Leading Through Directors

A fictional commercial insurer put its AI transformation under a vice president with three directors: Data Platform, ML Platform, and AI Applications. The VP had told each director that their area was theirs. Over one quarter her office logged the 24 decisions the directors made on their own authority. Nine were reopened in her weekly staff meeting, and six were reversed. The ML Platform director’s two-week freeze on serving changes during a migration lasted three days, until a claims executive mentioned it to the VP in a corridor. The Data Platform director’s refusal to build yet another one-off extract was undone by a single email. Both directors then stopped announcing decisions and began asking the VP first, so the VP, who had wanted fewer questions, was answering more of them than before.

Team boundaries and decision classes were already settled, along the lines of team design and interaction and operating models and decision structure. What was missing was the working layer between the VP and her directors: what each director may decide, which business result each answers for, how the VP learns what was decided, and where a dispute goes. This post builds that layer, then asks what the VP remains accountable for once she has built it.

Why delegated work gets overruled

The log gives the reasons. Of the six reversals, four happened because nobody had defined whose decision it was. The director believed the call was theirs, and the VP, when she heard about it from someone affected, believed it was hers or at least shared. One reversal followed a constraint the director could not have known, a regulatory filing window. One was simple preference: the VP would have chosen differently. Four of six is 67%, and none of those four turned on a mistake in the decision itself; the authority was undefined.

Several conditions produce that pattern, and each needs its own repair. When authority is stated only as ownership, the VP and the director fill the gaps differently. When the director’s goals describe activity (ship the platform, keep it available), the VP cannot tell whether a decision serves the business, so she intervenes wherever a stakeholder complains. When the only way she learns of a decision is through the person it affects, her first act is a reversal, because the alternative is to look uninformed. And when stakeholders have no path other than the VP, every disagreement with a director arrives at her desk already framed as an appeal. The four repairs are decision rights, goals tied to business results, an operating review, and an escalation path, and they work as a set because each one removes a reason for the VP to step in.

Writing down what a director may decide

A decision right names a class of question, the scope it covers, the limit on the holder, and the escalation when the holder cannot or should not decide. A delegation level adds the degree: whether the VP decides after consulting, decides jointly, advises, or only asks afterward. Bain’s RAPID framework contributes a vocabulary for who else is involved: one person recommends, one decides, people whose agreement is required (used sparingly) must agree the proposal is feasible, those who carry it out perform it, and others provide input. Its point for this case is that a decision with an unrecorded participant is a decision waiting for a reversal. The claims executive who mentioned the freeze in a corridor had a stake that nobody had named in advance.

The table shows the charter the VP and the ML Platform director wrote. Dollar limits are hypothetical. The table is a lookup, so the columns are the ones the director needs when a question arrives.

Decision area	Director’s right	Condition or limit	Escalate to the VP when
Serving maintenance windows and change freezes	Decides; VP asks about it afterward (level 6)	Affected business owners give input at least five business days ahead; a freeze that covers a date on the shared quarter-close calendar needs that owner’s agreement	The business owner and the director disagree
Evaluation tooling and vendor renewals inside approved budget lines	Decides (level 7)	Annual value up to \$150,000 and a term of 12 months or less	Either limit would be exceeded
Retiring a platform capability that other directors’ teams use	Recommends; VP decides after consulting the affected directors (level 3)	A migration plan with a date for each consumer	Every case
Ranking of requests from the three directors’ teams	Decides	Published ranking criteria	Two directors dispute the ranking and cannot settle it within five business days
Releasing a model into a regulated claims decision	Owns the release process; approval sits with the model owner and the risk function, outside this charter	Not applicable	The process blocks approvals the business needs
Hiring	Decides inside approved headcount and pay bands (level 7)	Approved headcount	Any increase in headcount, which also needs the budget owner

Levels follow reversibility as well as trust. A maintenance window can be changed next week, so the VP leaves it with the director and checks afterward. A capability that three other teams depend on is hard to bring back once retired, so she decides before it happens. This is the reasoning behind a two-way door decision: reversibility sets how much review is worth its delay. The delegation practices in Fully Delegated, Then Overruled apply between any manager and report; the charter here applies them to a director’s whole span.

A charter cannot list every case, so each director also needs the VP’s intent. US Army doctrine builds its approach to decentralized command on this idea: mission orders that enable “disciplined initiative within the commander’s intent.” Army guidance describes intent as a shared picture of the purpose, the priorities, and the desired end state, so subordinates can act sensibly when the plan no longer fits. The VP’s version is a few sentences per director. For ML Platform: claims operations must keep serving through the quarter-end close; a model change should reach users within one release cycle; and no change may alter a regulated decision without the model owner’s approval. With that, a director facing a case the table does not mention can reason from purpose instead of escalating. Without it, the sensible response to uncertainty is to ask, which is how the insurer’s directors ended up pre-clearing everything.

