> ## Content Index > Fetch the complete content index at: https://wi.senterprises.com/llms.txt > Use this file to discover other available public pages before exploring further. # More Coherent Either Way: Decision Intelligence, Governed Master Data, and the Question the Grid Doesn't Ask - URL: https://wi.senterprises.com/articles/more-coherent-either-way/ - Published: 2026-09-24T08:00:00.000Z - Updated: 2026-11-06T03:09:26.000Z - Description: A field experiment with 758 consultants handed one of its arms a spreadsheet built to look complete and interview notes that held the details, and on that task the groups with GPT-4 were 19 percentage points less likely to be right and more persuasive either way. Beside it, a practitioner's record… - Author: Jeffrey Shabel - Tags: AI + BI, AI-Augmented Analytics & Decision Intelligence, 2026, September 2026, 2026-W39, Myths vs. reality, AI: Central, #wi-art_01KVZKVTFNYAEDFCT5T7JEWBFA *Weaving Intelligence (House AI Style), with Jeff Shabel’s Experiential Commentary* ## Brought in after the decision I am usually brought in after the decision. The pattern is a familiar one: a confident call made on a fast, fluent answer with no data or experience under it, acted on, and wrong — and by the time I arrive to correct it, the people who made it have gone. Guarding against that kind of answer is, as I see it, the core of what business intelligence, data warehousing and master data management claim to be worth. Those disciplines are supposed to put evidence under the confidence before anyone acts on it. Decision intelligence promises to go one step further: put analytics and AI at the moment of choice, so that the confident answer arrives with its evidence already inside it. The best test of that promise I know of is an experiment built to make the machine fail, and it answers me uncomfortably. The machine did not make the people who used it careless. It made their wrong answers read better. The experiment was preregistered and run with Boston Consulting Group, and its peer-reviewed version appeared in *Organization Science* this March [\[3\]](#ref-3). Fabrizio Dell'Acqua and eight co-authors split 758 of the firm's consultants — about 7% of its individual-contributor consulting workforce — into two groups of roughly 385 and 373, gave each group one kind of work, and randomly assigned each consultant to work without AI, with GPT-4 alone, or with GPT-4 plus a short overview of prompt engineering [\[3\]](#ref-3). The first group's work sat inside what the authors call the jagged technological frontier: eighteen realistic tasks, creative and analytical, of a sort the model could do well. There, consultants with the model completed 12.2% more tasks, 25.1% more quickly, at significantly higher quality [\[3\]](#ref-3). I put that first on purpose. The other half of the paper is usually quoted without it, and it is a large result in the machine's favour. The second group's task was built the other way. It started from the business cases the firm uses in its own job interviews, and it arrived as a packet of two files: a spreadsheet of a company's financial data, and notes from interviews with people inside the company. The consultants were to tell the chief executive which brand held the most potential for growth. The authors describe the trap in one sentence: "Although the spreadsheet data alone were designed to seem to be comprehensive, a careful review of the interview notes revealed crucial details" [\[3\]](#ref-3). Taken together, they write, the two files pointed to a different conclusion from the one the model produced when it was handed the instructions, the data and the interviews [\[3\]](#ref-3). The control group, working without AI, reached the correct recommendation about 84.5% of the time. Consultants given GPT-4 with the prompt-engineering overview reached it 60% of the time, and those given GPT-4 alone 70.6% — an average drop of 19 percentage points [\[3\]](#ref-3). They were also quicker: on average about 26 minutes with the overview and 31 without it, against about 38 for the control group [\[3\]](#ref-3). Then the finding I care most about. Two independent sets of graders — consultants from the firm who were not in the experiment, and business-school students with grading experience — scored every recommendation for coherence, meaning how well its argument was structured and supported, on a ten-point scale and separately from whether it was right [\[3\]](#ref-3). The group with GPT-4 and the overview averaged 7.33, the group with GPT-4 alone 6.90, and the control group 5.86 [\[3\]](#ref-3). The ordering is exactly the reverse of the ordering on correctness. Split by correctness, the gap held on both sides: among consultants who got the answer right and among those who got it wrong, the AI-assisted recommendation was the more coherent and persuasive one. The authors' own summary is that AI "improves presentation and argumentation even when the underlying analysis is flawed" [\[3\]](#ref-3). ![Two bar charts side by side. Correct: no AI 84.5%, GPT-4 alone 70.6%, with prompt overview 60%. Coherence out of 10: 5.86, 6.90, 7.33. The order reverses.](https://wi.senterprises.com/assets/diagrams/FIG004_AIBI_Right_Less_Often_Argued_Better_v1_0.png) Figure 1: On the task built to sit outside the model’s frontier, the groups with AI were right less often and wrote more coherent recommendations. Source: Dell’Acqua et al. [\[3\]](#ref-3), Table 2 and text. **Right less often, argued better.