Two examples sit side by side in this article. The first is a case from our own archive, about a deal team that stopped building investor lists by hand. The second is a picture, labeled as one, about software that decides who may read a client’s confidences. Both exist to save a revenue producer’s time. They differ in what happens when the software is wrong, and that difference is the whole reason that we have a Citizen AI → Production AI promotion pipeline in our governance framework.
The Easy Button, Before AI
Our case story, Building the Easy Button · An AI Success Story in CRE Capital Markets, describes a commercial real estate capital markets firm where deal teams built buyer lists the old way. They copied the list from a similar past deal, then adjusted it from memory: this investor likes medical office near hospitals, that one wants new builds in growing suburbs. The firm’s CRM held years of notes and outside data kept piling up, and no person could weigh it all. So the firm built a model that scored every potential buyer for each new property, using observed deal activity as well as stated preferences.
They copied the list from a similar past deal, then adjusted it from memory: this investor likes medical office near hospitals, that one wants new builds in growing suburbs.
As the case reports it, the model ran alongside the manual process for three weeks. Brokers kept making their own lists while the model made its own. The overlap was large and the model missed fewer buyers, and by the end of the first month the teams had dropped the manual lists. The firm reported $4.5 million in annual efficiency savings and a six-point gain in retention worth $3.2 million in revenue, against a $2.1 million build over twelve months, with payback in under four months of live operation. Those are the firm’s own figures as the case gives them. It is one firm, we have not audited the numbers, and the result is seen in hindsight.
Why That One Was Easy to Prove
Two things made that case easy to prove, and neither was the algorithm. The first is that a wrong answer was cheap. A poor recommendation meant an irrelevant email, and the deal teams stayed in the loop. The second is that the parallel run handed the firm a control group by design. The old process kept running, so the new one could be compared with it on the same deals. Proof was almost a byproduct.
Most AI work now lacks that luxury, as we wrote in The Control Group Is Gone, where the problem is measuring a tool after nobody will work without it. The easy button had a second advantage that is easy to miss. Because a miss cost little, the model could earn its place in the work gradually.
The Delicate Version
Now picture a different kind of help for the same kind of person. An account rep at a services firm hears something from a client in confidence: a leadership change not yet announced, a budget cut, a plan to move some work to another vendor. After every call the rep writes follow-up notes, and the firm’s habit is to copy the account team. Everyone knows the risk, and today it is handled by memory and care. The rep writes two versions, one full and one stripped down, or copies fewer people and hopes colleagues are not surprised, or skips the team note. All of it costs time, and the temptation to cut corners is strongest when the rep is busiest.
Careless addressing is a familiar way for confidences to escape. One research firm writing about the email security market reported that misdirected email was the most common incident type reported to the UK data regulator in the last quarter of 2024, at about one in five. We could not confirm that figure on the regulator’s own page, it covers one quarter in one country, and it concerns personal data and not client strategy. We cite it only as a sign that mistakes in addressing are common, not as a measure of this risk.
Where the Pipeline Says Stop
Suppose a rep builds a helper that drafts those follow-up notes. That is the open lane doing its job, the one we described in The Best of Both Worlds: people experiment with assistive tools, and the useful ones surface through a front door. A drafting helper that the rep reads before sending is a time saver with a person in the loop. Colleagues hear about it and want it too.
Then the use changes. The rep wants the tool to recognize which passages are confidential and to send the right version to the right people without being asked. That tool is no longer drafting. It is deciding who sees what, in a setting where a mistake damages a client relationship. This is the moment our Citizen AI to Production AI pipeline exists for, and the helper is flagged for promotion. In the terms of our Governance Freeway, a production tool earns an owner, an audit record, testing before release, a fallback, and a review date before it touches live work.
What the Promoted Version Does
Here is the design we would want, as a picture of the idea and not a description of a deployed system. An AI sensitivity flag reads each follow-up note and marks passages that look like client confidences: plans, numbers, names, and intentions the client asked to keep private. The full note goes to the client and the rep, one to one. For everyone normally copied, the system produces a separate, scrubbed summary through a pre-built secondary lane: the meeting happened, the topics in general terms, the next steps and who owns them, and nothing the client said in confidence. Routine communication still flows, and the confidential detail stays between the rep and the client.
Two design choices carry the weight, and both are our judgment. First, the system errs toward flagging, because a false alarm costs the rep a moment and a miss costs the client’s trust. Second, a flag in doubt goes to the rep for a decision, and the system does not guess quietly. The rep remains accountable for what the client said and who may know it.
Promotion Is the Unlock
Many companies treat a governance review as a brake on automation. This picture shows the opposite. Without promotion, nobody would let that tool run on client communications, and nobody should, because the stakes are too high for an experiment that no one owns. With promotion, the capability becomes something the firm can automate at all. The gate is what turns a clever helper into a permitted one. It is the idea behind the freeway comparison in our governance work: lanes and on-ramps are the reason traffic can move fast.
Many companies treat a governance review as a brake on automation. This picture shows the opposite.
That is also where we think our approach differs. The unsanctioned helper and the promoted tool might use the same model. What changes is that the second has an owner, evidence, and a way to be pulled back, so the question “is this safe to automate?” has an answer the firm can show.
How We Would Prove It Before It Acts
Here the easy button’s lesson turns over. The old case could lean on a parallel run that created a control group. This one has a harder problem, because its best outcome is something that does not happen. A leak that never occurs leaves no trace, so you cannot count the saves. You can, however, assemble the evidence before the tool is allowed to act.
The method is what machine learning teams call a shadow deployment: the new system runs on live work alongside the current process without affecting anyone, and its outputs are compared before the cutover. Here, for a few weeks, the flag would read real follow-up notes while reps carried on as usual and no scrubbed summary was sent. A reviewer would then check a sample against what the flag said. The measures are set in advance: how many confidential passages the flag missed, which is the number that decides everything; how many harmless passages it flagged, which measures the nuisance; how often reps would have overridden it; and how long a note takes with and without it. The promotion threshold, and the trigger for pulling the tool back after launch, are written down before the first note is read.
That is the Prove step of Project Vista applied to delicate work: the project carries the result it was approved to produce, and it ends in a verdict to scale it, rework it, or stop. It also differs from the approach in the Control Group article, which is about measuring a tool after it has become part of the work. Here the proof comes before the permission.
What This Will Not Do
The second example is a picture. It has no measured results and we claim none; its point is the path from helper to permitted tool, not a figure. The easy button numbers come from one firm’s own reporting, and the two examples are not like for like, since one measured time and revenue after the fact and the other would measure errors before launch. A shadow run can share blind spots with the reviewer who checks it. A detector that does well on the notes it saw may stumble on a new client or an unusual phrase. No sensitivity flag replaces the rep’s judgment, or the confidentiality terms in a client’s contract, which are questions for counsel. And the thresholds we describe are our reasoning, not a published standard; the shadow deployment practice itself is described mainly by vendors in their own glossaries, and we have not seen it measured for this use.
Where to Start
Look at the helpers your people have already built and ask one question of each: does it only draft, or does it decide who gets what? The ones that decide are your promotion candidates, and the delicate ones are where a good gate pays for itself fastest. If you want a second set of eyes, our independent assessment starts with that inventory, and our 90-day pathway lays out the evidence folder you would want behind it. If you would rather talk it through first, you can reach us through the contact form on our About page.
The ones that decide are your promotion candidates, and the delicate ones are where a good gate pays for itself fastest.
