Building your annual operating plan with AI
The plan is due, and by the time it is signed off the market will have moved. The numbers coming up from the teams are gamed, and you are refereeing spreadsheets instead of shaping the bet. Here is how to make AI build a plan that is a real bet on where your unit should go, not a tidied-up version of last year.
Using AI for the annual operating plan does not mean asking it to write commentary on last year's numbers. It means naming the strategic bet first, then making AI interrogate the plan before it drafts: which assumptions are load-bearing, where the team is sandbagging, what would have to be true for each target. The Havruta Methodology (formerly the Think Partner Methodology) installs that discipline and anchors the machine in your unit's real figures and prior cycles, so the output is defensible rather than a plausible average. The plan then reads as a forward bet you can defend, not a reconciled spreadsheet. This is not a planning template. It is how a leader reasons with AI using the 4-Lines and the Flip on a plan they have to stand behind.
The situation
The annual operating plan is due, and three things are working against you at once. The cycle is too slow to be useful: by the time the plan is signed off the market has moved, and you are committing to twelve months when you cannot honestly see past six. In 2025, most organisations reported they could forecast only about six months ahead (OneStream, 2025), which makes a stable twelve-month horizon a fiction the plan is built on.
Then there is the planning itself. You spend the weeks as a referee, not a strategist, chasing spreadsheets across functions and reconciling conflicting assumptions, with forecasts shaped more by internal politics than by data. The numbers coming up are gamed: sandbagging, overpromising, or pure guessing, and you are forced to untangle them rather than shape the bet. A CFO interviewed in 2025 described exactly this, "spreadsheets flying across departments, conflicting assumptions, and forecasts often shaped more by internal politics than data," and the referee role it forces on the leader (CIO.com, 2025). The gaming is not new. It was named in management literature two decades ago: managers present a plan "less ambitious than one they know they could probably fulfill" (Steele and Albright, 2004).
And you cannot stress-test the result against a changing world. Most finance leaders say they struggle to rapidly model the implications of decisions, respond to external events, and run contingency planning for disruption (Deloitte CFO Signals, reported in CIO.com, 2025). The plan that survives all this is usually backward-looking: a tidied-up version of last year's spreadsheet rather than a bet on where the unit should go.
What commodity AI does with it
Drop last year's figures into an assistant and ask it to "build our annual operating plan" and it returns a tidy artefact in seconds: revenue lines extended, narrative commentary under each function, a set of initiatives, a phased timeline. It reads like a plan. But it has answered the easy question, what last year looks like rounded forward, and skipped the hard one, what your unit should bet on next. It cannot see your real constraints, your team's true capacity, or where the market is turning, so it produces something that could belong to any company in your sector.
Worse, it invents support, and it does so most where you most need help. A peer-reviewed study in 2025 found that across literature reviews generated by a leading model, roughly one in five citations were fabricated, and fabrication ran far higher on specialised, lower-familiarity topics than on generic ones (JMIR Mental Health, 2025). That maps directly onto planning: the model is at its most confident and least reliable exactly on the niche, low-data parts of your business, the new line, the unproven market, where the plan most needs judgement. Treated as a vending machine, AI gives you last year warmed over, with a few confident inventions baked in.
The Flip: make it interrogate the plan, not extend the spreadsheet
The move that changes the task is to stop asking the machine to produce the plan and start making it pressure the plan. Before any number is touched, name the Goal: what decision this plan must enable, what your unit is actually trying to win. That single discipline, naming the bet first, is what converts the cycle from backward-looking reconciliation into forward strategy. It is the antidote to High-Speed Waste, the reflex of scaling planning activity before anyone has defined what the planning is for.
Then comes the Flip: instead of you dumping figures and asking for commentary, the AI interrogates you first. Which assumptions are load-bearing? Where is the team likely sandbagging? What would have to be true for this target to hold? That line of questioning surfaces the gamed numbers and the unexamined politics before they are baked into the plan, which is the referee problem solved at the source. And if the first draft comes back generic, that is not a verdict on AI; it is the Mirror Principle at work. Generic output signals generic input, so you sharpen the Persona, the Goal, and the grounding rather than concluding the machine cannot plan.
What the machine must ask before it drafts
This is the part that proves it is reasoning about your unit and not a plausible average. Before it drafts a single line of the plan, it has to request what last year's spreadsheet cannot tell it:
-
What is the actual bet this year: the one decision this plan exists to enable, and what the unit is trying to win?
-
Which assumptions are load-bearing: the two or three that, if wrong, break the whole plan?
-
Where are the submitted numbers likely gamed: which targets look less ambitious than the team could fulfil?
-
What are the unit's real figures, prior cycles, and verified constraints to reason from, rather than industry averages?
-
What external shifts would force a re-plan, and what is the contingency if the six-month view changes?
Without those answers, a plan is a guess in a nice format. With them, the same line lands differently, funded, cut, or stress-tested, because it is now measured against your reality. That request is Ground Truth, the methodology's discipline of anchoring AI in your verified unit rather than its general knowledge, and it is the direct fix for the fabrication problem: a model held to your real numbers cannot invent support on the parts of the business it knows least. The standing version of this lives in the Brain Pillar: a persistent planning brain holding your assumptions, prior cycles, and strategy, so each year compounds on a living substrate instead of restarting from a blank spreadsheet, and so you can re-plan at speed when the market moves.
