November 2026

Bill McDermott
Chairman & CEO
ServiceNow

Re: The next CRM frontier: agents with sagacity

Dear Bill,

In your July interview with LEADERS, you asked a question that deserves a practical answer: “What does your team know that AI doesn’t?” (LEADERS, July 2026).

You also emphasized the enduring need for people who anticipate consequences, build trust, and exercise judgment when stakes run high (LEADERS, July 2026).

That question motivates a collaboration we would like to explore with ServiceNow: testing how selected elements of experienced practitioners’ judgment can help AI agents choose better next actions while preserving decisions that require human accountability.

Core proposal

ServiceNow already connects enterprise context with governed action, and your Context Engine announcement describes intelligence that compounds through human and agent decisions (ServiceNow, May 2026).

Our proposal targets a specific contribution within that ambition: capturing the cues, interpretations, exceptions, and reasons behind consequential sales decisions, then testing whether those records improve both sales and agent performance.

Scenario: A buyer asks for a discount, and a standard agent retrieves the pricing policy to prepare an approved offer.

Practitioner Insight: An experienced account leader recognizes the buyer may lack internal sponsorship and needs help building a coalition before price matters. They test that diagnosis before offering concessions.

We call this capacity sagacity: discerning what matters in a particular situation, anticipating consequences, and choosing a defensible course when evidence remains incomplete. It includes knowing when to act, when to wait, and when to seek another person’s judgment.

We propose sagacity engineering as the disciplined work of making such judgment explicit, testing its usefulness, and defining its limits. The goal should not require turning every expert intuition into an automated rule.

Sagacity engineering and the five elements

Records of calls, emails, opportunities, and workflows provide essential evidence. Our working hypothesis concerns what those records may omit: the cue an expert noticed, the explanation they rejected, the question that changed their interpretation, or the condition that would have led them to choose differently.

This knowledge can travel informally through experienced teams, often under the label “tribal knowledge.” For this collaboration, we propose treating it as practitioner knowledge that needs provenance, challenge, and revision, not automatic deference.

An expert can carry hard-won insight and outdated assumptions into the same conversation. Capture should therefore begin an evaluation, not confer authority.

To capture practitioner knowledge for evaluation without automatic deference, we propose preserving five structured elements for each consequential decision:

• Situation: What happened, and what information reached the practitioner at that moment?
• Recognition: What did the practitioner notice, and what did it suggest?
• Choice: Which action followed, and which alternatives did the practitioner reject?
• Outcome: What happened next, including surprises and mistakes?
• Boundary: Under what conditions should an agent reuse the guidance, question it, or escalate?

Proposed next steps for decision-making

Our proposed contribution would connect these records to specific selling behaviors and examine their effects through execution and buyer response. We do not yet know whether this adds anything to what Context Engine already records. Your team can see inside those decision traces; we cannot. We expect the difference to lie here: a trace records the action taken and its outcome, but rarely what the practitioner noticed beforehand, the alternatives they rejected, or the conditions under which they would have chosen differently.

We propose a focused, low-risk evaluation framework to determine value:

1. Select scope: Choose one consequential sales motion, a small group of experienced practitioners, and an agreed set of held-out decision cases. Your team selects the cases.
2. Blind comparison: Compare your own agent, grounded in Context Engine, with the same agent given the five-element records as well. Evaluators blind to the versions assess recommendation quality, assumptions, concessions, and escalation timing.
3. Clear thresholds: Fix measures and pass marks before the first case runs. Begin with historical cases whose outcomes are already known, chosen because the outcome reveals the right call. Each agent sees the deal only as it stood at the moment of decision. If that stage passes, run a time-limited test on live deals over a period your team chooses, measured on discount depth, slip rate, win rate, and time to close. Stop or redesign if no clear value is shown or risk is introduced.
4. Governed capture: Practitioners take part by consent, can review what is recorded in their name, and can withdraw it. Records belong to the organization whose people created them, serve only this evaluation, and are never used to assess an individual’s performance. Customer details are de-identified, and records are returned or deleted at the end, as you choose.
5. Governed execution: Ensure customer control, explicit permissions, traceable reasoning, and human decision rights remain paramount.

We commit in advance to publishing the result and its method, whichever way it falls. If the records make no measurable difference, the finding will be that your traces already capture practitioner judgment, and that is worth knowing. If they do, you will know exactly what your traces are missing. The opportunity then extends beyond preserving what an expert once knew to learning which judgments deserve reuse as circumstances change.

Bill, would you join a 30-minute working session, or designate an executive from your CRM or agent team to explore this with us? We would use that conversation to select one decision worth improving, establish the baseline, and agree what evidence would justify proceeding.

Sincerely,