Get FP&A best practices, research reports, and more delivered to your inbox.
Bottom line: 93% of mid-market finance leaders are already using or trialling AI somewhere in budgeting and 78% expect it to make this season easier — but the teams furthest along report the longest hours. That is a selection effect, not a warning: hard budgets drive teams to AI. What genuinely improves is the mix of work, not the volume of it.
In our August 2026 survey of 273 finance leaders, all Director level or above at companies from 101 to 5,000+ employees, only 7.3% were not using AI for budgeting at all. That makes adoption a settled question and turns the useful one into: what has it actually changed?
Are mid-market finance teams actually using AI for budgeting?
Yes, almost universally, though depth varies a lot. A third are using it for a few tasks, roughly another third across the process, and a small leading group has moved to agents.
- A few tasks — 33.3%
- Across the process — 29.7%
- Just exploring — 23.1%
- Deployed agents — 6.6%
- Not at all — 7.3%
The gap worth noticing is between expectation and deployment: 42.5% say AI will definitely make this budget season easier, while 6.6% have actually deployed agents. Most teams are betting on a capability they have not yet operationalised.
Has AI made budget season easier?
Not on hours, and the direction surprises people. Overtime and burnout both rise with AI maturity, step by step.
Read this carefully, because the obvious interpretation is backwards. Teams adopt AI because their budget is already painful — more entities, more contributors, tighter deadlines. AI maturity is therefore a marker of a hard budget, not a cause of one. Nothing in this data supports "AI makes budgeting worse".
What the table does show, in the last column, is the one thing moving the right way. Busywork share falls as maturity rises: 65.1% of teams merely exploring spend half or more of the season on busywork, against 50.6% of teams using AI across the process. AI is currently reallocating finance time rather than reducing it.
What finance leaders most want AI to do
Asked which part of budgeting they would most want AI to take, the answer is not the analysis. It is the plumbing.
- Pulling and consolidating data — 34.4%
- Building the first-pass budget — 23.8%
- Modelling scenarios — 23.8%
- Writing variance narratives — 15.8%
- Chasing inputs from stakeholders — 2.2%
That top answer lines up exactly with the season's worst pain: consolidating data and wrangling versions, named as the most painful part by 37.4%. It also explains the hours finding. Most AI deployments target drafting and analysis, while the actual bottleneck is data movement — so the work being automated is not the work causing the pain. More on that gap in automating FP&A workflows.
AI budgeting software compared
"AI-powered" is now on every vendor page, so the useful question is what the AI concretely does and against which data.
Pricing is quote-based across most of this category; confirm current figures with any vendor. For a wider shortlist see the best AI FP&A tools and budgeting software for mid-market companies.
What to ask a vendor about its AI
Four questions separate a real capability from a demo.
- Does it touch consolidation? If the AI only drafts commentary, it is not addressing where a third of teams say the pain is.
- Can it show its work? Ask for a number, then ask which transactions make it up. Commentary you cannot trace is not usable in a board pack.
- Whose permissions does it use? An agent on a shared service account can reach everything that account can reach, which is the wrong answer once department heads are involved.
- What happens when it is wrong? Ask how the output is checked and how a correction persists rather than being re-explained each month.
That last point is the practical difference between a prompt and a repeatable workflow. Anthropic's announcement of Skills covers the underlying idea, and we go deeper in Claude skills for finance. For how agents fit a finance team more broadly, see AI agents in finance and FP&A.
Where AI is actually working today
Separating the parts that reliably work from the parts still in trial matters more than a capability list, because the failure mode is deploying AI against the wrong step.
Working reliably: drafting variance commentary from real drivers, answering ad-hoc questions against governed data, flagging anomalies against prior periods, and summarising a pack for a specific audience. These share a shape — the data is already structured and the output is a draft a human edits.
Still uneven: building a first-pass budget. Named by 23.8% as what they most want, but an agent needs your driver logic, your headcount plan and your commercial assumptions to do it, and most teams have not encoded those anywhere an agent can read.
Barely started: chasing inputs from stakeholders, which only 2.2% nominated — probably because everyone understands it is a human problem wearing a workflow costume.
What the leading teams do differently
The 6.6% who have deployed agents are not using better models. Three things separate them.
They encoded their conventions. Which date field revenue is reported on, how the chart of accounts maps to reporting lines, what commentary register the CFO wants. An agent without those guesses, and it guesses silently.
They put the checks in. Row counts, period-closed confirmation, a tie-out — the same instinct as a check tab in a model. This is what makes output trustworthy enough to send without re-deriving it.
They stopped re-prompting. A correction made twice becomes a stored instruction rather than a monthly habit. That is the difference between a prompt and a repeatable workflow, and it is why the teams furthest along spend less of their season on busywork. See Claude skills for finance.
What to expect in year one
Set expectations against the data rather than the pitch. Adoption will not give you back hours in the first season — the teams furthest along work the most, and complexity is why they adopted. What you should expect is a shift in what the hours contain.
A realistic first year: commentary drafting moves from writing to editing, ad-hoc questions stop requiring an analyst, and one or two recurring summaries run without being asked for. Consolidation improves only if you fixed the data layer, which is a different project. Sequencing is covered in FP&A implementation steps, and the category-wide picture in the state of AI in finance.
Where Aleph fits
Aleph puts the agent on top of a governed layer rather than beside it: it reads live ERP and CRM actuals with each user's own permissions, drafts variance commentary with the driver named and drillable, and answers questions in the spreadsheet your team already works in. Because the consolidation is handled underneath, the AI is working on the part that actually costs you time.
See the Aleph agent and budget planning, or read the full budgeting benchmark report for the rest of the survey data.
Get FP&A best practices, research reports, and more delivered to your inbox.
