What Happens After You Engage Us? Our AI Implementation Process From Scoping to Handover

Our AI Implementation Process Stage by Stage

Five stages carry the build from scope to steady state, and one signed artefact closes each of them before the next stage opens.

Five standing artefacts carry the process end to end: the opportunity map, the eval-set, the governance evidence pack, the dashboard set and the runbooks. The same governance evidence pack the engineering team writes to serves as a reusable baseline for procurement and for any internal assurance, audit or regulatory review; it can be supplemented for client, sector and jurisdiction-specific requirements rather than assumed to satisfy every reviewer.

Structuring AI delivery into staged decisions with gates between them is a documented discipline, not house style. A peer-reviewed stage-gate model screens candidate AI solutions through defined development stages and decision gates to manage uncertainty, per Leppinen, Salo and Compare (2026; published online 2025) in the EURO Journal on Decision Processes — a study whose scope is AI solutions for preventive maintenance, supporting structured evidence and continuation/termination decisions under constraints rather than a general conclusion about all delivery. At least 30% of generative AI projects face abandonment after proof of concept by the end of 2025, per Gartner’s July 2024 forecast, and the stage map above exists to make that pilot-to-production boundary an explicit gate rather than a cliff. The build runs under a named framework, the Production Path, taking each engagement from Discover to Prototype to Productionise, governed throughout by the Production Controls, covering evaluation, observability and governance.

What Enters and What Leaves Each Stage

Each stage takes named inputs and returns a signed artefact: Discover takes approved sample or production-shaped data and returns a scoped opportunity map, Prototype takes authorised real-data access and returns a frozen eval-set, and Productionise completes the governance evidence pack that was initiated in Discover. The same artefact set serves procurement, engineering and any later review, so no compliance-only version gets built afterwards.

Discover takes sample data and stakeholder access, and it exits through the scope sign-off gate with a signed opportunity map.

Prototype takes real data against the scoped opportunity, and it exits through the validation gate with a frozen eval-set.

Productionise takes the validated build, and it exits through the acceptance gate with a governance evidence pack, dashboards and rehearsed rollback paths.

Why Bespoke-Only Delivery Breaks Down

Bespoke-only delivery breaks down because it reinvents its failure modes on every project instead of reusing repeatable stage gates that make outcomes predictable. Stage gates in an AI project work like acceptance tests in software delivery: nothing ships until the gate passes. Repeatable stage gates, not bespoke artistry, make outcomes predictable, and a process reinvented on every project reinvents its failure modes right along with it. A one-pass, waterfall-style rollout skips those same gates and pays for the skip later, when a regulator or a budget owner asks for evidence that was never produced.

A stage-gate decision process brings consistency to AI development and reduces the risk of premature fixation on a single candidate solution, per Leppinen, Salo and Compare (2026; online 2025) in the EURO Journal on Decision Processes (a preventive-maintenance study) — support for consistency and structured decisions, rather than a blanket case against bespoke delivery. Background estimates that Amrit and Narayanappa (2025) discuss — roughly 13% of business machine-learning projects reaching production (a figure traced to an earlier VentureBeat article (2019)) and prototyping plus deployment consuming up to 90% of project time (attributed to Flaounas) — point to effort living after the demo, where one-off delivery carries no repeatable gate; these are cited estimates, not results of the paper’s own study. In the version 0.1 MIT NANDA report (State of AI in Business 2025, an observational report), about 5% of custom enterprise generative-AI tools reached production; repeatable prerequisites, not artistry, are what this page argues separate that group from the rest.

Infrastructure Prerequisites

Infrastructure prerequisites cover environment access, an integration surface into live systems, and telemetry hooks. Discover works with approved sample or production-shaped data, and authorised access to real systems is arranged for Prototype, so the prototype runs against production-representative data rather than a sandbox copy that never resembles what ships.

Human Capital Prerequisites

Human-capital prerequisites name one client owner per gate, a reachable sponsor, and an operations owner for handover. Every gate carries a named approver on the client side before delivery starts, so no decision waits on an absent stakeholder once a stage reaches its checkpoint. Synthesised.ai builds to ISO/IEC 42001, the NIST AI Risk Management Framework, IMDA AI Verify and DISR guidance, and shows no client logos or testimonials until written consent exists.

Data Prerequisites

Data prerequisites cover sample data, access approvals, and data handling set against applicable privacy law and voluntary guidance from Discover onward: Singapore’s PDPA and, for covered Australian entities, the Privacy Act 1988 and APPs, alongside Australia’s DISR Guidance for AI Adoption (published 21 October 2025 as updated guidance that evolves the earlier Voluntary AI Safety Standard). Data-handling controls attach at the first stage, so the governance evidence pack records provenance from the moment the first record moves rather than reconstructing it after the fact.

