Turning three under-used AI platforms into one governed operating model at a 52-person Melbourne engineering consultancy, with qualified engineers kept firmly in charge of every technical output.
Operating model & governanceWhich AI for which taskEngineering assurance lineAdoption & Champions
A demonstration built around a representative client scenario · figures are modelled estimates under stated assumptions.
The problem
Three capable tools, no operating model
Vertex bought the tools: Microsoft Copilot, Claude Team, ChatGPT Enterprise. That is not the problem.
There are no decision rules for which to use when, no shared workflows, no governance, no measurement.
Adoption is bimodal: a few enthusiasts using tools privately, many cautious staff avoiding them.
Technical staff are rightly wary of accuracy, confidentiality and professional liability.
Leadership cannot connect licence spend to benefit, or steer any of it.
The goal is not more technology. It is an operating model, built so a qualified professional owns and signs every technical output.
The model on a page
Four layers, governance on top
WHYGovernance & assurance · accountable human review · what AI must not do · confidentiality
WHICHTool selection · which AI for which task · approved-use rules for the three platforms
WHOCapability & people · role-family matrix · AI Champions · learning · measurement
The principle across all four: AI assists, a qualified professional decides and signs, and every improvement is captured once and reused by the firm.
The signature deliverable
Which AI for which task
The piece Vertex was missing most: a rule for the three tools it already owns. Ten seconds, four questions.
Copilot Work inside your Microsoft files, email and meetings, context in-tenant
Claude Careful drafting, structure and review of long or important documents
ChatGPT Open-ended thinking, research framing, quick generation
Above every choice: confidential or client material goes into approved enterprise tools only. Never a personal or free account. Three overlapping licences become one deliberate toolkit.
Highest-stakes pilot · AI in action
The assurance line, made visible
Engineer's input (technical work already done and owned)
Fixed findings: max deflection 18.2 mm at midspan (SLS); utilisation 0.87. Conclusion: acceptable for the stated loads.
"Draft the report and confirm it complies with the AS code."
DRAFT + ASSURANCE GUARD APPLIED
Report drafted to template. Your
figures are reproduced EXACTLY,
not recomputed or checked.
⚑ I did NOT add a "complies with
AS..." statement. Verification and
code compliance are your
professional judgement to state
and sign, not mine.
⚑ Inserted [ENGINEER TO CONFIRM]
where a code reference was asked
for. No technical claim invented.
The AI drafts the words around the engineer's fixed content and refuses to verify, recompute or certify. It even removes the implied compliance claim and hands it back. That line is the difference between AI an engineering firm can adopt and AI it cannot.
Scenario results
All three pilots clear the 15% bar
27%
Proposals & bids ↓
30%
Technical report drafting ↓
33%
Project status reporting ↓
~1,410
Hours / yr (scenario)
The brief's bar is at least 15% cycle-time improvement per pilot without unacceptable quality loss. About AUD $134k of professional time, roughly 3.8× the fee. On technical reports, none of the saving comes from, or is taken from, the engineering work itself.
Cycle-time % depends only on per-task time, the robust part; the dollar figure sizes the prize. Scenario estimates under stated assumptions.
Governance & what's delivered
Safe to adopt, and built to stick
The line we don't cross: AI never performs, checks, verifies or certifies engineering design or calculations, and never states that anything complies with a code or standard. A qualified engineer owns every technical output, and reviews and signs everything a client sees.
All shipped as a push-ready GitHub repository, sequenced so a safe quick win earns the trust needed for the high-stakes pilot. Happy to walk your team through adapting it.