How do you scale AI beyond the pilot?

How do you scale AI beyond the pilot and into the way a business actually runs? A Microsoft post published on 12 November gives some useful clues. One of their examples is Northern Trust, which used to spend about two and a half weeks preparing data each time it onboarded a client onto one of its platforms. An AI agent has cut that to three days. What I like in this case is where the agent lives, inside Excel and SharePoint, the tools the analysts already used. It’s one way to do it, not the only way, and I’ll come back to that.
The post comes from Rodrigo Kede Lima, President of Microsoft Frontier Company. Microsoft launched that business in July, with a stated investment of $2.5 billion and 6,000 industry and engineering experts who work inside customers’ organisations to build and improve AI systems. He says it has run nearly 500 projects with more than 200 customers across 28 countries. It’s a vendor post, so it deserves a sceptical read, but the question underneath it is one every leader will recognise: how do we move from AI that impresses to AI that changes how the business actually runs?
Why it matters now
Microsoft published research from IDC alongside it, based on more than 5,200 senior IT and business leaders across 28 industries. It found that 66% report substantial or transformative results from AI, but only 48% are scaling that value across multiple business units. That suggests a lot of businesses have good results in pockets that haven’t yet spread. IDC also found that 71% expect AI spending to rise by at least 10% over the next year, so the pressure to show a return is only going up.
Two caveats are worth holding on to. Microsoft commissioned the research, and the returns are reported by the leaders themselves. Microsoft’s own write-up is clear that the findings show association rather than cause.
It’s rarely the technology
Rodrigo’s argument is that the technology is ready, and the hard part is everything around it: working out where AI adds most value, building in trust and governance, redesigning how teams operate and helping people work differently. It’s close to something I say in talks, that AI adoption is about habits, workflows, governance and culture, not just tools. It’s a vendor’s view, so test it against your own experience. Mine, from working with mid-sized businesses at ramsac, is that it holds up.
Four habits worth borrowing
Microsoft says the same pattern shows up across its customers, and none of it needs 6,000 engineers.
Start with the value, not the technology. Northern Trust can say what changed in days. KT, a South Korean telecoms company, says its network agent now spots problems in two to three minutes rather than more than ten. Both began with something they could measure.
Redesign the process, not just the task. Dropping AI into one step of a clumsy process mostly gives you a faster clumsy process. 3M’s work on customer order management is a better illustration, with an agent handling credit checks, delinquency assessments and system updates, people approving along the way and a dashboard showing what’s happening. Microsoft describes the benefits there as expected, though, so that one is still a plan rather than a result.
Work in short cycles with your own people. Microsoft’s closing point is that the gap between impressive and useful closes one workflow at a time, with people who know the business sitting alongside people who can build.
Put security, governance and cost control in from the start, so you don’t have to retrofit them once something is working. My own rule of thumb is to govern what matters and not slow what works, and it sits comfortably alongside the GRAICE framework from the Global Council for Responsible AI, which is built around governance that's practical and measurable and supports innovation instead of stifling it. A one-page set of rules covering what data AI can touch, who approves what and who owns it beats a fifty-page policy nobody reads. IDC found that 87% of its top group, which it calls Frontier Firms and defines as the top 10% on its AI maturity scale, have a process for governing agents, against 51% of the slowest adopters.
An agent is one option, not the starting point
Every example in Microsoft’s post is an agent, which makes sense given what Microsoft is selling, but the habits above don’t depend on one. For one workflow the right answer might be an agent. For another it might be Microsoft 365 Copilot used well with proper training, a simple automation, or tidying up your data and permissions first so that anything you build later can be trusted. Choosing the tool comes after you’ve chosen the workflow and the measure, not before.
The gaps nobody budgets for
Two numbers stood out for me. Only 20% of organisations have a comprehensive, organisation-wide plan for building AI skills, even though 67% say their workforce already reports high trust in AI. Among Frontier Firms, 80% have such a plan, against 3% of the slowest adopters. The second gap is measurement, with 61% of organisations highly confident they can measure AI’s business value, compared with 93% of Frontier Firms.
The skills point isn’t only about teaching people which buttons to press. KOHLER’s Chief Digital Officer made the case that the skilling companies tend to miss is business skilling, learning how to break a process down and turn it into an AI workflow. That’s the muscle that decides whether a pilot spreads.
One more number needs care. 90% of Frontier Firms spent more on AI than they planned. I wouldn’t read that as advice to overspend. It may simply mean that firms which learn quickly find more worth doing than they budgeted for, and the research can’t tell us which way round it works.
What to do next
