The 2027 budget round is under way in most organisations, and AI turns up in almost every proposal. Some of it will deliver measurable results next year, but far from all of it. Below is a simple matrix and a set of questions your leadership team can use to sort AI investments before the money is allocated.
Why AI proposals need a filter
Adoption is growing fast. According to Statistics Norway (SSB), 30 per cent of Norwegian enterprises with ten or more employees used AI in 2025, up from 20 per cent the year before. Among enterprises with more than 100 employees, the figure was 58 per cent. SSB has scheduled new figures for 2026 on 25 September.
The effect on the bottom line is lagging. In McKinsey's State of AI 2026, published in August, 37 per cent of respondents attribute at least some EBIT impact to AI. That is about the same as last year. Only around 6 per cent qualify as "high performers", which McKinsey defines as organisations attributing at least 5 per cent of EBIT to AI. About one in five also say AI operating costs, including token costs, have limited their use.
Two more numbers come up in a lot of budget meetings. MIT's NANDA initiative reported in August 2025 that around 95 per cent of the generative AI pilots in its study had no measurable impact on the P&L. The sample was limited (52 interviews, 153 survey responses and a review of more than 300 publicly disclosed initiatives), so the figure works best as a warning sign. Gartner predicts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.
The matrix: two questions for every proposal
Ask two questions about every AI proposal, then place it in one of four quadrants.
Axis 1: Value and measurability. Can whoever made the proposal say which process changes, what it costs today and how you will measure the change within twelve months? Money, hours, lead time and error rates count. "Better customer experience" without a target does not.
Axis 2: Data dependency and risk. How much has to be in place in terms of data, integrations and access control before the solution works? And what happens when it gets something wrong? A tool that drafts text for an employee to review is low risk. An agent that updates customer records or sends emails on its own is high risk.
That gives you four quadrants:
- High measurability, low data dependency: go ahead.
- High measurability, high data dependency: budget it as a project, including the data work.
- Low measurability, low data dependency: give it a small, time-boxed budget.
- Low measurability, high data dependency: say no, or park it.
Quadrant 1: Measurable and light on data
This is where proposals that apply off-the-shelf tools to tasks with a known cost belong. Examples include meeting notes, first drafts of customer service replies, translation and coding assistants for developers. The gain is measured in time, and the solution needs little more than licences, training and an AI policy.
Budget licences and training together. A tool rolled out without training tends to be used by a handful of enthusiasts, and then the gain never shows up in the numbers. Set a target for usage and time saved, and follow it up after three months.
Quadrant 2: Measurable, but the data has to be in order
These are often the most valuable proposals: AI connected to your CRM, order system, document archive or customer dialogue, where the gains can be large. They are also the ones that most often fall apart, because the data is not ready when the project starts.
Treat them as digitalisation projects where AI is one component. Split the budget into two phases: first data, integrations and access control, then the AI feature. That way you can stop after phase one without wasting the money, because cleaner data pays off regardless.
The MIT report found that solutions bought from vendors or built with external partners succeeded about twice as often as purely internal builds (around 67 versus around 33 per cent). That is an argument for buying or partnering where good solutions exist, and building yourself where you have something nobody else has. [QA: add your own example from a client meeting where the data work had to come before the AI feature]
Quadrant 3: Unclear gain, light on data
Proposals land here when they could be useful but nobody can say how much. It might be an internal AI tool for ideation, or a test of a new AI feature in marketing. The risk is low, so there is no need to say no.
Give them a fixed budget in money and time, for example eight weeks, and require the trial to end with an answer: continue, move to quadrant 1 with a measurable target, or stop. Without an end date, trials like these tend to linger as fixed costs.
Quadrant 4: Unclear gain, heavy on data
This is the AI noise. Typical signs are big words ("autonomous agents across the entire value chain"), a vendor that cannot point to comparable solutions in production, and a plan that assumes data you do not have. Gartner warns of "agent washing", where existing assistants, RPA tools and chatbots are rebranded without real agentic capabilities. Gartner estimates that only about 130 of the thousands of vendors marketing agentic AI are the real thing.
The idea may be sound but early. Ask for the proposal to come back once the data work in quadrant 2 is done, or once whoever proposed it can measure the gain.
Ten questions for the budget meeting
Ask these questions about every AI proposal. The answers decide which quadrant it lands in.
- Which process changes? Name it, and say who owns it today.
- What does the process cost now? In hours, money or errors. Without a baseline you cannot measure the improvement.
- How will we measure the effect, and when? A number and a date, ideally within twelve months.
- What data does the solution need, and is it ready? Who has checked?
- What happens when the solution gets it wrong? Who notices, and what does a mistake cost?
- What are the running costs? Licences, model usage (tokens), operations and maintenance. Ask for an estimate for years two and three, not only for the launch.
- Are we buying, building or doing it with a partner? And why that choice?
- Who will use it, and what training will they get? Changing how people work is part of the cost.
- Which requirements apply? Privacy, information security and upcoming AI regulation.
- What is the stopping point? Which result, by which date, makes us shut it down?
McKinsey finds that the best-performing organisations are more likely to have redesigned their workflows around AI, instead of adding AI on top of existing processes. Questions 1, 2 and 8 are the ones that reveal whether a proposal does that.
I work with organisations that want to move beyond individual use of AI chat to AI as part of their operations, and much of that work starts with sorting proposals this way.
Sources
- Bruken av KI har skutt fart det siste året, Statistics Norway (SSB), 24 September 2025
- The state of AI in 2026: On the road to ROI, McKinsey, 25 August 2026
- MIT report: 95% of generative AI pilots at companies are failing, Fortune (via Yahoo Finance), 18 August 2025
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner, 25 June 2025
Transparency note: This article is based on the author's own analysis and experience. AI has been used as an editorial aid for language and structure.
