Task statement 4.5 · Workflow integration and solution design

4.5 · Communicating Value and Limitations

In Wrenfield Housing, the course example, the operations team has finished a six-week pilot of its redesigned repairs process (4.2–4.4). The operations director must now present the results to the board and to the tenants' panel. A slide that claims too much is soon disproved, and a slide that only lists risks gets the change dropped.

Why it's tested

The tempting slide sells the pilot, because the people who ran it want the change approved. A line such as "Claude has eliminated triage errors" is disproved by the first error a board member hears about, and after that every other figure on the slide is doubted. The opposite habit fails too. A slide that lists what Claude cannot do, with nothing on how each risk is handled, reads as a reason to stop. Neither slide lets the audience decide, because neither says what changed, by how much, and what keeps the remaining risk in check.

KnowledgeValue is a measured change against a baseline

A stakeholder is anyone who decides about a change or whose work or service it touches: at Wrenfield, the board, the call handlers on the repairs line and the tenants' panel. For a stakeholder, value means a measured change against a baseline, the same measure taken on the same work before the change. Anthropic's advice on adopting Claude says the teams that succeed track concrete metrics, such as productivity gains, time savings and quality improvements. It adds that this evidence is what proves the return to stakeholders. "Much faster" gives the board nothing to check, while "median 6 to 4 working days, over 610 requests" can be checked. This rules out value stated as an impression or as another organisation's result. Each value line on a slide needs four parts:

  1. The measure: one thing you can count, such as working days from report to booked visit.
  2. The baseline: the same measure on the same kind of work before the change.
  3. The result: the pilot figure, with the number of cases and the number of weeks.
  4. What else changed: anything else that could explain the result, named and not credited to Claude.

KnowledgeLimitations, stated plainly, each with its safeguard

A limitation is something Claude cannot do, or cannot be relied on to do alone, and Anthropic's own pages state them plainly. The Claude Help Center says Claude can occasionally produce incorrect or misleading responses, and that users should not rely on it as a singular source of truth. That is why drafts are checked for invented detail (2.2) and reviewed at a level that fits the stakes (2.4). Each model also has a knowledge cutoff, so it may not know about events after that date. Finally, Claude has no context about your organisation beyond what the conversation, your attached files and your connectors give it. A connector, a link that lets Claude reach an app such as Google Drive, works only with the user's own permissions there. A limitation stated alone sounds like a warning, but stated with the safeguard that handles it, it becomes part of the design. This rules out hiding the limitations, and also listing them bare.

Interactive · what backs each line of the results slide

Five lines of the repairs pilot's results slide on the left, each with an arrow to what backs it on the right. Two value lines are backed by a baseline from the six weeks before the pilot. Three limitation lines are each backed by a safeguard. A line with nothing behind it is dashed and labelled no baseline, missing or no safeguard. ON THE SLIDE WHAT BACKS IT Value · speed Repairs booked much faster Report to booked visit: 6 → 4 days BACKED BY NOTHING BASELINE NO BASELINE 6 weeks before vs 6-week pilot nothing anyone can check Value · triage quality Eliminated triage errors Recategorised after booking: 7% → 4% BACKED BY NOTHING BASELINE NO BASELINE 41 of 590 → 24 of 610 jobs one error disproves it Limit · mistakes NOT ON THE SLIDE Claude can make mistakes BACKED BY NOTHING SAFEGUARD MISSING NO SAFEGUARD a call handler confirms each category Limit · access NOT ON THE SLIDE Sees only what it is given BACKED BY NOTHING SAFEGUARD MISSING NO SAFEGUARD call handlers book in the repairs system Limit · knowledge cutoff NOT ON THE SLIDE May not know recent changes BACKED BY NOTHING SAFEGUARD MISSING NO SAFEGUARD current triage guide in the Project A line with nothing behind it is the line a stakeholder stops trusting first.

Step through the pilot's results slide, from the first draft to the version the board saw, and see what backs each line.

First draft: two claims and no limitations. “Much faster” gives the board nothing to check, and a single recategorised job disproves “eliminated triage errors”.

Baselines added: each value is now a before-and-after figure from the same kind of work. The slide still says nothing about what Claude cannot do.

Limitations added: the three limits come from Anthropic's own pages, but nothing says how each is handled, so they read as reasons to stop.

Each limitation with its safeguard: the board can see what changed, by how much, and what keeps the remaining risk in check.

Dashed and labelled NO BASELINE, MISSING or NO SAFEGUARD: nothing backs the line. Solid and labelled BASELINE or SAFEGUARD: a stakeholder can check it.

