Most teams pick their first AI project badly. They pick the work that annoys them most, or the work a vendor demoed well. Across sixteen interviews with senior leaders between April and June 2026, the work that produced real results had a consistent shape, and it had nothing to do with which function it belonged to.
What kind of work is worth pointing AI at?
Messy data comes in.
The work requires pulling numbers or records out of many disconnected systems, reconciled by hand today. This is what AI handles well and what people are slowest at.
Judgment is required on the way out.
The work needs scenarios, trade-offs and a call at the end. That keeps a person at the centre rather than replaced, and it’s why this work survives automation rather than being erased by it.
It has been deferred for years.
This is the work that nobody has the hours for. It’s the analysis that never gets run, the scenario that never gets tested, the customer who never gets answered.
That third condition is the one most teams skip, and it’s the one that decides the size of the result. Work that already happens can only get faster. Work that has never happened is new capacity, and the step-change results in this research all came from there.
Why is FP&A the clearest example?
FP&A meets all three conditions at once, which is why every finance leader in the research named it first.
Consolidating numbers from many disconnected systems is most of the job. Sensitivity analysis, scenarios and trade-offs are the part only a person can sign off. And forecasting is the work finance teams defer indefinitely because closing the books consumes the month.
It isn’t that FP&A is rote work ready for automation. Bookkeeping, AP and payroll are already fairly optimized, and several finance leaders treated them as a second-order target. FP&A is the answer because it’s where the deferred work sits.
What are finance teams already doing?
Three results from the research, all achieved by small teams rather than transformation programs:
- Board and investor reporting moved from around the tenth of the month to two or three days after close.
- A seven-day financial forecasting job was rebuilt in roughly two hours, by a curious staffer rather than a project team.
- A company of about fifty people that historically barely forecasted now runs a rolling daily revenue forecast.
One CFO described an AP manager approaching retirement who became the team’s biggest advocate, because accruals that used to take a week of spreadsheet work now take a few minutes.
Notice who did this. Small teams and curious individuals with permission from leadership. Not a central program and not a consultant.
What does the same shape look like outside finance?
The same three conditions appear in every function, and leaders outside finance pointed straight at their own version. A shared service centre. HR processes. Manufacturing scheduling. Underwriting.
Two examples from the research show how far the shape travels.
Operations
A residential real-estate operator rebuilt the front of its leasing process around AI and now does 40% of its leasing after hours. That isn’t the old work sped up. It’s demand that used to ring into a closed office and go unanswered or pile up.
Internal IT and support
One of the most advanced organizations in the research runs production agents with actual jobs. One triages roughly 95% of the support tickets coming in from thousands of internal users, a job a person used to do. Others took over routine internal communications that never happened reliably when they depended on busy people. As that leader put it, a lot of the work he could never get people to do consistently is now done by agents.
Accounting, finance, HR, IT and operations all have work with this shape. What travels is the shape, not the finance label. If your highest-pain function isn’t finance, look for messy data coming in, judgment required on the way out, and a backlog everyone stopped noticing.
Why do these workflows stall after one person builds them?
Because they never leave that person’s machine. This was the fastest-growing theme in the research, and it came from the most advanced organizations rather than the ones falling behind.
Someone builds a genuinely good AI workflow. A reporting pipeline, a forecasting model, an analysis assistant. Then it lives, permanently, on one laptop. One CFO at a venture-backed software company described it precisely: they hadn’t worked out how to get these tools off people’s local drives, so sharing still took real effort from IT.
Single-player AI works. Shared, governed, maintained AI barely exists.
The prototype that saves one analyst a day a week never becomes the system that saves the whole team a day a week. That gap is where most functional teams are stuck right now, and it isn’t a tools problem.
How do you get past it?
By teaching the AI the business. The most advanced organization in the research built an internal library of its own acronyms, models and business rules, so AI outputs are grounded in how the company actually works. Their phrase for it was “taking the AI to school”.
That’s the difference between a clever demo and something a team can rely on. Someone has to build that layer, maintain it, and know what good looks like well enough to catch output that is confidently wrong.
It also has to be somebody’s job. In most mid-market teams it currently isn’t anybody’s.
What question should a functional manager ask?