Build our annual operating plan. Here are last year's actuals and the targets each function submitted.
Before I extend anything, what is the bet this year? What is the one decision this plan has to enable, and what are we trying to win that last year did not?
We are shifting the mix toward the newer line. The bet is whether it can carry a quarter of revenue without breaking delivery.
Then the submitted numbers are the wrong starting point. Two questions. Which assumption is load-bearing for that shift, the demand or the delivery capacity? And the sales target on the new line looks flat against last year's growth rate. Is that conservatism, or is the team sandbagging because the bet is theirs to miss?
Delivery capacity is the constraint, not demand. And yes, that target is sandbagged. They are protecting themselves.
Then the plan is a capacity bet wearing a revenue target. Give me the real delivery numbers and last year's ramp, and I will build it around the capacity assumption, flag the sandbagged line with the true range, and attach the re-plan trigger if demand or capacity moves past six months. That is a bet you can defend, not a reconciled spreadsheet.
What you walk away with
Not a forecast extended forward. A plan built as a bet, where every line carries the reasoning that placed it: the assumption it rests on, whether the submitted number was sandbagged and the true range, and the trigger that would force a re-plan. It survives a board member asking what would have to be true for this to hold, and it tells you which numbers to watch rather than burying the bet under reconciled totals. Reclaimed planning weeks, the Hidden Headcount you spent refereeing spreadsheets, go back into shaping the bet.
- Stress-tested
The capacity bet on the new line
Load-bearing assumption Grounded in real delivery numbers Re-plan trigger attached - Flagged, not accepted
The submitted sales target
Sandbagged True range shown - Held, pending trigger
The cost line carried forward
Last year warmed over Not yet a bet
The 4-Lines you can run yourself
-
Act as a sceptical board member who will probe every soft number, then as my planning partner. Hold both seats.
-
Goal: an annual operating plan built as a defensible bet on where my unit should go, each line tied to a load-bearing assumption with a re-plan trigger, not last year's spreadsheet extended forward.
-
Ask me detailed questions and for supporting data before you draft anything: what the bet is, which assumptions are load-bearing, where the submitted numbers are likely sandbagged, my unit's real figures and prior cycles, and what would force a re-plan. Challenge anything vague and do not conclude until you have what you need.
-
Ask one question at a time, step by step.
Frequently asked questions
How do you use AI for the annual operating plan?
Not to draft commentary on last year's numbers. Name the strategic bet first, then make AI interrogate the plan before it writes it: which assumptions are load-bearing, where the team is likely sandbagging, what would have to be true for each target. Anchor it in your unit's real figures and prior cycles so the output is defensible rather than a tidy average. The plan then reads as a forward bet, not a reconciled spreadsheet.
Why does generic AI fail at building an operating plan?
Because a plan's job is judgement about your unit, and a model writing from general knowledge cannot make that judgement for your market, your team, or your numbers. Ungrounded, it invents support and degrades most on the specialised, low-data parts of the business where you most need help. It returns fluent commentary that says nothing true about your unit, which is last year warmed over with better formatting.
How do I stop my teams from sandbagging the plan?
Make AI play the sceptical board member before the numbers are baked in. Have it press each bottom-up figure: what would have to be true for this target, where is this less ambitious than the team could fulfil, which assumption is doing the heavy lifting. Sandbagging and consensus-gaming are structural and long-standing (Steele and Albright, 2004), so the method interrogates the assumptions rather than accepting the submitted numbers.
Can AI build a plan when I can only see six months ahead?
It can help you re-plan at speed rather than forecast further. The honest planning horizon has shrunk, with most organisations able to forecast only about six months out (OneStream, 2025), so a once-a-year cold start produces a plan that is stale on arrival. A standing planning brain that holds your assumptions and prior cycles lets you re-run the plan when the market moves, instead of starting from a blank spreadsheet each time.
What does it mean to ground AI in your unit's data?
It means giving the model your real numbers, prior plans, and verified assumptions to reason from, rather than letting it answer from its general knowledge. Ungrounded models fabricate support and are unreliable on the specialised parts of your business (JMIR Mental Health, 2025). Grounding makes the commentary defensible: every figure traces to something true about your unit, not to a plausible industry average the model produced.
Where do we start?
With the Eye-Opener Workshop, a half-day where your leadership team sees the shift on its own real planning question. A Strategic Briefing maps the right entry point.
References
- OneStream. "From 2025 to 2035: How Today's FP&A Trends Are Shaping the Future of Finance." OneStream, August 2025.
- Gross, G. "How AI is replacing the painful, manual process of building an annual operating plan." CIO.com, 9 September 2025. Cites Deloitte CFO Signals and quotes Aarif Nakhooda, CFO, CoreAI.
- JMIR Mental Health. "Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models." JMIR Mental Health, 12 November 2025.
- Steele, R., & Albright, C. "Games Managers Play at Budget Time." MIT Sloan Management Review, 15 April 2004.