The AI success formula: what high-performing organisations do differently

Written together with Engage Process, whose process-mapping practice informed the framework set out below.

Artificial intelligence is in many ways still sitting solidly at the top of the hype, and with access to AI by workers increasing by 50% last year, it is also not showing any signs of slowing down. What is changing is the maturity of the technology, and with that we can start assessing what results organisations can potentially achieve.

In a 2026 report by Deloitte, organisations are reporting some strong results, such as a 66% increase in productivity and efficiency, a 53% improvement in decision-making capability, and cost reductions of up to 40%.

However, not every project is successful. Over 88% of organisations are now using AI in one or more business functions, but only 23% say they are scaling it to the wider enterprise, as reported by McKinsey. In another report, Gartner cites that 40% of agentic projects are likely to be cancelled by next year due to escalating costs and unclear returns.

Based on the bright spots it is clear that the technology holds a lot of potential to have a positive impact. So what separates the 23% that scales successfully from the others who are stuck?

01 · What the research tells us about AI success

Studies published by Deloitte, McKinsey, BCG and Gartner provide us with a useful insight into what works well. In the projects we have been involved in directly we see the same patterns. The good news is that most studies and our own observations are consistent, and that gives us a clear view of what factors build success.

Success factor 1: workflow redesign over tool deployment

McKinsey's reports have been very clear about their position: "The greatest value comes from redesigning workflows, not deploying more tools." Essentially, organisations that stick AI onto existing processes without reengineering achieve modest results at best. In contrast, those who redesign work from the ground up before applying new technologies achieve far greater benefits.

Deloitte's research backs this view up. Their study found a direct correlation between the organisations who redesign from the ground up and the results they report, where larger changes directly result in greater benefits. Those who just applied AI to existing approaches reported some benefits, but they remain modest.

Our own experience extends this view. Gysho's earlier internal experiments augmented project management functions with AI bots. This was somewhat useful, but inevitably we still had to do the work ourselves and the cost did not justify the tools. Once we started with a clean sheet and redesigned our project management process completely, we managed to almost double the team's capacity and increased project control across the board.

Success factor 2: process understanding precedes technology choice

Choosing the right technology is as important as thinking about how to use it. Fed by hype, many organisations decided to purchase licences to Copilot, ChatGPT or Claude for their employees. That works for experimentation, but not for sustained results and integration with workflows.

BCG analysed more than 10,000 workers in 11 countries and showed that successful AI deployments did not start with a technology purchase or experiment. They started with getting an understanding of how they work and the friction that exists, and used that as a framework to find what technology was the right tool to solve it. Those purchases translate into investments to solve friction, instead of hopeful experiments.

Success factor 3: clear business outcomes defined upfront

A third factor links to clear business outcomes. Organisations who define the success metrics before purchases consistently report better results than those who do not.

Focusing on the success metrics forces a transparent logic onto any AI decision. The hype is taken out of the conversation and an investment is put into a clear frame of ROI. That in itself leads to a better understanding of what you need, better decisions, and a clearer focus on what the project needs to deliver.

Success factor 4: governance built in, not bolted on

Sustained success relies on governance by design, rather than an afterthought. AI agents process incredible amounts of sensitive data and, in some cases, act autonomously to take decisions that have real world impact. In one example an agent deleted a full customer database, tried to hide it, and eventually apologised but could not recover the information.

Any project that uses AI needs clear guardrails that determine what data it is allowed to use, how much autonomy it gets and when to escalate. Projects that include that governance by design have a far lower risk of failures, and run at greater scale with better outcomes.

In a project in 2025, we built a semi-autonomous system which performed compliance reviews on an airport construction project. Reliability was critical, since the output of this system would determine if the airport was built to the right standards. Any deficiencies could lead to disaster. The guardrails ensured that AI was able to work effectively and loop in human engineers when needed, leading to reliable and repeatable results.

02 · The common thread: process mapping as the foundation

The four success factors share one common pattern. They all require process mapping at some stage, whether it is to understand how we work and where friction exists, which metrics matter, or at which points governance is critical to process reliability.

Redesigning work requires an understanding of what the current flow looks like.
Measuring success needs an understanding of the metrics that matter and how to influence them.
Effective governance is achieved by knowing the risk points and deploying effective controls against them.

That makes process mapping the perfect exercise to apply AI success factors from the start.

While this may all seem intuitive, very few organisations we have worked with prioritise process mapping in their AI projects. They start from a great idea or shiny new technology and figure things out on the fly. Process mapping is seen as a time-consuming and costly exercise which slows down innovation. In reality, the projects where we did the mapping from the start actually reached completion in less time and delivered targeted results.

03 · A process-first framework for AI implementation

Based on the studies we collected and the observations from projects we have completed, we created a framework that ensures success factors are embedded from the start. This framework fits any scenario, whether it is for a purpose-built solution or something bought off the shelf.

