a purged staff list lead back to its summary; a deactivated account gets a clear message rather than a retry.
• Figures are formatted the same way everywhere: compact pounds for modelled figures, one decimal place for shares, ranges such as £4.7m–£7.2m and 2.3–3.7×.
• On phones, buttons are easier to tap and wide tables scroll sideways with a shaded edge.
• Exported reports: tables scroll on phones, the floating buttons (Markdown, Present, PDF) no longer cover the text, and icons replace emoji.

24 September 2026

• Demo data in one command. Developers can fill a development database with three fictional clients, a team of five and realistic engagements (npm run demo:seed). See the Quick start.
• Short words in title rules match whole words only. Credit Analyst and Audit Analyst were wrongly classed as IT support because they contain "it analyst"; they're now an operations analyst and an auditor. The HR and MI rules had the same weakness and were fixed too.
• Exposure anchors use the AIOE table's own percentiles. 32 roles quoted a percentile 0.1 away from the published table; exposure for those roles moves by a rounding step at most.

23 September 2026

Versioned reference data and prompts

• Everything the model runs on is data, not code: occupation tables, AI exposure, pay, title rules and roles, scenario rates, adoption and investment, locations, the offshore view, value chains, benchmarks, Copilot fit, industry packs and sources. Administrators edit a draft, check it and publish numbered versions on Admin → Reference data. See Reference data.
• Each staff list is pinned to the version it was scored with, and offers Rescore with version N when a newer one is published.
• The four AI prompts are versioned too, on Admin → Prompt Lab: draft, preview, test beside the live version, publish. See Prompt Lab.

Sign-in

• Sign out in the account menu, and, in development only, Sign in as to switch between test users.
• An expired sign-in reloads the page to sign in again.

Security

• Job titles reach the AI classifier as quoted data, and a batch of answers that looks steered is used for its staff list but never shared through the classification cache.
• AI spend counts the model's thinking, and Ask this workforce has a daily limit per person.
• Reference data and prompt drafts can't be overwritten or published from a stale copy: if a colleague saved in between, the page asks whose version to keep.

Quality

• Tables no longer widen the page on phones.
• This user guide.

22 September 2026: first release

The first release of Workforce Workbench brought the earlier Workforce Impact (COGS) analysis tools together as one Davies platform:

• Engagements and access: leads, analysts and viewers per engagement; access to names and emails granted per person; administrators; enforced by the database.
• Personal data: names and emails stored only by choice, encrypted per engagement and destroyed on purge; automatic purging after a retention period once an engagement closes.
• Staff-list upload in the browser: column mapping, row filters, manager-email resolution, Microsoft 365 tenants from mail domains, resumable upload.
• Classification by override, industry pack (general, insurance and claims, banking), shared cache, AI or rules, with a review queue and engagement-wide overrides.
• The AI impact model anchored to published research, with provenance on every figure; the organisation analysis; the Copilot business case and rollout planning by wave and tenant.
• Comparison and benchmarks across staff lists and engagements.
• Davies-branded deliverables: the HTML report with presentation mode, PowerPoint decks, the live-formula finance workbook, the Copilot allocation workbook, the Copilot business-case report and CSV extracts.