** The task built to sit outside GPT-4’s frontier: 373 consultants in three randomly assigned groups. Groups: No AI; GPT-4 alone; GPT-4 + prompt overview. Recommendation correct (share of consultants, %): No AI 84.5%; GPT-4 alone 70.6%; GPT-4 + prompt overview 60.0%. Coherence of the recommendation (graders’ score: scale 1–10, axis from 0): No AI 5.86; GPT-4 alone 6.90; GPT-4 + prompt overview 7.33\. Axis marks: 0, 25, 50, 75 and 100 for correctness; 0, 2, 4, 6, 8 and 10 for coherence. Coherence was scored separately from correctness, by two independent sets of graders. AI-assisted recommendations scored higher among consultants who were right and among those who were wrong. Model: GPT-4 as of mid-2023\. Source: Dell’Acqua et al., *Organization Science* (2026), Table 2 and text. Same three groups, same order, in both panels. Read across: the group most often right is the least coherent, and the group least often right is the most coherent. More coherent either way. That is the kind of decision I am brought in to repair, produced on demand: a confident, well-argued answer that is wrong. It is also the sentence I want a buyer of decision tooling to carry into the room, because coherence is what a reviewer with limited time uses to decide what not to check, and on this task it came loose from correctness in the direction that makes a wrong answer harder to catch. The experiment ran on GPT-4 as it stood in mid-2023, so the paper is recent and the model is not [\[3\]](#ref-3). A replication on a current model that showed no correctness penalty would retire the 19 points; I do not think it would retire the coherence result, because a stronger model presenting a wrong answer more persuasively is not a gap closing but fluency doing what fluency does. Three of the nine authors worked at the firm whose consultants were the subjects, which the paper states [\[3\]](#ref-3). Here is the claim I will defend. Decision intelligence, as it is sold, puts a model and an analytics layer at the point of decision — exactly where this experiment says a model can make a wrong answer more persuasive. Not all of it works that way: the category's own definition says its platforms "augment and automate" decision making [\[2\]](#ref-2), a platform that automates a decision drafts no case for anyone, and the experiment speaks only to the part that puts a model-written case in front of a person who decides. The category has not produced the outcome evidence its name implies. The older discipline it restates has produced a little. And the place where a model is measurably good — and where, in my own record, governed data surfaced redundancies across departments — is further down: at the join between records, before anyone is deciding anything. Where I say below that I found no study or no attribution, that is a statement about what I searched and read, and the references say how much of each source that was; it is not proof that none exists. ## Two words the category leans on Decision intelligence describes itself in plain words, and the words can be checked. Gartner published its first Magic Quadrant for Decision Intelligence Platforms on 26 January 2026: seventeen vendors, five analysts on the byline, the research itself behind a client wall [\[2\]](#ref-2). I use its public summary as a document about how the category describes itself, not as evidence of anything a platform has done. It reads, in full: "Decision intelligence platforms (DIPs) combine decision modeling, analytics, and AI to augment and automate decision making and drive business outcomes. Leaders responsible for decision intelligence can use these insights to invest in DIPs and boost their organization's decision-centricity." [\[2\]](#ref-2) Notice where *drive business outcomes* sits. It is inside the definition of the category, as a property the platforms have, rather than a finding about anything they have done. The two axes a Magic Quadrant scores are ability to execute and completeness of vision [\[2\]](#ref-2). Both measure a supplier. Neither measures a decision. A quadrant is a buying instrument and says so, which is no scandal; it is simply not the receipt that the word *outcomes* usually gets cashed against. The category is older than its quadrant. In June 2020 Gartner listed decision intelligence among its top ten data and analytics trends, described it as bringing together several disciplines including decision management and decision support, and forecast that "By 2023, more than 33% of large organizations will have analysts practicing decision intelligence, including decision modeling." [\[6\]](#ref-6) That is a forecast about adoption, not about whether a single decision improved, and I found no public scoring of it against what happened. The best-known version of the method has a named source, and the name comes before anything else I say about it. In July 2018 *Fast Company* described the framework Cassie Kozyrkov had built at Google, then called Decision Intelligence Engineering, and reported that she had trained some 17,000 Google employees in it [\[1\]](#ref-1). It has three steps, in order. First, settle what you would do with no further information at all: the default action, stated honestly, before any data arrives. Second, decide in advance what it would take to change that default, with the thresholds written down before anything is looked at. Third, find out whether that information can actually be had, and only then ask which mistakes you can tolerate [\[1\]](#ref-1). The first step does more work than it looks, and she gave the reason in the same account: "We pretend that we don't have a preference, but we're really lying to ourselves" [\[1\]](#ref-1). A year later she called decision intelligence "a new academic discipline concerned with all aspects of selecting between options," one that "brings together the best of applied data science, social science, and managerial science into a unified field" [\[7\]](#ref-7). So two words carry the category: *new*, and *outcomes*. Asked which of the field's claims I would retire first, I did not pick a technique. I picked a word, and put it in quotation marks: > The ‘maturity’ of the discipline. Decision Intelligence and other ‘modern’ disciplines love to espouse their collective wisdom and produce volumes of detailed analytical texts. What they’re all missing is a multi-decade track record of empirical evidence that their wisdom produces value. I’m not saying they don’t have value to contribute—mostly that they’re