Step 1: map the current state

Start with the basics: understand how teams work today. This step is not about just documenting, it is about discovering where teams experience friction or gaps and where the opportunities for AI automation exist.

What effective mapping reveals:

What steps teams truly take, both inside and outside of software, giving a true representation of workflows.
Where decisions are made and what input is needed for them.
Whether those decisions follow the same patterns, or require human judgement.
Which steps are repetitive and simplistic, such as manual data entry.
At which points the process runs into bottlenecks that need problem solving.
Where handoffs between teams create delays.

Published example, UK public sector: adult social care assessment

Bradford Council, Norfolk County Council and West Northamptonshire Council faced a common challenge: their adult social care "front door" processes were overwhelmed. Residents seeking support faced long wait times, while social workers spent excessive time on administrative tasks rather than direct care.

Before evaluating any technology, the three councils mapped their end-to-end assessment workflows together. They discovered that initial information gathering, eligibility screening and routing were highly repetitive and rule-based, while complex case assessment required human judgement.

This process understanding led them to co-design an AI digital assistant (now called AIDA) that handles initial conversations, gathers information, and performs preliminary eligibility checks. The AI routes complex cases to social workers with all relevant information pre-populated.

The outcome: social workers now spend significantly more time on direct client care and complex decision-making. The process redesign, enabled by AI but driven by process understanding, has improved both efficiency and the quality of citizen interactions.

Step 2: design the future state

After the current state is mapped, teams can start redesigning work to solve bottlenecks and find useful cases for AI deployment. The critical part is to start with a redesigned process, before thinking about what technology is needed to serve it.

Quite often, once organisations have mapped a current state, they transition to purchasing a solution. Yet the redesigned flow is the ultimate solution design, dictating exactly what a technology needs to deliver. That, in turn, can even save money by procuring something which does the problem solving really well instead of doing many things at a high cost.

Key questions at this stage:

If we removed this bottleneck, what would the workflow look like?
Which steps could be fully automated without human intervention?
Where could AI augment human capability rather than replace it?
What would "good" look like in terms of speed, accuracy and consistency?
How do we measure success?

Step 3: build the business case

With a redesigned workflow, teams can now determine which benefits they can achieve if implemented. This unlocks a cost/benefit analysis, where an investment in AI can be offset against the expected results once completed. It derisks AI projects and sets a clear objective for execution.

AI business cases are more complex than traditional software purchases:

What are the baseline costs for licensing and platforms?
How many tokens will the solution use, and does this ongoing expense outweigh the current cost?
Does the process change justify completely custom development, or are you served best with an off-the-shelf solution?
Which AI models do you need, and is there potential to leverage cheaper low-end inference?
What does the change journey look like, which partners do you need, and how are you going to upskill teams?
Are there any governance costs you need to include to stay compliant and in control?

The process drives the insight needed to make the end-to-end calculation and ensures the AI project does not simply become a cash burner.

Step 4: choose your technology path

Once the business case is proven, organisations can start looking at the best technology fit, driven by a clear solution design and transparent requirements. At this stage we recommend approaching the selection process with an open mind. There are a myriad of AI solutions, ranging from ready built to completely customised. The best fit is dictated by your process requirements.

Off-the-shelf AI works well when:

Your processes are standard and align with vendor capabilities.
You are willing to adapt workflows to fit the tool.
Speed of deployment is the primary driver.
The problem is well solved by existing products.

Bespoke AI makes sense when:

Your processes are unique or highly regulated.
Off-the-shelf solutions require too many concessions.
Integration with legacy systems is complex.
Long-term flexibility and ownership are priorities.

In recent months the cost of software development has decreased dramatically, changing the balance between choosing to invest in a perfectly fitting bespoke solution, versus trying to adapt to a rigid off-the-shelf package. Additionally, in line with the benefits of deep integration and full redesign, AI shows greater results when it is perfectly tuned to your internal process.

Step 5: iterate and measure

When the right technology has been selected, it is key to get started. Rapid prototyping works well in AI projects since it allows teams to test their process logic, review AI's outputs and iterate until the system works as expected.

This iterative "build early, fail fast" approach benefits the typical AI project today. We see many organisations get started with AI and gain new insights as they start building and prototyping with the technology. Those new insights change the final design over time, and the rigidity of waterfall methods is at odds with those insights.

As teams progress to delivering a production solution, they should ensure the business metrics defined in process design are measured from beginning to end. They will not only show you if the project is delivering on expectations, they are also a critical safeguard against excessive token costs versus output.

Note that results are never measured by number of executions or tokens. They are measured in the organisation where the output is received.

