Before Your Team Approves the Next AI Budget, Ask This
Most companies are getting close to, or are already in, the 2027 budgeting process, so we wanted to help teams focusing on next year's AI strategy think through a few considerations.
Most AI budgets get approved because the pitch is sound, the vendor sounds confident, and the number seems reasonable, not because anyone checked whether the AI actually has what it needs to do the job well. That's the gap this piece walks through, and it's fixable, as long as you know what to look for before the deck gets presented, not after.
Before your group votes on the next AI proposal, ask this: has anyone actually mapped the workflow and figured out what the AI needs to know to do it well, or did you just like the demo and the price?
The audit needed here is simple to describe, even if it takes real work to do. Someone maps the specific workflow the AI will touch, writes down what it needs to know to do that work well, and figures out who currently holds that knowledge, on paper or only in someone's head. Skip that step, and your team is really just approving the pitch and the number, which tells you nothing about whether the project will work.
ROI takes 28 months? Why?
Gallagher surveyed more than 1,200 businesses this year and found that companies tracking AI's return on investment expect it to take an average of 28 months to pay off. That's a discouraging number, but the cause isn't what you'd expect.
Sixty-three percent of those companies have already put AI to work somewhere in the business, up from 45% the year before, so adoption isn't the bottleneck. Gallagher's researchers point to the real driver: putting AI to genuine use means redesigning how processes and roles actually work, and few companies had done that groundwork first. A separate MIT study, on a different set of companies, found the same pattern: 95% of AI pilots showed no measurable effect on the bottom line, traced to a “learning gap”, the AI never got the operational knowledge it needed to be useful. Two research teams, two years apart, land on the same story.
The one check teams skip most often: what a person actually knows
This becomes most visible, and most costly, when a decision touches headcount, and a few numbers are worth thinking through before signing off on any cut tied to an AI rollout. Through others’ mistakes, you have a chance to learn the lesson the easy way, for a change.
Robert Half found that 29% of companies that laid off workers and cited AI as the reason have already rehired for those same roles. CNBC reported in July that Ford spent three years hiring back 350 engineers after assuming AI could handle its design specs without experienced oversight. It couldn't, and those engineers now run quality troubleshooting. IBM handed its HR function to AI, which correctly resolved 94% of requests, but the remaining 6% needed judgment the system lacked, often an ethical call nobody had written a rule for. IBM's response was to triple entry-level hiring for 2026, reasoning that skipping that level now leaves no experienced pipeline in a few years. Commonwealth Bank of Australia made a similar reversal.
of AI-cited layoffs already rehired into the same roles.
engineers rehired after three years without experienced oversight.
of HR requests needed judgment the system didn't have.
Look at who tends to get called back and a pattern shows up. It's disproportionately people who knew which client relationships needed careful handling, or who carried the real history of how a process had evolved. None of that lived anywhere but in their memory, so cutting them meant losing the knowledge, and rehiring meant paying a premium to get it back.
This is the audit question in its most concrete form: before approving a headcount reduction tied to an AI rollout, has anyone documented what the departing employee knows that isn't captured anywhere else, like a wiki or a shared drive? If the answer is no, the team is approving a decision it doesn't have enough information to make safely.
Why this check gets skipped
Most leaders do care about getting this right. What gets in the way is more mundane: audit work is slow, it certainly isn’t sexy, and it tends to lose to a budget conversation already moving fast.
One recent workforce survey found that executive expectations for AI increasing entry-level hiring dropped from 40% two years ago to a mere 15% now. But when researchers checked what companies were doing to prepare, the numbers were thin: roughly a quarter had run any skills audit, and fewer than a third were funding real retraining. Companies are moving faster to cut jobs than to figure out which ones actually need to change.
Meanwhile, AI is creating new roles that pay well: positions like AI operations lead and quality steward are already appearing in job postings, and people who work well alongside AI are earning wage premiums well over 50% in some sectors. A C-suite approving cuts without an audit typically never hears about any of this, because nobody wants to own up to it.
A two-minute gut check before you vote
Before the full audit process, here's a faster way to check whether your current proposal is ready for a vote. Read these five questions out loud in the meeting and see how many the room can answer with genuine confidence.
Can someone in this meeting describe, step by step, the specific workflow this AI will touch, not just the department or team it lives in?
Has anyone actually talked to the person or team doing this work today about the judgment calls they make that aren't written down anywhere?
If this proposal includes cutting or reassigning any roles, has anyone documented what those employees know that doesn't exist in any system or shared drive?
Do we have a specific, written definition of what “working correctly” looks like for this AI, or are we planning to figure that out after it's already live?
If this doesn't work as expected six months from now, do we know exactly who would still be around, with the right knowledge, to fix it?
If two or more of these come back as “not sure,” that's a useful signal on its own: the proposal needs the fuller audit below before it goes to a vote.
What the audit actually looks like
A full audit doesn't take long. A short, structured review before the vote usually covers four things worth checking.
Map the specific workflow the AI will touch, in enough detail that you can point to the exact steps where a person is currently exercising judgment, not just following a script.
Name who holds that judgment today, and check whether it's written down anywhere a system could use it, or only exists in that person's head.
Decide in advance what “good enough” means for the parts AI will handle, so nobody is guessing six months later whether the remaining piece is a small fraction of the job or a third of it.
If the proposal includes a headcount change, have the audit happen first and the personnel decision come second.
Every one of the costly examples earlier in this piece skipped at least one of these four steps.
Why this approach shortens the timeline instead of extending it
In Rampwell's own engagements, we aim to show real, visible value in about eight to ten weeks. That comes from doing the audit first, writing down the rules the business runs on, its history, and who in your company has the expertise, before anything goes live.
The audit output is a documented context layer, the facts, rules, knowledge and history behind how a business runs, connected to whichever AI tools the team already uses. Building that layer is what we do at Rampwell. It belongs to the client once the engagement ends, sits in their own systems, and stays useful even as the AI model changes.
The same MIT research mentioned earlier found that AI projects built with an outside partner succeed roughly twice as often as ones companies build entirely alone. That gap has less to do with technical skill than with sequencing: bringing in outside help tends to force the audit before the team votes, rather than months later, after a rehiring cycle finds out what got missed.
For your next AI team meeting
So here's what I'd suggest: the next time an AI proposal comes up for a vote, take one extra meeting and ask the question that matters most: have we actually audited this, or are we just approving it? Every costly surprise in this piece traces back to a team that skipped that question and paid to find out the answer later.
Gallagher, 2026 AI Adoption and Risk Benchmarking Survey: ajg.com
MIT NANDA, The GenAI Divide: State of AI in Business 2025 (full PDF): nanda.media.mit.edu
Robert Half data, via AZFamily: azfamily.com
CNBC, “Employers who laid off workers citing AI are already starting to regret it”: cnbc.com
The HR Digest, “Companies are rehiring people they replaced during AI layoffs”: thehrdigest.com
Accenture data, via Computer Weekly: computerweekly.com
Deloitte, 2026 State of AI in the Enterprise: deloitte.com
Gloat, AI Workforce Trends 2026 (citing PwC's Global AI Jobs Barometer for the wage-premium figure): gloat.com
Note: the wage-premium figure is sourced via Gloat's citation of PwC's Global AI Jobs Barometer. Primary report for cross-checking: pwc.com