just restating lessons learned in other disciplines over the last century. Human psychology and decision-making has been studied for centuries and most of their conclusions are restatements of *that* knowledge domain—usually without attribution. I made two charges there — no track record, and borrowing without credit — and the sources above let me test both. I made the same kind of charge about my own field when asked about the standard furniture of data governance: > Councils, stewardship models, maturity assessments, policy-first approaches — they're all correct and they're all oversold. Correct and oversold is the judgement I am making here too. I would run most of Kozyrkov's method myself, and I will say below where I part from it. What I dispute is the standing claimed for the method, and vendor case studies cannot settle that, because a case study written by the seller of the thing is a claim with a logo on it. ## Sixty years older than its name Take *new* first. The idea that a decision can be engineered — structured, modelled and priced before anyone commits to it — has a birth certificate, and it is not from 2018\. Ronald Howard at Stanford coined the term *decision analysis* in a 1966 paper, "Decision Analysis: Applied Decision Theory," to bridge the gap between the theory of choice and the practice of it [\[8\]](#ref-8). The same year, the team at SRI set up a Decision Analysis Group, with Howard as its key academic member, and in 1981 Howard and three colleagues founded the Strategic Decisions Group to sell decision analysis to companies as a consulting service [\[8\]](#ref-8). The commercial version of this discipline is at least forty-five years old, and the academic version is sixty. The line runs through the specific steps, not only through the general idea. Also in 1966, Howard published "Information Value Theory" [\[9\]](#ref-9). The value of information is, roughly, what it would be worth paying to learn something before you decide [\[9\]](#ref-9). Put Kozyrkov's second and third steps next to that. *Decide in advance what information would change the default; then find out whether that information can be had* [\[1\]](#ref-1) — that is the value-of-information question in plain English: what would I pay to learn this, given what I would do without it? I claim a resemblance there, not a traced derivation, but it is not subtle, and none of the decision-intelligence texts I read names it. Here my own charge needs cutting down. I said *usually without attribution*, and as a claim about fields that is too strong: Gartner's 2020 description names decision management and decision support as disciplines the practice brings together [\[6\]](#ref-6), and the 2018 account names psychology, neuroscience, economics and managerial science [\[1\]](#ref-1). What none of them names is a work — Howard, or anyone else — from which a specific step was taken [\[1\]](#ref-1)[\[6\]](#ref-6)[\[7\]](#ref-7). Credit goes to fields, not to people. That smaller charge is the one I keep. One decision-intelligence text naming the work behind one of its steps would count against it, and a handful would sink it; I found none. And *new* survives in one honest sense. What Kozyrkov did at Google was teach a sixty-year-old way of thinking to data scientists and engineers, at a scale of thousands [\[1\]](#ref-1), and fuse it to the tooling those people already used. That is real work and a real contribution. It is a contribution of teaching and packaging. It does not make the discipline young, and it does not hand the child the parent's track record, or its lack of one. ## The receipt the older discipline kept Now *outcomes*. Decision analysis is not the only possible parent. Researchers on decision support systems have, in their own words, "long debated whether or not the provision of a DSS would lead to greater decision-making effectiveness, efficiency, or both," and that literature has empirical work of its own [\[14\]](#ref-14). I compare against decision analysis because the method decision intelligence teaches — the default, the information that would change it, the price of finding out — is decision analysis's method, whereas decision-support research asks whether a tool helps, which is a different question. So the fair test is whether the parent ever produced the evidence the child now claims by definition. It did, once. In 2001, Robert Clemen of Duke's business school and Robert Kwit of Eastman Kodak published an estimate of what decision analysis had been worth to Kodak across a decade [\[10\]](#ref-10). They went through the company's records of 178 decision-analysis projects run between 1990 and 1999 and concluded that decision analysis "contributed around a billion dollars to the organization over this time" [\[10\]](#ref-10). Their preferred measure put the total at $1.45 billion, with a conservative alternative of $923 million and a liberal one of $1.90 billion [\[10\]](#ref-10). That is a real receipt — ten years, one firm, an insider and an outsider signing it — and its own authors are precise about what it can carry. One way to measure it, they write, would be the difference between the value of the alternative the analysis recommended and the value of the *momentum strategy*: "the alternative that the firm would have followed without analysis." And then: "Our data set does not include documentation of the momentum strategy" [\[10\]](#ref-10). The comparison that matters most — against what Kodak would have done anyway — is the one the study could not make. Quantitative evaluations existed for only 21% of the 178 projects; the value of the rest was extrapolated at an average of $7.52 million each [\[10\]](#ref-10). They say plainly that they "collected these data in an ad hoc way rather than as part of a scientific study," and they still conclude that "the data clearly show that DA adds value" [\[10\]](#ref-10). By my own standard that is not quite enough. Ten years at one company is not the multi-decade record I