GUIDE: HOW TO IMPLEMENT AI IN BUSINESSHow to start with AI in your business without pilots that go nowhere: a stage-by-stage guide

Why do so many businesses struggle to get started with AI?

The problem is rarely access to the technology. It is turning use into value.

• McKinsey’s 2026 global survey of 1,719 respondents found nearly nine in ten report regular use of AI in at least one business function, but only 37 per cent attribute any EBIT (earnings before interest and tax) impact to it, about the same as a year earlier. [1]
• In Gartner’s survey of 2,002 CIOs and technology executives, only 13 per cent say their enterprise has realised significant or transformational value from AI, and “the likelihood of use case success remains one-in-five.” [2]
• Among Australian small and medium businesses, 43 per cent reported some level of AI adoption across December 2025 to February 2026, and 19 per cent said they simply don’t know how to use AI in their business. [3]

The organisations that do get value behave differently. McKinsey’s high performers are twice as likely to say senior leaders show commitment and that they have defined processes to measure impact, and nearly three-quarters fundamentally redesign workflows, against one-quarter of everyone else. [1] BCG’s 2026 study of 1,330 senior leaders found leaders “run AI as an enterprise-wide program”, start with clarity on where AI creates value, and measure impact in the P&L. [4]

That is why this guide is organised by stage, not by tool.

What are the stages of implementing AI in a business?

Durations are indicative for a mid-sized enterprise with one priority use case; regulated sectors and slow data access add time.

Where are you now?

0Get leadership aligned1Find where AI pays off2Get your data and people ready3Run one pilot that proves ROI4Put it into production safely5Scale and keep improving

Next use case

Stage 0 Get leadership aligned

Goal

Agree why you are using AI and what success means

Owner

CEO or executive sponsor

Typical duration

1 to 2 weeks

Deliverable

A one-page AI charter

Gate: Are we committed, with a sponsor, budget and two or three outcomes?

Go to stage 0Example: ISB

Get your AI roadmap

quantana.com.au

Stage navigator: six stages of implementing AI, from leadership alignment to scale, with a decision gate between each. Choose a stage to see its owner, typical duration, deliverable and exit decision.

Stage	Owner	Typical duration	Deliverable	Exit decision
0. Get leadership aligned	CEO or executive sponsor	1 to 2 weeks	A one-page AI charter	Are we committed, with a sponsor, budget and two or three outcomes?
1. Find where AI pays off	Executive sponsor, with business unit heads and finance	2 to 4 weeks	A ranked shortlist and one business case	Which one use case goes first, and what is its stop rule?
2. Get your data and people ready	Data lead and HR or capability lead	4 to 10 weeks (overlaps stage 3)	Clean data, an evaluation set, trained users	Is the data fit for purpose, and are the users ready?
3. Run one pilot that proves ROI	Business owner of the process	8 to 12 weeks	Results against baseline and a go, change or stop call	Go to production, change scope, or stop?
4. Put it into production safely	Business owner, with risk and technology	4 to 8 weeks	A governed production release with a named owner	Has risk signed off, and is there a fallback?
5. Scale and keep improving	Executive sponsor and finance	Ongoing, reviewed quarterly	A live AI portfolio and the next wave	Keep, improve or retire each AI system?

Stage 0: How do you get leadership aligned on AI?

AI starts as a business decision, not an IT project.

In Gartner’s April 2026 survey of 782 infrastructure and operations leaders, those with at least one AI success most often credited integrating AI into existing workflows and securing full support from business executives. [5]

What to do

• Write down two or three outcomes AI should move this year, in P&L terms: cost per transaction, cycle time, revenue per customer, error rate.
• Name an executive sponsor with budget, and a senior owner for AI governance. The National AI Centre lists that owner and an AI policy as the first steps for any organisation. [6]
• Agree an interim AI policy: approved tools, data that must never go into public tools, and who to ask.
• Run a short leadership session so the executive team shares a realistic view of what AI can and cannot do.
• Decide what you will not use AI for, and how much risk you accept.

Mistakes to avoid: starting with a tool purchase; handing AI wholly to IT; goals like “use more AI” that nobody can measure.

Deliverable: a one-page AI charter with outcomes, sponsor, governance owner, budget envelope and interim policy.

