Write the Options Memo, Then Make AI Attack Every Option

Workflow area: Solution Evaluation · Tier: Core · Hands-on time: ~35 minutes

Options memos fail two ways. The advocacy memo picks a winner in the author's head and arranges evidence around it; the false-balance memo pads a foregone conclusion with strawman alternatives nobody costed honestly. An AI red-team pass counters both — the model attacks each option with equal energy, has no career stake in any of them, and will happily articulate the case against the option you secretly prefer. Your job is to make it attack hard, then adjudicate which hits land.

When to use — and when not to

• Use it for any decision with materially different solution pathsbuild vs buy vs configure vs wait, replatform vs remediate, single vendor vs best-of-breed.
• Use it when you notice you already have a favoritethat is precisely when the red-team pass on your preferred option earns its keep.
• Use it before the memo reaches reviewa red-teamed memo anticipates the steering committee's hardest questions instead of collecting them as actions.
• Skip it for decisions already made upstreamif the sponsor has signed the vendor contract, an options memo is theater; write the implementation risk assessment instead.
• Skip it with no evidence basered-teaming assumptions you invented five minutes ago polishes fiction; gather the fit-gap, cost, and vendor evidence first.
• Skip letting the AI picka model recommendation is a fluency artifact, not a judgment; the memo's recommendation section is written by you and owned by the governance forum.

The recipe

1. 1.

   Frame the options honestly. Define three to five genuine options — including "wait" with its real meaning (defer with a review date, and what deteriorates meanwhile). For each, gather the evidence you have: fit-gap results, indicative costs, vendor claims tables, capacity constraints. Note where evidence is thin; the red team will find it anyway, and pre-declaring it is cheaper.

2. 2.

   Draft the comparison skeleton. Have the AI structure the memo: context, decision to be made, options with a consistent per-option template (description, cost range with basis, time to value, key dependencies, exit cost if reversed), and a criteria matrix scored provisionally from your evidence. Rewrite the problem statement and criteria weights yourself — those encode the business's priorities, not the model's.

3. 3.

   Run the red-team pass, one option at a time. For each option:

   Copy

   You are a hostile but fair reviewer on an architecture and
   investment board. Attack the following option as presented.
   
   Option: [PASTE the full option section, evidence included]
   
   Produce:
   1. ASSUMPTION ATTACKS — every stated or implied assumption,
      with the strongest plausible argument it is wrong and
      what evidence would settle it.
   2. COST ATTACKS — omitted cost lines (integration, change
      management, ongoing support, exit costs), optimistic
      estimates, and which numbers lack a stated basis.
   3. FAILURE SCENARIOS — the three most plausible ways this
      option fails 18 months in, and the earliest observable
      warning sign of each.
   4. SECOND-ORDER EFFECTS — what this option makes harder
      later: lock-in, skill atrophy, roadmap dependence.
   5. THE STEELMAN AGAINST — one paragraph: the best case for
      rejecting this option, written to persuade.
   Attack the option actually on the page, not a strawman.
   Do not soften. Do not rank options — attack this one only.
   

   Run it separately per option so the model cannot slide into premature comparison.

4. 4.

   Adjudicate the hits. Sort every attack into three bins: FIXABLE (the memo gets a costed mitigation or the missing cost line), FATAL-IF-TRUE (needs evidence before the decision — becomes a pre-decision action), and NOISE (plausible-sounding but wrong for your context — record why, because a committee member will raise it too). Sample adjudication:

   Copy

   ATTACK (Buy option, cost): "Estimate omits data migration
   from the legacy vendor's proprietary schema."
   → FATAL-IF-TRUE. Action: migration sizing from vendor
   before decision. Owner: DK. Due: 22 Aug.
   
   ATTACK (Build option, failure): "Two named engineers are
   single points of failure; attrition kills velocity."
   → FIXABLE. Mitigation added: pairing + documented ADRs,
   costed at 15% velocity overhead in the estimate.
   
   ATTACK (Wait option): "Defers DPDP compliance exposure."
   → NOISE for this scope — compliance is handled by the
   existing remediation program; noted in memo §2.
   

5. 5.

   Write revisit triggers into the memo. Whatever is chosen, name the conditions that reopen the decision: "revisit if vendor's promised API slips past Q2," "revisit if build velocity falls below X by month 3," "the wait option expires on [date] or on [event], whichever first." Triggers convert a one-shot bet into a governed position — and they are what makes "wait" a real option instead of a euphemism for never.

6. 6.

   Ship with the red team visible. The memo's final form: your hand-written recommendation and rationale, the per-option analysis with mitigations incorporated, an appendix listing the material attacks and their adjudication, and the revisit triggers. Committees trust memos that show the case against the recommendation; that trust is the point.

Make it yours

• Two-model red teaming: run the attack prompt in a second model (or a fresh session with no drafting context) — a red team that watched you write the memo inherits your framing.
• Persona passes for stakeholder-shaped attacks: rerun step 3 as the CFO (cost attacks sharpen), the CISO (data-path and vendor-risk attacks), and the operations lead (workflow and adoption attacks) — three cheap passes cover the committee's actual seating chart.
• Decision-matrix templates as the spine: keep your org's standard criteria matrix as the skeleton the AI fills, so options memos are comparable across the portfolio.
• Lightweight mode for small decisions: for a two-option, low-stakes call, collapse to the STEELMAN AGAINST for each option plus revisit triggers — ten minutes, and still more rigor than most memos ever get.

Quality check

• Options are realeach could genuinely be chosen; "wait" has a definition, an expiry, and a cost.
• Every option was attacked with equal forceincluding, especially, the one you prefer; the adjudication log proves it.
• FATAL-IF-TRUE items are closed or escalatedno decision request ships with an unresolved fatal attack; it ships with the evidence action instead.
• Costs survived the omission sweepintegration, change management, run costs, and exit costs appear for every option, or their absence is explained.
• Revisit triggers are observabledates and measurable events, not sentiments like "if things change."
• The recommendation is humanwritten by you, owned by the governance forum, with the AI's role confined to drafting and attacking.