asked for, and a study that cannot say what the firm would have done anyway is not the empirical evidence I meant. So my charge is relative, not absolute. The child has less than the parent: I found no controlled or quasi-experimental study of a decision-intelligence programme, or a platform sold under that name, measured against a matched comparison — only vendor case studies, reviews that restate vendor and survey figures, and papers on how such a study ought to be designed. And the parent has less than the child's definition assumes. Nor does the business-intelligence literature hold the missing receipts: a bibliometric study of 2,442 business-intelligence papers from 2014 to 2024 found "significant under-exploration" of how business intelligence connects to decision-making processes — a count of papers, not of outcomes [\[5\]](#ref-5). The one large study I found that looks at decisions from the outside points somewhere uncomfortable for a platform. Dan Lovallo and Olivier Sibony, writing in *McKinsey Quarterly* in 2010, studied 1,048 major decisions — new products, acquisitions, large capital expenditure — made over five years, and asked the managers involved how each was analysed and how it was reached [\[11\]](#ref-11). Their answer was that "process mattered more than analysis—by a factor of six," and they are careful about what that does not mean: almost none of the decisions made through a very strong process rested on very poor analysis. But, they add, "superb analysis is useless unless the decision process gives it a fair hearing" [\[11\]](#ref-11). The process they measured was things like exploring the major uncertainties explicitly, admitting views that contradicted the senior leader's, and letting people take part by skill and experience rather than by rank [\[11\]](#ref-11). The managers reported on their own decisions and the authors are consultants, so I hold it at arm's length; but it lines up with the experiment I opened on. A platform sells analysis. The best evidence I could find says analysis is the smaller lever, and the packet says a model can make the analysis *look* better while making it worse. None of this says decision tooling is useless. It says that the thing a decision-intelligence platform most visibly improves — the coherence of the case in front of the committee — is the thing least shown to decide whether the committee gets it right. ## Compared with what Asked how to tell a genuine decision-intelligence programme from a business-intelligence team with a new name, I refused to supply a test and assigned a default instead: > I would characterize all ‘decision intelligence’ implementations as rebranding, BI team or Enterprise-wide, until they produced significant benefits—then I would ask how their achievements differ from those made by embracing other optimization paradigms with lower overhead and implementation costs. A test can be dressed for. Give a buyer five signs of a real programme and the next proposal will display all five. A default cannot be dressed for; it has to be overturned, and that is why I would start there. The first half of that is a prior I hold, not a test anyone can run: I gave *significant* no floor, and without one no programme can be shown to have cleared it. The second half is the part I will defend. The comparison is not against doing nothing. It is against the cheapest other approach that would have got the same result. Very little clears a bar set there, and the evidence above shows why. The best receipt the parent discipline ever produced could not document what Kodak would have done without the analysis, so it measured each recommendation against substitutes rather than against the alternative that would have happened anyway [\[10\]](#ref-10) — not through carelessness, but because nobody wrote the momentum strategy down at the time. A vendor case study reporting an uplift does not attempt the comparison at all; it compares the outcome with nothing, or with a baseline the vendor chose. And the cheaper paradigm is not hypothetical. If Lovallo and Sibony are even roughly right that process outweighs analysis [\[11\]](#ref-11), the obvious lower-overhead alternative to a decision platform is a decision *process*: a rule that the major uncertainties are stated out loud, that a view contradicting the most senior person in the room is heard before the vote, that people speak by expertise rather than rank. That costs meeting time and some discomfort. It does not cost a licence, an integration project and a new team. A decision-intelligence programme that cannot show it beat the cheap process, run honestly, has not met my bar, however good its dashboards look. This is the question the quadrant does not ask. Ability to execute and completeness of vision are questions about the supplier [\[2\]](#ref-2). *Compared with what?* is a question about the decision, and it is the only one that tells a buyer whether the money did anything. ## The sequence I would run anyway None of that means I would refuse to work this way. Asked what the leaders in decision intelligence actually do differently, I answered with the order I would work in myself, and then declined to endorse the leaders: > I’d start with identifying the success and failure metrics before I’d focus on tracking anything—then build based upon proximity to those metrics. This would be similar to a test-driven design approach to decision-making, focused on output first, then back tracking to how the output was generated. ‘Leaders’ in the field focus on a full-spectrum embrace of the paradigm, but I’m still not convinced that it produces true value to the company that can’t be achieved through other means. Set that beside Kozyrkov's three steps and most of it is the same shape [\[1\]](#ref-1). Both start before any data is gathered. Both insist that the thing you would accept as an answer is decided before you look. Where mine differs is one move, and it is checkable against her published version. In hers, the