Goals a director can answer for

The executive team gave the VP one number for the year: cut mean handling effort per commercial claim from 5.0 hours by 10%, which is 0.5 hours per claim. The directors had received goals by function. Data Platform would deliver the new claims feed with agreed availability. ML Platform would meet serving availability and latency targets. AI Applications would ship a triage assistant by the third quarter. All three goals were met. Mean handling effort fell by 0.336 hours, or 6.7%.

The gap is visible once the business goal is broken into factors. Suppose the assistant applies to the 70% of claims whose type has feed and model coverage, handlers use it on 40% of those eligible claims, and it saves 1.2 hours each time it is used. The expected saving per claim is 0.70 × 0.40 × 1.2 = 0.336 hours, which is 6.72% of 5.0. Every factor had a number, but only two had an owner. Coverage belonged to Data Platform and quality to ML Platform. The use of the assistant in claims handlers’ daily work belonged to no one’s goal, because shipping the assistant was an output, not a change in how claims were handled (the output and outcome distinction). What good goals provide is line of sight from a director’s work to the result the executives asked for.

Factor	What it measures	Owner	Starting value	Year-end target
Eligible share	Claims whose type has data and model coverage	Data Platform director	70%	80%
Adoption	Eligible claims where the handler uses the assistant	AI Applications director, jointly with the head of claims operations	40%	55%
Saving per assisted claim	Hours saved when the assistant is used	ML Platform director	1.2 hours	1.2 hours, held

The targets multiply to 0.80 × 0.55 × 1.2 = 0.528 hours, or 10.56% of 5.0, which meets the goal with a small margin. Moving adoption alone would require 0.5 ÷ (0.70 × 1.2) = 59.5%, a larger leap, and the table shows why a split across directors is more realistic. The same arithmetic shows how a miss spreads. If adoption reaches only 45% while the other two factors hit target, the product is 0.80 × 0.45 × 1.2 = 0.432 hours, or 8.6%. Two directors are on target and the company is not, and the review knows which factor to discuss.

Four cautions follow. The product is a planning model that assumes the saving per assisted claim is the same across claim types; the outcome itself is the measured mean handling effort from claims operations’ own data. Raising adoption can be bought at the price of quality, so the adoption goal travels with a guardrail, such as the error rate in assisted claims, as described under primary and guardrail metrics. Adoption depends on claims operations, which is outside the VP’s organization, so the goal is written as joint, with the head of claims operations named beside the AI Applications director. The VP’s part is the negotiation with that peer. Assigning the factor to a director who lacks the lever to move it would not create accountability, only a score. Finally, a goal needs a label as either a commitment or a stretch. Google’s public OKR guidance calls a 60% to 70% grade the sweet spot for ambitious objectives and states that OKRs are not synonymous with employee evaluations. A director who scores 0.6 against a stretch has not failed, and a director who misses a commitment has not been stretched, so the two outcomes call for different conversations.

The director operating review

The weekly staff meeting is where decisions were being reopened, because it was the one place the VP heard about them. A separate monthly review for each director gives her an earlier, structured way to learn what was decided and keeps the staff meeting for coordination among the directors. The review runs about 90 minutes, so three directors cost the VP 4.5 hours a month. It follows a fixed order.

1. The factors from the goal table, with what moved each one and what the director believes will move it next. This is not a status deck; a factor that did not move is the agenda.
2. The decision log since the last review: each decision taken under the charter, the right it used, and what the director knew at the time. The VP asks questions here, and her disagreement is recorded without being enforced.
3. Open escalations and how long each has waited.
4. Proposed changes to the charter, wider or narrower, with the evidence behind each.
5. What the director needs from the VP: budget, a conversation with a peer, a decision above their limit.

The second item carries most of the design. A review that checks decisions after they are made, for reversible ones, is a different instrument from an approval that happens before. DORA’s research on change approval found that heavyweight external approvals, such as a change advisory board approving each release, have a negative effect on software delivery performance, and it points to peer review and automated checks close to the work. Those findings concern software changes. The trade is general in one respect: a gate in front of every decision spends the VP’s judgment at the price of every director’s waiting time, which is the decision latency the insurer had been creating. Decisions with a high cost of reversal still go to her first, as the retirement row in the table does.

When the VP disagrees with a decision that sits inside the charter, she has four honest options. She can ask the question that would change her mind. She can say she would have chosen otherwise and let it stand. She can point to a limit the decision crossed or an effect on another director that the director could not see, and reverse on that basis. Or she can conclude that her own intent has changed, say so, and rewrite the charter or the intent; in that case the director has not erred. Preference alone does not justify a reversal, because a delegation that disappears whenever the VP disagrees is an advisory arrangement. The test in each case is whether the reason would have been stated in the charter if she had known it then.

