Case blocks from our engagements

How to read these blocks

We publish case blocks so a sponsor can see the grain of an engagement before they book a scoping call. Every block follows the same pattern: the situation as we found it, the sequence of sessions, and the contents of the handover pack. Sectors and roles are real categories of work. Organisation names, product names and any performance claims supplied by a client stay out of the public record.

Numbers you will see (twelve people, twenty-two prompts, four workflows) are counts of what we taught, wrote or documented together. They tell you about the size of the room and the volume of material we left behind. They do not stand in for a return on investment, a ticket-time reduction, or any other figure we have not independently verified and have no licence to publish.

If you recognise your own function in one of these write-ups, treat it as a starting point for a conversation rather than a template we will copy onto your team. We still begin with a task audit of the work you actually run, because a logistics reply library and a finance commentary pack share a method and almost none of the wording.

Regional customer support team, logistics sector

A regional support function handling shipment queries across several time zones had already issued licences for a general assistant. People were pasting tickets into the tool and sending the first draft that came back. Tone drifted between desks, escalation notes missed the fields the operations team needed, and a handful of replies had quoted tracking details that were not in the source ticket. The sponsor asked us to build a shared way of drafting that still left a human holding the send button.

We ran a six-week cohort online with twelve people. Week one was a task audit of the macros they already used: delay notices, customs queries, damaged-goods first replies, and internal escalations. Weeks two and three were spent rewriting those macros as instructions that named the source fields, the forbidden claims, and the reading age of the customer. Week four introduced a review checklist that had to be completed before anything left the queue. Week five practised escalation notes against anonymised tickets the team brought in. Week six was documentation: we sat with two team leads and turned the working prompts into a library organised by macro category.

The handover contained a 22-prompt library tied to their existing categories, a nine-point review checklist, and a short routine that said when a reply must be written without the tool (legal holds, injury claims, and anything involving a named public official). At the 30-day follow-up the leads were still editing the library themselves. We did not measure handle time; we counted how many of the original macros had a written instruction and a named reviewer.

Finance operations, manufacturing

A finance operations team in a manufacturing group spent the last week of every month turning spreadsheet exports into commentary for a leadership pack. The numbers were already in the workbook. The bottleneck was the narrative: variance explanations, one-line summaries for non-finance readers, and a consistent way of flagging items that needed a human judgement rather than a generated sentence. They did not want a bot writing the pack. They wanted help turning a known export into a first draft that a controller could edit.

We ran a two-week embedded sprint on-site with one team of seven. The first two days were a task audit of the monthly pack: which tables were stable, which commentary lines repeated, and which lines should never be drafted by a model because they implied a cause the workbook did not contain. We then rebuilt four recurring outputs with the team at their own desks: a variance paragraph template grounded in named cells, a rolling commentary skeleton, a checklist for numbers that must be copied rather than regenerated, and a short list of lines the controller would always write by hand. Afternoons were supervised practice on the previous month’s export. Evenings we wrote the workflow pack from the notes on the wall.

The handover was four documented workflows and an explicit “do not automate” list covering causal claims, forward-looking statements, and any figure that had been adjusted outside the workbook. We also left a review step that required the controller to initial three checks: source cell, period, and whether the sentence asserted a reason. Programme parameters: seven people, ten on-site half-days, four workflows, one follow-up call at 30 days. We do not publish the pack’s page count as a proxy for value; the team already knew how long the month-end took.

Bid and proposal team, professional services

A proposal team was under pressure to reuse past submissions without copying stale client names, outdated case figures, or paragraphs written for a different jurisdiction. They had a share drive full of winning documents and a general assistant that was happy to blend them. The risk was not speed. The risk was a paragraph that looked polished and pointed at the wrong source. The sponsor wanted a drafting habit that always named the file a sentence came from, and a hygiene pass before anything went to a partner.

