NEWMake Viral Reels turns any link into a ready-to-post reel.Try it free →
AI for Enterprise · Playbook

How to build a 2-year AI strategy when you can't predict the next 12 months

A CIO I spoke with recently asked a question that most AI roadmaps quietly dodge: how do you plan two years ahead when the next twelve months are already hard to see?

You cannot say which models your teams will be using. You cannot say what a unit of intelligence will cost. You cannot even say whether the systems we call AI today will still be large language models by 2030. I have not seen a forecast I would bet a budget on, and I do not think chasing one is a good use of a leadership team's time.

A better starting point is this: stop making the model the strategy. Build the parts of your system that stay useful even if the model you use today is gone next year.

Why a model-first plan falls apart

Twelve months is a long stretch in this field. Prices drop. New names appear. Some of the tools you hear every week will fade. Systems that are not language models at all will start showing up in real products.

If your two-year plan is "we will standardize on Model X," that plan is already fragile. The model is the part that changes fastest. A strategy built around it has to be rewritten every time a new release lands. That is why "which model should we pick?" is the wrong first question. The useful question is: what are we putting in place that still matters if that model is replaced?

The habit of always picking the biggest model

A lot of teams are doing the opposite right now. For every task, they reach for the newest, largest model they can get. I call this intelligence maxing. It feels careful. The most capable model should give the best result. In practice, it is often the wrong tool for the job.

A smaller model, or no model at all, often does the work better and at a lower cost. If a rule, a lookup, or a fixed workflow can settle a request, you do not need a frontier model to "think" about it. You need the rule.

Here is a simple, illustrative cost picture, not a forecast or a measured result from a client. Imagine a support workflow that handles 100,000 tickets a month. If every ticket goes through a frontier model at a few cents each, you are looking at thousands of dollars a month for work that may not need that model. If 80 of every 100 tickets are routine, a smaller model or a rules engine can handle those, and you reserve the expensive model for the harder 20. Same answer for the customer. A very different bill, and a system that is easier to control.

The exact numbers will vary. The habit is the problem. Defaulting to the biggest model is not a strategy. It is a quiet tax.

A demo is not the system

AI also makes it easy to confuse a polished prototype with something you can run every day. A demo in a meeting can look finished. Production is a different test. The plain parts still decide whether it works: the data it can see, the context around each request, what it costs to run, the controls around it, and the people who have to use and trust it.

So build around what will still be there in two years, not around the layer you already know will change.

Five things that still matter when the model changes

These five layers stay valuable as models improve, disappear, or get replaced by something new.

1. Your data and company knowledge

This is the asset a model vendor cannot ship you: your contracts, incident history, product rules, and customer language. If a system cannot see the knowledge that makes your company yours, it will give you generic answers.

What to do: decide which sources are allowed into AI systems, who owns them, and how they stay current. Write down where the important knowledge lives, and make it available without pasting it into a prompt by hand.

2. Your workflows

AI is most useful when it sits inside work that already happens: a claim, a ticket, a month-end close, a design review. A chat box on the side of a process is easy to ignore. A step inside the process is hard to ignore.

What to do: pick two or three high-volume workflows. Map the steps, the handoffs, and the points where a person currently waits or retypes. Put the model in those spots, with a clear input and a clear output. Leave the rest of the process alone until that one step is reliable.

3. Your ability to judge the output

If you cannot tell a good answer from a bad one, you cannot improve the system, and you cannot defend it to a risk team. A fixed set of examples is how you stay in control when the model underneath changes.

What to do: for each workflow, write a small set of real cases: the input, the acceptable output, and the failures you will not tolerate. Run those cases every time you change a model, a prompt, or a tool. Keep a person on the cases the system is unsure about, and record why.

4. Your controls

This is access, logging, retention, and the rules for what the system is allowed to do. Controls are not a late compliance chore. They are how you can put the system in front of customers and staff.

What to do: set who can use it, what data it can see, and what actions it can take without a person. Log prompts, outputs, and tool calls. Make it possible to turn one feature off without taking the whole product down.

5. Human judgement

People still decide what "good" looks like, when to override the system, and when a result should never ship. That judgement is a design input, not a backup you add after something breaks.

What to do: name the owners. Give them a way to correct the system, and a way for those corrections to feed back into the workflow and the example set. A tool that staff do not trust will not get used, no matter how strong the model is.

A setup that does not lock you to one vendor

Once the layers above are clear, the technical shape should make it easy to change models without rewriting the business.

  • Route first. Send each request to the cheapest system that can do it well: a rule, a small model, or a larger model.
  • Keep a thin interface. Your apps should call "the assistant" or "the classifier," not a vendor model name buried in ten services.
  • Make providers easy to swap. Prompts, tools, and output shapes should live outside any one vendor's library, so you can change the model without a rewrite.
  • Test on every change. The same example set should run when you switch providers, upgrade a model, or edit a prompt. If quality drops, you see it before users do.
  • Keep company context in your own layer. Retrieval, permissions, and memory should be yours. The model reads what you give it. Do not store company knowledge inside the model.
  • Cap cost and blast radius. Set a budget per workflow. If a new model costs more, the router should have to earn that spend. If something fails, one workflow should fail, not all of them.

If you cannot swap the model in a week, you are more locked in than you think.

A 90-day starting plan

You do not need a finished two-year architecture on day one. You need a first version of the durable layers, and one workflow that proves them.

Days 1 to 30: choose and prepare. Pick one workflow with clear volume, a measurable outcome, and a person who owns it. Inventory the data it needs, and who is allowed to see it. Write 20 to 50 real examples of good and bad outputs. Agree the controls: access, logging, and what the system is not allowed to do.

Days 31 to 60: build a thin slice. Stand up a path that can use more than one model. Put the router in front, even if it only has two routes to start. Run the example set every day. Put a person on the uncertain cases. Measure time, cost, and error rate against how the work is done today.

Days 61 to 90: make it ordinary, then decide. If the slice is better than the current process, move it into the real workflow, with the controls on. Write down what you would reuse for a second workflow: the data path, the example set, the router, the review step. Only then set the next nine months of work. The two-year strategy is this pattern, repeated, not a bet on a single model.

At the end of 90 days you should be able to answer four questions: which workflow improved, what it costs, what happens when the model changes, and who is accountable.

What to take into the next two years

You will not get a reliable forecast of models, prices, or architectures. You do not need one. Build around your data, your workflows, your tests, your controls, and your people's judgement. Keep the model easy to replace. Start with one workflow in 90 days, then repeat.

The question to keep asking is not "which model should we pick?" It is "what are we building that still helps, no matter which model wins?"

If you want more writing like this, the Hyperautomation AI Report is a free newsletter, three times a week, with practical notes on building and using AI. You can subscribe on the newsletter page.

Talk to me about AI in your company

If you're working through an AI rollout, tell me what you're trying to do. I read every message and reply by email.

Or connect with me on LinkedIn.