Annual Planning for 2027 with AI: Data, Scenarios and OKRs

Annual planning with AI is not about asking a model for your 2027 plan. It is about stripping out the three weeks of mechanical work that eat the planning cycle — consolidating data from five sources, building scenarios by hand, rewriting the same objective fifty times — so leadership can spend that time on the one thing it cannot delegate: deciding where the money and the people go.

We are publishing this in September on purpose, because that is when the cycle starts in most companies and because arriving in November with unconsolidated data is the usual reason plans close in January. We plan our own businesses this way. What follows is the process, the tooling and the limits, without the hype.

Which parts of the planning cycle can be automated?

The mechanical parts, which are more than half. A typical cycle has five blocks and only the last one is irreducibly human:

The practical consequence is counterintuitive: the biggest value of AI in planning is not brilliant analysis, it is plumbing. When the finance lead stops spending ten days reconciling exports from the ERP, the CRM and three spreadsheets, ten days appear for thinking. That is the real return.

How do you prepare the base data without losing three weeks?

With an agent that treats data collection as a scheduled job, not an annual project. The difference between companies that close the plan in November and those that close it in February is rarely analytical capability — it is that some have the data ready and others start looking for it in October.

1. Inventory your sources. ERP, CRM, billing, payroll, marketing tools. For each one, note who maintains it and how often it updates.
2. Automated extraction into a single format and cadence. Here the agent is unbeatable: it does not get tired, does not skip a month and leaves a trace of what it did.
3. Inconsistency detection before analysing anything: duplicate customers, categories that changed mid-year, revenue booked into two periods. It is the work nobody wants to do and that invalidates the plan when skipped.
4. One base dataset from which every scenario derives. If each department brings its own version of the truth, the planning meeting becomes an argument about numbers instead of a conversation about decisions.

If your cash forecast is also rebuilt by hand every month, this is the moment to fix it, because it feeds the plan directly — we cover it in AI agents for treasury and cash forecasting.

How do you build scenarios that support decisions?

With three scenarios, not ten, and by naming the two or three variables that actually move the outcome. AI helps build them quickly and test sensitivities, but the value lies in the discipline of the format, not in the number of simulations.

• Base case: observed trends continuing, no heroics. This is the one you budget on.
• Stress case: what happens if the critical variable drops — the customer worth 30% of revenue, the channel bringing half the leads, the margin on the flagship product. The question is not "how much less would we earn", it is "which month do we hit minimum cash".
• Expansion case: what growing 30% would require, expressed in people, capacity and capital, not in optimism.

One warning we always give: a language model does not predict your revenue. It can structure the model, run the calculations with the right tools, flag inconsistent assumptions and write the narrative — but the assumptions are yours and so is responsibility for the number. When someone presents an "AI-generated" forecast without being able to explain where each assumption comes from, the problem is not the model, it is the process.

How do you prioritize the 2027 initiative list?

With the same discipline we apply to automation cases: estimated value, effort and risk, plus a hard cap on simultaneous initiatives. AI contributes at three specific points, and none of them is the final call.

First, estimation: it pulls the history of what similar initiatives actually cost last year — not what was budgeted — and puts it next to the new estimate. It is the most effective antidote to the systematic optimism of proposals. Second, dependencies: reading all fifty proposals together, it spots that three depend on the same team and two overlap. Third, coherence: it flags which initiatives do not contribute to any stated objective, and in every portfolio there are a few.

For automation initiatives, the estimating frame is the one we already use: a custom SMB agent runs €3,000-15,000 and 4-8 weeks, plus 10-20% a year in maintenance. How that translates into expected return is in the ROI of artificial intelligence in business.

Is AI useful for writing OKRs?

It is better as a critic than as an author. An objective written by a model reads well and commits to nothing; an objective written by the team and subjected to criticism genuinely improves. The pattern that works: the team writes, the agent asks.