Applicable privacy law governs personal-data handling: Singapore’s Personal Data Protection Act, and, for covered Australian entities, the Privacy Act 1988 and the Australian Privacy Principles. Australia’s DISR Guidance for AI Adoption is voluntary AI-governance guidance rather than a binding privacy law or standard. Delivery builds to the voluntary guidance and takes account of applicable legal obligations.

Stage One: Discover

Maps the client’s data estate before scoping starts.

Documents the workflows the build touches.

Scores each candidate opportunity against measured volumes.

Scopes the build against the highest-scoring opportunity.

Defines the eval baseline the Prototype gate measures against later.

RAND’s 2024 root-cause research, drawn from 65 practitioner interviews, names misunderstood problem intent and inadequate data among the five leading causes of AI-project failure, and both are Discover-stage defects, caught by scoring opportunities against measured volumes before sign-off. Gartner’s July 2024 forecast attributes post-proof-of-concept abandonment to poor data quality and unclear business value first, the two questions Discover answers before a line of production code gets written. The eval baseline Discover defines becomes the reference point the Prototype validation gate measures against.

Discover Step Detail and Owners

Discover pairs an Synthesised.ai build lead with a named client sponsor and data owner across five steps: map, document, score, scope, baseline. Each step names one owner on each side once scoping starts.

Failure Mode: Scope Without Evidence

Scope agreed from opinion rather than measured volumes drifts the moment build work starts, and a prototype built against a drifting target moves its own eval baseline every week the disagreement continues. The correction anchors every opportunity score to measured workflow volumes before the scope sign-off gate passes, closing the same gap RAND names as its first root cause: misunderstood problem intent.

Stage Two: Prototype

Builds the working system against the scoped opportunity.

Connects real data, not sample data, into the build.

Validates the build against the frozen eval-set.

Reviews results with client stakeholders before the scope locks.

Locks the production scope once validation passes.

In the version 0.1 MIT NANDA report, about 5% of custom enterprise generative-AI tools reached production, and the validation gate exists to help a prototype earn its way into that group rather than celebrate a passing demo. At the decision gates of a stage-gate process, the continuation or termination call gets made by identifying non-dominated candidate solutions under time and budget constraints, the mechanic behind validating before scaling, per Leppinen, Salo and Compare (2026; online 2025), whose study concerns preventive maintenance. A background estimate discussed by Amrit and Narayanappa (2025), attributed to Flaounas, puts prototyping and deployment at up to 90% of machine-learning project time, and freezing the eval-set here fixes the reference point before that expensive stage begins. The frozen eval-set becomes the process’s first telemetry artefact too, versioned next to the code rather than filed in a wiki nobody reads.

Prototype Step Detail and Owners

Prototype runs on the Synthesised.ai build team against a named client data owner and sponsor across five steps: build, connect, validate, review, lock. The validation gate schedules the Productionise decision before the demo happens.

Failure Mode: The Demo That Never Hardens

A prototype praised in the demo room and never funded for production runs for months with no frozen eval-set behind it, and the target drifts every week that continues. The correction schedules the Productionise decision at the validation gate itself, so the next step gets booked, not hoped for, once the prototype passes. With about 5% of custom tools reaching production in the version 0.1 MIT NANDA report, an unhardened demo is a common outcome here.

Stage Three: Productionise

Productionise attaches eval-loops, observability and the governance evidence pack, so the system that ships is the same one whose evidence supports any later internal assurance, audit or regulatory review (phase timings are illustrative and agreed per engagement). Productionise runs on the Synthesised.ai build team with a named client operations owner across five steps.

Instruments telemetry and per-tenant cost attribution.

Attaches eval-loops to every sprint from this point on.

Configures rollback paths and rehearses them before cutover.

Completes the governance evidence pack in-sprint, building on the provenance record initiated in Discover, not after go-live.

Rehearses the cutover against production infrastructure before the real one runs.

MLOps research identifies deployment and monitoring, not modelling, as the binding constraint on production machine learning, and with a commonly cited estimate (discussed by Amrit and Narayanappa (2025), traced to an earlier VentureBeat article) of roughly 13% of business ML projects reaching production, the Productionise discipline is where much of the outcome gets decided. RAND names missing deployment infrastructure as a root cause of AI-project failure, the exact gap the observability layer, the rollback paths and the evidence pack close at this stage. ISO/IEC 42001:2023 sets the international AI management-system requirements the governance evidence pack builds to, the anchor for production-ready delivery beyond the engineering team’s own sign-off. The NIST AI Risk Management Framework’s MEASURE and MANAGE functions describe the test-then-monitor loop the eval-loops and telemetry instrument, and its four functions, GOVERN, MAP, MEASURE and MANAGE, map onto the Production Controls: evaluation, observability and governance. The governance evidence pack is initiated in Discover and completed in-sprint during Productionise, not assembled after go-live, and each framework relationship is stated as builds-to.