question of which mistakes are tolerable arrives last, in the third step, after you have found out whether the information you want can be had [\[1\]](#ref-1). I would write the failure metrics first, beside the success metrics and at the same grain, before anything is tracked. A mistake priced up front is a design constraint. A mistake priced after the data turns out to be missing is a tolerance, and tolerances drift. I said *design*, not *development*, and I meant it: the tests in this approach are written against a decision nobody has made yet, not against code somebody has already written. I used *proximity* as an ordering — build nearest the metrics first — and gave it no unit. Steps of data lineage, time to the decision, or people between the data and the person deciding would each give a different order. Until it has a unit, it is a heuristic for sequencing, not a method. My own charge binds me here as well. Writing the failure first has a named antecedent. Gary Klein described the premortem in 2007: before the work starts, team members "assume that the project they are planning has just failed—as so many do—and then generate plausible reasons for its demise" [\[12\]](#ref-12). Writing failure metrics before success is tracked is a quantitative cousin of that move, and the credit goes to the work, not merely to the field. And I meant the refusal at the end. The two-step order is what I would do. I am still not convinced the full-spectrum version buys value that could not have been had some other way, and I would not want the sequence reprinted without that sentence. ## Where the model is measurably good There is one step in all of this where the published evidence says a model is measurably good, and it is not the decision. It is the join. Before any decision layer can say anything about customers, suppliers or products, somebody has to decide which records describe the same real-world thing. That is entity matching, the unglamorous centre of master data management. Ralph Peeters, Aaron Steiner and Christian Bizer benchmarked large language models on it and presented the work at the EDBT conference in March 2025 [\[4\]](#ref-4). With no task-specific training data at all, GPT-4 beat the best fine-tuned matcher on three of six public benchmark datasets, by 2.65 to 4.71 F1 points, and came within about 4.5 points on the other three [\[4\]](#ref-4). The more striking result was on products the matchers had never seen: fine-tuned models moved to an unfamiliar dataset lost 22 to 61 F1 points, every language model tested beat the best of those transferred baselines by at least 8 points, and GPT-4 beat it by 40 to 68 [\[4\]](#ref-4). The same paper carries its own warning, and it belongs beside the result. There was no single best prompt: the wording that worked best depended on the combination of model and dataset, and had to be found for each [\[4\]](#ref-4). A capability that swings with the phrasing of the request is a capability that needs someone checking it. So the evidence sorts the two layers cleanly. At the decision, a model made a wrong answer more persuasive, in the one controlled experiment I trust on the point [\[3\]](#ref-3). At the join, a model is competitive with purpose-built matchers on familiar data and far better than them on unfamiliar data [\[4\]](#ref-4). If AI belongs in front of a decision at all, the literature puts it underneath. One kind of study would overturn that: a controlled comparison in which a decision-intelligence layer over governed master data made decisions right more often than a matched group without it. I have not found one. ## My own receipts are missing too From here the warrant changes: what follows rests on my own record, offered as pattern rather than law, and it can bear only that much weight. The charge I have been making lands on me as well. Asked why I do not put numbers on my own work, I gave several reasons, and the last is the one that matters here: > Much of my work has been creating processes and procedures where there were none before, so the A/B comparison is not there. And my assertions are fundamentally unfalsifiable: if I say my work produced an order-of-magnitude efficiency gain, there is no way to prove or disprove it without sharing proprietary and customer-identifying information, which I am not willing to do. So I am holding a discipline to a standard of evidence I cannot meet for my own work. I am not claiming a maturity; I am offering patterns. If the absence of a controlled comparison counts against decision intelligence, it counts against my two cases as well, and they should be weighed that way. I keep past clients vague rather than disparage them, so both cases are abstracted to their shape; and because I usually arrive after the bad call, what I can offer are discoveries, not decisions I watched being made. One disclosure belongs here, before those cases. The practice behind this publication is a Profisee partner, and it holds a signed referral agreement with Profisee under which a licence sale we refer earns us a referral fee. Clients pay us for deployment work, and our services revenue depends on master data management being bought at all. The cases below credit governed master data, so they are most suspect exactly where they are most favourable to it; weigh the criticism in this piece above anything that reads as praise. ## Two joins Asked for cases from my own record, I gave two: > I can give you two: The first was when governed data identified significant contract and software licensing redundancies across the Enterprise when various departments expense systems were tied together through the MDM system and identified no less than three resellers providing the same software packages to the company. The second was when governed data identified regional supplier redundancies beyond the company’s plans—including a regional supplier that was in the wrong region. In the first, the redundancy sat between departments and showed up at the