She is writing the same monthly board update she has written for fourteen months. The format was last meaningfully changed by a board member who left nine months ago. She has not asked anyone whether they read it. She has not asked anyone whether they would notice if it stopped. She does not know who the report is for, in the sense of whose decision it changes, and she has never made the time to find out. It takes her four hours every fourth Friday and an hour of low-grade anxiety the following Tuesday wondering what the silence means.
That is where the eliminate question lives. Before any AI tool gets pointed at a recurring report to make it faster, the prior question has to be asked: what is the report actually for?
What does it mean to eliminate a report rather than automate it?
Eliminating a report means deciding the report should not exist at all, rather than producing the same report more cheaply. Automation makes a task cheaper to keep, which is the wrong move when the task itself was the problem. If a recurring report informs no specific decision and the recipient cannot name an action that would change if the numbers moved, the right move is to stop producing it.
This is the spine of the EAD-Do framework recast for AI, which sits inside the AI for your own work cluster. Eliminate is the first step for a reason. Apply it before you reach for the tool. The four hours come back, and the model never had to be involved.
Why do reports survive past their useful life?
Reports survive because they accumulate by addition and almost never by subtraction. A board member asks a question, a section appears in the pack to answer it, the board member leaves nine months later, the section stays. A bank covenant requires a quarterly KPI sheet, the covenant gets renegotiated out of the loan, the sheet stays. Nobody told the writer they could stop, so the writer kept going.
A founder got nervous about cash in 2020 and built a weekly liquidity update. The cash position normalised, the update stayed. The same pattern shows up across firm after firm and reporting line after reporting line, layered over years.
The HBR Decision-Driven Organization frame, written by Blenko, Mankins and Rogers at Bain in 2010, makes the point cleanly. Every recurring artefact in an organisation should map to a specific decision and a specific decision-recipient. The McKinsey follow-on a decade later sharpened the same point into a four-decision taxonomy. Both pieces argue that the test is not whether a report is informative, but whether anyone changes their behaviour because of it. Reports that pass the informativeness test and fail the behaviour test are habits, not needs.
The reason AI helps here is unglamorous. The model is good at running the boring audit you have been avoiding. Paste a list of every recurring report you produce with its recipient and approximate prep hours, ask it to map each to a specific decision the recipient takes because of it, and flag the orphans. The prompt is two sentences. The output names the reports nobody can defend.
What patterns show up after the audit?
Three. Decision reports inform a real decision, the recipient can name the action that changes if the numbers move, and the report stays (often shorter and less frequent than today). Relationship reports do not change a decision but maintain trust with a partner, investor, or key client, and they stay too, named honestly as relationship work. Orphan reports inform no decision, no recipient can defend them, and they get killed.
The orphan category is usually the largest of the three, and it is the one founders consistently underestimate. A typical UK services SME audit finds that a meaningful share of recurring reporting work, often a third or more, falls into orphan territory. The reports survived because the cost of asking “should this still exist” felt higher than the cost of producing it for another month. Compounded over a year, the maths is brutal in the other direction. A three-hour monthly report killed in February saves thirty-three hours by Christmas, and a stack of orphans across the firm saves a working week per quarter.
What does the worked example actually look like?
A quarterly board pack of eighteen pages, four hours of prep, a one-hour board meeting where the chair asks two questions both answered on page two. The audit kills six pages outright as orphans, compresses eight into appendices supplied on request, and leaves a one-page memo summarising the decisions the board needs to take, the recommendation, and the risks.
Total prep drops to forty-five minutes. Total reading time for the board drops from twenty minutes to four. The two questions still get asked and still get answered. This shape comes from the Amazon six-pager pattern, documented in the AWS Startups blog and elsewhere, adapted by Drew Houston at Dropbox in his memo-first meeting culture. Houston has been public about why he made the move: the writer thinks more clearly, the reader engages rather than nodding through a deck, and the meeting is shorter and better. The appendices live one click away for anyone who wants the detail. They almost never get clicked. That, on its own, is the audit running in real time.
When should you not cut, and when should you ask before you cut?
Two zones stay outside the cut. The first is the UK fiduciary boundary: Companies House annual filings, the seven directors’ duties under the Companies Act 2006 (set out by the Chartered Governance Institute and named UK law firms), and ICAEW corporate reporting obligations where the firm is in scope. These exist because the law and the profession require them, not because anyone asked. They stay, in full, regardless of what your audit says.
The second is the contractual boundary. If an investor’s term sheet, a lender’s covenant, or a client’s master services agreement specifies a recurring report, that report stays until the contract is renegotiated. The audit is still useful inside this boundary because the report can usually be compressed to the minimum the contract actually requires, with everything else stripped. Many founders quietly produce more than their lender asked for because nobody read the covenant carefully when the report was first set up.
Outside those two zones, the cut is yours to make. The one-sentence note to the recipient is the right shape: “I’m pausing the monthly X report from next month, let me know if you want it to continue.” Silence is the answer in the typical case. The hour you save every month is the dividend. If a recipient asks for it back, you reinstate it knowing it now passes the decision test, which is a strictly better position than where you started. The next post in this cluster, on auto-summarising every meeting, is where the surviving reports become candidates for AI-drafted compression, not before.
