A working playbook for the weeks after the board hands you an AI headcount number.
The mandate has landed. Cut 15% with AI. Reduce operating cost 20% by year-end. You own the number now, and you own everything that breaks if it is executed wrong.
You have two options: work backward from the number and hope, or respond with a plan the board cannot dismiss and your operation can survive. This framework is the second option, broken into five steps you can run starting this week. Each step tells you what to do, what to ask, and what to produce, so that by the end you are holding a counter-proposal instead of a guess.
One principle runs through all five steps: never argue with the mandate on instinct. Convert it into a data question, then answer the question. Instinct gets overruled in budget meetings. Evidence gets a seat at the table.
Step 1: Clarify the Mandate
Most mandates arrive as a percentage and a deadline, and almost nothing else. Executing an ambiguous mandate is how you end up accountable for a target nobody actually defined. Before you plan anything, pin the mandate down in writing.
Ask the mandate-giver these questions. Record the answers.
| Question | Why it matters |
|---|---|
| What exactly is the target: headcount, operating cost, or cost per transaction? | These are three different mandates with three different plans. Cutting heads without changing the work reduces capacity, not unit cost. |
| What baseline is the percentage measured against, as of when? | A 15% cut from January’s headcount and from today’s are different numbers. Undefined baselines get redefined later, against you. |
| Which operations are in scope, and which are explicitly out? | Scope creep turns a 15% mandate on one function into a moving target across three. |
| Where did the number come from? | Analyst benchmarks (the McKinsey 30%-of-work-hours figure, industry studies) describe the economy, not your operation. Knowing the source tells you what you are validating against. |
| What must not break? Which SLAs, compliance obligations, and customer outcomes are non-negotiable? | This becomes your risk boundary in Step 5. Get it stated now, while nobody is defending a decision yet. |
| What is the real deadline, and is there flexibility for a data-gathering phase? | Most mandates carry one to two quarters. A 90-day baseline fits inside almost all of them. |
Output of Step 1: a one-paragraph restatement of the mandate in your own words, sent back to the mandate-giver for confirmation. “Confirming: we are targeting a 15% reduction in operating cost for claims processing against the Q2 baseline, by year-end, without breaching our 48-hour SLA.” If they confirm it, you have a defined target. If they correct it, you just avoided planning against the wrong one.
Step 2: Establish Ground Truth
You cannot decide what AI should absorb until you know what the work actually is. Not what the org chart says, not what job descriptions say, but what your people actually do all day at the task level. This is the step most leaders skip, because the data has historically taken consultants 12 to 18 months to assemble. It is also the step everything else depends on.
Measure these five things for the in-scope operation:
- Time allocation per role, at the task level. Not “the team uses the claims system” but “this role spends 47% of its time on core processing, 19% on internal communication, and the rest on rework, status updates, and tool switching.”
- Task inventory, including the unofficial work. The workarounds, the duplicate entry, the spreadsheet nobody admits to. The work that is invisible on the org chart is still work, and it is often the most automatable.
- Volume and repetition per work type. How many of each item, how often, how consistent. AI economics depend on volume.
- Exception and judgment share. What portion of each role’s time goes to rules-based, repeatable work versus interpretation, exceptions, and relationship handling. This line decides everything in Step 3.
- Knowledge concentration. Who holds process knowledge that exists nowhere else, and which tasks they uniquely handle. This is the institutional knowledge a wrong cut destroys, and it does not come back at rehiring time.
A note on method: manager estimates and employee surveys will not carry this. People are reliably wrong about their own time allocation, and the errors compound when the answers feed a headcount decision. You need measured activity data. This is what the Summit Trails 90-day assessment produces: individual-level activity capture across the operation, classified at the task level, delivered as the Ground Truth AI² Report™. However you source it, the standard is the same: every number in your counter-proposal must trace back to measured activity, because every number will be challenged.
Output of Step 2: an activity baseline for the in-scope operation that you would be willing to hand to the board as an appendix.
Step 3: Segment the Work
With a baseline in hand, sort every significant work type into three buckets. The mandate can only be met responsibly from the first bucket, partially from the second, and never from the third.
| Bucket | Criteria | What happens to it |
|---|---|---|
| Automatable | High volume, rules-based, low judgment, structured inputs, low error cost, measured exception rate under control | AI absorbs it. This is where the mandate is met. |
| Augmentable | Judgment work with automatable components: AI drafts, retrieves, or pre-processes; a human decides | AI reduces the time per item, not the role. Capacity gain, not headcount, and often the larger dollar figure. |
| Human-critical | Exception-heavy, compliance-bound, relationship-driven, or dependent on concentrated institutional knowledge | Explicitly protected, in writing, with reasons. This list is as important to your credibility as the automation list. |
Two rules for the sort:
- Segment tasks, not roles. No role is “automatable.” Roles are bundles of tasks, and the bundle is what changes. The moment you sort by role instead of task, you are back to arithmetic on the org chart.
- Score against value and risk, not feasibility alone. Not everything automatable is worth automating, and the highest-impact opportunities often carry the highest risk. Rank the automatable bucket by volume times time saved, discounted by error cost and exception rate.
Output of Step 3: a segmented work map with, for each automatable item: the hours per week it consumes today, its volume, its exception rate, and its rank.
Step 4: Build the Counter-Proposal
Now convert the segmentation into the plan you bring back to the board. It is a counter-proposal, not a refusal: it takes the mandate seriously enough to answer it with evidence, and it replaces the top-down percentage with a bottom-up number, whether that number turns out to be higher, lower, or the same.
The arithmetic, in four lines:
- Total automatable hours per week (from Step 3), converted to FTE-equivalents.
