← Back to Outcomes Now

Where AI Actually Sits in Your Business
— and What the Gap Costs

Adoption Gap Diagnostic · Baseline, week two of two
MERIDIAN CIVIL
Civil infrastructure · Victoria
120 people · est. 1998
ILLUSTRATIVE COMPOSITE — fictional company; every figure anchored to a cited national pattern (provenance at foot of page)
Prepared forManaging Director
Cohort measured44 office staff
MethodAnonymous survey, workflow interviews, one measured hour
StatusBaseline signed-off pending 4 questions
What producing this took from Meridian's people — under three hours, total: 40 minutes with the Managing Director · a 5-minute anonymous survey (44 staff) · three 45-minute workflow conversations · one hour reconstructing a single variation claim with one estimator. No systems access, no downtime, no one pulled off a tender.

1 · What's already happening (measured, not assumed)

Leadership's guess going in: “maybe eight people use it.” The anonymous survey came back: nineteen.

19/44
office staff used AI on real work in the past month — more than double what leadership estimated (8)
7
use it weekly or more on personal accounts, with company documents, outside any oversight
0
workflows where anyone could say what a good AI output looks like, in writing

Survey was anonymous and scored the conditions around people, not the people — stated up front, 91% response rate.

2 · The conditions read — why usage isn't becoming capability

ConditionWhat we foundEvidenceBinding?
Knowing howAdequate in pockets. Three estimators self-taught to a high standard; most others one prompt deep.MEASUREDNo
Able toTools present (two paid seats, many personal accounts). No shared setup, no company knowledge connected — answers are generic because the AI can't see Meridian's data.MEASUREDPartly
Feeling safeDominant blocker. Survey verbatims: “not sure I'm allowed to put client docs in”, “feels like cheating”, “don't want to be the one who caused a leak”. 14 of 19 users hide or downplay their use.MEASUREDYES
Room in the dayEstimating team runs at tender-deadline load; new methods abandoned under pressure (“no time to do it the new way twice”).MEASUREDYES
Environment allows itNo stated permission line, no worked examples, no checking routine. A confident wrong answer in a variation claim would currently reach the client unchecked — confirmed by the estimating manager.MEASUREDYES

Read in one line: your people are further ahead than you thought, and the business is further behind than they are. More training targets the two conditions that aren't the problem.

3 · What the gap costs, per month

WhereBasis$/monthStatus
Variation claims — drafting & assembly hours carried by senior estimatorsOne measured hour, extrapolated across 11 claims/month (see split-hour sheet)$14,300ESTIMATE
Tender responses — repeat content rebuilt from scratch each submissionInterviews, 3 recent tenders reviewed$9,800ESTIMATE
Site reports & client correspondence — office rework of field notesSampled week, 2 project admins$6,200ESTIMATE
Paid AI seats currently delivering the above: unusedLicence billing$400MEASURED
Recoverable, at conservative uptakeAll figures become measured in week 3 if you proceed; every estimate is flagged until then≈ $30,700ESTIMATE

4 · Still open — the four questions only you can answer

#QuestionWhy it matters
1Which client contracts restrict where project documents can be processed?Sets the permission line staff are guessing at GAP
2Who do you trust to say what a “good” variation claim looks like — and will you give them the hours to write it down?The checking standard everything hangs on GAP
3Estimating’s deadline load: what comes off their plate during the 16 weeks?New habits don't form at 100% load GAP
4Is the exit horizon 3–5 years?Decides how hard we document for due diligence GAP

The verdict, in one paragraph

Meridian doesn't have an AI adoption problem — it has an ownership vacancy. Nineteen people already use the tools; nobody owns what happens after that. The blockers are permission, deadline load, and a missing definition of “good” — none of which training touches. Fix those three in the estimating workflow first (fastest payback, highest judgment content), and the ≈$30,700/month stops leaking. If we proceed and fewer than 60% of the cohort are genuinely applying it in live work by the program's end, remediation is at our cost — your decision stays defensible either way.

Whatever happens next, this document is yours. Proceed, and week 3 replaces every estimate above with your measured numbers. Pause, and you still hold the census, the conditions read, and the gap figure — evidence for whatever you decide, with whoever you decide it. Conclude the gap isn't worth closing, and this page is how you'll defend that call too. There is no step in this process where stopping costs you what you've already gained.

Reading this as an owner of a 20–300 person firm? Your version of this page will differ in scale, not shape: the hidden-user count, the three binding conditions, and the monthly figure move — the pattern doesn't. The question the diagnostic answers is only ever which numbers are yours.

Where the composite's numbers come from — 19/44 quiet users: national surveys find 21–30% of employees use AI without telling their employer (Jobs & Skills Australia; Employment Hero) and 70% of businesses have little visibility of it (Josys) · permission-anxiety verbatims: UTS Human Technology Institute SME survey · sub-30% sector adoption: National AI Centre SME Pulse 2026 · deadline-reversion and training-decay patterns: MYOB Business Monitor; MIT NANDA 2025 · dollar figures: award and market senior-estimator rates applied to reconstructed hours. In a live diagnostic, every one of these becomes your own measured number.