Wide adoption, thin returns: why strategy matters
AI use in finance is now common. Gartner put adoption at six in ten finance functions in November 2025. With the tools this widely spread, the debate in most functions has moved on from whether to how. The returns have lagged the rollout though. In Gartner's data, 91 percent of finance organisations reported only low or moderate impact initially. What separates the functions that get value from the rest is rarely the technology but a clear strategy for its use. That strategy need not be long or elaborate, but it does have to make six deliberate choices, and each shapes the ones that follow. The six are set out below and developed in turn through the rest of this article:
- A clear view of where the function is going and what success will be measured by.
- Clear ownership within finance and alignment with the organisation's wider AI and data strategy.
- Delivery through small proven steps rather than one sweeping programme.
- A hard focus on return on investment.
- Efficient governance and accountability from the start.
- Predictable roadblocks anticipated rather than discovered.
Define the destination first
The first strategic step is to define what a programme of AI is setting out to deliver, and what success looks like. No single destination is an exact fit, a Series B scaleup and a listed group need different things from finance. Nevertheless, the direction most finance functions are working towards is broadly shared:
- A transactional layer, invoicing, payments and matching for example, that largely runs itself.
- A close measured in a day or two rather than over a week.
- Profit and loss and cash positions visible close to real time rather than weeks in arrears.
- Accurate rolling forecasts of both profit and cash.
- One trusted set of numbers, drawn from a data platform that the operational systems feed automatically, rather than a patchwork of exports and spreadsheets.
- Roles that have moved up with the work: the data specialist supervises the automation and owns data quality, the analyst supplies business context, the manager tunes the process and tells the story, senior leaders set strategy.
With the destination clear, the priorities, the sequencing and the choices of what to buy, assemble or build all become easier to make.
Establish ownership and the data foundation
Finance rarely adopts AI alone. Many organisations now run their AI plans on a hub and spoke model. A central AI function, the hub, takes the decisions that should only be taken once, sanctioning the platforms, setting the data and security rules, fixing the procurement route and defining the proof a use case must show before it scales. The functions, the spokes, lead delivery in their own domains, because a reconciliation problem is understood by the people who reconcile. Where such a plan exists, finance fits within it without being run by it. Getting the balance right matters, because centralising everything creates a bottleneck, while leaving every team to itself duplicates effort and risk.
For the federated approach to work, the finance side still needs an owner, usually the CFO. This role cannot be handed sideways to IT or down to an enthusiastic junior, and it is a working one. AI tool use has to be understood personally, since a leader who has never reviewed an agent's output cannot judge the claims made about it. The owner also picks the initial use cases, gives cover while results still lag the effort, and sets the standard of proof, which is a demonstrable return on investment. An engaged senior champion also pushes the mandate down to every level and gathers ideas from the top and the floor alike. A programme run in a silo stalls, and where adoption works best for finance the CFO, the CIO and the CTO own it together. However, where no central AI strategy exists, finance should not wait. Action is better than inaction in this case, and finance is well equipped to find its own path, since controls, approval routes and spend visibility are disciplines it already runs. A finance led implementation can then become the template the rest of the organisation adopts.
Whatever the ownership model, the work rests on one shared foundation, a robust data layer supported by digitised workflows. AI at scale cannot run on spreadsheets and manual processes, and a model placed on top of the general ledger returns aggregated insight at best. The target is a single platform holding granular financial and operational data together, with the key streams, from the CRM, payroll, timesheets and the operational systems, named and classified the same way throughout. Without that consistency there is no trusted version of the numbers, and no AI programme delivers its full benefit. Happily, AI now helps build this layer itself, mapping fields, standardising names and cleaning records, so what was once a multi year data programme can be advanced use case by use case.
Choose good over perfect
Wherever ownership lands, more programmes die waiting for the flawless dataset or the fully redesigned process than from any technology choice. The working alternative is practical. Take one process, bank reconciliations for instance, state honestly where it stands today, set a specific target such as ninety percent matched automatically, deliver the improvement, measure what changed, and move to the next process. The work has to run inside one hard constraint though. A finance function cannot pause its obligations while it modernises, the close, the audit and the reporting calendar carry on regardless, so anything that disrupts them carries a risk that efficiency gains will struggle to repay. In effect that rules out one sweeping rollout. The safer pattern is to land a limited number of controlled use cases, ideally starting in a sandbox where nothing touches live systems, prove the value in weeks, then repeat across the function. Each use case follows the same path: it runs with a person in the loop first, proves its accuracy, and gains autonomy step by step. Small starts work because the route to scale is set from the outset. Perfection is also less necessary than it once was. Earlier waves of automation made an existing process faster and cheaper. AI can improve the process itself. It has also made automation cheap enough that a pilot on an imperfect process often reveals where redesign is genuinely needed faster than months of mapping would, provided the process meets a sensible threshold of being good enough. Fundamentally, starting early pays, since the maturity built over a cycle or two puts the finance functions that experiment far ahead of those that wait. Nevertheless, staying adaptable matters as much as starting early. No single model stays ahead for long and pricing shifts almost monthly, which makes flexibility a design requirement. The stack should be able to switch models without a rebuild, locked into none of them. The simplest system that meets the target also beats the elaborate one: marginal extra benefit is rarely worth permanent extra upkeep. Small steps, sandboxes and flexible tooling all serve the same end, a stream of improvements that demonstrably pay for themselves.