- Discounted for reality: ramp time, exception handling that stays human, review overhead. A flat 20 to 30% discount on theoretical capacity is honest; your exception data will tell you where in that range you sit.
- Augmentation gains stated separately, as capacity and cost-per-transaction improvements, not as headcount. Mixing them inflates the number and someone will catch it.
- The result is your evidence-based equivalent of the mandate number, phased across quarters.
Phase it. A credible counter-proposal never delivers the whole number on day one:
- Phase 1 (quarter 1): automate the top-ranked, lowest-risk work types. Prove the capacity release against the baseline. Resize nothing yet.
- Phase 2 (quarters 2 to 3): expand to the next tier, begin resizing around the work that remains, redeploy protected knowledge holders onto exception handling and oversight.
- Phase 3 (quarter 4 onward): reassess against measured results, not projections, and adjust the target with the board.
Headcount change is the result of removing work, in that order. Cut first and the work does not disappear; it lands on whoever is left, unit costs rise, and you become part of the 55% of companies that regret AI-driven layoffs.
The One-Page Response Outline
This is the document structure for the board meeting. One page, five blocks, every number traceable to the baseline.
- The mandate, restated. Target, baseline, scope, deadline, as confirmed in Step 1. One sentence.
- What the data shows. Two or three findings from the activity baseline: total automatable hours identified, where they concentrate, the one finding nobody expected. (“31% of processing time in claims intake is re-keying data between two systems.”)
- What AI can absorb. The evidence-based number, phased by quarter, with the discount stated. This is your answer to the mandate.
- What we are protecting, and why. The human-critical list with one-line reasons tied to revenue, SLA, or compliance. This block is what separates a plan from a cut.
- The decision requested. Approve Phase 1, agree on the measurement baseline, agree on the checkpoint where the target is revisited against actuals.
Step 5: Present Risk-Adjusted Scenarios
Do not walk in with a single recommendation and defend it. Walk in with three scenarios and let the data argue. Boards respond to options with priced risk far better than to a subordinate saying no.
| Scenario | What it is | Risk profile |
|---|---|---|
| A. The mandate as issued | Full target, original deadline, executed as arithmetic | Priced honestly using your data: which human-critical work would have to go, which SLAs are exposed, what the Klarna-style rehiring scenario costs. Not a strawman. A real option with its real price tag. |
| B. The evidence-based target | Your Step 4 number, phased | The recommendation. Every line traceable to the baseline. State what it delivers by the original deadline and what it delivers in four quarters. |
| C. The staged commitment | Phase 1 only, with a data checkpoint before further commitment | The lowest-risk path if the board doubts the data. It converts disagreement about the target into an experiment with a measurement date, which is an argument you can win. |
If your data says the mandate number is achievable, say so, and Scenario A merges into B. That outcome is just as valuable: you now execute with confidence instead of hope. The point of the framework is not to fight the number. It is that nobody, including you, knows whether the number is right until the ground truth exists.
Talking Points: Mandate Claim to Data-Backed Response
| When they say | You respond |
|---|---|
| “McKinsey says 30% of work hours are automatable.” | “Across the economy, yes. Our activity data shows what the number is for this operation: [X]% of measured hours are automatable at acceptable risk. That is the number we can commit to and defend.” |
| “Competitor X already cut 20% with AI.” | “We do not know what work their 20% contained, and neither do they, publicly. We know ours: here is the automatable share, task by task. Matching their percentage instead of our data is how 55% of companies ended up regretting AI layoffs.” |
| “Why should we wait 90 days for more analysis?” | “It is not analysis, it is the baseline the whole plan is measured against. Ninety days fits inside the mandate deadline, and it replaces guessing with evidence on a decision worth [X]% of operating cost. Cutting wrong costs more than measuring first.” |
| “This looks like slow-rolling the target.” | “Phase 1 starts immediately. The staging is not delay, it is sequencing: remove the work first, then resize around what remains. Cut first and unit cost goes up while capacity goes down. The phased path hits the target without the rebound.” |
| “The number is not negotiable.” | “Then the plan matters more, not less. Here is what hitting the full number requires cutting, priced in SLA exposure and institutional knowledge. If the board accepts that price, we execute. But it should be accepted knowingly.” |
| “Can’t we just buy an AI tool and start?” | “A tool needs to be pointed at the right work. Every automation candidate in this plan traces to measured activity data. Buying before knowing is how AI initiatives automate what sounds impressive instead of what moves cost.” |
The Sequence, in One Look
| Step | You produce | Timeframe |
|---|---|---|
| 1. Clarify the mandate | Confirmed restatement: target, baseline, scope, risk boundary, deadline | Week 1 |
| 2. Establish ground truth | Activity baseline: time allocation, task inventory, volumes, exception share, knowledge map | 90 days, parallel to everything below |
| 3. Segment the work | Automatable / augmentable / human-critical map, ranked by value and risk | As baseline data lands |
| 4. Build the counter-proposal | One-page response outline with phased, discounted, evidence-based numbers | Week of baseline delivery |
| 5. Present scenarios | Three risk-priced options plus talking points | Board meeting |
The mandate is real. The number, until the ground truth exists, is a guess, and you are the one accountable for the difference. Run the sequence and you walk into the board meeting with the only thing that changes that conversation: evidence.
The framework runs on one input you probably do not have yet: the activity baseline in Step 2. That is what the Summit Trails 90-day assessment delivers. Book a 30-minute strategy call, bring your mandate, and we will walk through what the baseline would show for your operation and how it maps to the steps above.
Use this framework with: how to respond to an AI headcount mandate, how to challenge AI headcount targets with data, and why companies regret AI-driven layoffs.