We started with a half-day Fluency Clinic so the whole group of fourteen could practise source-grounded drafting on a redacted past bid. That clinic surfaced the actual failure modes: fabricated project dates, blended case studies, and boilerplate that still named a previous buyer. We then ran the six-week cohort. Early weeks covered task audit and rewriting. Middle weeks were spent building a routine for pulling only from an approved excerpt pack, with a verification step for every figure and every named project. Later weeks documented a boilerplate hygiene guide: how to strip client-specific facts, how to mark a paragraph as reusable, and how to refuse a model output that could not point at a file.

The handover contained the drafting routine, a verification checklist (citation, number, quotation, jurisdiction, client name), and the hygiene guide. Fourteen people attended; the cohort produced one shared excerpt index rather than fourteen personal prompt stashes. We recorded the number of past submissions they chose to bring into the excerpt pack (nineteen) because that is a programme parameter. We did not record, and will not publish, win rates.

Public-sector communications unit

A communications unit needed a policy conversation before anyone opened a tool. Managers were being asked whether staff could use general assistants for press summaries, internal explainers, and first drafts of web copy. The unit already had a records rule and a clearance path. What they did not have was a plain description of where a model is useful, where it invents, and how a usage note should be written so people can follow it on a Tuesday afternoon.

We delivered a 90-minute Leadership Briefing for 25 people, then a Fluency Clinic for 16 of the hands-on staff. The briefing covered what current tools do with a supplied source, where fabricated quotations appear, and how to judge a short pilot (did people still check the source, did anything go out uncleared, did the usage note get used). The clinic used the unit’s own public pages as grounding material. Participants practised summaries that had to quote a paragraph, headlines that had to stay inside the source, and a refusal path for requests that would have meant pasting unpublished material into a general tool.

The handover was a draft internal usage note, written in the room with the sponsor, and a one-page explainer colleagues could circulate. We did not write the unit’s official policy; that remains their document. Programme parameters: 25 in the briefing, 16 in the clinic, one follow-up hour to edit the draft after legal-ops comments. No campaign metrics appear here because we were not engaged to measure them, and we do not invent them after the fact.

HR and people operations, technology company

A people-operations group wanted help with three recurring writing jobs: job descriptions that stayed inside an agreed grade language, summaries of interview notes that did not invent competencies the notes never mentioned, and answers to policy questions that had to be grounded in their own handbook. They had already seen what happens when a general assistant is asked for a “better” job ad: extra requirements appear, tone shifts, and the handbook is ignored in favour of a fluent paragraph.

Nine people joined a six-week cohort. The task audit put job descriptions in the “assist with review” pile, interview summarisation in the same pile with a stricter check, and anything touching individual performance or medical information in the “not suitable” pile. Sessions then practised grounded drafting against the handbook PDF they already used, with a rule that every policy answer must point at a section heading. Interview note work used a template that only allowed phrases present in the typed notes. We added a fairness review step: a second person checks that the summary does not introduce criteria that were not asked in the interview.

The handover was a grounded-answer workflow, a job-description instruction set aligned to their grade language, the fairness review step, and a short list of people-data tasks we recommended they never paste into a general tool. Group size nine; six live sessions; one 30-day follow-up. We do not publish time-to-hire or similar figures. Those numbers, if they exist, belong to the organisation and are not part of our delivery record.

What we count as a good outcome

We look first at whether people are still using the written routine at the 30-day follow-up. If the prompt library has been edited by the team, that is a better sign than if it is still exactly as we left it. A document that only we can maintain is a failed handover, even if the sessions felt busy. The follow-up is a working session, not a satisfaction survey: we open the routine, watch one live task, and patch whatever has drifted.

We also look at whether fewer outputs are leaving the building unreviewed. That is a habit we can observe in the room and in the checklist ticks people keep. It is not a claim about quality scores, conversion, or cost. When a sponsor asks for those figures we say plainly that we do not collect them and we will not put invented percentages on a slide. Programme outcomes depend on your team’s context; the examples on this site describe our delivery format.

The third mark is a written document the team maintains itself: a routine, a checklist, a do-not-automate list. We write the first version with them. They own the second. We do not publish client names, logos, or client-supplied performance figures, and we do not ask for testimonials to display here. If you want to know whether a format would fit your function, send us the work you are trying to shorten and we will tell you which programme, if any, matches it.

Start with a scoping note Read the method