• Is this key result measurable with data that already exists? If you have to build a system to measure it, it is not a good key result for January.
• Is it a result or an activity? "Launch the new portal" is a task; "halve order-status enquiries" is a result.
• Who controls it? An objective depending on three departments has no owner.
• Does it contradict another objective? Growing volume and raising margin at once with no extra resources is usually a hidden conflict.

Add a coherence check across levels: team objectives should add up to the company's, and that check is exactly the kind of dull, systematic task where an agent skips no line. What to measure once the company runs on agents is broken down in AI-First company KPIs.

How much should you budget for AI in 2027?

Four lines, and everyone forgets all but the first:

1. New agent development: €3,000-15,000 per agent for an SMB. Budget by concrete cases, not by "AI project"; a global number with no cases behind it is a number that gets cut entirely at the first review.
2. Maintenance of what is already live: 10-20% a year of accumulated development. This line grows every year and is what breaks budgets from the third year onward.
3. Licences and consumption: per-user assistants and model usage. Prices change often, so confirm current rates when budgeting instead of carrying over last year's.
4. Training: the cheapest line and the first to be cut. In Spain part of it can be offset against the FUNDAE training credit, which cuts the net cost significantly.

One structural recommendation: leave 15-25% of the AI budget unallocated. Enough changes in a year that October's opportunities were not in the previous December's plan, and a fully committed budget forces you to choose between skipping the process and letting the opportunity pass.

What should you never delegate to AI?

Capital and people allocation. It is the decision that defines the year, it depends on information that sits in no system — which team can absorb more pressure, which customer is about to leave, which bet you actually believe — and it is what leadership is paid for.

• The objectives. They can be polished with help; they get chosen in a room with people.
• Headcount decisions. Growing, reorganizing or cutting is not a model output.
• Communicating the plan. An annual plan delivered in obviously generated prose destroys credibility faster than a mediocre plan explained in person.
• Commitment. Nobody commits to an objective they have not argued about. Planning exists to align intentions, and that does not automate.

A September-to-December calendar

1. September: set up automated data collection and close the base dataset. Do only this and you will already have gained two weeks in November.
2. October: year analysis, scenarios and a first initiative portfolio with estimates grounded in real history.
3. November: prioritization, budget by line and draft team objectives, with the coherence review.
4. December: close, communicate and set the quarterly review dates. A plan with no review date is a statement of intent.

If you want to build the automatable part with a team that operates these systems every week — starting from a diagnosis of where your data lives rather than from a tool — that is how we work at our AI consulting practice.

Frequently asked questions

Can I just ask a model to write my annual plan?

You can, and you will get a correct, generic, useless document. An annual plan is worth something because of your business's specific assumptions and the commitment of whoever will execute it, and neither comes out of a prompt. What is worth delegating is data consolidation, variance analysis and criticism of the objectives your team writes.

What data do I need as a minimum?

Revenue and margin by business line, fixed and variable costs, headcount and cost per person, and operating metrics for your two main processes. That is enough to build useful scenarios. If you lack detail by business line, that is the cycle's first fix, because without it every scenario comes from the same aggregate and distinguishes nothing.

Is it safe to put financial data into a model?

It depends on the configuration, not on the vendor in the abstract. The minimum: a data processing agreement, no training on your data, limited retention and control over who has access. In many cases the simplest answer is to have the agent work with aggregated or anonymized data, which is usually enough for planning.

How long does the automated part take to build?

Data collection and consolidation takes three to six weeks depending on how many sources there are and what state they are in. It pays for itself the same year, because the same setup serves the monthly report and the quarterly review, not only the annual plan.

What if my company is too small to have a planning cycle?

Then the cycle is one afternoon in December, and it is still worth doing. Consolidate the year's numbers, write three scenarios on one page and pick three objectives. AI contributes mostly the consolidation and the critique of those three objectives — which is where small companies most often go wrong, by setting ten.