Deploying AI is the easy part: who runs it once it goes live?

Friday, 4pm

Picture an AI system your organisation switched on earlier in the year. It drafts replies to customers, triages requests or summarises cases for staff. On a Friday afternoon it starts getting things wrong.

The model came from one supplier and the platform from another. The data sits with a third team. The project that built it closed months ago and its people have moved on. Everyone owns a piece. Nobody owns the whole.

Two pieces of research published on 12 November 2026 suggest this is not an unusual picture. It may be one of the main reasons AI programmes fall short of what was promised.

What the research found

HFS Research, working in collaboration with Rackspace, surveyed 502 senior AI decision makers between August and September 2026. All worked at enterprises with at least $100 million in annual revenue, and all were already running AI pilots or more advanced programmes. Rackspace sells managed AI services, and the report recommends an accountable operator of the kind it offers, so treat this as vendor research. The findings describe organisations already active with AI, not the market as a whole.

The headline numbers:

• 33% have met or exceeded the business case for their most significant AI programme, falling to 28% where AI runs across several functions.
• 96% report significant governance gaps, and 78% see governance failures as a bigger threat to scaling AI than technology failures.
• 44% have no single, consistent approach to running AI in production.
• Gains so far are mostly about efficiency: 58% cite workforce productivity and 42% cost reductions, against 28% reporting better customer experience and 21% revenue growth.

Phil Fersht, founder and chief executive of HFS Research, said: "Enterprise AI has an accountability problem that nobody in the industry wants to admit."

Genpact's Autonomy in Practice report, released the same day, surveyed more than 2,000 executives at companies with revenues of $1 billion or more. Genpact also sells AI operations services, and its findings are self-reported. It found 79% have AI agents in production, but only 20% use them broadly and consistently across functions. Only 15% have fully tied AI investment to profit and loss outcomes, and only 9% have completed significant operating model change. Its chief technology and innovation officer, Sanjeev Vohra, said: "Enterprises have become effective at deploying AI agents. Running them is the hard part."

What it means for leaders

Both reports describe the same gap. Organisations have become good at starting AI and much less good at running it.

That will sound familiar to anyone who has worked in service management. It is the moment a project hands over to operations and discovers that nobody agreed who would support the thing, how it would be measured, or what happens when it fails. It is a question of ownership and operating discipline far more than a question of models.

John Sansbury makes the point about adopting service management in The Journey to Service Excellence:

"The adoption of good service management practice has no end, so treating it as a project may create a false impression within your enterprise and an expectation that resources may at some point be released and repurposed."

Applying that principle to AI, the lesson is plain. A live AI system is a service. It needs people and budget for as long as it runs, not just until launch. John is equally firm about ownership: "Regardless of the size of your enterprise, there can only be one owner for each process you choose to formally adopt." The same logic holds when part of the work is done by a machine.

Practical steps

1. Plan for the day after go-live. Before approving any AI deployment, ask who will run it, with what budget and for how long. If the answer is the project team, ask what happens when the project closes.
2. Give each live AI service one owner. One person, senior enough to act across suppliers and internal teams, who answers for whether it delivers. A committee can advise. It cannot own.
3. Define working in business terms. Uptime and usage are outputs. Agree the outcome that should change, such as faster resolution for customers or fewer errors in casework, and check it regularly.
4. Treat AI failures as incidents. Decide in advance who gets called, who can pause the system, how customers are looked after in the meantime, and how lessons feed back into improvement.
5. Bring suppliers under one remit. Several suppliers is normal. Several suppliers who each assume someone else is responsible is the problem the research describes.

Where training fits

Most of these steps need no new technology. They need leaders and managers who understand how services are owned, run and improved, and who can apply that thinking to AI. That is a capability gap, and training and enablement can close it: for executives deciding what to fund, for the owners who will answer for each service, and for the operational teams who keep AI services working day to day.

Honest limits

Both surveys come from firms that sell AI operations services, and both rely on what executives say about their own organisations. Neither shows that clearer ownership causes better returns; HFS itself notes that its figures describe different groups and do not establish cause. Ownership will not fix poor data, weak security or a use case that was never worth doing. Those need technical work. Where AI makes decisions about people or customers, take proper legal advice on your obligations rather than relying on a general article.

What the research does show is a pattern worth checking in your own organisation. If you cannot say who runs each AI system on a Friday afternoon, that is the place to start.

Tags: ai governance, ai adoption, service management, leadership, accountability, operating model

Sources

• https://www.globenewswire.com/news-release/2026/11/12/3401574/0/en/new-hfs-research-and-rackspace-study-finds-only-one-third-of-enterprises-realize-expected-ai-value.html

• https://media.genpact.com/2026-11-12-79-of-Enterprises-Report-AI-Agents-in-Production,-but-Only-15-Have-Fully-Tied-AI-Investment-to-P-L-Outcomes