Productionise Step Detail and Owners

Productionise runs on the Synthesised.ai build team with a named client operations owner across five steps: instrument, attach, configure, produce, rehearse. The operations owner joins at this stage, so handover becomes a transfer rather than a surprise.

Failure Mode: Late Governance Evidence

Governance evidence assembled after go-live can stall a tender or a review, the exact point a missing artefact gets noticed. The correction initiates governance artefacts in Discover and completes them in-sprint, so the evidence pack already exists the moment procurement or a reviewer asks for it. ISO/IEC 42001 frames the kind of evidence an organisation-level management system retains; producing it in-sprint makes the evidence available for review, though whether any given audit passes depends on the applicable requirements.

Stage Gates and Verification

Five artefacts carry the evidence trail end to end: the opportunity map, the eval report, the cutover checklist, the dashboard set and the handover pack. A stage-gate process makes the continuation decision at each gate an explicit, evidence-based choice rather than a default to proceed, the peer-reviewed basis for a gate that returns a stage instead of skipping it, per Leppinen, Salo and Compare’s 2026 (online 2025) research. Regression-testing each change against a frozen eval-set before it merges is the illustrative engineering practice that stops a passing gate from silently regressing later; the controls in place are agreed for each engagement.

How Verification Failures Are Handled

A failed gate returns the stage, never skips it. The stage reworks against the same named metric and re-verifies before delivery continues, so a missed benchmark costs rework rather than a silent pass into the next stage. The rework path runs four steps: identify the failed metric, return the stage, re-run the gate, and record the re-verification in the evidence pack.

Avoiding Common Failure Modes in Delivery

Gartner’s July 2024 forecast expects at least 30% of generative AI projects to face abandonment after proof of concept by the end of 2025, driven by poor data quality, inadequate risk controls, escalating costs or unclear business value, the four skips this section closes. MIT NANDA’s version 0.1 report estimates about 95% of organisations obtained no measurable P&L return from generative-AI pilots, with the divide sitting at the pilot-to-production boundary these failure modes guard. RAND reports more than 80% of AI projects failing by some estimates, twice the rate of non-AI IT projects, and its five root causes cluster around the ownership, data and infrastructure gaps addressed below.

Human Error: Ownership Gaps

Ownership gaps appear when a gate has no named approver and the sponsor is absent. The mitigation names one owner per gate at contract stage and confirms sponsor availability at scoping, so every verification decision carries an accountable person before delivery starts. The mitigation runs alongside a wider discipline: keeps sponsor and gate-approver ownership assigned through delivery supports the sponsor and gate-approver structure this process assumes, keeping ownership assigned once delivery gets underway.

System Latency: Silent Quality Drift

Silent quality drift appears when model output degrades undetected between releases. The mitigation runs drift detection and eval-set versioning next to the code, so a quality regression surfaces as a failed check rather than a customer complaint.

External Variables: Vendor and Regulatory Shift

External shift appears when a model gets deprecated or a framework updates under a live system. The mitigation arms rollback paths and refreshes the evidence pack on a set cadence, so a vendor or regulatory change becomes a scheduled task rather than an incident. The framework set that refresh checks against stays fixed: ISO/IEC 42001, the NIST AI Risk Management Framework, and Singapore’s IMDA AI Verify governance testing framework and toolkit, alongside Singapore’s PDPA and Australia’s DISR Guidance for AI Adoption named earlier.

Methodology Details

The build runs under a named framework, the Production Path, moving each engagement from Discover to Prototype to Productionise, governed throughout by the Production Controls, covering evaluation, observability and governance. Discover runs 2 to 4 weeks, Prototype runs 4 to 8 weeks and Productionise runs 8 to 16 weeks, with the exact length set by scope at the sign-off gate.

88% of organisations report regular AI use in at least one business function, up from 78% a year earlier, per McKinsey’s State of AI 2025 survey (November 2025); that reflects broad use rather than enterprise-wide scaling — McKinsey notes most organisations had not scaled AI and only about a third had begun scaling — so the differentiator becomes whether delivery follows a repeatable, documented process that actually reaches production. For the full delivery range this process feeds into, spans every AI service from strategy through automation catalogues every engagement Synthesised.ai runs today, from strategy through automation.

Worked Example Slot

A worked stage example publishes here once client evidence lands. No invented case data appears on this page ahead of that confirmation.

Where Your Team Plugs In

The client side supplies four recurring roles across the sequence: a sponsor, a data owner, an operations owner, and a gate approver named at each checkpoint. Scope it now. The shipped system is carried into steady-state operation by the same eval-loops, observability layer and governance evidence pack, rather than dying in a wiki nobody reads; phase timings are illustrative and agreed per engagement.