Consider an enterprise where the introduction of AI has gone well. The investment was justified, the technology selected and deployed, and the necessary governance, security and assurance arrangements established. Safety engineering is treated as an ongoing discipline rather than a certification obtained once and filed away. Employees have been trained, business functions are using the systems, and the promised benefits are beginning to materialise. The CIO and the leadership team have good reason to regard the programme as a success.
What, then, is unfinished?
Not simply the technology, which will keep improving, or the list of applications still to deploy. What is unfinished is the enterprise's understanding of what it has introduced. AI can change not only how existing work is performed, but which work is worth doing, how decisions are formed, what capabilities people need, and which services or business models become viable. And the enterprise is not making these discoveries in isolation: customers, competitors, regulators, employees and technology providers are revising their own expectations at the same time.
Enterprises cannot suspend their investments, contracts, workforce plans or strategies while these questions are being resolved. They must continue operating, making commitments and delivering results. A successful AI programme demonstrates that an organisation has achieved what it set out to do under the conditions in which it was implemented. It does not establish that the assumptions behind its operating model, division of responsibilities or future strategy will remain appropriate. Indeed, the experience of using AI may reveal possibilities that were never part of the original plan.
Commitments outlive their assumptions
Experienced executives hardly need reminding that enterprise transformation rarely moves in a tidy line from strategy to implementation to benefits. Technology moves on, people find their own ways of working, and the assumptions behind a business case seldom survive unchanged. Organisations have built considerable expertise in navigating this through ERP programmes, outsourcing, cloud adoption and the disruption of the pandemic. Those lessons remain valuable.
Long-term outsourcing agreements show the problem plainly. Winning a contract, and making the commercial and operational promises needed to secure it, is hard. Living with those promises for five, seven or ten years is harder. Customer expectations change, the economics shift, new technologies appear, and the people who made the original decisions may no longer be in their roles.
AI adds particular pressures to this familiar situation. An enterprise may commit to productivity gains based on an anticipated division of work between people and technology, only to find that a newer capability makes that division needlessly conservative. A more demanding application may show that human verification is more involved than expected, changing the economics of the business case. A service provider may introduce AI into a contracted process while the customer is still deciding what authority it is prepared to delegate and what evidence it needs to retain.
None of this automatically invalidates an investment or contract. Enterprises can renegotiate and adapt. The difficulty is that conditions change at different speeds. Commercial commitments, workforce structures and accountability arrangements stay relatively fixed while the capabilities available to the business, and its expectations of them, evolve. This was true before AI. What is new is how many of those conditions can now change within the life of a single transformation.
From supporting decisions to shaping them
Sophisticated analytical assistance is becoming widely available, and that is one of the most promising developments. An employee who once relied on reports or specialist advice can now ask an AI system to compare documents, explain an unusual result, explore alternatives or develop a reasoned recommendation. A small organisation may gain capabilities that once required a much larger team.
Enterprise systems have supported decisions for decades; business intelligence, forecasting and optimisation already went far beyond presenting raw information. What is changing is the breadth of systems that can take part in constructing explanations, interpretations and recommendations across ordinary business activity.
That matters for the person making the decision. A fluent recommendation can frame which assumptions and alternatives receive attention before the recipient has formed an independent view. Formal authority may remain with the human, while the work of reaching a judgment changes. An enterprise that set out to improve access to information may discover that it has also altered how decisions take shape.
An experiment by researchers at Harvard Business School, MIT and the University of Washington illustrates the distinction. In a study of 228 evaluators assessing 48 real innovation submissions, AI recommendations without explanations improved decisions against an expert benchmark. Adding persuasive narratives increased agreement with the AI, including when it was wrong, particularly when it favoured rejecting promising ideas. The finding concerns a specific innovation-screening task, not enterprise decision-making in general. It does show that an explanation that encourages acceptance is not necessarily one that improves judgment.
The wider evidence is mixed. A meta-analysis in Nature Human Behaviour, covering 106 experiments, found that human-AI combinations performed better on average than humans alone, but worse than the better of the human or the AI working independently. Decision tasks tended to show losses where content-creation tasks showed gains. Much of the underlying research predates the latest models, so these are not permanent limits. They do show that adding AI to a decision process does not reliably improve it simply because both participants are individually capable.
Governance and safety engineering remain essential, but the consequences of introducing AI into decisions extend beyond the systems themselves. The CIO can enable dependable services and define conditions for responsible use; business leaders must also discover how those services alter judgment, responsibility and the choices their organisations are equipped to make.
As AI becomes a routine participant in professional work, the original plan to assist decisions may give way to a larger question: how should decisions themselves be organised when some of the reasoning is performed and presented by a machine?
When new capabilities change the purpose of a plan
The opportunities available to an enterprise can also expand during the life of its transformation. This is not merely a risk to the original business case; it may be the most valuable consequence of delivering it.
Earlier enterprise technologies were not static, and organisations retain considerable discretion over when to introduce new capabilities into production. A supplier's announcement does not immediately change an enterprise's operating environment. But new capabilities change the choices open to the business even before it adopts them. Employees encounter them outside approved systems. Competitors begin offering services that were previously too expensive to offer. Customers start asking why an established process still takes as long as it does.
Research on AI-assisted professional work shows why capability boundaries matter. In an experiment with 758 Boston Consulting Group professionals, participants using GPT-4-era AI completed 12.2% more tasks and worked 25.1% faster on tasks selected to fall within the system's capabilities, with improvements in measured quality. On a deliberately selected task outside that range, AI assistance made participants less likely to reach the correct answer.
The technology was from 2023 and cannot establish how today's systems perform. The lasting point is the distinction between tasks where assistance is productive and tasks where reliance is misplaced. That boundary can change as the technology improves, but the enterprise's understanding of where it lies does not necessarily change at the same pace. A process designed around yesterday's limitations may become needlessly restrictive, while a new application may introduce difficulties absent from the original deployment.
