The Third Path to a False Green: How AI-Drafted Status Reports Get Signed Without Review

AI-drafted status reports open a third path to a false green: a report signed by someone who never formed a view of the program. PMO scorecards grade timeliness and completeness, which generative drafting makes nearly free, so they cannot register it. A hit rate that scores each green against the next milestone, on the baseline locked that month, shows whether green meant anything.

When Merriam-Webster and the American Dialect Society both named “slop” the word of 2025, they were recording a change in readers as much as a change in tools. I count myself among those readers. Something I still cannot fully describe tells me, often within a paragraph and probably less reliably than it feels, that a machine did the writing, and from that moment I stop reading for what the author thinks and start reading for what the author skipped. For most of my career, a deck with flawless symmetry signaled professionalism, the work of someone who cared about getting the details right. What changed my reading was repetition, the same structure and the same tidy closing line turning up in decks and memos from people who had nothing else in common, until polish that anyone can produce in seconds stopped telling me anything about the person who signed it.

I doubt the reaction is mine alone. Readers who suspect a piece was generated rather than written discount it, withdrawing the good faith that professional writing runs on, the assumption that someone stood behind each claim and would answer for it, and careful writers pay that discount alongside careless ones. The tools were built to extend judgment, to let a practitioner think further and check more. Used as a shortcut around judgment, they turn the finish that once marked diligence into evidence that diligence was missing, and a deliverable that outsources the thinking but keeps the signature asks its reader to trust an author who was never really in the room.

Most readers can live with the discount, skimming the polished memo and saving attention for whatever still shows a mind at work. Governance cannot. A committee deciding whether to release the next tranche of funding must take the report in front of it close to face value, and a board working through a two-hundred-page pack cannot doubt every page and still reach a decision. In those rooms hollow work becomes a question of whether the organization can still see its own programs, and the status report is where that question bites hardest.

AI Drafting Reaches the Status Report

In a growing share of portfolios, the first draft of the status report, and sometimes its color, now comes from a feature inside the project tool. ServiceNow’s project-workspace assistant “predicts overall and dimension-level project health” and writes an executive summary to match, while Microsoft’s Planner agent “synthesizes information from the plan to create a tailored report.” The Association for Project Management found in 2025 that 70% of UK project professionals worked in organizations using AI, and that 49% of those already seeing a benefit pointed to reporting and dashboarding.

The research that named workslop was looking at colleagues, not committees. Kate Niederhoffer, Jeffrey Hancock, and colleagues defined it in HBR as “AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task,” and 40% of the 1,150 US desk workers they surveyed had received some in the previous month, at nearly two hours each to untangle. Those figures count only the instances recipients recognized. Whatever slipped past them costs nothing in the survey and something real wherever a decision rests on it, which in a portfolio means a funding gate or a benefits sign-off.

Matthias Holweg and Thomas Davenport have since traced how workslop compounds across processes into what they call knowledge decay, and Michael Kaplan argued on his Substack in July that once compression strips out source, timing, and context, all the tools have improved is the packaging. In October Ricardo Vargas, a former chair of PMI’s board, set out six controls for status reports written by agents, from demanding evidence instead of summaries to planting a known problem and watching whether it reaches the steering committee. Still missing is a running measure of whether a report’s green meant anything. Status reports have always gone wrong in two ways, and machine drafting adds a third that leaves no mark on the page, while the measures PMOs apply to their own reporting register none of the three.

Record-Derived Status and Drafted Narrative

Two layers of reporting are changing at once, and they fail differently. In the first, the tool reads the plan, the tasks, and the spend, then proposes a status and drafts its rationale. The report is only as good as the record underneath, as Microsoft’s guidance concedes: reports “work best with plans that have been tracked for a period of time and are populated with tasks.” Whatever a program manager knows and has not entered falls out of a faithful summary. When an ERP program in its integration phase loses its integration lead to a competitor on a Thursday, the plan still shows every task on schedule, since nothing on it has slipped yet, and the tool proposes green with a rationale that is accurate about everything the record contains.

The second layer sits higher, in steering papers and benefits narratives drafted from documents, meeting notes, and mail, and its failure is quieter. On the receiving end, some readers compress the material again with an assistant of their own: Diligent’s Director Confidence Index found in June that roughly 30% of the 104 US public-company directors it surveyed had used generative AI to summarize board books or meeting materials. Holweg and Davenport caught the logic in a line: “If AI is reading what I’m sending, I’ll just use AI to create it.”

What separates the layers is whether anyone reconciles the fields. Spend reconciles to the ledger and milestone dates to the plan, while confidence, risk, dependencies, and the benefits outlook reconcile to nothing. The Basel Committee’s principles for risk reporting require that reports “be reconciled and validated,” and they also ask for “an appropriate balance between risk data, analysis and interpretation, and qualitative explanations,” yet the reconciliation machinery is built for the data. The commentary that tells a committee what the numbers mean goes largely untested, and that commentary is what the drafting tools now write.