Decision before moving on: are the executives committed to a small number of outcomes, with a named sponsor and budget? If not, stop here. Nothing later will survive without it.

How it looked in practice

For the Indian School of Business, Quantana started with design sessions that captured what the leadership team wanted from a next-generation online learning experience, then delivered it in phases. ISB’s Director, Praveen Garimella, said: “Whatever was on our mind was very nicely and rapidly captured through the design sessions we have had.” AI-generated content later cut course development from months to minutes.

Stage 1: How do you find where AI will pay off?

Look for value in two places.

Cost opportunities, where AI removes manual effort, delay or error. Growth potential, where AI helps you win, keep or serve more customers. The best first candidates usually do both.

McKinsey’s respondents most often report cost reductions from AI in supply chain management, service operations and manufacturing, and revenue gains in marketing and sales, product development and software engineering. [1]

What to do

• List high-volume, repetitive processes and the people who run them: intake, triage, document review, reporting, customer queries, proposals.
• Record a baseline for each: volume, time per item, cost, error or rework rate, and revenue at stake.
• Sort candidates into cost, growth or both. Ask where delays cost you sales, not just staff hours.
• Score the shortlist on value, data readiness, feasibility, time to value and risk. Our AI roadmap guide has a weighted matrix you can copy, and how we work shows how to size each one in dollars.
• Choose one first use case, and write its target and stop rule before anyone builds anything.

Mistakes to avoid: picking the most impressive idea rather than the most measurable; skipping the baseline; starting five pilots at once. Gartner found many failed AI initiatives “expected too much, too fast”. [5]

Deliverable: a ranked shortlist with baselines, and a one-page business case for the first use case, including a written stop rule.

Decision before moving on: which single use case goes first, who owns it in the business, and what result would make you stop? Organisations with rules for when a pilot should not move to production are 2.3 times more likely to be top performers, yet only 31 per cent of CIOs have them. [2]

Not sure where to start? AI Compass does stages 0 and 1 with you and gives you your AI roadmap in 14 days, for one fixed fee, never per user: From A$15,000 + GST. Get your AI roadmap.

How it looked in practice

Charantra’s deal teams relied on email, spreadsheets and relationship-dependent buyer tracking, which caused stalled deals, poor pipeline visibility and revenue leakage. That is both a cost problem and a growth problem. Quantana centralised deal operations on one AI-enabled platform, with document OCR (optical character recognition) and AI synthesis that compresses days of contract reading into minutes, semantic buyer matching with explainable fit scores, and alerts on stalled deals.

Stage 2: How do you get your data and people ready for AI?

Fix only the data and skills your first use case needs. Cleaning everything first is how AI efforts lose a year.

Data. In Gartner’s April 2026 survey, 38 per cent of leaders who faced setbacks said poor data quality or limited availability directly caused AI project failure. [5] Check that the data for your first use case is accurate, complete, owned, reachable through a governed interface (not an export), and classified so you know whether it may go to a hosted model.

If it lives in an ERP system, plan a secure, read-only connection rather than bulk extracts. Our QSAP connects AI assistants to SAP S/4HANA with each user’s SAP authorisations respected and every request on record. Our AI in SAP guide covers the options.

People. The same Gartner survey found 38 per cent blamed persistent skills gaps. [5] BCG’s 2026 AI at Work survey of 11,749 workers found 74 per cent of frontline employees now use AI regularly, but 66 per cent of those saving significant time get limited or no guidance on how to use it. [7]

What to do

• Map where the first use case’s data lives, who owns it and how good it is.
• Fix access and quality for that data only; classify it by sensitivity.
• Build an evaluation set of real examples and edge cases now, before the solution.
• Train the people who will use and oversee the system on real tasks from their own work.
• Tell staff what is changing and why; involve the people who do the work in redesigning it.

Mistakes to avoid: a data lake project with no use case; training that stops at a generic webinar; leaving HR out until launch. Our article on what CHROs ask about AI covers the people questions that come up most. AI Academy runs C-suite masterclasses, hands-on team bootcamps, prompt engineering workshops and responsible AI playbooks.

Deliverable: data that passes agreed quality checks, an evaluation set, and a trained pilot user group.