KnowledgeThe same facts, a different emphasis for each stakeholder

Every stakeholder should get the same facts, because a figure that changes between audiences destroys trust in all of them. What changes is which facts lead, because each group asks the pilot a different question. Fitting the wording to each reader is covered in 2.5. A group whose question goes unanswered tends to resist a sound change:

The board

Asks: should we extend this to every estate, and what could go wrong?

Lead withThe measured results, the risks with their safeguards, and the decision asked.

Staff whose work changes

The call handlers ask: what does my day look like, and what do I still decide?

Lead withThe steps that changed, the time saved, and the category decision that stays theirs.

Tenants' representatives

The tenants' panel asks: are repairs quicker, and what happens to tenants' details?

Lead withThe booking time, no personal data in prompts, and a person confirming every category.

SkillWorked example: the results slide that claimed too much

Wrenfield ran the redesigned process on the Millbrook estate for six weeks, on 610 new repair requests. For each request, Claude suggested a category from the triage guide, using no tenant names or addresses. The call handler confirmed or changed the category before booking the visit in the repairs system. The baseline was the six weeks before, with 590 requests. The operations director's first draft of the results slide read well and claimed too much:

# before: draft slide, "Pilot results"
Claude has eliminated triage errors. Repairs are booked much faster.
Limitations: Claude can make mistakes. It has a knowledge cutoff.
# after
Report to booked visit: median 6 → 4 working days (590 vs 610 requests)
Recategorised after booking: 41 of 590 (7%) → 24 of 610 (4%)
Open jobs 640 → 598: not credited; damp reports fall after the winter peak
Claude can be wrong → a call handler confirms every category (38 changed)
Claude sees only what it is given → bookings stay in the repairs system
Claude may not know recent changes → it works from the current triage guide in the Project

The symptom was a slide no board member could check: the 24 recategorised jobs disprove "eliminated", and "much faster" has no number. The cause was writing from the pilot's impression rather than from its figures. The fix states each value as a before-and-after figure and refuses to credit Claude with the fall in open jobs, because damp reports fall after the winter peak whatever the process. It also puts a safeguard against each limitation. A direct link from Claude to the repairs system stays a job for IT's developers. The 38 categories that call handlers changed stay on the slide on purpose, because they show the check working. The operations director checked the final slide, as the AI policy requires. The tenants' panel then saw the same figures, led by booking time and by how tenants' details are kept out of Claude.

For every line on the slide, ask: can I point to a before-and-after figure, or a named safeguard, behind it? Yes → keep it. No → measure it against a baseline, add the safeguard, or cut the line.
Takeaway
  • State value as a measured change against a baseline from the same work, with the number of cases and weeks, and name any other cause of the result.
  • Take Claude's limitations from Anthropic's own pages, and pair each one with the safeguard that handles it.
  • Give every stakeholder the same facts, but lead with each group's question: the board's decision, the staff's changed work, the tenants' service and data.

check yourself

0 / 3 answered

A marketing team used Claude for eight weeks to draft product descriptions, and its manager must ask the leadership team to keep paying for it. The finance director has said that only results measured on the team's own work will count. The team logged how long each description took, both before and during the trial.Which way of stating the value best meets the finance director's condition?

Select one.

An HR team piloted Claude for drafting answers to staff questions about company policies. Before the change goes wider, the HR lead presents to the staff representatives, who have asked how each of Claude's limitations will be handled in practice. The HR lead is choosing which statements to put on the limitations slide.Which TWO statements best present the limitations to the representatives?

Select two.

A finance manager showed the accounts payable clerks the same deck the board had approved for a Claude pilot on invoice queries. Its main slide read "1.5 full-time roles of capacity released". Afterwards the clerks raised concerns with their union, and few of them used Claude in the first month.What is the most likely root cause?

Select one.

sources

  1. Claude Help Center, "Claude is providing incorrect or misleading responses. What's going on?": Claude can occasionally produce incorrect or misleading responses; do not rely on Claude as a singular source of truth; scrutinise high-stakes advice. Checked 2026-10-15.
  2. Claude Help Center, "How up-to-date is Claude's training data?": each model has a knowledge cutoff and may not be aware of events after it. Checked 2026-10-15.
  3. Claude Help Center, "Introduction to prompt design": Claude has no context about you, your task or your organisation. Checked 2026-10-15.
  4. Claude Help Center, "Use connectors to extend Claude's capabilities": connectors let Claude access your apps and services; Claude inherits each person's permissions from the connected service. Checked 2026-10-15.
  5. Claude Blog, "How enterprises are driving AI transformation with Claude": successful teams track concrete metrics (productivity gains, time savings, quality improvements), and this evidence proves the return to stakeholders. Checked 2026-10-15.
  6. Wrenfield's repairs pilot, its figures and the first draft of the results slide are this course's example.