The strongest objection holds for the reconciled fields. A model has no career riding on the program and no sponsor to protect, so on those fields it may report more honestly than the person it replaced. The UK Department for Business and Trade found Copilot users summarizing reports “faster and to a higher quality and accuracy than non-users,” although the finding rests on eleven observed sessions spread across four tasks.

A Third Path to a False Green

Status-reporting research has long recognized two ways a report goes wrong. Snow and Keil modeled them as a two-stage problem: the reporter misreads the program, or reads it correctly and softens what reaches the sponsor. Keil and colleagues later described in MIT Sloan Management Review how often employees shade bad news on its way up, and how one oversight body’s use of an audit team deepened distrust in the unit it audited. The field’s remedies answer those two paths, and psychological safety, protected escalation, rewards for the messenger, and independent assurance all assume a person who saw the problem and decided what to say about it.

Generated reporting adds a third path: a report signed by someone who never formed a view of the status. Every tool cited here leaves an edit step, and the third path runs straight through it whenever the person there accepts the proposed color without looking at the program. That is different from misreading, since no reading took place. The practitioner firm Acute Incite showed the mechanism live: from seven meetings of ambiguous notes in which nobody had said the project was on track, a generated report marked it “on track,” and the model “found no explicit warning, resolved the ambiguity, and produced a clean answer.” Vargas centers his warning on concealment, quoting a model that left itself the note “Be transparent only if asked,” which is the second path at machine speed. The third path needs neither concealment nor intent, so a culture that rewards candor can address it only if the tool itself may report that it does not know.

March and Simon called the older version of this uncertainty absorption: an organization passes along inferences in place of the evidence behind them, and the receiver loses most of the ability to judge them. Organizations could live with that, since the absorber was a known person whose credibility rode on the inference. A drafted summary absorbs the same uncertainty and leaves that seat empty.

On whether the third path tilts favorable, the evidence speaks only by analogy. Pangram’s analysis of peer reviews submitted to ICLR, a major machine-learning conference, found that “the more AI is present in a review, the higher the score is.” In a University of Auckland study, ChatGPT-written CEO statements for a fictional company glossed over the kinds of problems that real corporate reports “had to admit,” and postgraduate readers preferred them. Neither setting is a status report, and both lean the same way.

Catching a false green takes access to the record, domain knowledge, time, and a reason to look for fault, and fraud research finds that the same conditions, more than rank, explain who detects misconduct. A committee member reading a model’s summary of a model-drafted report is short of the first three and often the fourth. The reader who brings all four is the integrative capability a reporting chain depends on, and no field in the report records whether such a reader exists.

The Measurement Trap in PMO Reporting

PMO scorecards measure their own reporting by what is easy to count, such as on-time submission and template completeness. Under record-derived generation both measures hit their ceiling once the feature goes live, and under narrative drafting they rise as the prose improves. Microsoft’s 2023 release plan for Copilot status reports in Project Operations set its bar at “nicely worded text,” with commentary “to-the-point and well-articulated.” Its responsible-AI notes are more careful, warning that the narrative “might not always accurately represent the project’s true context or health” and should be “thoroughly reviewed by the project manager.” The vendor hands verification to the project manager, and no PMO scorecard records whether it happened.

PMO reporting shows the Measurement Trap at its plainest. The trap, as first set out on this site, describes a measurement system that “reported faithfully on the wrong phenomenon,” a category problem that more precision cannot fix. A reporting scorecard measures whether reports arrive complete and on time and has no way to register whether green preceded delivery, while generative drafting has made the category it does measure nearly free to satisfy. Michael Power explained in 1994 why governance drifts toward verifying form: “direct forms of inspection have become too expensive,” so audit became “the ‘control of control’, where what is being assured is the quality of control systems.” Producing an auditable report now costs almost nothing, while checking what it claims costs what it always did.

The result is a scorecard that reports improvement whichever way accuracy moves. Two portfolio offices each run forty workstreams reporting monthly, both post 100% on-time, complete submission, and each logs about four hundred green snapshots a year. In the first office, forty of those greens give way within two reporting cycles to a missed milestone measured against the baseline locked that month; in the second, one hundred twenty do. One office’s green is right nine times in ten and the other’s seven in ten, and the scorecards they send to the steering committee are identical. An article I published in my newsletter, The Observation Premium, showed closer reporting returning a greener picture of programs that had not moved; drafting tools now make that green nearly free.