Measure AI as an investment
AI is an investment and should be measured like one, with a baseline, a target and actuals reported against both. Usage statistics and pilot numbers say something while AI fluency is being built, but count for little over the longer term. The measures that matter are the function's own operating markers, a faster close, cheaper invoice processing, sharper forecasts, fewer errors and reworks. Movement alone is not a return, though. The hours redeployed and what they cost, the working capital released and the outside spend avoided all have to be quantified and netted against the full running cost of the AI itself. The resulting figure is the programme's profit and loss, and it is the number the board should see.
For meaningful return, the place to point the programme first is usually not the reporting layer, which is typically well controlled and streamlined, but the manual grind of accruals, reconciliations and invoice chasing that eats junior time. The gains in this unglamorous territory are large and, with tooling that costs significantly less than a systems change, the investment case is rarely difficult to make. Take a simple example. A business with £20m of invoiced annual revenue that cuts debtor days from 58 to 49 releases roughly £490,000 of cash. Targets of that kind make a programme governable, and crucially without them every review is an exchange of opinions. KPMG's May 2026 survey found only 29 percent of finance functions track where their AI fails, and a function that does not track outcomes will not know which successes are worth repeating. A useful record for each AI use case is four short entries, the status before, the status after, what changed, and the value delivered. That evidence decides whether an initiative deserves to scale. Whether it is allowed to, though, is largely a matter of governance.
Design governance from the start
Governance should not be seen or implemented as a brake on AI adoption. Rather, it is needed to remove the ambiguity that stalls progress, by clearly defining what AI may do on its own, who owns the output and where a person must check the work. In finance that clarity is non negotiable and a short focused policy is all it takes to begin with. An LLM can help produce a workable first draft from a set of agreed principles, covering what may run unsupervised, the data and security standards a tool must meet, and where human review is required. One requirement is specific to finance. At the outset, anything that touches the general ledger keeps a person in the loop, with the controls and audit trails intact, the outputs explainable, and nothing posting as a black box.
The governance framework should tighten as the stakes rise. While a use case runs in a sandbox, on copied data and unable to touch a live system, oversight stays light and experimentation stays cheap. Once its numbers are to be relied on and reported, the controls firm up and accountability comes with them. Every initiative that graduates from pilot needs its own named owner and a clear understanding of what the AI decides and what a person decides. That shows everyone what good looks like, who is responsible, and when it is safe to scale, so decisions are made quickly rather than renegotiated each time. The same discipline extends to standing agents as they arrive, each running under its own login with the least access the task requires, and it extends outward to the regulators, whose expectations of finance are now taking written form. Both are covered in depth in the governance article. None of it lands, though, unless the people running it have been brought along from the start.
Anticipate the roadblocks
Programmes rarely stall at the top or the bottom. The executive team is supportive and the wider team is curious. The block typically sits in the middle, with the controller and the team leads asked to redesign the work on top of a demanding day job. The first thing to give them is time. A team with no slack cannot redesign its own work, so capacity comes first, and the quickest way to create it is to take the most repetitive tasks off the team before asking anyone to build anything new. How to do that across the function is covered in detail in the implementation roadmap and later articles, but the first steps are simple and quick to value. An agentic workbench such as Cowork, for example, pays back immediately when deployed on routine workflows.
Working culture is often another barrier. What is being asked of the team is a change of habits, not just of tools, and the change that matters most is in management style, from checking every output to equipping people to do more themselves. The quiet sceptics come round on results rather than arguments, and louder objections are heard and logically answered in the open. Winning the arguments is only half of it, though. The people already in the team need to be brought along rather than left to guess. That means spelling out the benefits of AI adoption, chiefly that the repetitive work goes first and more interesting work replaces it. The tools and training also need to arrive before the expectations do. Part of this conversation is also being honest about where the roles are heading. The team's shape shifts towards a smaller core of accountants, many moving into analytical roles, working alongside data scientists and closer to the wider business, so that people move up with the work rather than feel displaced by it.
None of these strategic choices has to be right first time. Made thoughtfully and recorded plainly, each one improves the odds and shortens the path of the next, and the team's fluency with AI builds with every cycle. The strategy sets the direction. How to deliver it is the subject of the implementation roadmap that follows, and the reward for starting now is simple. The finance functions that move first are the ones the rest spend years catching up with.