Build the As-Is/To-Be Gap Table Where Every Number Has a Source

Workflow area: Process Modeling & Mining · Tier: Core · Hands-on time: ~35 minutes

The gap table is where analysis turns into a case for money, which is why it attracts confident numbers of unknown parentage. "Current cycle time: 9 days" gets typed from someone's recollection, survives six document generations, and collapses in the finance review when nobody can say where it came from. AI makes this worse and better at once: worse, because a model asked to draft a gap analysis will cheerfully supply plausible currents and targets; better, because a model given your actual evidence and a strict labeling rule is an excellent bookkeeper of what is measured versus what is assumed. The discipline this playbook enforces is binary and visible: every quantified cell either cites its extract — mining output (pb-25), report figure, profiling result — or wears the ASSUMPTION label with a named owner who believes it. A table that shows its assumptions honestly survives review; a table that hides them survives only until the first hard question.

When to use — and when not to

• Use it when as-is and to-be both exist in some evidenced forma validated current-state map with data (pb-23/pb-25) and a target state someone accountable has endorsed; the table is the bridge between them.
• Use it to feed a business casethe deltas with sources become the benefit hypotheses of pb-09-style sizing, and the enabler column becomes the cost side's shopping list.
• Use it to force the to-be conversationan explicit target column with "TARGET UNSET" cells is the politest way ever invented to show sponsors they have not actually agreed what good looks like.
• Skip it before current-state validationa gap analysis from an unvalidated as-is measures the distance from a fiction; run the walkthrough first.
• Skip letting AI fill either state from general knowledge"typical industry cycle times" have no place in your table; benchmarks enter only as labeled external references with a citation, never disguised as your current state.
• Skip the assumption-free illusionif your table has zero ASSUMPTION labels, it is not evidence-complete, it is honesty-incomplete. Real analyses carry assumptions; credible ones label them.

The recipe

1. 1.

   Stage the evidence, labeled by kind. Collect the validated as-is model, mining or report extracts (with scope: window, filters, log version), profiling results, the endorsed to-be description, and any stakeholder-stated targets with who stated them. Label every input MEASURED (with source) or STATED (with speaker) before the model sees anything.

2. 2.

   Generate the table with the labeling rule as law. Paste into Claude or ChatGPT:

   Copy

   You are building an as-is/to-be gap analysis table.
   Process: [NAME + one-line scope]
   Evidence inputs (each labeled MEASURED with source, or STATED
   with speaker): [PASTE]
   To-be description: [PASTE]
   
   Output a table:
   # | Capability/dimension | As-is (current) | Source | To-be
   (target) | Source | Delta | Enabler (what closes the gap) |
   Owner | Confidence
   
   Rules — non-negotiable:
   1. Every quantified cell cites a provided source, or contains
      the value prefixed "ASSUMPTION:" with the reason it is
      believed. NEVER present an unsourced number as fact.
   2. If a target is not stated in the inputs, write "TARGET UNSET
      — decision needed" rather than proposing one.
   3. Compute deltas only where both sides are quantified on the
      same basis (same measure, same scope); otherwise write
      "NOT COMPARABLE: [why]".
   4. Enablers must be concrete (system change, policy change,
      staffing, AI/automation candidate) — no "optimize" verbs.
   5. Use no knowledge beyond the inputs. No industry benchmarks.
   6. After the table: ASSUMPTIONS REGISTER (every assumption, why
      held, how it could be tested) and DECISIONS NEEDED (every
      TARGET UNSET and NOT COMPARABLE).
   

3. 3.

   Audit the cells like a finance reviewer. Good output looks like this:

   Copy

   3 | Claim approval cycle time | median 1.8 days; 6% tail >20
   days (MEASURED: mining view V-12, closed cases Apr–Jun, log
   v2026-07-14) | ≤2 days for 95% of cases (STATED: ops director,
   workshop 18 Jul) | tail reduction from 6% to 5pp target gap —
   NOT COMPARABLE: target is percentile-based, current tail is
   share-based; restate one | region-East escalation rework +
   auto-routing | Ops director | MED
   
   4 | Re-keying into tracker | 100% of claims re-keyed
   (MEASURED: walkthrough table rows 4–5) | 0% — portal-to-core
   integration | ASSUMPTION: integration feasible in current
   platform (IT architect, unverified) | integration build | IT
   lead | LOW
   

   Chase every Source cell to its extract; verify deltas recompute; confirm no number appeared that you did not supply. Any cell that fails becomes ASSUMPTION or gets deleted — the audit is the analysis.

4. 4.

   Route the two registers to their owners. DECISIONS NEEDED goes to the sponsor and process owner — unset targets and incomparable measures are their calls, and getting them made is the fastest value this table delivers. The ASSUMPTIONS REGISTER gets triaged: which assumptions are cheap to test (a mining query, one profiling run, a phone call) before the analysis ships, and which are accepted with a named believer and a review date.

5. 5.

   Publish with challenge instructions. Circulate the table with its registers attached and one sentence up front: "every number is sourced or labeled — challenge any line." Version it against its evidence set (log versions, extract dates), because a gap table's numbers age; regenerating from refreshed extracts is cheap, and a dated table quietly becomes fiction.

Make it yours

• Spreadsheet + LLM as the working format: keep the table in a sheet with a Source column of real links; use the model to draft and to re-audit ("list every row whose source cell is empty or vague") on each revision.
• Mining-fed pipelines: where Celonis/Signavio runs (pb-25), pipe verified view figures straight into the as-is column with their scope strings — the strongest possible Source cells.
• Signavio journey/process models as the to-be anchor: link target-state rows to the governed to-be model so the gap table and the process repository cannot drift apart.
• Portfolio roll-up: with consistent columns across processes, a second prompt aggregates gap tables into a heatmap of biggest evidenced deltas — a prioritization view for pb-36-style AI-opportunity scanning that inherits the same source discipline.

Quality check

• Zero unlabeled assertionsevery quantified cell is sourced or wears ASSUMPTION; scan the table specifically for confident numbers with vague sources, the classic smuggling route.
• Sources resolve and scopes matchsampled Source cells land on real extracts, and comparisons share basis and scope; mismatched bases are marked NOT COMPARABLE, not averaged over.
• Deltas recomputethe arithmetic between current and target checks out by hand for every row that will appear in a business case.
• No foreign numbersnothing in the table originates from the model's general knowledge; benchmarks, if any, are labeled external with citations.
• The assumptions register is triagedtestable assumptions were tested or scheduled; accepted ones carry a named believer and a review date.
• Unset targets reached the decision makersevery TARGET UNSET line has a decision forum and date, because a gap without an agreed target is an opinion about direction.