Decision before moving on: is the data good enough to test the use case honestly, and do the pilot users know how to work with it?

How it looked in practice

A global enterprise with testing laboratories across several regions held information on lab health, performance and utilisation in manual spreadsheets. Quantana is designing and building a centralised platform that puts structured daily data entry, dashboards and automated escalations in place first, with single sign-on and role-based access, and an AI insights layer for downtime analysis and trend forecasting on top. The work is in progress.

Stage 3: How do you run an AI pilot that proves ROI?

A pilot is an experiment with a result, not a demo.

Run it inside the real workflow, with real users and real data, and measure it against the baseline you set in stage 1. Our article on starting with one factory shows how one manufacturer chose a single-site, single-metric pilot designed to prove ROI in about 90 days.

Expect variation. Gartner found only 28 per cent of AI use cases in infrastructure and operations fully met ROI expectations, and 20 per cent failed outright, largely because they were over-ambitious or poorly scoped. [5]

What to do

• Fix scope, users, duration, success metric and stop rule in writing before the build starts.
• Redesign the workflow around the AI where it helps, not just bolt it on. In BCG’s survey, people at companies pursuing workflow redesign were 24 percentage points more likely to see measurable business improvement. [7]
• Keep a person reviewing outputs during the pilot, and log every override.
• Measure weekly against the baseline: time, cost, quality, adoption and running cost per task.
• Hold a formal review at the end, with finance in the room.

Pilot ROI and stop rule

Monthly saving at target$0

Months to recover pilot0

Running cost share9%

Pilot result: actual time reduction 30%

StopChangeGo

Change scope. Above the stop rule but short of target: change scope once. Net saving at this result: $11,600 a month.

At 30 per cent actual reduction: Change scope. Net monthly saving $11,600. At target: $16,000 a month, pilot cost recovered in about 2.5 months.

Your stop ruleIf handling time falls by less than 20 per cent, or accuracy on the evaluation set is below 90 per cent after 8 weeks, we stop.

Items a month 4,000Minutes per item today 12Loaded cost per hour $55AI running cost per item $0.40Pilot cost $40,000Target time reduction 40%Stop below 20%Minimum accuracy 90%

Read how one manufacturer scoped a 90-day pilot

quantana.com.au

Pilot builder: set monthly volume, minutes per item, hourly cost, AI running cost, pilot cost, target reduction, stop threshold and minimum accuracy. Tiles show the monthly saving at target, months to recover the pilot cost and running cost as a share of saving. A result slider shows whether a pilot result means stop, change scope or go.

Measure	Value
Monthly saving at target	$16,000
Months to recover pilot cost	2.5
Running cost as share of saving	9%
Stop rule	If handling time falls by less than 20 per cent, or accuracy on the evaluation set is below 90 per cent after 8 weeks, we stop.

Mistakes to avoid: a sandbox that never touches real work; moving the goalposts mid-pilot; counting time saved that is not redeployed to something measurable; letting a pilot run for a year without a decision.

Deliverable: pilot results against baseline, evaluation scores, user feedback, and a recommendation: go, change or stop.

Decision before moving on: did the pilot meet its target? If yes, plan production. If close, change scope once. If not, stop and pick the next use case from your shortlist.

How it looked in practice

Greenpark Hotels faced an overload of daily emails that delayed responses and affected revenue. Quantana built an AI email agent for one well-defined process: it sorts messages into bookings, escalations, FAQs and priority queries, drafts brand-consistent responses, and integrates with the CRM and property management systems. Audit trails show response times, conversion rates and SLA performance, so the result can be measured. The outcomes: faster responses, increased conversions and no more manual email sorting.

Stage 4: How do you put AI into production safely?

Production needs what a pilot can skip: monitoring, logging, access control, a human override, a fallback and an accountable owner.

In Australia, the National AI Centre’s Guidance for AI Adoption sets out six essential practices: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control. [6] Its getting-started actions include an AI register, risk screening, testing before deployment, monitoring after it, and human override points where people can “pause, override, roll back or shut down” AI systems. [6]

Choose where each workload runs by the sensitivity of its data. Public or low-risk work can use hosted models; confidential or regulated data may need a private cloud tenancy or on-premises deployment. Our private AI guide compares the options and maps APRA’s CPS 230 and CPS 234 to an LLM deployment.