A legitimate reversal has a procedure as well. She tells the director first, not the stakeholder. The director communicates the change where possible, because the decision came from the director’s role. And someone repairs the commitments the original decision created. In the insurer’s freeze, the director had already told two application teams to plan around the window, so the reversal required those promises to be unwound. The reversal rate of 25% (six of 24) then becomes a measure the VP can read by cause: how many reversals trace to unstated scope, new constraints, or preference.

Escalation that does not become the default

Escalation is how a decision moves to the holder of the right when it exceeds a limit, when directors disagree, or when time has run out. It has a cost that the VP’s calendar makes concrete. Suppose she sets aside 6 hours a week for decisions and each takes 1.5 hours with reading, conversation, and a recorded outcome. Her capacity is 4 decisions a week. In the quarter before the charter, about 7 a week reached her, a mix of referrals from the directors and stakeholders who went to her directly. That is 3 more than she could resolve each week, so the queue lengthened by 3 per week and after six weeks 18 were waiting. The directors’ new habit of asking first added to the queue it was meant to avoid.

Suppose the charter let the directors settle 60% of those requests. The VP then receives 7 × 0.4 = 2.8 a week, using 70% of her decision time. Even so, waiting is not zero. If requests arrive at random, the standard single-queue estimate of average wait is the utilization divided by the unused capacity: 0.7 ÷ (4 − 2.8) = 0.58 weeks, about 2.9 working days. At 3.6 requests a week the utilization is 90% and the same estimate is 0.9 ÷ (4 − 3.6) = 2.25 weeks, about 11 working days. A VP who runs her decision time almost full makes everything she touches slow, so the escalations she receives should be those that need her.

The charter makes the triggers explicit: a limit would be exceeded, two directors cannot settle a shared dependency, or a response target has been missed. A director escalates with a recommendation and the facts, never as an open question, and the VP replies in one of three ways. She decides. She returns it with a question that lets the director decide. Or she moves it to the body that holds the right, such as a governance body for a question that crosses her unit. The response target applies to her as well. She gives an expected date and names a deputy for when she is unavailable, since an escalation path that stalls teaches people to go around it.

The side door needs an answer too. When the claims executive raises a freeze with the VP directly, the useful reply is to ask whether it has been raised with the director, to tell the director that the question has arrived, and to ask for a response by a stated date. She decides herself only when the matter lies outside the director’s right. Handled this way, the executive learns that the director is the route, and the director learns that the VP will not undermine the route.

The VP can also read the escalations as evidence about the charter. If she upholds the director’s recommendation in 9 of 10 escalations (90%), the limit is probably too tight: the director is right and the referral is a delay. If she upholds 3 of 10, the director may lack information or judgment, and coaching or a narrower right is the response. Neither reading is a verdict on a single case, but a pattern over a quarter shows where the limits sit.

What accountability means once authority is delegated

The difference between responsibility and accountability matters here because the VP is accountable to her executives for a result produced partly by decisions she did not make. A director is accountable for the outcomes inside the charter, and also for the soundness of the team boundaries and staffing that make those outcomes feasible. The VP is accountable for the charter itself: whether each director has the authority, resources, and information the goal assumes. She is also accountable for her own reversals.

After a missed goal she can ask four questions, in order. Was the goal reachable with the rights and resources the director held? Did the director decide within the charter and with the information available then? Was escalation used where the limits required it? And did a reversal, a late constraint, or a shortage in her own commitments contribute? A director who made a reasonable decision under the charter and missed a stretch goal has a different conversation from one who skipped a required consultation. An owner who cannot impose conditions, obtain corrections, or stop use has a title and no accountability. Goals, rights, and resources therefore have to be agreed together.

The VP, a quarter later

The VP wrote a charter for each director and a few sentences of intent, and she shared them with the business owners who had been going around her. She replaced her habit of hearing about decisions in the corridor with the monthly reviews, and she tracked a short set of signals: the reversal count by cause, the age of open escalations, the share of escalations she upheld, and the three factors behind the claims goal. When the ML Platform director proposed another freeze, the charter required input from affected owners five business days ahead. The claims executive raised the quarter-close date in that window, the director shortened the freeze, and the VP learned of the decision at the next review, where she asked one question about the rollback plan and let the decision stand.

Adoption remained the hard factor. The AI Applications director and the head of claims operations agreed a plan for handler training and a change to the claims workflow, and the VP carried the funding question to her own manager. Whether the factor reaches 55% will show in the claims data, not in the charter. What the structure provides is a way to see which factor is behind, who owns it, and which decisions belong to which director while the answer is still open.

References

• Bain & Company, RAPID: Bain’s tool to clarify decision accountability.
• US Army, Understanding mission command.
• US Army, Mission command requires sharp commander’s intent.
• Google re:Work, Set goals with OKRs.
• DORA, Streamlining change approval.