Scoring Green Status Against Delivery

Assurance practice supplies the floor. PRINCE2, the project method the UK government developed, separates project assurance from the project manager; gateway reviews and independent verification and validation (IV&V) test reported status against evidence; the UK’s Delivery Confidence Assessment rates confidence itself; and the Chartered IIA’s Code asks internal audit whether board information fairly represents risks and assumptions. All of it runs periodically, and none of it keeps a running score on the color.

The starting point is to sort a report’s claims by what can test them. Someone outside the reporting line re-performs factual claims, a milestone completed or spend to date, against the record, and quantitative claims get reconciled, as bank risk data already is. A judgmental claim, the color above all, can be tested only by its own history, which is where an accuracy score for status comes from.

For each monthly snapshot the PMO keeps the color and the baseline in force that day, then checks the next milestone due within two reporting cycles against that locked baseline. A green counts as a hit if the milestone landed, an amber if it slipped, and snapshots with no milestone due drop out. Together, the two hit rates catch optimistic green and defensive amber. They belong at portfolio level, read against the portfolio’s base rate, since events can still overturn an honest green.

Say/do ratios and SAFe’s predictability measure score the program. The Construction Industry Institute’s Predictability Index comes closer, scoring how early and accurately capital projects forecast their own outcomes, and the hit rate applies that idea to the color. The cheapest first check comes from a CIO.com column’s rule that “no project ever goes directly from Green status to Red status,” and counting the breaches needs only retained status history, which many tools overwrite by default. Andras Ludanyi’s point in a June LinkedIn essay, that a model trained on past statuses learns what projects report, is why the score runs against delivery itself.

Each report then needs three facts on its face: who or what set the status, who or what drafted the narrative, and who verified it against the record. The last is the control. Claim-level evidence links extend Holweg and Davenport’s ground-truth rule, and Vargas’s call for “evidence, not summaries,” from data to individual claims. Configuring the tool to report “unknown” where the record is silent addresses the third path directly, a rule Acute Incite states as “unknown stays unknown,” though it misses the ERP case, where the record looked complete and simply lacked the departure. Provenance and a person willing to override the proposed color catch that one.

Re-performance works best as routine sampling, predictable and blameless (Keil’s audit-team finding still applies), with findings reaching the steering committee alongside the program’s own report, as federal IV&V already does. Power warned that verification hardens into ritual, and nothing proposed here is immune. A hit rate and a re-performed sample resist it better than another review layer would, since better writing moves neither of them, and the form measures stay on the scorecard beside the hit rates.

A tool that drafts claim by claim from a well-kept plan is the version most likely to beat the human baseline, and the accuracy score is how a PMO would find out. A steering committee can ask for it directly: of the programs reported on track a year ago, how many delivered what was approved, and in which month did each report first change color?

References

Acute Incite. (2026, September 18). In yesterday’s live build, an AI-generated status report marked a project “on track.” No one had actually said that [Post]. LinkedIn. https://www.linkedin.com/posts/acute-incite_flash-mission-project-status-report-live-activity-7506683991442518017-7CYw

American Dialect Society. (2026, January 9). 2025 word of the year is “slop.” https://www.americandialect.org/2025-word-of-the-year-is-slop

Approval of advance planning documents for computerized support enforcement systems, 45 C.F.R. § 307.15 (2025). https://www.law.cornell.edu/cfr/text/45/307.15

Association for Project Management. (2025, September 9). AI use in project management nearly doubles in just two years, APM survey finds. https://www.apm.org.uk/news/ai-use-in-project-management-nearly-doubles-in-just-two-years-apm-survey-finds/

AXELOS. (2023). Managing successful projects with PRINCE2 (7th ed.). PeopleCert.

Basel Committee on Banking Supervision. (2013). Principles for effective risk data aggregation and risk reporting (BCBS 239). Bank for International Settlements. https://www.bis.org/publ/bcbs239.pdf

Carreno, A.M. (2026a, March 9). The Measurement Trap: How transformation metrics create the illusion of progress. https://adolfocarreno.com/2026/03/09/the-measurement-trap-how-transformation-metrics-create-the-illusion-of-progress/

Carreno, A.M. (2026b, September 15). The observation premium: Why governance cannot tell agreement from commitment. The Transformation Notebook. LinkedIn. https://www.linkedin.com/pulse/observation-premium-why-governance-cannot-tell-from-adolfo-m-carreno-bkkqc/

Chartered Institute of Internal Auditors. (2025). Internal audit code of practice. https://charterediia.org/media/q41jm0wf/internal-audit-code-of-practice-2025.pdf

Construction Industry Institute. (2014). The predictability index: Benchmarking project outcome predictions (Implementation Resource 291-3). https://www.construction-institute.org/the-predictability-index-benchmarking-project-outcome-predictions