APRA’s April 2026 letter expects regulated entities to keep “an inventory of AI tooling and AI use cases” and “human involvement for high-risk decisions”. [8] And from 10 December 2026, organisations covered by the Privacy Act must explain in their privacy policy when a computer program makes, or substantially and directly helps make, decisions that could significantly affect individuals. [9]

What to do

• Add the system to your AI register, with owner, purpose, data classes, model and risk tier.
• Choose the deployment pattern (hosted, private cloud or on-premises) by data class.
• Set access by role, log prompts, sources, outputs and actions, and define what counts as an AI incident.
• Re-run evaluations before go-live and on every model, prompt or data change.
• Document the fallback process and test the override.
• Tell customers and staff where they are dealing with AI.

Mistakes to avoid: treating governance as a sign-off at the end; production with no monitoring; one person’s login reused by an AI agent; no plan for when the model provider changes or fails.

Deliverable: a production release with monitoring, logging, human override, fallback, risk sign-off and a named owner.

Decision before moving on: has risk signed off at the right tier, and can the business keep running if the AI system is switched off?

General information, not legal advice. Regulatory position as at 8 November 2026, with links to the primary sources.

How it looked in practice

CasX.ai is an autonomous pharmacovigilance platform that Quantana engineered with dual AI agents. One handles case intake and processing; a quality control agent checks regulatory compliance before anything is submitted, with full audit trails and 100 per cent audit traceability. A related design choice: for TracQ’s acquisition workflow, Quantana kept financial calculations deterministic and separate from AI reasoning, so ratios and valuations stay exact.

Stage 5: How do you scale AI and keep improving it?

Scaling is a portfolio discipline.

Each quarter, review every AI system on measured value, running cost and risk, then keep, improve or retire it, and pull the next use case from your shortlist. Organisations with guidelines for abandoning low-value AI already in production are 2.6 times more likely to be top performers, yet only 17 per cent of CIOs have them. [2]

Watch running costs from day one. About one in five organisations already limit AI use because of operating costs, including tokens. [1] Our guide to reducing AI costs covers model choice, caching, routing and when self-hosting breaks even.

Plan the controls before you hand AI more authority. BCG found 42 per cent of companies expect to give AI agents genuine decision-making autonomy by 2030, but only 5 per cent have the full set of controls in place today. [4] And redesign roles, not just tasks: Gartner found organisations with explicit plans to redesign roles for a human-agent workforce are 2.4 times more likely to be top performers, yet only 4 per cent of CIOs have such plans. [2]

What to do

• Track value at three levels: use case (time, cost, quality), adoption (active users, override rate) and enterprise (EBIT contribution, revenue).
• Report cost per completed outcome, not cost per token.
• Reuse what you built: data connections, evaluation sets, governance patterns and training.
• Run a quarterly keep, improve or retire review with the sponsor and finance.
• Refresh training as tools and roles change.

Mistakes to avoid: scaling a pilot that never met its target; letting each team buy its own tools with no register; ignoring model updates that change behaviour; measuring activity instead of outcomes.

Deliverable: a live AI portfolio with value and cost per system, and the next wave of use cases.

Decision at every review: keep, improve or retire each system, and which use case comes next.

How it looked in practice

For a US insurance services firm, Quantana’s work began with an AI-powered commercial submission and quoting platform, contracted in July 2026 for a four-month first phase followed by a roadmap of further phases. A second, connected platform, a product configuration and underwriting workbench, followed in September 2026. Both run under a governed model with versioning, review comments, exception tracking and an auditable history of every change. The engagement is in progress.

How long does it take to start with AI, and what does it cost?

Stages 0 and 1 usually take three to six weeks in-house; AI Compass covers them in 14 days. A first use case typically reaches production three to six months after that, depending on data readiness and approvals.

Budget in three parts: the discovery work (mostly people time), foundations (data, integration and governance, often the largest item) and running costs. BCG found corporate AI spending reached 3.3 per cent of revenue in 2026, with more than 80 per cent of it sitting outside the enterprise IT budget, so make sure finance can see all of it. [4]

To put a dollar figure on a single use case before you commit, try the value sizer.