Two beliefs run the same building at once. Whatever we buy will be outdated before we finish rolling it out; we are already behind, and we lose another month every month we deliberate. Executives state both in one conversation, then settle the tension by doing nothing — the single course that guarantees both fears come true. The useful question is narrower than act or wait: what shape of commitment survives a technology that reprices itself this fast?
Timing the First AI Commitment
Should You Wait for AI to Settle,
or Start Now?
Which parts of a first AI commitment carry forward through a model change, which part you are knowingly buying twice, and how to write the difference into the deal.
Built from real buyer questions in our sales meetings
Direct Answer
Start now, and buy for portability. Pace of change makes long monolithic programs dangerous and small portable ones cheap, so the obsolescence fear argues for a shape of commitment rather than for delay. Buy capability you can carry forward — your own cleaned data, your own hardware, a model you can swap — and keep any single model provider out of a multi-year agreement. Structure the first purchase so the durable assets survive a model change.
The limit: portability postpones the hardware cost rather than removing it. Iternal states that added features will eventually require the customer’s equipment to be upgraded to handle them, and puts a reasonably powered local machine about a year behind what a data center can do. Your cleaned data and your workflows hold value through a model change; the machine underneath them ages on its own schedule. Fund the hardware leg as a recurring line.
Name the assets before you sign. Say which parts of the commitment you expect to survive a model change and which part you are knowingly buying twice. Then take three answers in writing: which upgrades come with the license, whether a different model can be loaded without reopening the agreement, and which future capabilities will require new equipment.
Waiting and pacing are different decisions. Pacing buys a small bounded first step that teaches the organization something usable whatever it purchases next. Waiting holds position until the market steadies, and returns the same result each year: thinner internal skill and a wider gap to close. For more information on how the capability distance between local and frontier models moves, visit the model market page; on how AI software is priced and owned, visit the licensing page.
Both Pressures Are Real, and One Room Holds Both
Buyers described the pace problem in operational terms, again and again. A complex integration project returns no productivity gain for two years while the technology moves underneath it. A stack chosen this week can be superseded by next week. One buyer put the arithmetic plainly: by the time a long project completes, the work is no longer worth what it cost.
The same executives described the opposite pressure with equal force. An organization not using large language models today is falling behind the curve. Competitors already run AI as a productivity multiplier. Both descriptions are accurate, and together they give one instruction: shorten the commitment and keep the learning.
Comfort is the wrong trigger. A technology leader who waits until AI feels settled deploys nothing at scale, because settling is not a state this technology enters. Spending a little now buys the one asset that compounds whichever model wins — your own competence.
What Carries Forward Through a Model Change
Obsolescence lands unevenly across a deployment. Part of what you buy expires with a specific model; part belongs to you whichever engine runs next year; one part ages on its own cycle whatever you decide. Separating the layers turns an unanswerable question about the future of AI into five answerable questions about your purchase order.
| Layer of the commitment | Carries forward? | What Iternal states |
|---|---|---|
| Your cleaned data | Yes | Blockify is model agnostic and platform agnostic: a cleaned data set can be refreshed for AirgapAI and the same package used with other systems. |
| Workflows and prompting skill | Yes | Most of adopting AirgapAI is setting up workflows, which carry the detailed prompt behind the scenes. Iternal teaches prompting as tool-agnostic enablement. |
| The license | Yes, within its terms | A one-time perpetual license per device, good for the life of that device, upgrades included and no annual maintenance fee. It can be reassigned to another user. |
| The model | Replaceable by design | AirgapAI supports bring your own model: any open-source model can run in it, several can sit on the machine with one active, and models swap at any point. |
| The hardware | Ages on its own cycle | Added features will eventually require an equipment upgrade, and a reasonably powered local machine runs about a year behind what a data center can do. |
Four layers travel with you; the fifth is the one every organization already replaces on a schedule. The data work is the investment, the model is a setting, and the hardware is a refresh cycle already in somebody’s budget.
The Six-Month Question Belongs in the Contract
One objection arrived almost word for word, from buyer after buyer: if I buy AI today it will be outdated in six months and I will look bad in my job. It is a procurement question wearing a technology costume. The software still runs in six months; what can look wrong is the size and shape of the commitment, and the career damage attaches to the purchase order rather than the product. Buyers said as much — money went unallocated precisely because the purchase might be invalidated.
Procurement is where that risk gets managed, and four terms carry almost all of the work:
- Size. Iternal’s motion is to start small, take a win at low cost, prove it works, then scale. A commitment small enough to be wrong is a commitment nobody gets fired over.