Not what could be faster. What isn’t happening at all.
Efficiency is a reasonable on-ramp. Faster reporting, automated templates, quicker review. Real wins, quickly measurable, low risk.
But every step-change result in the research came from the other question. Which customers go unserved. Which analysis never gets run. Which scenarios never get tested. Which internal work quietly doesn’t happen because it depends on someone who is already at capacity.
Asking that question well requires knowing the function. It isn’t a question a vendor or a consultant can answer for you.
Who builds this?
A doer inside the function, not a technologist. The AI champions leaders named were the people who know the process and now know where AI helps and where it doesn’t. One VP Finance described his champions plainly: they are the doers of the function.
That profile has a name in the research. The Functional AI Expert is a domain expert in accounting, finance, HR, IT or operations who designs AI-enabled workflows inside their own function. It was the single most requested profile across the sixteen conversations, named unprompted in more than ten of them.
Most teams don’t have one, and hiring for it is harder than it looks. You can teach a functional expert to use AI. You can’t teach a technologist the specifics of your close, your reporting lines, or how your team actually works. That comes from sitting in the seat.
This is also difficult because the screen is behavioural rather than credential-based, which is a problem for hiring managers and candidates alike: what leaders now look for, and what a candidate can actually show.
The teams making progress brought this person in, often for six to twelve months rather than as permanent headcount, then paired them with someone who could productionize the work so it didn’t end up back on one laptop.
Where to start
Three questions worth putting to your own team this month:
- Which AI workflows are running right now, and who built each one?
- Who else can use them, and what breaks if that person leaves?
- What isn’t happening at all in this function because nobody has the hours?
The answers usually surprise the manager more than the team.
Read the full research
What We Heard, and What the Frontier Is Saying covers all sixteen conversations: the diagnostic, the eight patterns, the full roles map, and a self-diagnostic you can run on your own function. Part II sets the findings against what the consultancies and the frontier labs published in the first half of 2026.
Download the full report
About this research
Clarity Recruitment is a specialist recruitment firm in Toronto working with mid-market and growth-stage companies across accounting, finance, HR, IT and operations. The listening tour was conducted between April and June 2026 across sixteen organizations, including seven finance leaders and nine leaders in people, operations, technology and the founder seat. All interviews were off the record and every finding published here is aggregated and anonymized.
Frequently asked questions
What work should a team use AI for first?
Work with three characteristics: messy data coming in from many systems, judgment required on the way out, and a long history of being deferred because nobody has the hours. The third condition matters most, because work that already happens can only get faster, while work that never happened is new capacity.
Why do finance teams start with FP&A?
Because FP&A meets all three conditions at once. It consolidates data from disconnected systems, it requires judgment across scenarios, and it’s the work finance teams defer while closing the books. The transactional layer of finance is already relatively optimized.
Will AI replace FP&A analysts?
The pattern in the research is the opposite. Analysts move from gathering data to analyzing it, which is the job that was always supposed to happen and rarely did. The shift is in who a team needs, not how many people it needs.
What are realistic AI use cases for a finance team?
Documented results from the research include board reporting moving from the tenth of the month to two or three days after close, month-end accruals dropping from a week of spreadsheet work to minutes, a seven-day forecasting job rebuilt in about two hours, and a fifty-person company moving from barely forecasting to a rolling daily forecast.
Does this only apply to finance?
No. The same shape appears in HR processes, shared service centres, manufacturing scheduling, underwriting and internal support. One organization in the research uses an agent to triage roughly 95% of internal support tickets. A real-estate operator now does 40% of its leasing after hours.
Why do AI workflows stay stuck on one person’s laptop?
Because nobody owns turning a personal prototype into a shared, governed, maintained system. Single-player AI works. Multiplayer AI requires someone to productionize it, and most mid-market teams have not staffed for that.
How do we get AI to understand our own business rules?
By building an internal library of your definitions, metrics, acronyms and rules so outputs are grounded in your business rather than generic. The most advanced organization in the research described this as taking the AI to school.
Who should own AI inside a functional team?
A domain expert within the function, not a central technical team. Across the research, the people succeeding with AI were the doers who knew the process first and learned where AI helps second.