Consider an enterprise that begins by using AI to help specialists locate and summarise material in a large document collection, to cut search time and improve access to institutional knowledge. That is worthwhile in its own right. As the technology develops, the same enterprise may find it can compare cases across years, identify patterns for expert investigation, explore alternative commercial arrangements, or offer customers a service that was previously too costly. An initiative justified by efficiency savings could end up supporting a different service proposition. Some of these possibilities will be anticipated by the roadmap; others will emerge from employees solving problems nobody thought to include in the business case.
Not every discovery will be favourable. Some uses will prove less reliable, less economical or less acceptable than expected, and others will require new skills or forms of oversight. Either way, the original investment case may stop describing the most consequential choices the enterprise faces, even when its initial promises have been fulfilled. A successful project can create the conditions for opportunities and questions that could not reasonably have been specified when it began.
The organisation changes the technology in use
The system of record and the system of work are not necessarily the same thing.
I recall an enterprise project-management platform introduced with substantial consulting effort and strong executive sponsorship. It became part of the formal reporting process, yet some project managers continued to organise their day-to-day work in Excel or Microsoft Project, entering information into the new platform largely to meet reporting requirements. The gap between the system the organisation had formally adopted and the tools in which its work actually took place is the point.
AI adds a new dimension. Generative systems invite experimentation. Employees can ask unexpected questions, combine materials and discover uses a central programme team never considered. Some will be valuable innovations; others will introduce errors, bypass established processes or cross boundaries that exist for good reasons.
Adoption figures therefore tell an incomplete story. An AI assistant might reduce the time needed to prepare a report while increasing the effort required to verify it. It might let a small team do work that once required external expertise, or let an organisation offer a service it could not economically provide before. The resulting value may appear in productivity, decision quality, customer experience or new revenue, and may be hard to capture with the measures used to justify the rollout.
Nor is the experience of using AI a reliable substitute for measuring its effects. In a randomised experiment with 16 experienced open-source developers and 246 tasks, the research nonprofit METR found that participants believed AI assistance had made them about 20% faster, while measured completion was 19% slower when AI was allowed. The study involved particular developers, repositories and early-2025 tools, and METR's later work suggests effects may have changed with newer systems, so it should not be generalised to current coding assistants.
Its significance here is narrower: perceived usefulness, enthusiastic adoption and measured performance are different things. A disappointing result on one productivity measure does not mean the technology creates no organisational value. A tool may change the kinds of work people undertake, the expertise they develop or the opportunities the organisation can pursue.
These effects are not determined by the model alone. They emerge from the interaction of technology, working practices, incentives, professional judgment and the organisation's willingness to adapt. Employees are not merely recipients of an AI transformation. Through their use of the systems, they help determine what it becomes.
The enterprise is not changing alone
The environment in which the enterprise operates is going through its own process of discovery. AI developers are still investigating the capabilities and limits of their systems, and providers are refining products, safeguards and commercial models. Regulators are taking different approaches across jurisdictions. Employees may welcome new capabilities while questioning what they mean for professional expertise and future roles. Customers may value faster, more accessible service in one context and insist on human involvement in another.
These responses do not divide neatly into enthusiasm and resistance. The same person may welcome AI assistance in their own work while being uncomfortable with its use in a sensitive decision affecting them. What is acceptable in one industry or transaction may be inappropriate in another, even where the underlying technology is identical.
Technical feasibility, regulatory permission, commercial viability and social acceptance are distinct conditions, and they do not change together. A capability can become practical before customers want it, or commercially attractive before the organisation knows how to use it responsibly.
A tightly regulated enterprise will face different questions from a small professional-services firm gaining sophisticated analytical capability. A company relying on outsourced operations must understand how its providers are introducing AI into services for which it remains accountable. A business using AI for creative work may find that customer expectations of authorship and authenticity matter as much as technical quality.
Consultants and technology providers are learning too. Their experience, methods and products can help enterprises decide well under uncertainty, and a credible strategy or contract can establish responsibilities and expectations. It cannot guarantee that every future use case, economic consequence or customer response has been anticipated.
None of this makes conventional disciplines obsolete. Strategy, change management, commercial governance, cybersecurity, assurance and safety engineering remain indispensable. But their task is not only to protect a fixed transformation plan. They must also support an organisation whose understanding of its opportunities, obligations and consequences keeps developing.
What does it mean to finish?
The most useful distinction may be between completing an AI programme and completing the changes it makes possible.
The first has meaningful boundaries. Systems can be deployed, controls tested, employees trained, contracts fulfilled and benefits measured. These are substantial achievements, and describing the organisation as unfinished should not diminish the work required to deliver AI successfully and responsibly.
The second has no comparable endpoint. An enterprise may find that AI changes how decisions are framed, that a previously impractical service has become viable, that employees have developed new ways of working, or that customers and partners value something different from what the strategy anticipated. Some discoveries will justify caution and restraint. Others will open opportunities for investment, differentiation or growth.
None of this requires enterprises to abandon long-term planning or avoid firm commitments. It does suggest that the assumptions behind those commitments deserve as much attention as the milestones used to measure their delivery. An organisation needs to understand not only whether its AI systems are being used successfully, but how their use is changing the work, decisions and opportunities for which those systems were introduced.
An unfinished AI enterprise is not necessarily an unsuccessful or inadequately governed one. It may have delivered everything it promised, established responsible operating practices and realised substantial benefits. Yet the experience of using AI can change its understanding of what work is worth doing, which capabilities matter and what opportunities lie ahead.
It may have completed the transformation it planned, only to discover that the more consequential transformation is still taking shape.