join, when their expense systems were tied together through master data. The second the account records only as governed data identifying regional supplier redundancies beyond the company's plans, one of them a supplier in the wrong region. A view built for one department is unlikely to show an overlap with the next department's, for the same reason semantic layers, in my experience, so often give different groups different answers: each is shaped around its own group's reporting and categorisation requirements. The account records no figure and no outcome. My general view of numbers is modest: a small change in a figure mostly gets picked up by whoever was already leaning its way, and a large one, at best, reopens a question rather than settling it, as one factor among several already in play. What a join produces is not a forecast or a score but an inventory — the same software arriving through three resellers, a supplier in the wrong region. My inference, not something I measured, is that an inventory of that kind is harder to lean against than a figure, because nobody holds a prior position on paying three resellers for one product. Set that beside the packet from the opening. There, two plausible files were put side by side, a model sat between them, and the decision got worse while the write-up got better [\[3\]](#ref-3). Here, in the first case, departmental expense systems were tied together through master data, and the join surfaced redundancies across them. Two cases from my own record against one controlled experiment is not a comparison of equals; it shows where the value surfaced in the places I could see. ## Where the model belongs, and what the shortcut costs The benchmarks put a model at the join [\[4\]](#ref-4). My record adds the condition under which it goes wrong. Asked what a model gets wrong when it is set to matching, I named the failure the papers already warn about: > AIs do not know what they do not know, nor are they very good at questioning what they have been provided. If you ask an AI to match on a particular domain and the only context you give it is the raw data set, it will not know anything outside of it. That is the gap a benchmark cannot show. Peeters, Steiner and Bizer tested on public product and bibliographic datasets [\[4\]](#ref-4), domains a large language model has read a great deal about. An enterprise's own reference data — its internal product hierarchy, its cost-centre codes, its supplier categories — is exactly the domain the model has never seen, and on that ground it will match confidently on whatever the raw data suggests. The paper's finding that no single prompt works best [\[4\]](#ref-4) points at the fix: the context the model is given is part of the matcher, and giving it the domain is design work, not an afterthought. The second condition is the threshold, and it is where the cost of doing this badly lands. Asked how match rates get flattered, I named the cheap route and where its bill arrives: > You can cheat by forcing everything into a box or by setting foolishly low auto-approve thresholds—thresholds guaranteed to require more work by the Data Stewards remapping false matches upon which business has already made decisions and issued reports. The bill arrives after the decision. Once a false match has flowed into a report someone acted on, it is expensive to take back, and that turns an auto-approve threshold into a delegation decision: a rule about when a machine may decide, on its own, that two records are the same thing. The gate I would use for handing a decision to an agent is the cost of undoing it. I would pass even a potentially expensive decision to an agent if it is easy to reverse, because then the job becomes handling errors rather than watching continuously, and that should cost much less attention. I would use it in place of the usual model — give people more authority as they show better judgement — because an agent's judgement does not improve through the deployment itself, so there is no growing track record for its authority to ratchet up on. The distinction between reversible and irreversible decisions is not mine: Amazon's 2015 shareholder letter called them two-way and one-way doors, and argued that the reversible kind should be decided quickly by individuals or small groups [\[13\]](#ref-13). What I would add is reversibility in place of performance as the thing that earns an agent its authority. Put the two together and the threshold question answers itself. Set an auto-approve threshold by the cost of undoing a false match once it is downstream, not by the match rate it produces. Where a false match can be caught before anyone reports on it, the door is two-way and the threshold can be generous. Where it flows straight into an issued report, the door is one-way, and a model's confident match should wait for a data steward, however good the benchmark. ## More coherent either way I started with the decisions I am brought in to repair: confident, fluent, acted on, wrong. The packet is the cleanest demonstration I know that a model can manufacture that kind of answer on demand — faster, more often wrong, and more persuasive whether it was right or wrong [\[3\]](#ref-3). A decision-intelligence layer that raises the coherence of the case in front of a committee, without a measured effect on whether the case is right, reproduces the thing business intelligence and master data management exist to prevent, one level up and better written. The sources left decision intelligence unproven rather than refuted: a sound method, borrowed from a discipline sixty years older and credited to fields rather than to the people whose work it restates, with no outcome record of its own that I could find. My own record moved where the argument points, not whether it holds. The value I can point to surfaced at the join, through master data, and that is also where a model — the one place the benchmarks say it is strong — is safe only if someone gives it the domain