Pick one workflow, ideally one that’s slow, repetitive and owned by a team that wants it fixed. Write down what it costs today in time or money, because without that baseline you can’t prove anything later. Give it a named owner, let a small team try the right approach, whether that’s an agent, Copilot with proper training or a simple automation, with a person approving the outcomes, and review it after 90 days. If it works, you have a template for the next workflow and evidence for the board. If it doesn’t, you’ve found out cheaply.
For your IT team
If your first answer involves an agent, it’s only as safe as the access it’s given, so give it its own identity and the minimum permissions it needs rather than letting it borrow a person’s account. Decide where it runs, which data it can read and change, and log what it does. Northern Trust built its agent inside Excel and SharePoint, so permissions and sharing settings in those places matter. Add a person’s approval before it changes records or moves money, along the lines of the approvals and monitoring dashboard 3M describes, record a baseline measure before go live, and set usage and spend alerts so costs don’t surprise you. IDC found that 65% of organisations have a process to govern agents but only 60% are highly confident about AI security and governance risks, so having a process isn’t the same as having control.
If you’d like a quick way to see whether your business is helping or hindering AI adoption, the AI enabler or blocker tool on this site is a good place to start: thoughtprovoked.co.uk/guides/ai-enabler-or-blocker. And if you’d like a second pair of eyes on your first or next workflow, that’s the kind of work our Ironbridge team does with clients at ramsac, from choosing where to start through to governance and measuring the result.
So if your pilot is impressive but nothing about how a team works on a normal Tuesday has changed, which workflow would you redesign first?
Further reading
-
From AI ambition to business impact, The Official Microsoft Blog: the post that prompted this piece, with the customer examples in full. https://blogs.microsoft.com/blog/2026/11/12/from-ai-ambition-to-business-impact/
-
Scaling AI value without losing control: Lessons from Frontier Firms, The Official Microsoft Blog: the IDC findings, including the Frontier Firm comparisons. https://blogs.microsoft.com/blog/2026/11/12/scaling-ai-value-without-losing-control-lessons-from-frontier-firms/
-
Microsoft Frontier Company: AI engineering that amplifies and protects your intelligence, The Official Microsoft Blog: explains what Frontier Company is and what Microsoft says it will do. https://blogs.microsoft.com/blog/2026/07/02/microsoft-frontier-company-ai-engineering-that-amplifies-and-protects-your-intelligence/
Frequently asked questions
Why do most businesses struggle to scale AI beyond initial pilots?
While 66 percent of organisations report transformative results from AI, research shows only 48 percent scale value across multiple business units. The technology itself is ready, but the challenge lies in redesigning operational workflows, adapting habits, establishing clear governance, and addressing the widespread failure to train staff in breaking business processes down into AI-driven workflows.
What practical steps should a business take to scale an AI workflow?
Organisations should pick a single slow, repetitive workflow owned by an engaged team and record its baseline costs. Rather than dropping AI into a clumsy task, redesign the wider process. Assign a named owner, run a 90-day trial using an appropriate tool with human approvals, and introduce simple governance from day one to maintain security and control costs.
Are dedicated AI agents always necessary to achieve business value?
No, AI agents are just one option rather than an automatic starting point. While agents excel at handling multi-step tasks across systems, simpler solutions like Microsoft 365 Copilot with proper training or straightforward automations may suit better. Organisations must define the specific workflow problem, required measurements, and data permissions before choosing which technology fits best.
What unexpected gaps do companies fail to budget for when expanding AI?
Leaders frequently overlook budgeting for measurement and business skilling. Only 20 percent of organisations have an organisation-wide plan for AI skills, particularly teaching employees how to break down processes into AI workflows. Additionally, many companies fail to establish baseline costs in time or money before starting, making it difficult to demonstrate genuine business returns later.
How should IT teams safely govern and secure AI agents?
IT teams should give every agent its own separate identity and minimum necessary permissions rather than using a staff member's credentials. They must define what data the agent can access, log activities, and require human approval before records change or funds move. Setting cost and usage alerts is also essential to prevent spending surprises as adoption grows.
You may also be interested in

Is Claude in Copilot compatible with UK GDPR?
A reported 9 December change could switch Claude on by default for UK Copilot tenants. Here’s what UK GDPR actually asks, what Microsoft’s EU Data Boundary move does and doesn’t mean, and an action list for leaders.

Your SharePoint agents have an expiry date. Here’s what to do about it
SharePoint agents built as .agent files stop working after March 2027, and Microsoft isn’t upgrading them for you. Here’s how to decide what each one becomes before the deadline arrives.
Never miss an article
Get new articles by email
Whenever I publish something new on AI, cybersecurity and cyber resilience, I'll send you a link. No newsletters, no selling, and one click to stop at any time.
Your address is used only to send you new articles. See the privacy notice.