04 · Published examples from UK local government

The public sector faces unique challenges in AI adoption: complex legacy systems, stringent compliance requirements, diverse stakeholder needs, and the imperative to serve all citizens equitably. Yet councils that start with process understanding are achieving remarkable results.

The three examples below are drawn from the Local Government Association's published case study bank. They are not Gysho projects, and we were not involved in them. We include them because each council started with process understanding rather than a technology purchase.

Technology Enabled Care, St Helens Borough Council

St Helens created an in-house Technology Enabled Care (TEC) hub using Microsoft Copilot Studio to help match residents with the most suitable care technology, automatically drafting personalised care plans to support discharge and reablement.

Before implementing the technology, the council mapped their care assessment and discharge workflows. They identified that matching residents to appropriate technology and drafting care plans were highly repetitive tasks that consumed significant social worker time.

The outcome: the team identified clear, cashable savings from the use of automated medication dispensers (around £160,000 saved for every 20 residents) and reinvested these savings to expand the council's AI capability. The next phase connects this work to a resident-facing digital front door, making TEC easier for residents to discover and access.

AI-powered case notes, Barking and Dagenham Council

Barking and Dagenham Council developed an in-house AI tool called BD Notes to help draft statutory case documents. Before building the tool, they mapped the document creation workflow for Education, Health and Care Plans (EHCPs), identifying which sections were formulaic and which required professional judgement.

The outcome: the time needed to draft sections of EHCPs fell from four to five hours to under five minutes, while officers remain fully responsible for reviewing and approving the final content. The council is now embedding BD Notes across 13 service areas by 2026.

Predictive homelessness prevention, Kent County Council

Kent County Council faced a reactive homelessness system: families often reached crisis point before receiving support, making intervention more difficult and costly.

The council mapped their homelessness prevention workflow and identified that early warning signals existed across multiple systems (housing benefits, social care, health services) but were never connected. The process problem was not a lack of data, it was a lack of integration and proactive triggering.

Using predictive analytics, Kent built a system that identifies households at risk of homelessness before crisis occurs, automatically triggering early intervention protocols. The AI model was trained on historical data, but the real innovation was redesigning the prevention process to act on predictions.

The outcome: families receive support earlier, preventing homelessness rather than responding to it. The council has reduced emergency housing costs while improving outcomes for vulnerable residents.

05 · Pitfalls to avoid

While there is clear upside to a best-practice AI deployment, the studies we reference in this article have also shown some pitfalls to avoid. We have listed the most common ones to look out for.

Skipping the process map and creating a baseline. The most often skipped and underestimated step is creating a process map. Without understanding your current workflow, you cannot know where AI will add value, what success looks like, or how to measure it. Process mapping is not optional: it drives your solution design and investment decisions.
Chasing innovation over practicality. More expensive models are not always better, and complexity does not equal usefulness. Focus on simple, practical solutions tied to business outcomes. Start with the problem, not the technology. Apply programmatic automation before AI, and use the simplest model feasible for your solution.
Treating governance as an afterthought. AI carries real risk depending on what data you use and where it is sent. If governance is not accounted for early, it will be harder to integrate correctly, becoming a liability or deployment blocker. Build compliance, security and ethical considerations into the process design from day one.
Confusing activity with outcomes. Many software solutions offer usage statistics out of the box, but aside from telling whether the solution is used, they do not answer the question of whether it is getting results too. Process metrics such as lead time, total cost or waiting times may not naturally sit inside the AI itself and exist outside of the software, measured in longer periods and within the documented processes.
Running pilots without production paths. Experiments are small-scale versions of a design that is made to fit production settings. Anything that does not start with that in mind is unlikely to withstand the proof-of-concept test and produce a viable production application.

Our vision

As AI comes down from the hype and becomes accepted technology, the organisations that focus on thoughtful design and deployment are the ones that will see the greatest success. As the studies and our own experiments show, the AI paradox is that its success starts like any other software project: map how you work now, redesign services for how you want to work in the future, then choose the technology to fit.

References

Deloitte (2026) State of AI in the Enterprise. Available at: deloitte.com.
McKinsey & Company (2025) The State of AI in 2025: Economic Potential and Beyond. Available at: mckinsey.com.
Boston Consulting Group (2025) AI at Work 2025: Momentum Builds, but Gaps Remain. Third edition, June. Survey of more than 10,600 leaders, managers and frontline employees across 11 countries. Available at: bcg.com.
Gartner (2025) 'Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027', Gartner Newsroom, 25 June. Available at: gartner.com.
Local Government Association (2025-2026) Artificial Intelligence Case Study Bank. Available at: local.gov.uk. Primary source for the Bradford, Norfolk, West Northamptonshire, St Helens, Barking and Dagenham, and Kent examples.

Council project names (AIDA, BD Notes, TEC Hub) are as reported in the LGA case studies and may be updated by the councils over time.