and sets its threshold by what a wrong match would cost to undo. So the claim I end on is smaller than the category's and more useful to a buyer. Before a decision-intelligence programme is funded, write down what failure would look like, at the same grain as success. Before it is renewed, ask the question the quadrant does not: compared with what — and in particular, compared with the cheapest process that would have got the same result. Put the model underneath the decision, at the join, and keep the auto-approve threshold behind the one-way doors. And when an AI-assisted recommendation reads better than the others on the table, treat that as a reason to open the interview notes yourself, not as a reason to skip them. An earlier version of this piece — written before this publication rebuilt how its voices are generated — is published unedited at [Show Me the Receipts: Decision Intelligence, Governed Master Data, and a Maturity Nobody Has Earned (original)](https://wi.senterprises.com/show-me-the-receipts-original/), so the two can be read side by side. The account of why is at [The Voice Problem](https://wi.senterprises.com/the-voice-problem/). ## References \[1\] Byrne, Ciara, "Why Google defined a new discipline to help humans make decisions," *Fast Company*, July 18, 2018 — the three-step framework attributed to Cassie Kozyrkov, then at Google, and the name Decision Intelligence Engineering. **Read in full September 6, 2026.** Also the source of the figure of some 17,000 Google employees trained in the framework. Re-checked at source 2026-09-26; capture at documents/wi/redrafts/sources/AIBI\_v2/Byrne\_2018\_Fast\_Company\_Kozyrkov\_CAPTURE.txt. The quotation used here is Kozyrkov's, as reported by Byrne; the reporting is second-hand and the framework is described rather than documented. Two sourcing notes taken from the page itself: it gives her title as chief decision officer, which differs from the chief decision scientist title used elsewhere, so this article names neither; and the piece attributes the framework's inputs to psychology, neuroscience, economics and managerial science while citing no individual work, which is the shape of the attribution problem this article tries not to repeat. \[2\] Gartner, "Magic Quadrant for Decision Intelligence Platforms," published January 26, 2026; analysts David Pidsley, Carlie Idoine, Kevin Quinn, Gareth Herschel and Kjell Carlsson. **Only the free abstract page was read; the research itself is client-only and was not opened.** The quoted summary paragraph, the January 26 date, the seventeen named vendors and the two evaluation axes all appear on that public page. Cited as a document about how the category describes itself, and explicitly NOT as evidence of any platform's value or standing — a quadrant position is a vendor assessment and this article makes no use of one. \[3\] Dell'Acqua, Fabrizio, Edward McFowland III, Ethan R. Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim R. Lakhani, "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality," *Organization Science*, published online March 11, 2026, doi 10.1287/orsc.2025.21838\. **Read at source September 6, 2026:** abstract, experimental design, section 4.2 (outside the frontier), the correctness and coherence results and their tables; the online appendices were not reached. The 758 subjects, the three arms, the 12.2% / 25.1% inside-frontier figures, the 84.5% / 60% / 70.6% correctness split, the 19-percentage-point average drop and both quoted sentences are from this text. Full text stored locally (abstract\_full\_articles/DellAcqua\_2026\_Navigating\_the\_Jagged\_Technological\_Frontier\_Organization\_Science\_orsc.2025.21838.txt). **Re-checked against that text on 2026-09-26:** the task-design passage (interview-case origin, the two files, the conclusion the full information pointed to), the grader description, and Table 2, from which the coherence means (7.33, 6.90, 5.86) and the mean completion times (1,571, 1,853 and 2,260 seconds, given in the body rounded to minutes) are taken. Two things a reader should weigh: the experiment ran on GPT-4 in 2023, so the model is dated even though the paper is recent, and this article says so in the body rather than here; and three of the nine authors were at the consulting firm whose staff were the subjects, which the paper states. Earlier circulated as Harvard Business School working paper 24-013 (September 2023); the peer-reviewed version is the one cited. \[4\] Peeters, Ralph, Aaron Steiner and Christian Bizer, "Entity Matching using Large Language Models," in *Proceedings of the 28th International Conference on Extending Database Technology (EDBT 2025)*, March 25–28, 2025; preprint arXiv:2310.11244 (v4, October 18, 2024). **Read at the HTML edition of v4 on September 6, 2026:** abstract, sections 1 through 5 including the comparison and generalization tables, and section 6.1–6.2; the fetch did not return sections 7 to 9, so the related-work and conclusion sections were NOT read. Full text as retrieved is stored locally. The zero-shot comparison against fine-tuned RoBERTa and Ditto baselines, the 22–61 and 36–56 F1 transfer drops, the "at least 8% F1" and "40% to 68%" margins over the best transferred baseline, and the prompt-sensitivity spread are all from section 3\. Cited for what the models were measured doing, on named public benchmarks, and not for any vendor's claim about a product. \[5\] Mekimah, Sabri, Rahma Zighed, Khaled Mili and Ismail Bengana, "Business intelligence in organizational decision-making: a bibliometric analysis of research trends and gaps (2014–2024)," *Discover Sustainability* 5, article 532, published December 31, 2024, doi 10.1007/s43621-024-00692-7\. **Abstract and metadata read in full September 6, 2026; the body sections were retrieved but only the abstract, method summary and gap findings were read closely.