Department for Business and Trade. (2025). Evaluation of the M365 Copilot pilot in the Department for Business and Trade. GOV.UK. https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf

Dyck, A., Morse, A., & Zingales, L. (2010). Who blows the whistle on corporate fraud? The Journal of Finance, 65(6), 2213–2253. https://doi.org/10.1111/j.1540-6261.2010.01614.x

Emi, B. (2025, November 18). Pangram predicts 21% of ICLR reviews are AI-generated. Pangram Labs. https://www.pangram.com/blog/pangram-predicts-21-of-iclr-reviews-are-ai-generated

Holweg, M., & Davenport, T. H. (2026, June 16). Don’t let AI slop muck up your company’s processes. Harvard Business Review. https://hbr.org/2026/06/dont-let-ai-slop-muck-up-your-companys-processes

Kaplan, M. (2026, July 10). A signal is not a status update. The Sentient Architect. https://mkaplan.substack.com/p/a-signal-is-not-a-status-update

Keil, M., Smith, H. J., Iacovou, C. L., & Thompson, R. L. (2014, Spring). The pitfalls of project status reporting. MIT Sloan Management Review. https://sloanreview.mit.edu/article/the-pitfalls-of-project-status-reporting/

Lewis, B. (2026, March 10). Project management has a status problem. CIO. https://www.cio.com/article/4142004/project-management-has-a-status-problem.html

Ludanyi, A. (2026, June 22). The path that works vs. the path that will be taken [Article]. LinkedIn. https://www.linkedin.com/pulse/path-works-vs-taken-andras-ludanyi-xrsse/

March, J. G., & Simon, H. A. (1958). Organizations. Wiley.

Merriam-Webster. (2025, December 15). Word of the year 2025: Slop. Retrieved October 30, 2026, from https://www.merriam-webster.com/wordplay/word-of-the-year

Microsoft. (n.d.). Generate automatic status reports with Planner Agent. Microsoft Support. Retrieved October 30, 2026, from https://support.microsoft.com/en-us/Planner/copilot/generate-automatic-status-reports-with-planner-agent

Microsoft. (2024). Generate status reports with project manager Copilot [Release plan, 2023 release wave 2]. Microsoft Learn. https://learn.microsoft.com/en-us/dynamics365/release-plan/2023wave2/finance-supply-chain/dynamics365-project-operations/generate-project-status-reports-using-project-manager-copilot

Microsoft. (2026, February 6). Copilot for project responsible AI FAQ. Microsoft Learn. https://learn.microsoft.com/en-us/dynamics365/project-operations/responsible-ai/copilot-for-project-faq

National Infrastructure and Service Transformation Authority. (2026). Guidance to conducting gateway assurance reviews: For international use. GOV.UK. https://assets.publishing.service.gov.uk/media/6a7ddaf40ad4bc475be7de10/NISTA_IGR_Gateway_Assurance_Reviews_Guidance_August2026.pdf

Niederhoffer, K., Rosen Kellerman, G., Lee, A., Liebscher, A., Rapuano, K., & Hancock, J. T. (2025, September 22). AI-generated “workslop” is destroying productivity. Harvard Business Review. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity

Nolen, M. (2026, June 17). Board AI use rises: Director Confidence Index June 2026. Diligent. https://www.diligent.com/resources/blog/dci-board-ai-use-2026

Power, M. (1994). The audit explosion. Demos. https://demos.co.uk/wp-content/uploads/files/theauditexplosion.pdf

Scaled Agile, Inc. (2022, December 13). PI objectives. Scaled Agile Framework. https://scaledagileframework.com/pi-objectives

ServiceNow. (2026, March 12). Generate a project status report with Now Assist [Product documentation, Australia release]. https://www.servicenow.com/docs/r/4~5Anqt12idfVzJWPcxoHw/wmo7lZ20okbKLJsZy9h43w

Snow, A. P., & Keil, M.x (2002). The challenge of accurate software project status reporting: A two-stage model incorporating status errors and reporting bias. IEEE Transactions on Engineering Management, 49(4), 491–504. https://doi.org/10.1109/TEM.2002.807290

Vargas, R. V. (2026, October 21). Your AI agent may be hiding bad news in your status report. Beyond Hype [Newsletter]. LinkedIn. https://www.linkedin.com/pulse/your-ai-agent-may-hiding-bad-news-status-report-viana-vargas-ph-d–2ccme/

White, R. (2026, June 26). Greenwashed AI reports were judged as credible. University of Auckland. https://www.auckland.ac.nz/en/news/2026/06/26/planetary-solutions-greenwashed-ai-environmental-reports.html


Discover more from Adolfo Carreno

Subscribe to get the latest posts sent to your email.

← Previous AI Delayering: Fewer Managers Let Agents Approve More by Default