- Ownership. AirgapAI is a one-time perpetual license per device with no monthly fee and no annual maintenance, upgrades included for the life of the device. A purchase that looks dated still runs, and still costs nothing further.
- Substitution. Bring your own model support keeps the engine a setting instead of a bet. The model you regret is the cheapest part to replace.
- Extraction. Data cleaned for one system stays usable by others, so the preparation effort remains yours when the platform changes.
Turn each term into a written answer before the purchase order leaves the building:
Pin it down: questions for your evaluation
-
Which published upgrades are included for the life of the device, and which future capabilities will require new equipment?Separates the software entitlement you already own from the hardware refresh you will fund later.
-
Can we load a different model into the same deployment without reopening the agreement?Whether a change of model is a settings change or a commercial event.
-
When we replace the machine, can the license be retracted and reassigned?Whether the license survives your hardware refresh cycle or repeats with it.
If You Are Not Ready, Readiness Is the First Project
A second position deserves more respect than it usually gets: hold off until the guardrails exist and the organization is confirmed ready. IT leaders said they must experiment first so they do not deploy AI the wrong way. Putting a general-purpose assistant at the fingertips of every associate does carry downside.
Treat readiness as work rather than as a waiting condition. The organizations that got moving completed a short list first, none of it requiring a long-term platform choice:
- Write the policy first. Iternal tells clients with no AI policy or culture statement to start there before any AI project.
- Frame the governance. A detailed exploration produces a governance framework, risk management, an ROI and business case, training and a pilot charter — what a board asks for before a deployment rather than after.
- Train people before you deploy to them. AI Academy teaches prompt engineering interactively: the learner gets the requirement, writes the prompt, and is scored immediately.
- Rehearse on material that cannot hurt you. A first exercise can run on public or mock data resembling the real environment, sensitive material moving in once the flow is proven.
- Prepare the data whatever you choose later. Blockify gets an organization’s data AI-ready whichever platform gets picked afterwards, and sits early at the chunking stage so it works alongside the guardrails you add.
Every item on that list survives a model change, and none of them commits you to one.
Ownership Changes the Arithmetic of Waiting
A subscription makes hesitation look rational: every month of delay is a month unpaid, so waiting shows a visible return. Ownership inverts that math. AirgapAI is bought once as a perpetual license per device and owned for the life of that device, upgrades included and no annual maintenance fee, which lets finance amortize one figure across the replacement cycle instead of defending a renewal every year. Iternal licenses it that way for a practical reason too: many customers cannot reach an external network to verify a subscription. That difference decides what a six-month technology swing costs you. For more information visit the licensing page.
Answered elsewhere
- Which problem to point AI at first, and how to score the candidates — see choosing your first use case.
- What a paid engagement covers and how long it runs — see the engagement page.
- How the capability distance between local and frontier models moves — see the model market page.
- How AI software is priced across per-seat, per-device, perpetual and subscription models — see the licensing page.
- Getting a first project moving with no AI specialists on the payroll — see starting without in-house AI skills.
- The full order of the first decisions, and the gate in front of them — see the starting-sequence overview.
FAQ
FAQ: Acting Now Versus Waiting
The model will be superseded; the commitment around it does not have to be. Cleaned data, workflows and a perpetual per-device license carry forward through a model change, and AirgapAI supports bring your own model, so swapping the engine is a setting rather than a repurchase. The hardware genuinely ages: Iternal states added features will eventually require an equipment upgrade.
Waiting removes the wrong risk. It protects you from a purchase that looks dated and costs you the internal skill that takes longest to build, because an organization that never runs AI never learns to implement it. Pace the commitment instead: small, portable, and no multi-year tie to one model provider.
Pace punishes long programs specifically. Buyers described multi-year integrations that return no productivity gain while the technology moves underneath them, and stacks superseded within weeks of selection. The same pace makes a small bounded first step cheap to be wrong about. Match the length of the commitment to the speed of the market.
It changes the sequence rather than the shape. Start with the assets that carry forward whatever you buy next: a written AI policy, a governance framework, training, and data prepared to be usable by any platform. Arriving late is recoverable; arriving late having learned nothing is the expensive version.
Five things, none of which locks you to a platform: a written AI policy or culture statement, a governance and risk framework with a business case, training for the people who will use it, a first exercise on public or mock data, and data preparation that stays usable later. Run readiness as a project with an end date, because comfort never arrives on its own.
Make the First Commitment Small and Portable
The decision in front of you is the length of the commitment rather than the date on the calendar. Prepare data any platform can use, keep the model replaceable, own the license outright, and budget the hardware refresh you were always going to fund. The next model release then arrives as an upgrade instead of a write-off.