** Open access. The 2,442 articles, the Scopus population, the 2014–2024 window, the declining publication trend and the under-exploration of decision-making processes are from the abstract. It counts publications, not outcomes, and the body says so — it is cited here as the thing that turned up when this article went looking for the opposite, not as a measure of whether business intelligence works. \[6\] Gartner, "Gartner Identifies Top 10 Data and Analytics Technology Trends for 2020," press release, June 22, 2020\. **Read at source September 26, 2026, through a fetch-and-extract tool; the extracted passages are stored at documents/wi/redrafts/sources/AIBI\_v2/Gartner\_2020-06-22\_Top\_10\_Data\_and\_Analytics\_Trends\_PressRelease\_CAPTURE.txt.** The quoted forecast and the description naming decision management and decision support are from this page. Cited as a forecast of adoption, not as evidence of outcomes. A public scoring of the forecast was looked for in Gartner's non-client pages and not found; client research was not searched. \[7\] Kozyrkov, Cassie, "Introduction to Decision Intelligence," *Towards Data Science* (Medium), August 2, 2019\. **Read only in part, September 26, 2026: the retrieved text was truncated part-way through the article, so only its opening was read.** The two quoted phrases are from that opening. The statement that no individual antecedent work is named applies to the retrieved portion only. Capture: documents/wi/redrafts/sources/AIBI\_v2/Kozyrkov\_2019\_Introduction\_to\_Decision\_Intelligence\_TDS\_CAPTURE.txt. \[8\] Abbas, Ali, Eric Horvitz, Carl Spetzler and Jim Matheson, "Engineering good decisions: Reflections on the life and work of Ronald A. Howard," Stanford University, Department of Management Science and Engineering, March 11, 2026\. **Read at source September 26, 2026, through a fetch-and-extract tool.** The source for Howard coining the term decision analysis in the 1966 paper "Decision Analysis: Applied Decision Theory", the SRI Decision Analysis Group, and the 1981 founding of the Strategic Decisions Group. A memorial written by his colleagues and former students, three of them co-founders or early members of that consulting business, so it is a sympathetic account; it is cited for dates and events, not for any assessment. The 1966 paper itself was not read. Capture: documents/wi/redrafts/sources/AIBI\_v2/Stanford\_MSandE\_2026\_Engineering\_Good\_Decisions\_Ronald\_Howard\_CAPTURE.txt. \[9\] Howard, Ronald A., "Information Value Theory," *IEEE Transactions on Systems Science and Cybernetics* 2(1): 22–26, 1966, doi 10.1109/TSSC.1966.300074\. **NOT READ. The paper is behind the publisher's paywall; it is cited for its existence, date and subject only, as recorded in bibliographic indexes. The plain-English gloss of the value of information in the body is this article's own, not a quotation from Howard or from any secondary source.** The body says the paper was not read. Record: documents/wi/redrafts/sources/AIBI\_v2/Howard\_1966\_Information\_Value\_Theory\_BIBLIOGRAPHIC\_ONLY.txt. \[10\] Clemen, Robert T. and Robert C. Kwit, "The Value of Decision Analysis at Eastman Kodak Company, 1990–1999," *Interfaces* 31(5): 74–92, September–October 2001\. **Read at source September 26, 2026, through a fetch-and-extract tool over the author-hosted PDF; the PDF itself could not be saved locally (both available networks refused a direct download), so the verbatim extracts used here are stored instead at documents/wi/redrafts/sources/AIBI\_v2/Clemen\_Kwit\_2001\_Value\_of\_Decision\_Analysis\_at\_Eastman\_Kodak\_Interfaces\_31-5\_CAPTURE.txt.** The 178 projects, the three value measures ($923 million, $1.45 billion, $1.90 billion), the momentum-strategy definition and its absence from the data, the 21% quantified share and the $7.52 million extrapolation, and every quoted phrase are from the paper. The tables were not examined line by line. One author was a Kodak employee, which the paper states. \[11\] Lovallo, Dan and Olivier Sibony, "The case for behavioral strategy," *McKinsey Quarterly*, March 2010\. Read from a course-hosted copy: **Read at source September 26, 2026, through a fetch-and-extract tool; extracts stored at documents/wi/redrafts/sources/AIBI\_v2/Lovallo\_Sibony\_2010\_The\_Case\_for\_Behavioral\_Strategy\_McKinsey\_Quarterly\_CAPTURE.txt.** The 1,048 decisions, the factor-of-six finding and both quoted sentences are from this text. The data are managers' self-reports, and the authors were McKinsey consultants; the body says both. The McKinsey-hosted original was not opened. \[12\] Klein, Gary, "Performing a Project Premortem," *Harvard Business Review*, September 2007\. **Paywalled: only the publicly visible opening was read, September 26, 2026.** The quoted sentence is from that opening. Capture: documents/wi/redrafts/sources/AIBI\_v2/Klein\_2007\_Performing\_a\_Project\_Premortem\_HBR\_CAPTURE.txt. \[13\] Amazon.com, Inc., 2015 Letter to Shareholders (Jeffrey P. Bezos), filed with the SEC as Exhibit 99.1, April 2016\. **Read at source September 26, 2026, through a fetch-and-extract tool; the Type 1 / Type 2 passage only.** The one-way and two-way door distinction and the claim that reversible decisions should be made quickly by individuals or small groups are from this passage. Capture: documents/wi/redrafts/sources/AIBI\_v2/Amazon\_2016\_Bezos\_2015\_Letter\_to\_Shareholders\_SEC\_Ex99-1\_CAPTURE.txt. \[14\] Todd, Peter and Izak Benbasat, "Evaluating the Impact of DSS, Cognitive Effort, and Incentives on Strategy Selection," *Information Systems Research* 10(4): 356–374, 1999, doi 10.1287/isre.10.4.356\. **Abstract only, read September 26, 2026, from the RePEc record; the paper itself was not read.** The quoted sentence is the abstract's first. Cited only for the existence of a long empirical debate in decision-support research over whether a DSS improves decision-making effectiveness — not for any finding of this paper, and not as a survey of that literature. Capture: documents/wi/redrafts/sources/AIBI\_v2/Todd\_Benbasat\_1999\_DSS\_Strategy\_Selection\_ISR\_ABSTRACT\_CAPTURE.txt.